# BRAIN NETWORKS FOR STUDYING HEALTHY AND PATHOLOGICAL AGING MECHANISMS AND INTERVENTION EFFICACY

EDITED BY : Christos Frantzidis, Ana B. Vivas and Panagiotis D. Bamidis PUBLISHED IN : Frontiers in Aging Neuroscience

#### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88966-122-0 DOI 10.3389/978-2-88966-122-0

#### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

#### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

#### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

#### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# BRAIN NETWORKS FOR STUDYING HEALTHY AND PATHOLOGICAL AGING MECHANISMS AND INTERVENTION EFFICACY

Topic Editors:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece Ana B. Vivas, International Faculty of the University of Sheffield, Greece Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

Previous studies showed that both healthy and pathological aging are associated with changes in brain structure and function of the mature human brain. The most prominent anatomical alteration are changes in prefrontal cortex morphology, volume loss and reduced white-matter integrity and hippocampal atrophy. Cognitive decline affects mainly the performance of episodic memory, speed of sensory information processing, working memory, inhibitory function and long-term memory. It has been also proposed that due to the aforementioned changes the aging brain engages in compensatory brain mechanism such as a broader activation of cortical regions (mainly frontal) rather than specialized activation. Evidence suggests that similar changes occur with pathological aging but to a greater extent. In this case information flow is disrupted due to neurodegeneration, functional activation of posterior (occipito-temporal) regions is decreased and as a consequence the brain fails to process sensorial input in the ventral pathway and cognitive deficits appear.

In the last years, functional alterations associated with aging have been studied using the mathematical notion of graph theory that offers an integrative approach since it examines different properties of the brain network: 1) Organization level 2) amount of local information processing, 3) information flow 4) cortical community structure and 5) identification of functional / anatomical hubs. So, graph theory offers an attractive way to model brain networks organization and to quantify their pathological deviations.

Previous studies have already employed this mathematical notion and demonstrated that age-related neurodegeneration is often accompanied by loss of optimal network organization either due to diminished local information processing or due to progressive isolation of distant brain regions. They have also found that changes in network properties may be present even in the preclinical phase, which could be taken as a biological marker of disease.

Citation: Frantzidis, C., Vivas, A. B., Bamidis, P. D., eds. (2020). Brain Networks for Studying Healthy and Pathological Aging Mechanisms and Intervention Efficacy. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88966-122-0

# Table of Contents

*08 A Triple Network Connectivity Study of Large-Scale Brain Systems in Cognitively Normal APOE4 Carriers*

Xia Wu, Qing Li, Xinyu Yu, Kewei Chen, Adam S. Fleisher, Xiaojuan Guo, Jiacai Zhang, Eric M. Reiman, Li Yao and Rui Li

*17 Patterns of Longitudinal Neural Activity Linked to Different Cognitive Profiles in Parkinson's Disease*

Atsuko Nagano-Saito, Mohamed S. Al-Azzawi, Alexandru Hanganu, Clotilde Degroot, Béatriz Mejia-Constain, Christophe Bedetti, Anne-Louise Lafontaine, Valérie Soland, Sylvain Chouinard and Oury Monchi

*29 Global Efficiency of Structural Networks Mediates Cognitive Control in Mild Cognitive Impairment*

Rok Berlot, Claudia Metzler-Baddeley, M. Arfan Ikram, Derek K. Jones and Michael J. O'Sullivan


Hui He, Cheng Luo, Xin Chang, Yan Shan, Weifang Cao, Jinnan Gong, Benjamin Klugah-Brown, Maria A. Bobes, Bharat Biswal and Dezhong Yao

*61 Neuroanatomical and Neuropsychological Markers of Amnestic MCI: A Three-Year Longitudinal Study in Individuals Unaware of Cognitive Decline*

Katharina S. Goerlich, Mikhail Votinov, Ellen Dicks, Sinika Ellendt, Gábor Csukly and Ute Habel


Heather T. Whittaker and Jason D. Warren

*90 Resting State fMRI Reveals Increased Subthalamic Nucleus and Sensorimotor Cortex Connectivity in Patients With Parkinson's Disease Under Medication*

Bo Shen, Yang Gao, Wenbin Zhang, Liyu Lu, Jun Zhu, Yang Pan, Wenya Lan, Chaoyong Xiao and Li Zhang

*100 Age-Related Differences in Reorganization of Functional Connectivity for a Dual Task With Increasing Postural Destabilization* Cheng-Ya Huang, Linda L. Lin and Ing-Shiou Hwang

*120 MEG Beamformer-Based Reconstructions of Functional Networks in Mild Cognitive Impairment*

Maria E. López, Marjolein M. A. Engels, Elisabeth C. W. van Straaten, Ricardo Bajo, María L. Delgado, Philip Scheltens, Arjan Hillebrand, Cornelis J. Stam and Fernando Maestú


Ravi Rajmohan, Ronald C. Anderson, Dan Fang, Austin G. Meyer, Pavis Laengvejkal, Parunyou Julayanont, Greg Hannabas, Kitten Linton, John Culberson, Hafiz Khan, John De Toledo, P. Hemachandra Reddy and Michael W. O'Boyle


Jose A. Santiago, Virginie Bottero and Judith A. Potashkin

*182 Racial Differences in Insular Connectivity and Thickness and Related Cognitive Impairment in Hypertension*

Ganesh B. Chand, Junjie Wu, Deqiang Qiu and Ihab Hajjar

*192 Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation*

Giovanni Assenza, Fioravante Capone, Lazzaro di Biase, Florinda Ferreri, Lucia Florio, Andrea Guerra, Massimo Marano, Matteo Paolucci, Federico Ranieri, Gaetano Salomone, Mario Tombini, Gregor Thut and Vincenzo Di Lazzaro

*210 Corrigendum: Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation*

Giovanni Assenza, Fioravante Capone, Lazzaro di Biase, Florinda Ferreri, Lucia Florio, Andrea Guerra, Massimo Marano, Matteo Paolucci, Federico Ranieri, Gaetano Salomone, Mario Tombini, Gregor Thut and Vincenzo Di Lazzaro

*211 Inhibition of PirB Activity by TAT-PEP Improves Mouse Motor Ability and Cognitive Behavior*

Ya-Jing Mi, Hai Chen, Na Guo, Meng-Yi Sun, Zhao-Hua Zhao, Xing-Chun Gao, Xiao-Long Wang, Rui-San Zhang, Jiang-Bing Zhou and Xing-Chun Gou

*219 Age-Related Decline in the Variation of Dynamic Functional Connectivity: A Resting State Analysis*

Yuanyuan Chen, Weiwei Wang, Xin Zhao, Miao Sha, Ya'nan Liu, Xiong Zhang, Jianguo Ma, Hongyan Ni and Dong Ming

#### *230 Altered Neuronal Activity Topography Markers in the Elderly With Increased Atherosclerosis*

Takashi Shibata, Toshimitu Musha, Yukio Kosugi, Michiya Kubo, Yukio Horie, Naoya Kuwayama, Satoshi Kuroda, Karin Hayashi, Yohei Kobayashi, Mieko Tanaka, Haruyasu Matsuzaki, Kiyotaka Nemoto and Takashi Asada


Evatte T. Sciberras-Lim and Anthony J. Lambert

*260 Disconnectivity Between Dorsal Raphe Nucleus and Posterior Cingulate Cortex in Later Life Depression*

Toshikazu Ikuta, Koji Matsuo, Kenichiro Harada, Mami Nakashima, Teruyuki Hobara, Naoko Higuchi, Fumihiro Higuchi, Koji Otsuki, Tomohiko Shibata, Toshio Watanuki, Toshio Matsubara, Hirotaka Yamagata and Yoshifumi Watanabe

*266 Beta-Band Functional Connectivity Influences Audiovisual Integration in Older Age: An EEG Study*

Luyao Wang, Wenhui Wang, Tianyi Yan, Jiayong Song, Weiping Yang, Bin Wang, Ritsu Go, Qiang Huang and Jinglong Wu

*277 Altered Functional and Causal Connectivity of Cerebello-Cortical Circuits Between Multiple System Atrophy (Parkinsonian Type) and Parkinson's Disease*

Qun Yao, Donglin Zhu, Feng Li, Chaoyong Xiao, Xingjian Lin, Qingling Huang and Jingping Shi

*288 Does Aerobic Exercise Influence Intrinsic Brain Activity? An Aerobic Exercise Intervention Among Healthy Old Adults*

Pär Flodin, Lars S. Jonasson, Katrin Riklund, Lars Nyberg and C. J. Boraxbekk

*301 Balance Training Enhances Vestibular Function and Reduces Overactive Proprioceptive Feedback in Elderly*

Isabella K. Wiesmeier, Daniela Dalin, Anja Wehrle, Urs Granacher, Thomas Muehlbauer, Joerg Dietterle, Cornelius Weiller, Albert Gollhofer and Christoph Maurer

*314 Open- and Closed-Skill Exercise Interventions Produce Different Neurocognitive Effects on Executive Functions in the Elderly: A 6-Month Randomized, Controlled Trial*

Chia-Liang Tsai, Chien-Yu Pan, Fu-Chen Chen and Yu-Ting Tseng

*330 Acute Stress Affects the Expression of Hippocampal Mu Oscillations in an Age-Dependent Manner*

Samir Takillah, Jérémie Naudé, Steve Didienne, Claude Sebban, Brigitte Decros, Esther Schenker, Michael Spedding, Alexandre Mourot, Jean Mariani and Philippe Faure

*351 Cognitive Training Enhances Auditory Attention Efficiency in Older Adults* Jennifer L. O'Brien, Jennifer J. Lister, Bernadette A. Fausto, Gregory K. Clifton and Jerri D. Edwards


Elena Solesio-Jofre, Iseult A. M. Beets, Daniel G. Woolley, Lisa Pauwels, Sima Chalavi, Dante Mantini and Stephan P. Swinnen

*498 The Aerobic and Cognitive Exercise Study (ACES) for Community-Dwelling Older Adults With or At-Risk for Mild Cognitive Impairment (MCI): Neuropsychological, Neurobiological and Neuroimaging Outcomes of a Randomized Clinical Trial*

Cay Anderson-Hanley, Nicole M. Barcelos, Earl A. Zimmerman, Robert W. Gillen, Mina Dunnam, Brian D. Cohen, Vadim Yerokhin, Kenneth E. Miller, David J. Hayes, Paul J. Arciero, Molly Maloney and Arthur F. Kramer


Vasiliki I. Zilidou, Christos A. Frantzidis, Evangelia D. Romanopoulou, Evangelos Paraskevopoulos, Styliani Douka and Panagiotis D. Bamidis

*565 Greek Traditional Dances: A Way to Support Intellectual, Psychological, and Motor Functions in Senior Citizens at Risk of Neurodegeneration* Styliani Douka, Vasiliki I. Zilidou, Olympia Lilou and Magda Tsolaki

# A Triple Network Connectivity Study of Large-Scale Brain Systems in Cognitively Normal APOE4 Carriers

Xia Wu1,2 , Qing Li <sup>1</sup> , Xinyu Yu<sup>1</sup> , Kewei Chen<sup>3</sup> , Adam S. Fleisher 3,4 , Xiaojuan Guo<sup>1</sup> , Jiacai Zhang<sup>1</sup> , Eric M. Reiman<sup>3</sup> , Li Yao1,2 and Rui Li <sup>5</sup> \*

<sup>1</sup> College of Information Science and Technology, Beijing Normal University, Beijing, China, <sup>2</sup> State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China, <sup>3</sup> Banner Alzheimer's Institute and Banner Good Samaritan PET Center, Phoenix, AZ, USA, <sup>4</sup> Eli Lilly and Company, Indianapolis, IN, USA, <sup>5</sup> Center on Aging Psychology, Key Laboratory of Mental Health, Institute of Psychology, Chinese Academy of Sciences, Beijing, China

The triple network model, consisting of the central executive network (CEN), salience network (SN) and default mode network (DMN), has been recently employed to understand dysfunction in core networks across various disorders. Here we used the triple network model to investigate the large-scale brain networks in cognitively normal apolipoprotein e4 (APOE4) carriers who are at risk of Alzheimer's disease (AD). To explore the functional connectivity for each of the three networks and the effective connectivity among them, we evaluated 17 cognitively normal individuals with a family history of AD and at least one copy of the APOE4 allele and compared the findings to those of 12 individuals who did not carry the APOE4 gene or have a family history of AD, using independent component analysis (ICA) and Bayesian network (BN) approach. Our findings indicated altered within-network connectivity that suggests future cognitive decline risk, and preserved between-network connectivity that may support their current preserved cognition in the cognitively normal APOE4 allele carriers. The study provides novel sights into our understanding of the risk factors for AD and their influence on the triple network model of major psychopathology.

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Ramesh Kandimalla, Texas Tech University, USA Neha Sehgal, Wisconsin Institute for Discovery, USA

#### \*Correspondence:

Rui Li lir@psych.ac.cn

Received: 06 July 2016 Accepted: 16 September 2016 Published: 28 September 2016

#### Citation:

Wu X, Li Q, Yu X, Chen K, Fleisher AS, Guo X, Zhang J, Reiman EM, Yao L and Li R (2016) A Triple Network Connectivity Study of Large-Scale Brain Systems in Cognitively Normal APOE4 Carriers. Front. Aging Neurosci. 8:231. doi: 10.3389/fnagi.2016.00231 Keywords: Alzheimer's disease, APOE4, Bayesian network, connectivity, fMRI, triple network model

### INTRODUCTION

The apolipoprotein e4 (APOE4) gene has been well established as a susceptibility gene for sporadic and late-onset familial Alzheimer's disease (AD; Poirier et al., 1995; Reitz and Mayeux, 2010; Kandimalla et al., 2013; Tai et al., 2014). Epidemiologic evidence has clarified that APOE4 decreases the age-at-onset of AD in a gene dosage-dependent manner (Corder et al., 1993; Breitner et al., 1999). Neuroimaging studies have demonstrated that APOE4 carriers exhibit elevated medial temporal lobe (MTL) atrophy (Agosta et al., 2009; Fleisher et al., 2009a,b; Wolk and Dickerson, 2010), and recent studies have shown that the APOE4 allele is associated with Cerebrospinal fluid (CSF) biomarkers including Aβ42, tau (Kandimalla et al., 2011) and ubiquitin levels (Kandimalla et al., 2014). Thus the APOE4 allele has been suggested as an important factor that leads to lower cognitive performance, or the progression to mild cognitive impairment (MCI) and AD (Barabash et al., 2009; Sasaki et al., 2009).

Functional neuroimaging connectome studies of AD have proposed a disconnection hypothesis of the disease. Many studies have consistently reported that the cognitive impairment in AD and the cognitive decline in its preclinical stage were largely due to the disruptions of the brain networks (Stam et al., 2007; Lo et al., 2010; Wang et al., 2013). For example, as one of the most relevant networks in AD, various studies have shown that the default mode network (DMN) exhibited a disruption in functional connectivity in AD (Greicius et al., 2004; Rombouts et al., 2005; Celone et al., 2006; Petrella et al., 2007; Wu et al., 2011), and even at early stages of the disease such as MCI (Lustig et al., 2003; Rombouts et al., 2005; Celone et al., 2006; Petrella et al., 2007; Qi et al., 2010; Li et al., 2013). In addition to the DMN, other networks have also been found to show alterations in AD. For example, the salience network (SN), whose connectivity showed negative correlation with DMN has been linked to AD (Zhou et al., 2010; Balthazar et al., 2014). These alterations in functionally coordinated brain systems can occur long before disease onset in cognitively normal people with various risk factors for AD (Poirier et al., 1995; Kivipelto et al., 2001; Song et al., 2015). For example, Westlye et al. (2011) demonstrated a negative correlation between DMN synchronization and memory performance in healthy APOE4 carriers. Besides, the functional alterations in the DMN and SN connections were also demonstrated in the elderly APOE4 carriers (Machulda et al., 2011). These evidences suggested that the presence of APOE4 gene is accompanied by brain network alterations that are closely relevant to AD progression.

Recently, a triple network model of major psychopathology has been proposed by Menon (2011). The triple network model consists of the central executive network (CEN), SN and DMN. These three networks are generally referred to as the core neurocognitive networks due to their involvement in an extremely wide range of cognitive tasks (Greicius et al., 2003; Greicius and Menon, 2004; Menon and Uddin, 2010; Menon, 2011). Specifically, the CEN and SN typically show increased activation during stimulus-driven cognitive or affective processing, while the DMN shows decreased activation during tasks in which self-referential and stimulusindependent intellectual activity is not involved (Greicius et al., 2003; Greicius and Menon, 2004). The triple network model suggests that the aberrant internal organization within each functional network and the interconnectivity among them are characteristic of many psychiatric and neurological disorders. Recently the triple network model has been widely applied to elucidate the dysfunction across multiple disorders, including schizophrenia, depression and dementia (Menon and Uddin, 2010; Menon, 2011; Zheng et al., 2015; Yuan et al., 2016). However the triple network interactions in elderly APOE4 carriers who are at high risk to AD have not yet been explored.

In the present study, we investigated the APOE4 mediated modulation of the within-network functional connectivity and the between-network connectivity of the three core networks included in the triple network model in cognitively normal individuals carrying a family history of AD and at least one copy of the APOE4 allele using functional magnetic resonance imaging (fMRI). A group independent component analysis (ICA) approach and Bayesian network (BN) approach were used to separate the functional connectivity networks from the fMRI dataset and to determine the between-network effective connectivity, respectively.

#### MATERIALS AND METHODS

### Participants

fMRI data from 29 cognitively normal right-handed volunteers (8 males and 21 females, ages between 50 and 65 years) who were the subjects in our previous study (Fleisher et al., 2009b) were included in this work. They were divided into two groups: the high-risk group and the low-risk group. The high-risk group included 17 subjects who had a significant family history of dementia in a first-degree relative and at least one copy of the APOE4 allele. The other twelve participants who had neither a family history of dementia nor a copy of the APOE4 gene were regarded as the low-risk group. Notably, there were no significant differences in age, gender and education level between these two groups (all ps > 0.05). The two groups were matched on general cognitive function as evaluated by Folstein Mini Mental State Exam (p = 0.39). The study was conducted according to Good Clinical Practice, the Declaration of Helsinki and US 21 Code of Federal Regulations (CFR) Part 50-Protection of Human Subjects, and Part 56-Institutional Review Boards and was approved by the Institutional Review Board of the University of California, San Diego. Written informed consent for the study was obtained from all of the participants before protocol-specific procedures were performed, including cognitive testing.

All scans were performed on a General Electric Signa EXCITE 3.0 T short bore, twin speed scanner with a body transmit coil and an 8 channel receive array. High-resolution structural brain images were acquired with a magnetization prepared from threedimensional fast spoiled gradient sequence acquisition (FSPGR: 124 axial slices, 1 mm×1 mm in-plane resolution, 1.3 mm slice thickness, Field of View (FOV) = 256 mm<sup>2</sup> × 256 mm<sup>2</sup> , TR = 7.8 ms, TE = 3.1 ms, flip angle = 12◦ ). Blood oxygen level dependent (BOLD) data were acquired using echo planar imaging sequences (35 slices, perpendicular to the axis of the hippocampus, 6 mm in-plane resolution, 0 spacing, FOV = 220 mm<sup>2</sup> × 220 mm<sup>2</sup> , TE = 30 ms, TR = 2500 ms, voxel size = 3.4 mm<sup>3</sup> × 3.4 mm<sup>3</sup> × 6.0 mm<sup>3</sup> ).

#### Data Preprocessing

For each participant, the original first five-time functional images were discarded to allow for equilibration of the magnetic field. All of the preprocessing steps were performed using the Statistical Parametric Mapping program (SPM8<sup>1</sup> ). They included within-subject inter-scan realignment, between-subject spatial normalization to a standard brain template in the Montreal Neurological Institute (MNI) coordinate space, and smoothing by a Gaussian filter with a full width at a half maximum of 8 mm. Following this, the linear trend with regard to time was removed by linear regression via the Resting-State fMRI Data Analysis Toolkit (REST<sup>2</sup> ).

<sup>1</sup>http://www.fil.ion.ucl.ac.uk/spm <sup>2</sup>http://restfmri.net

After the preprocessing, we employed the Group ICA and BN to learn the functional interactions of the triple network model. Group ICA was first used to isolate the three brain networks for examination of the functional connectivity changes within each network in the high risk group. The BN was then used to show the directed causal effects between these three networks in the high risk group. Thus, the study was developed to delineate the influence of APOE4 on the triple networks in both within-network connections and betweennetwork interactions.

#### Group Independent Component Analysis

Group ICA is widely used to separate patterns of task-activated neural networks, image noises, and physiologically generated independent components (ICs) in a data-driven manner. The preprocessed data of all participants were entered into the Group ICA program in the fMRI Toolbox (GIFT<sup>3</sup> ) for the separation of the three networks included in the triple network model and the determination of networks for BN analysis. The Group ICA program included two rounds of principal component analyses (PCA) for reduction of fMRI data dimensions, ICA separation and back-reconstruction of the ICs (Calhoun et al., 2001). The optimal number of principal components, 31, was estimated based on the minimum description length (MDL). In the first round of PCA, the data for each individual subject were dimension-reduced to the optimal number temporally. After concatenation across subjects within groups, the dimensions were again reduced to the optimal numbers via the second round of PCA. Then, the data were separated by ICA using the Extended Infomax algorithm (Lee et al., 1999). After ICA separation, the mean ICs and the corresponding mean time courses over all of the subjects were used for the back-reconstruction of the ICs and time courses for each individual subject (Calhoun et al., 2001).

Finally, the ICs that best matched the CEN, DMN, and SN for both the low- and high-risk groups were selected separately. Following this, one-sample t-test (p < 0.001, corrected by family wise error (FWE)) was performed to determine the CEN, DMN, and SN functional connectivity for the lowrisk and high-risk groups respectively. Between group withinnetwork functional connectivity difference was determined by two-sample t-test (p < 0.05, corrected by false discovery rate (FDR)).

#### Bayesian Network Analysis

BN analysis can be used to learn the global connectivity pattern for complex systems in a data-driven manner, and has been applied in our previous studies of AD and MCI (Wu et al., 2011; Li et al., 2013). Here, we employed the Gaussian BN method to characterize the large-scale networks in terms of directed effective connectivity among CEN, DMN and SN.

To establish the effective connectivity pattern of the three networks for the low- and high-risk groups separately, we defined the region of interest (ROI) mask as each of the three one-sample t-test network map (p < 0.001, FWE corrected).

<sup>3</sup>http://icatb.sourceforge.net/

The averaged time series over these voxels in every subject was extracted and then entered into the BN analysis for the construction of an effective connectivity pattern of the three core networks.

A BN model is a directed acyclic graph that encodes a joint probability distribution over a set of random variables. The directed arcs in the graph denote the conditional dependence relationships between nodes, which are qualified by the conditional probability of each node given its parents in the network. Specific to our BN model, we have three nodes in total, which represent the three core networks in the triple network model, and the arcs connecting them represent the directed effective connectivity between these functional networks. The time series of each node was calculated as the mean time series in each network ROI, and was assumed to follow a linear Gaussian conditional distribution. To learn the effective connectivity of the triple network model, we employed the Bayesian information criterion (BIC)-based learning approach. The BN model that maximized the BIC score among the space of possible candidates was selected as the best fit network. We used the L1-Regularization Paths algorithm (Schmidt et al., 2007) and the Maximum Likelihood Estimation (MLE) implemented in the collections of Matlab functions written by Murphy et al.<sup>4</sup> to learn the structure and parameters of the BN model, respectively, for the high- and low-risk groups.

#### Effective Connectivity Comparison Between the High- and Low-Risk Groups

To examine the effective connectivity difference of CEN, DMN and SN between the high- and low-risk groups, we adopted the randomized permutation procedure. We used the differences of the connection weight coefficients between the two groups as the statistical measure. The reference distribution is obtained by calculating all possible values of the test statistic under rearrangements of the group labels on the observed fMRI datasets. The statistics for the real two group samples were calculated first. Then, at each iteration of the test process, the subject-group membership was randomly assigned for each subject. A BN model for each rearranged group was constructed, and the differences of the connection weight coefficients between the two rearranged groups were calculated. We ran a total of 1000 permutations and assessed the sample distributions for these statistics. Finally, for each of the connections presented in the BN model for the two risk groups, type I errors of having between-group differences were estimated.

#### RESULTS

#### Functional Connectivity of CEN, DMN and SN

**Figure 1** shows the three networks included in the triple network model in the low and high-risk groups detected by Group ICA

<sup>4</sup>https://code.google.com/p/bnt

corrected by family wise error (FWE)). Bar at the right shows T-values.

(one-sample t-test, p < 0.001, FWE corrected). In both groups, the CEN includes the dorsolateral prefrontal cortex and the lateral posterior parietal cortex. The DMN includes the posterior cingulate cortex, medial prefrontal cortex, bilateral inferior parietal cortex, inferior temporal cortex and the hippocampus. The SN includes the dorsal anterior cingulate cortex and the fronto-insular cortex.

#### Within-Network Functional Connectivity Difference Between Groups

To compare the within-network functional connectivity difference of the CEN, DMN and SN between the lowand high-risk groups, we performed a two-sample t-test (p < 0.05, corrected by FDR) on individual maps of the three networks between the two groups. **Figure 2** displays the functional connectivity differences between the low and high-risk groups.

Within the CEN, the angular gyrus displayed increased functional connectivity in the low-risk group compared with the high-risk group (''LR > HR''), whereas the inferior parietal lobule displayed increased functional connectivity in the high-risk group compared with the low-risk group (''HR > LR''). Within the DMN, the right medial frontal gyrus displayed increased functional connectivity in the low-risk group compared with the high-risk group (''LR > HR''), whereas the left middle frontal gyrus displayed increased functional connectivity in the high-risk group compared with the low-risk group (''HR > LR''). Within the SN, the regions including the right middle temporal gyrus, right middle frontal gyrus and the anterior cingulate cortex displayed increased functional connectivity in the low-risk group compared with the high-risk group (''LR > HR''). In contrast, the regions including the left middle temporal gyrus, posterior lobe of the cerebellum and the supplemental motor area displayed increased functional connectivity in the high-risk group compared with the low-risk group (''HR > LR''). Details on these regions with betweengroup functional connectivity differences are listed in **Table 1**.

### BN-Based Effective Connectivity of CEN, DMN and SN

**Figure 3** shows the effective connectivity of the CEN, DMN and SN in the low-risk group and high-risk group learned using Gaussian BN approach. In accordance with the triple network model (Menon, 2011), **Figure 3** demonstrates consistently in the two groups that the DMN together with CEN receive connections from SN. It is important to note that the SN plays as a special node that does not receive but only generates connections in the model in both groups. Furthermore, the result of the

TABLE 1 | Brain regions that showed functional connectivity differences between the low and high risk groups (two sample t-test, p < 0.05, corrected by false discovery rate (FDR)).


mediates the activity of the CEN and DMN in both groups. The numbers on the connections represent the BN connectivity weights between brain networks.

random permutation test indicates that there is no significant difference among the effective connectivity coefficients of these three networks between the low- and high-risk groups (all ps > 0.05).

## DISCUSSION

The focus of the present study was to explore the possible impairment of the within-network functional connectivity and the between-network effective connectivity of the large-scale triple networks in cognitively normal individuals with a family history of AD and at least one copy of the APOE4 allele. Group ICA of the triple network model found that a couple of brain regions in the three networks showed significantly altered functional connectivity in the high-risk individuals, while the BN analysis of the model did not find significant between-group difference in the causal connections among the three functional networks.

We first compared the within-network functional connectivity between the low-risk subjects and the highrisk subjects. The results demonstrated that a number of brain regions, including the medial prefrontal gyrus from the DMN, the angular gyrus from the CEN, the anterior cingulate, the right medial temporal and the right middle frontal gyri from the SN displayed significantly decreased functional connectivity in APOE4 carriers. The medial prefrontal gyrus is a critical area of the DMN (Greicius et al., 2003), and plays a central role in a variety of cognitive functions, especially memory (Euston et al., 2012) and executive function (Dalley et al., 2004) that are vulnerable to cognitive aging and AD (Greicius et al., 2004; Burke and Barnes, 2006; Li et al., 2013). Various studies of the DMN in AD have repeatedly reported functional connectivity disruption in this region (Greicius et al., 2004; Rombouts et al., 2005; Qi et al., 2010; Wu et al., 2011; Wang et al., 2013). Recently, Song et al. (2015) also demonstrated APOE effect on the medial prefrontal regions in the DMN using seed-based functional connectivity analysis. The angular gyrus is functionally related to associative memory (Ben-Zvi et al., 2015), visuo-spatial attention (Cattaneo et al., 2009), and language ability (Bernal et al., 2015). Agosta et al. (2012) have reported decreased functional connectivity of the angular gyrus from the fronto-parietal CEN in AD. Disrupted functional connectivity of the SN was associated with cognitive and emotional deficits, and has been found in advanced aging and MCI patients (He et al., 2014; Uddin, 2015; Lu et al., 2016). Recently, Joo et al. (2016) and Wang et al. (2015) investigated the functional disruptions in these functional networks, and found that greater reductions of inter-network connectivity were associated with lower cognitive performance in different levels of cognitive impairment. Thus the result here indicated that the functional connectivity in the triple networks was different between the high- and low-risk groups, which may be related to the presence of APOE4 and a family history of dementia. We speculate that these AD-like functional connectivity disruptions in the triple network model may suggest risks of future cognitive decline or the progression to MCI or AD for the APOE4 carriers.

In contrast with the decreased functional activation compared with the low-risk group, we found that the high-risk group also showed increased functional activation in the frontal gyrus, parietal lobe, temporal gyrus and the cerebellum. It is consistent with several recent neuroimaging studies of APOE effects on brain connectivity. For example, Machulda et al. (2011) found increased SN connectivity by calculating the functional connectivity of the anterior cingulate seed in APOE4 carriers. Westlye et al. (2011) and Song et al. (2015) demonstrated increased DMN synchronization in APOE4 carriers. Similarly in AD patients, increased functional activation compared with that in healthy controls has also been reported (Wang et al., 2007; Qi et al., 2010; Zhou et al., 2010; Li et al., 2013). These increases have been usually interpreted as a compensatory reallocation or recruitment of brain resources (Cabeza et al., 2002), which may be a protective factor to keep retain a normal cognitive level in individuals at high risk for AD.

We also employed a BN approach to model and compare the effective connectivity patterns between the CEN, SN and DMN in the low- and high-risk groups. The BN learning approach revealed same-directed connections and network features in these two groups; the SN node does not receive but only generates connections to CEN and DMN. The BNbased directed connectivity pattern in both groups is consistent with the triple network model of major psychopathology suggested by Menon (2011), in which the information transfer occurs only from the SN to the CEN and DMN. It is also consistent with the study of Uddin et al. (2011), in which they employed Granger causality analyses to model the effective connectivity of the triple network with development, and found consistently that the fronto-insular cortex in the SN significantly influence the functional activity of regions in the DMN and CEN. Moreover, a recent study of Liang et al. (2015) demonstrated that the topological organization of the triple network changes with cognitive task loads. By comparing the effective connectivity coefficients between these two risk groups via the random permutation test, however, we found no significant difference in the directed connectivity of the three networks between the low- and high-risk groups. It suggested that although the APOE4 carriers might demonstrate AD-like functional connectivity changes in each of the three networks, the interactions between them could retain a normal process as in non-APOE4 carriers. This interesting finding may be explained first by the methodological difference. The functional connectivity stresses the temporal correlation between different regions, while the effective connectivity refers explicitly to the causal influence that one system exerts over another (Friston, 2011), which is in accordance with the inherent meaning of the triple network model. Second, the BN-based directed connectivity reflects how these three networks in the model cooperate with each other to execute tasks. It essentially demonstrated an organizational architecture of these

#### REFERENCES


functional networks. We propose that the stable effective connectivity architecture of the triple networks may be a crucial factor, together with the increased within-network functional connectivity, that enables individuals at high risk for AD to retain a normal cognitive level. Finally, it might be related to the complexity of brain network itself in response to the APOE4 effect. We speculate that the within-network regional connectivity alterations might emerge earlier than betweennetwork changes, and the further deterioration of withinnetwork connectivity may gradually lead to disruptions in interactions between networks for the APOE4 carriers. For example, Zhu et al. (2016) recently reported more changes of within-network connectivity than between-network connectivity in AD and MCI. Further studies would be required to investigate the dynamic changes of the directed connectivity architecture of the triple networks in APOE4 carriers through a longitudinal study.

In summary, we have explored the functional connectivity and effective connectivity of the three networks included in the large-scale triple network model in individuals with low and high risk for AD. The results demonstrated aberrant withinnetwork functional connectivity that suggests future risk of cognitive decline or progression to AD, and preserved betweennetwork effective connectivity that may support their current preserved cognition in the cognitively normal individuals who have a family history of AD and at least one copy of the APOE4 allele.

#### AUTHOR CONTRIBUTIONS

XW, LY and RL: designed and wrote the article; KC, ASF and EMR: carried out the experiment and collected the data; QL and XY: analyzed the data; XG and JZ: participated in the discussion and criticized the manuscript.

#### ACKNOWLEDGMENTS

This work was supported by the National Institute on Aging (k23 AG024062), the Funds for International Cooperation and Exchange of the National Natural Science Foundation of China (61210001), the General Program of National Natural Science Foundation of China (61571047, 31200847), and the Key Laboratory of Mental Health, Institute of Psychology, Chinese Academy of Sciences (KLMH2014ZG03, KLMH2015G06).


in a Japanese community. Int. J. Geriatr. Psychiatry 24, 1119–1126. doi: 10. 1002/gps.2234


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Wu, Li, Yu, Chen, Fleisher, Guo, Zhang, Reiman, Yao and Li. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution and reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Patterns of Longitudinal Neural Activity Linked to Different Cognitive Profiles in Parkinson's Disease

Atsuko Nagano-Saito1 †, Mohamed S. Al-Azzawi 1 †, Alexandru Hanganu2, 3 , Clotilde Degroot <sup>1</sup> , Béatriz Mejia-Constain<sup>1</sup> , Christophe Bedetti <sup>1</sup> , Anne-Louise Lafontaine<sup>4</sup> , Valérie Soland<sup>5</sup> , Sylvain Chouinard<sup>5</sup> and Oury Monchi 1, 2, 3, 4 \*

<sup>1</sup> Centre de Recherche de l'Institut Universitaire de Gériatrie de Montréal, Montréal, QC, Canada, <sup>2</sup> Departments of Clinical Neurosciences and Radiology, University of Calgary, Calgary, AB, Canada, <sup>3</sup> Hotchkiss Brain Institute, Calgary, AB, Canada, <sup>4</sup> Department of Neurology and Neurosurgery, McGill University, Montreal, QC, Canada, <sup>5</sup> Centre Hospitalier de l'Université de Montréal, Montréal, QC, Canada

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Claire O'Callaghan, University of Cambridge, UK Panteleimon Giannakopoulos, University of Geneva, Switzerland David Eidelberg, Northwell Health, USA

\*Correspondence:

Oury Monchi Oury.Monchi@ucalgary.ca † These authors have contributed equally to this work

Received: 19 August 2016 Accepted: 04 November 2016 Published: 23 November 2016

#### Citation:

Nagano-Saito A, Al-Azzawi MS, Hanganu A, Degroot C, Mejia-Constain B, Bedetti C, Lafontaine A-L, Soland V, Chouinard S and Monchi O (2016) Patterns of Longitudinal Neural Activity Linked to Different Cognitive Profiles in Parkinson's Disease. Front. Aging Neurosci. 8:275. doi: 10.3389/fnagi.2016.00275 Mild cognitive impairment in Parkinson's disease (PD) has been linked with functional brain changes. Previously, using functional magnetic resonance imaging (fMRI), we reported reduced cortico-striatal activity in patients with PD who also had mild cognitive impairment (MCI) vs. those who did not (non-MCI). We followed up these patients to investigate the longitudinal effect on the neural activity. Twenty-four non-demented patients with Parkinson's disease (non-MCI: 12, MCI: 12) were included in the study. Each participant underwent two fMRIs while performing the Wisconsin Card Sorting Task 20 months apart. The non-MCI patients recruited the usual cognitive corticostriatal loop at the first and second sessions (Time 1 and Time 2, respectively). However, decreased activity was observed in the cerebellum and occipital area and increased activity was observed in the medial prefrontal cortex and parietal lobe during planning set-shift at Time 2. Increased activity in the precuneus was also demonstrated while executing set-shifts at Time 2. The MCI patients revealed more activity in the frontal, parietal and occipital lobes during planning set-shifts, and in the parietal and occipital lobes, precuneus, and cerebellum, during executing set-shift at Time 2. Analysis regrouping of both groups of PD patients revealed that hippocampal and thalamic activity at Time 1 was associated with less cognitive decline over time. Our results reveal that functional alteration along the time-points differed between the non-MCI and MCI patients. They also underline the importance of preserving thalamic and hippocampal function with respect to cognitive decline over time.

Keywords: Parkinson's disease, functional magnetic resonance image, Wisconsin Card Sorting Task, mild cognitive impairment, longitudinal study

### INTRODUCTION

In Parkinson's disease (PD), cognitive deficits are frequently present even early in the course of disease development (Foltynie et al., 2004). Mild cognitive impairment (MCI) has been conceptualized as a stage when cognitive deficits don't impede on daily activities (Litvan et al., 2012). The prevalence of MCI in the early stages of PD is estimated to be between 25 and 40% (Dalrymple-Alford et al., 2011). The evolution of cognitive deficits in PD remain poorly understood and medication treatment of cognitive deficits in PD yields very modest results (Goldman and Holden, 2014). Using functional Magnetic Resonance Imaging (fMRI) we recently studied the effect of MCI in PD on the patterns of fronto-striatal activations while performing the Wisconsin Card Sorting Task (WCST). When planning a set-shift in our WCST PD patients without MCI (PD non-MCI) revealed patterns of activation similar to healthy individuals in our previous studies (Monchi et al., 2004, 2007), with significant activation in the ventrolateral prefrontal cortex (PFC) and caudate nucleus. In contrast, PD patients with MCI (PD-MCI) had no significant activation in these regions (Nagano-Saito et al., 2014). Similar results have been reported by other studies that analyzed resting state networks (Baggio et al., 2015) or fMRI in PD-MCI and PD non-MCI patients while performing a working memory task (Lewis et al., 2003; Ekman et al., 2012; Nagano-Saito et al., 2014), suggesting that PD non-MCI have compensational patterns that allow them to maintain the cognitive function at the same level as healthy individuals, while MCI in PD is associated with the loss of compensational patterns and development of specific functional brain abnormalities in the cognitive cortico-striatal loops with decreased cortico-cortical and cortico-subcortical connectivity compared to PD non-MCI, including the caudate nucleus and the PFC. This hypothesis has been suggested by a previous longitudinal study using Positron Emission Tomography in PD patients while performing a visual sequence learning-task (Carbon et al., 2010). They observed that while hippocampus was not significantly solicited in healthy individuals, the PD patients who did decline in their performance between the baseline and repeated measurement, showed significantly increased regional cerebral blood flow in the hippocampus. In order to confirm whether functional patterns from our previous study are specific for MCI in PD, a longitudinal fMRI study has to be performed, that would help us understand the functional changes in these loops over time.

The aim of the present study was to use fMRI to longitudinally follow up the brain neural activity of non-demented PD patients with and without MCI while they were performing the WCST. We also aimed to find out which patterns of neural activity at Time 1 would be the most predictive of cognitive decline over time. We applied our previously developed WCST fMRI protocol (Monchi et al., 2001, 2004) in PD-MCI and PD non-MCI patients at two time points. We expected to find preserved activation of the cognitive cortico-striatal loop usually associated with planning a set-shift (Monchi et al., 2001, 2007; Nagano-Saito et al., 2014) in the PD non-MCI group over time. By contrast, we expected that PD-MCI patients would show a modification over time of activity beyond the cortico-striatal loops. Based on our previous studies showing that the hippocampus is one key region involved in cognitive function in PD patients (Nagano-Saito et al., 2004, 2005, 2014), we also predicted that recruitment of hippocampus activity at Time 1 would correlate with preserved cognition at Time 2.

#### MATERIALS AND METHODS

#### Subjects

Twenty-seven non-demented PD participants at stages I and II of Hoehn and Yahr were recruited for the study. All participants were assessed by the movement disorders neurologists (A-LL., SC., VS.) and diagnosed as PD with the UK brain bank criteria for idiopathic PD (Hughes et al., 1992). All patients were responsive to dopamine medication. Patients with other concurrent major neurological or psychiatric conditions were excluded. The demographic information of the patients is given in **Table 1**. All patients provided written informed consent, which was approved by the Research Ethics Committee of the Regroupement Neuroimagerie Québec.

Participants were studied twice at 19.8 ± 2.7 months apart. In each session (at baseline Time 1 and follow-up at Time 2) they underwent a comprehensive neuropsychological assessment and fMRI during which they performed a computerized version of WCST (Monchi et al., 2001, 2004). Participants were asked not to take any dopaminergic medication at least 12 h prior to the sessions. Based on the neuropsychological assessment, participants were divided into two groups: those with MCI and those cognitively intact (non-MCI) at Time 1. Three patients with non-MCI at Time 1 turned into MCI at Time 2. Therefore, for the analyses to investigate the group difference, 24 subjects (mean age, 60.33 ± 5.8 years; 11 males and 13 females; 12 MCI, and 12 non-MCI) were included. For group-combined analyses, all 27 subjects (mean age 60.3 ± 5.5 years; 12 males and 15 females) were included.



The p-values indicate the group difference based on chi-squared test (sex and handedness) or student t-test (others). Legend: MoCA, Montreal cognitive assessment; UPDRS, Unified Parkinson's disease Rating Scale; BDI, Beck Depressive Inventory II. Individual MoCA scores are presented in Supplementary Table 3.

Inclusion criteria for MCI were based on the Movement Disorder Society Task Force guidelines for Parkinson's disease (Level I and II) (Litvan et al., 2012), based on five cognitive domains (Supplementary Table 1) and were the same as our previous study (Nagano-Saito et al., 2014). Objective evidence of cognitive decline was set with performance >1.5 standard deviations below standardized mean of the same age-group on two or more subtests within a cognitive domain.

Demographically, no significant differences were observed between the groups with respect to age and the motor section of the Unified Parkinson's Disease Rating Scale at Time 1 and 2. Group demographic characteristics are listed in detail in **Table 1**.

#### Neuropsychological Assessment

A screening test, the Montreal Cognitive Assessment (MoCA) (Nasreddine et al., 2005), was administered at the beginning of each scanning session. The same comprehensive neuropsychological battery previously used in Jubault et al. (2009), Hanganu et al. (2013), Nagano-Saito et al. (2014), was performed by a licensed neuropsychologist (Dr. BMC) to assess five domains of cognition, attention and working memory, executive functions, language, memory and visuospatial functions. Details of the tasks used are given in Supplementary Table 1.

### Cognitive Task during fMRI

A computerized version of the WCST (Monchi et al., 2001, 2004) was administered using stimulus presentation software. Participants were fully trained on the task prior to the scanning sessions. On each trial of the task, participants were asked to match a new test card to one of the four fixed reference cards based either on the color, shape, or the number of the stimuli in each reference card.

In WCST the classification rule was not given to the participant. Instead, s/he had to find it using the feedback (positive or negative) that followed each trial. On each experimental trial, participants had to find the proper classification rule, and apply it as long as a positive feedback preceded their response. A bright screen indicated a correct classification and a dark screen indicated an incorrect classification. On each control trial, the test card was identical to one of the four reference cards, therefore participants only had to select the twin reference card. On the control trials, the screen maintain its original brightness throughout the feedback period.

The first period of each trial started with the presentation of a new test card, at which point the participant choose one of the four reference cards. Response time was measured for each selection. The second period of each trial started as soon as the subject made a selection and consisted of feedback conveyed through a change in screen brightness lasting 2.3 s.

Each functional MRI run contained blocks of each of the four trial classifications (color, shape, number, and control) presented in random order. In WCST trial blocks, six consecutive correct matching responses were required before a change in classification rule could occur. Control blocks contained eight trials.

To evaluate the pattern of activation during the different stages of the WCST, four experimental and two control time periods were defined as follows: (1) Receiving negative feedback (RNF): the screen darkens indicating an incorrect response: a set-shift is therefore required and must be planned; (2) Matching after negative feedback (MNF): execution of the set-shift; (3) Receiving positive feedback (RPF): the screen brightens indicating a correct response: the current matching criterion must continue; (4) Matching after positive feedback (MPF): selection using the same classification rule as the previous trial; (5) Receiving control feedback (RCF): original screen brightness is maintained; (6) Matching with control feedback (MCF): select reference card identical to test card.

### fMRI Scanning

Participants were scanned using the Siemens Tim Trio 3.0 T scanner at the Unité de Neuroimagerie Fonctionnelle of the Centre de Recherche de l'Institut Universitaire de Gériatrie de Montréal. Sessions began with high-resolution, T1-weighted, 3D volume acquisition for anatomical localization with resolution of 1 × 1 × 1 mm, followed by echoplanar T2<sup>∗</sup> -weighted image acquisitions with blood oxygenation level-dependent (BOLD) contrast (echo time 30 ms; flip angle 90◦ ; matrix size, 64 × 64 pixels; voxel size, 3.7 × 3.7 × 3.7 mm<sup>3</sup> ). Functional images were acquired over five runs in a single session. Volumes were acquired continuously every 2.5 s, for a total 155 volumes within runs, and contained 36 slices.

#### MRI Data Analysis Contrast Analyses

As was previously done with the same fMRI task, a General Linear Model data analysis was performed using fmristat (Monchi et al., 2001, 2004; Worsley et al., 2002; Jubault et al., 2009; Nagano-Saito et al., 2014). Briefly, the following contrasts were computed for each subject: 1. RNF vs. RPF: reflecting the planning of the set-shift, 2. MNF vs. MPF: reflecting the execution of the set-shift, 3. RPF vs. RCF: reflecting the maintaining of the set-shift, and 4. MPF vs. MCF: reflecting the matching according to the same rule. Next, the results of each run for each subject, were non-linearly transformed into standard proportional stereotaxic space (ICBM152 template) using anatomical MRI to template transformation parameters (Collins et al., 1994; Zijdenbos et al., 2002). In the second step, runs and subjects were combined using a mixed-effects linear model (Worsley et al., 2002). Both t-maps of inter-group (MCI and non-MCI of each time point) and within-group comparison (Time 1 vs. Time 2 of each group and combined groups) were generated. Statistical maps threshold was set at p < 0.05 correcting for multiple comparisons, yielding a threshold of t > 4.82 for a single voxel. Predicted peaks reaching p < 0.0001 (t > 3.87) uncorrected with a cluster size >40 mm<sup>3</sup> assessed on the spatial extent of contiguous voxels are also reported and identified with an asterisk (∗) in the tables. A region was predicted if it had been identified in our previous work using this task (Monchi et al., 2001, 2004; Jubault et al., 2009; Nagano-Saito et al., 2014) For within-group comparisons (Time 1 vs. Time 2), statistical maps threshold was set at p < 0.0001 (t > 3.87) uncorrected with a cluster size >40 mm<sup>3</sup> . Predicted peaks reaching p < 0.001 (t > 3.18) uncorrected with a cluster size >40 mm<sup>3</sup> are reported.

Additionally, in order to evaluate the common neural regions solicited by both groups (PD-MCI and PD non-MCI), we performed a conjunction analysis using a conjunction null hypothesis with a threshold of p < 0.0001 for each group.

#### Correlation Analysis

We wanted to separately address how individual cognitive performance ability affects the patterns of brain activity during the various stages of the WCST. To do this, we performed correlation analyses on the BOLD data while performing the WCST using cognitive scores of MoCA (Nasreddine et al., 2005) at the subject level. Secondly, in order to find which patterns of activity would be predictive of cognitive decline over time we correlated the BOLD data while performing the WCST at Time 1 with the difference of the MoCA scores between Time 1 and Time 2 at the subject level (Nasreddine et al., 2005). These analyses were performed across all participants combined (i.e., both MCI and non-MCI PD patients grouped). Finally, in order to look whether patterns of activity are associated with significant changes in performance of the WCST task, we correlated the BOLD data with the WCST accuracy and the WCST accuracy over time. These analyses were performed across all participants combined (i.e., both MCI and non-MCI PD patients grouped), at Time 1 and Time 2, separately.

Statistical maps threshold was set at p < 0.0001 (t > 3.87) uncorrected with a cluster size >40 mm<sup>3</sup> . Predicted peaks reaching p < 0.001 (t > 3.18) uncorrected are reported. A region was predicted if it had been identified in our previous work using the WCST. We especially expected involvement of the striatum, the thalamus, the hippocampus, and/or the PFC, because the activation of those regions was correlated with cognitive scores (Nagano-Saito et al., 2014).

#### RESULTS

#### General Cognitive Scale Measured by MoCA

The mean of MoCA is shown in the **Table 1**. A mixed-design repeated measures ANOVA (time × group) indicated main effect of group (F = 9.1, p = 0.006) but no effect of time and time × group interaction (p > 0.1).

#### Performance of WCST

The error rates and reaction times for the task are shown in the **Table 2**. A mixed-design repeated measures ANOVA for the error rate and reaction time (task × time × group) indicated a main effect of task (F = 98.8, p < 0.001). No other main effects (including the group effect) or interaction was observed (p > 0.1). The mean reaction times are shown in **Table 2**. A mixed-design repeated measures ANOVA for the reaction time in control, after positive and after negative feedbacks (task × time × group) indicated a main effect of task (F = 29.9, p < 0.001), and time (F = 4.70, p = 0.041). A trend of interaction was observed for the time × task × group (F = 3.0, p = 0.059). No other significant effect was observed (p > 0.1).

#### Imaging Analysis

Our previous study at Time 1 indicated no significant difference of other contrasts in the intergroup comparison. Thus, we selectively reported the comparison RNF vs. RPF (planning the set shift) and MNF vs. MPF (executing the set-shift), because our previous study at Time 1 indicated no significant difference of other contrasts in the intergroup comparison (Nagano-Saito et al., 2014).

#### Planning the Set-Shift

Similar to our previous study (Nagano-Saito et al., 2014), when planning the set shift at Time 1, the PD non-MCI group demonstrated significant activation in the PFC, precuneus, parietal cortex, occipital cortex and cerebellum (**Table 3**). Activation was also observed in the caudate when a lower threshold was used (p < 0.001 uncorrected). At Time 2, this group demonstrated significant activations in the posterior PFC, ventrolateral PFC, medial PFC, occipital, and parietal cortices. Activation was again observed in the caudate when a lower threshold was used (p < 0.001 uncorrected). Longitudinal comparison (Time 1 vs. Time 2) indicated decreased activity at Time 2 in the occipital area and cerebellum hemisphere, and increased activity at Time 2 in the right parietal lobe and the posterior prefrontal frontal cortex (Supplementary Table 2).

By contrast, the PD-MCI group demonstrated significant activation peaks in the dorsolateral PFC, posterior PFC, and the occipital cortex. At Time 2, PD-MCI revealed significant activation peaks in the dorsolateral PFC, ventrolateral PFC, medial PFC, posterior PFC, posterior parietal cortex, and visual occipital area (**Table 3**). Longitudinal analyses showed decreased


#### TABLE 3 | Significant activations in the contrasts for RNF minus RPF, and MNF minus MPF.


(Continued)

#### TABLE 3 | Continued


Legend. Results presented at t > 4.82; p < 0.05 corrected for multiple comparisons. \*predicted regions (t > 3.87; p < 0.0001). \*\*trends (t > 3.18; p < 0.001). t, t-value; L, left; R, right; DLPFC, dorsolateral prefrontal cortex (BA 46, 9/46); pPFC, posterior PFC (BA 6, 8, 44); VLPFC, ventrolateral prefrontal cortex and insula (BA 47/12/13); aPFC, anterior prefrontal cortex (BA 10); mPFC, medial prefrontal cortex (BA 6, 8, 32); Parietal, parietal cortex (BA 40, 7); Precuneus, precuneus cortex (BA 40, 7); Occipital, occipital or striate/ extrastriate cortices (BA 17, 18, 19); PMC, premotor area (BA 6).

activity at Time 2 in the cerebellum and increased activity at Time 2 in the visual occipital cortex (Supplementary Table 2).

Conjunction analysis over both groups revealed significant activations in the posterior PFC and occipital cortex at Time 1, as well as significant activation in the ventrolateral PFC, posterior PFC and occipital cortex at Time 2 (**Table 2**, RNF-RPF).

The localisation of the observed peaks are shown in **Figure 1A**. The results of the longitudinal comparison (Time 1 vs. Time 2) with all the subjects (n = 27) is also shown in the Supplementary Table 2.

#### Executing the Set-Shift

When executing the set-shift, the non-MCI group at Time 1 revealed significant activations in the dorsolateral PFC, premotor cortex, ventrolateral PFC, medial PFC, anterior PFC, parietal cortex, precuneus, the visual occipital areas, and cerebellum (**Table 3**). The pattern was similar to our previous study (Nagano-Saito et al., 2014). At Time 2, the PD non-MCI group revealed significant activation in the dorsolateral PFC, medial PFC, anterior PFC, parietal cortex bilaterally, precuneus and occipital visual area (**Table 3**). Longitudinally, there was reduced activation in the cerebellum and increased activity in the precuneus at Time 2 compared to Time 1 (Supplementary Table 2).

By contrast, the PD-MCI group demonstrated activation peaks in the middle occipital cortex, occipital cortex and cerebellum at Time 1, while at Time 2, positive significant activations were revealed in the parietal lobe, precuneus, visual occipital areas and cerebellum (**Table 3**). Longitudinally there were no significant differences between Time 1 and Time 2 in the PD-MCI group (Supplementary Table 2).

Conjunction analysis over both groups revealed significant activations in the anterior PFC at Time 1 and activations in the parietal cortex and precuneus at Time 2 (**Table 2**, MNF-MPF).

The localisation of the observed peaks are shown in **Figure 1B**. The results of the longitudinal comparison (Time 1 vs. Time 2) with all the subjects (n = 27) is also shown in the Supplementary Table 2.

#### Correlation Analysis

When planning the set-shift, no significant correlation was observed with the MoCA scores at Time 1. A significantly positive correlation was observed in the medial PFC at Time 2, overlapping with the more activated area at Time 2 (**Table 4**). When executing the set-shift, significant correlation was observed in the striatum and the left thalamus at Time 1 as well as with the precuneus at Time 2.

When we added the MoCA decline across the time points as a confounder, significant negative correlations were observed in the internal capsule, thalamus, and hippocampus during planning the set-shift (**Table 4**, **Figure 2**). During executing the set-shift, significant negative correlations were observed in the medial PFC (**Table 4**). They indicate that less cognitive decline occurs as more of these regions are solicited.

Correlations with WCST accuracy were also depicted. Planning the set-shift showed significant correlations with the ventral striatum, ventrolateral PFC and occipital area at Time 1, as well as with ventrolateral PFC, premotor cortex, parietal, and caudate at Time 2. Executing the set-shift showed significant




Legend. Results presented at t > 3.87; p < 0.0001, uncorrected; \*indicates predicted regions (t > 3.18; p < 0.001 uncorrected). t, t-value; L, left; R, right; DLPFC, dorsolateral prefrontal cortex (BA 46, 9/46); VLPFC, ventrolateral prefrontal cortex and insula (BA 47/12/13); aPFC, anterior prefrontal cortex (BA 10); mPFC, medial prefrontal cortex (BA 6, 8, 32); Parietal, parietal cortex (BA 40, 7); Precuneus, precuneus cortex (BA 40, 7); Occipital, occipital or striate/ extrastriate cortices (BA 17, 18, 19); PMC, premotor area (BA 6).

correlations between the WCST accuracy and activations in the medial PFC, dorsolateral PFC, caudate, hippocampus at Time 1, but no correlation were shown at Time 2.

When WCST accuracy over time was correlated with BOLD at Time 1, planning the set-shift showed significant correlations with the anterior PFC, while executing the set-shift showed correlations with dorsolateral PFC, medial PFC, occipital cortex, parietal, precuneus, and hippocampus, indicating that greater activity in these regions at Time 1 is indicative of preserved performance on the WCST over time.

#### DISCUSSION

We longitudinally followed up the brain neural activation of non-demented PD patients undergoing an fMRI session while performing the WCST with different levels of cognitive impairment. As predicted, the PD non-MCI and PD-MCI groups were different in their patterns of activity-change over time when performing the WCST.

#### PD Non-MCI Patterns over Time Patterns of Preserved Activity

At both time points, significant activation was observed in the frontal and parietal cortex, with a trend in the caudate, during planning the set-shift. Fronto-striatal activation in the cognitive loop has been reported to be preserved in non-MCI PD patients, but affected in MCI patients (Lewis et al., 2003; Ekman et al., 2012; Nagano-Saito et al., 2014: **Table 3**). Thus, our observation may indicate relatively preserved function in the cognitive loop at Time 2 in the non-MCI PD patients. This is in agreement with another longitudinal study showing stable activity across timepoints in non-MCI patients with the n-back working memory task (Ekman et al., 2014). During the execution of the setshift, significant activation was also preserved in the frontal and parietal cortex (**Table 3**). However, overall, the activation was relatively weak at Time 2, and intra-group comparison did not reach significance. Considering the fact that the activated regions during this period could correspond to the motor-loop (Monchi et al., 2004, 2007) this relatively weak activation may reflect the

disease progress affecting more motor-related function in non-MCI PD patients. Actually, slower RT was observed at Time 2 in non-MCI PD patients (**Table 2**), without a change in accuracy in the performance of the WCST.

#### Regions with Decreased Patterns of Activity over Time

Activation in the occipital visual cortex during planning the set-shift and the cerebellum during planning and executing the set-shift, was reduced during the time course, between Time 1 and Time 2 (Supplementary Table 2). It has been reported in non-demented PD patients that lowered regional cerebral metabolic rates in the occipital area correlate with the motor dysfunction (Bohnen et al., 1999) We also previously reported lower regional cerebral metabolic rates for glucose in the occipital area accompanied by dopaminergic availability in the striatum, along with motor dysfunction in non-demented PD patients (Nagano-Saito et al., 2004) Simultaneously, cerebellar activity is likely to contribute to pathophysiological change underlying PD (Martinu and Monchi, 2013) Thus, the reduced activation in the occipital area and cerebellum may reflect remote effects of the progress of the disease on dopamine projections in the striatum. However, they might also reflect a learning effect with reduction of visual attention and motor effort at Time 2. A longitudinal study with healthy volunteers would help to understand these observations.

#### PD-MCI Patterns over Time Patterns of Increased Activity

During planning the set-shift, PD-MCI revealed significantly increased activation in the frontal, parietal and occipital areas at Time 2 (**Table 3**). These regions were overlapped with the activation pattern of PD non-MCI. PD patients have been reported to often have fluctuations in non-motor functions (Witjas et al., 2002) This might explain the relatively higher performance at Time 2 observed in this group of patients. Actually, the error rate was smaller with MCI patients across time, although it did not reach significance. Nevertheless, RT became slower at Time 2. This may indicate that the relay of information from the cognitive loop to motor loop could not be recovered.

During executing the set-shift, more activation was observed in the parietal areas, precuneus and cerebellum, but not in the frontal regions at Time 2 (**Table 4**). Thus, the recovered activation is not likely to be supported by the cortico-striatal loops, but by alternative brain circuit.

### Medial PFC and Precuneus Patterns of Activation in Both Groups

We observed significant activation in the medial PFC during planning the set-shift in both groups and in the conjunction analyses only at Time 2 (**Table 3**). Moreover, in the group of all patients during planning a set-shift, the activity in the medial PFC showed a positive correlation with the MoCA scores at Time 2 and preserved WCST accuracy score over time (**Table 4**). In addition, when executing the set-shift, increased activity in the medial PFC correlated with preserved cognition over time (as measured by the MoCA), with WCST accuracy at Time 1 and with preserved WCST accuracy score over time (**Table 4**). The mPFC is considered to play an important role in learning association between events and in linking adaptive responses (Euston et al., 2012). The medial PFC activates both during externally guided and internally guided decision making tasks (Nakao et al., 2012). The recovery of the mPFC activity at Time 2 could therefore help in relating the cognitive processes required to select a response to the actual motor selection.

Another region that was recruited during executing the setshift at Time 1 in PD non-MCI only and at Time 2 in both groups which was confirmed by the conjunction analyses at Time 2 (**Table 3**) was the precuneus. In the all patients group the strength of the activation in the precuneus during executing the set-shift showed a significantly positive correlation with the MoCA scores at Time 2 and the preserved WCST accuracy over time (**Table 4**). This shows the importance of this region's function for cognition. The precuneus, as a part of the parietal cortex (Vogt et al., 2001), functionally connects to the medial PFC and superior frontal cortex (Laird et al., 2009; Margulies et al., 2009), and is considered to be involved in reaching to visual targets (Bernier and Grafton, 2010). More generally, precuneus and medial PFC are components of the dorsomedial motor stream. (Rizzolatti and Matelli, 2003; Binkofski and Buxbaum, 2013) Interestingly, in PD non-MCI the posterior PFC, which is considered as a part of the dorsomedial motor stream, was higher at Time 2, compared to Time 1, although the more anterior PFC showed tendency of decreasing of activity. Previous studies reported decreased resting state connectivity in bilateral PFC (as part of the dorsomedial motor network) and fronto-parietal areas in PD-MCI patients (Amboni et al., 2015; Baggio et al., 2015), along with a ordered connectivity reduction (HC > PD, non-MCI > PD-MCI). Thus, our results for the precunes and the medial PFC activation may reflect increased recruitment of the dorsomedial motor stream, compensating the cortico-striatal loop, and allowing a better performance as shown by the correlations with WCST accuracy.

### Patterns of Activity Predictive of Cognitive Evolution

When planning the set-shift, increased activity in the thalamus and the hippocampus correlated with preserved cognition over time as measured by the MoCA (**Table 4**, **Figure 2**). Furthermore, when executing the set-shift the increased activity in the hippocampus correlated with preserved WCST performance over time (**Table 4**).

We recently reported that cortical thinning occurs significantly faster overtime in the medial temporal lobe in PD-MCI vs. PD non-MCI, suggesting that temporal lobe atrophy could be used as a predictor of dementia in PD (Hanganu et al., 2014). Carbon et al. (2010), performed a longitudinal PET study in PD patients and controls, where participants performed a visual sequence-learning task (Carbon et al., 2010). The researchers observed that while the hippocampus was not significantly solicited in controls during the task, the PD patients whose performance did not decline between the two time points (i.e., the best performers) showed significant increase in rCBF in the hippocampus over time, as compared to the lower performers. They suggested that a compensatory hippocampal activation response may be a specific functional indicator of incipient cognitive decline in non-demented PD patients at early disease stages. Our present results are in agreement with this notion.

We have observed decreased thalamus activity in PD vs. controls in the context of two different set-shifting tasks (Monchi et al., 2001, 2007) hence, the thalamus activity level correlating with preserved cognition may reflect the importance of the preserved function of cognitive cortico-striatal loop. Furthermore, connectivity between the hippocampus and the posterior parts of the thalamus has been observed in animals (Herkenham, 1978; Wouterlood et al., 1990; Vertes et al., 2006) and in humans (Behrens et al., 2003). Our previous fMRI study, which examined the same task in healthy volunteers, showed that hippocampal deactivation was less accompanied by extensive activation in the thalamus during a set-shift, when dopamine was lowered (Nagano-Saito et al., 2008). A recent voxel based morphometry study reported reduced thalamic and hippocampal gray matter intensity in PD-MCI vs. PD non-MCI (Chen et al., 2016). Thus, the functional integrity of the thalamus, possibly connecting to the hippocampus, might be an important marker of cognitive preservation in PD.

### CONCLUSION

In conclusion, our results show differential neural changes in PD-MCI patients as compared with the PD non-MCI group over time. They indicate that, as long as the cognitive cortico-striatal loops are preserved, the circuit seems to be recruited. However, when the cortico-striatal loops cannot perform the demanded functional task, the extra circuits, including the mPFC, and the precuneus, could be recruited. Moreover, hippocampal compensation is used to maintain cognitive abilities over time. More work in this area is warranted to determine the precise phenotypes of such sub-groups of patients.

#### AUTHOR CONTRIBUTIONS

Research project: Conception: OM. Organization: BM, CB, AL, VS, and SC. Execution: CD, BM, AH, CB, AL, VS, and SC. Statistical Analysis: Design: OM and AN. Execution: MA and AN. Review and Critique: OM and AN. Manuscript: Writing of the first draft: MA and AN. Review and Critique: OM, AN, and AH.

### FUNDING

This work was supported by a Canadian Institutes of Health Research grant (MOP-81114), a psychosocial grant from the Parkinson Society Canada, the Canada Research Chair in nonmotor symptoms of Parkinson's disease, and the Tourmaline Oil Chair in Parkinson's disease to OM.

### REFERENCES


### ACKNOWLEDGMENTS

The authors would like to thank Holly Breton for thoroughly proof-reading the manuscript, all the participants for taking part in the study, as well as the Functional Neuroimaging Unit team of the Institut de Gériatrie de Montréal.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2016.00275/full#supplementary-material


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Nagano-Saito, Al-Azzawi, Hanganu, Degroot, Mejia-Constain, Bedetti, Lafontaine, Soland, Chouinard and Monchi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Global Efficiency of Structural Networks Mediates Cognitive Control in Mild Cognitive Impairment

Rok Berlot1,2, Claudia Metzler-Baddeley<sup>3</sup> , M. Arfan Ikram<sup>4</sup> , Derek K. Jones<sup>3</sup> and Michael J. O'Sullivan1,3,5 \*

<sup>1</sup> Division of Neuroscience, Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, UK, <sup>2</sup> Department of Neurology, University Medical Centre Ljubljana, Ljubljana, Slovenia, <sup>3</sup> Cardiff University Brain Research Imaging Centre (CUBRIC), School of Psychology, and the Neuroscience and Mental Health Research Institute, Cardiff University, Cardiff, UK, <sup>4</sup> Departments of Epidemiology, Radiology, Neurology, Erasmus MC, University Medical Center Rotterdam, Rotterdam, Netherlands, <sup>5</sup> Mater Centre for Neuroscience and Queensland Brain Institute, University of Queensland, Brisbane, QLD, Australia

Background: Cognitive control has been linked to both the microstructure of individual tracts and the structure of whole-brain networks, but their relative contributions in health and disease remain unclear.

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Veena A. Nair, University of Wisconsin−Madison, USA Stefano Delli Pizzi, University of Chieti-Pescara, Italy Roser Sala-Llonch, University of Oslo, Norway

> \*Correspondence: Michael J. O'Sullivan mike.osullivan@kcl.ac.uk

Received: 26 September 2016 Accepted: 21 November 2016 Published: 15 December 2016

#### Citation:

Berlot R, Metzler-Baddeley C, Ikram MA, Jones DK and O'Sullivan MJ (2016) Global Efficiency of Structural Networks Mediates Cognitive Control in Mild Cognitive Impairment. Front. Aging Neurosci. 8:292. doi: 10.3389/fnagi.2016.00292 Objective: To determine the contribution of both localized white matter tract damage and disruption of global network architecture to cognitive control, in older age and Mild Cognitive Impairment (MCI).

Materials and Methods: Twenty-five patients with MCI and 20 age, sex, and intelligence-matched healthy volunteers were investigated with 3 Tesla structural magnetic resonance imaging (MRI). Cognitive control and episodic memory were evaluated with established tests. Structural network graphs were constructed from diffusion MRI-based whole-brain tractography. Their global measures were calculated using graph theory. Regression models utilized both global network metrics and microstructure of specific connections, known to be critical for each domain, to predict cognitive scores.

Results: Global efficiency and the mean clustering coefficient of networks were reduced in MCI. Cognitive control was associated with global network topology. Episodic memory, in contrast, correlated with individual temporal tracts only. Relationships between cognitive control and network topology were attenuated by addition of single tract measures to regression models, consistent with a partial mediation effect. The mediation effect was stronger in MCI than healthy volunteers, explaining 23−36% of the effect of cingulum microstructure on cognitive control performance. Network clustering was a significant mediator in the relationship between tract microstructure and cognitive control in both groups.

Conclusion: The status of critical connections and large-scale network topology are both important for maintenance of cognitive control in MCI. Mediation via large-scale networks is more important in patients with MCI than healthy volunteers. This effect is domain-specific, and true for cognitive control but not for episodic memory. Interventions to improve cognitive control will need to address both dysfunction of local circuitry and global network architecture to be maximally effective.

Keywords: cognitive aging, cognitive control, mild cognitive impairment, tractography, neuroimaging, diffusion MRI, networks

#### INTRODUCTION

fnagi-08-00292 December 15, 2016 Time: 16:12 # 2

Cognitive or executive control describes the marshaling of cognitive resources in the face of complex or competing demands (Shenhav et al., 2013). Impairment of control is an important feature of dementia (Royall et al., 1998) and is associated with changes in brain structure. We have previously shown that alterations in a single portion of the anterior cingulum bundle predict variation of cognitive control in healthy older people (Metzler-Baddeley et al., 2012a). This observation fits with a key role for the dorsal anterior cingulate cortex (Shenhav et al., 2013). However, this is only one node of a widely distributed network that is activated by control tasks (Cole and Schneider, 2007). Alterations in brain structure occur at multiple levels with aging and early neurodegeneration. An alternative viewpoint, therefore, is that performance might depend on emergent properties of the whole network rather than any single tract. The relationship between alterations at the level of tracts and whole networks, and their relative contribution to cognitive performance in aging and neurologic disease, are not known.

Cognitive control and episodic memory have traditionally been associated with structures in the prefrontal cortex and medial temporal lobe, respectively (Alexander et al., 2007; Gläscher et al., 2012). This anatomical parcellation of function extends to key white matter connections. Cognitive control is exquisitely sensitive to microstructural differences in subsets of pathways within the cingulum bundle, including those likely to terminate in the dorsal anterior cingulate cortex (Metzler-Baddeley et al., 2012a). It is not, however, associated with variations in fornix microstructure, the principal correlate of verbal recall (Metzler-Baddeley et al., 2011). In Mild Cognitive Impairment (MCI), the prodromal stage of Alzheimer's disease, microstructure is compromised in the fornix and other temporal tracts and residual memory performance remains dependent on temporal lobe connections (Metzler-Baddeley et al., 2012b). Performance, therefore, has been linked with relative specificity to microstructure of white matter connections within relevant networks.

Graph theory provides a means to derive properties of the brain's global 'connectome', such as measures of efficiency of network structure and clustering of network nodes (Rubinov and Sporns, 2010). Global efficiency is inversely related to topological distance between nodes and is typically interpreted as a measure of the capacity for parallel information transfer and integrated processing (Bullmore and Sporns, 2012). The clustering coefficient is a measure more weighted to the local environment of each node, as it quantifies the extent to which neighboring nodes are connected to each other (Bullmore and Sporns, 2009). Reduced efficiency of network structure has been demonstrated in Alzheimer's disease and linked to performance in both memory and executive tasks (Lo et al., 2010; Reijmer et al., 2013). In MCI, similar alterations in structural network topology have been observed, though findings at this early stage of neurodegeneration are less consistent (Bai et al., 2012; Shu et al., 2012).

Previous neuroimaging studies have generally not considered both 'local' (nodes and connections) and 'global' (network topology) measures together. To date, diffusion MRI studies have tended to focus either on detailed tract reconstructions or whole-brain approaches. It remains unclear how microstructural changes in single tracts relate to global network topology, and how important such a pathway of effect might be in cognitive function and dysfunction. This is a particularly relevant question for cognitive control. The cingulate cortex and its connections harbor critical functional specialization, but the cingulum also provides a pathway of communication across large-scale networks whose topology might also relate to cognition.

The interplay between local tracts and global network properties – and the spatial scale of organization that is most relevant to performance – have important implications for treatment. Treatments based on noninvasive stimulation could target specific local alterations in function, or the restoration of more widespread patterns of network structure and function. For example, transcranial magnetic stimulation has been shown to normalize functional connectivity in depression (Liston et al., 2014), and transcranial direct current stimulation also influences resting-state networks (Peña-Gómez et al., 2012). This study combined investigation of critical tracts with global properties of structural networks. We determined whether network topology was altered in MCI and whether such alterations were an independent factor in cognitive performance. Mediation analyses were used to test the hypothesis that relationships between tract microstructure and cognition were mediated by alterations in global network topology.

#### MATERIALS AND METHODS

#### Participants

Twenty-five patients with MCI were recruited from the Cardiff Memory Clinic. Standardized assessment included clinical history, ascertainment of vascular risk status, neurological examination, basic hematology and biochemistry investigations, neuroimaging with CT or MRI and cognitive screening with the Addenbrooke's Cognitive Examination (Mioshi et al., 2006). Diagnosis of MCI was based on established current criteria (Albert et al., 2011). Objective memory impairment was confirmed by a score of >1.5 SDs below age-matched controls on

either the Addenbrooke's verbal memory subscore or the visual memory test from the Repeatable Battery for the Assessment of Neurological Status. All patients had a Mini-Mental State Examination score of ≥24 (mean 26, SD 1.7) and a Clinical Dementia Rating of 0.5. Seven patients had additional evidence of executive dysfunction (multidomain MCI), others had pure amnestic MCI. Consecutive patients, who were eligible and willing to take part, were recruited and assessed by a single neurologist (MJO).

The 20 healthy control participants were drawn from 46 individuals between the ages of 53 and 93 years, recruited for an aging study (Metzler-Baddeley et al., 2011). Among the 46 elderly participants, one withdrew and another did not complete the study due to ill health. One participant was excluded because of subsequent diagnosis of Parkinson's disease. Structural MRI scans (fluid-attenuated inversion recovery and T1-weighted) were inspected for overt pathology: three participants were excluded because of extensive white matter hyperintensities suggestive of significant cerebral small vessel disease (Fazekas grade 3) (Fazekas et al., 1993), and one participant was excluded due to severe motion artifact. From remaining 39 subjects, a matched control group was sampled. The control sample were matched for age and premorbid IQ using data from the National Adult Reading Test-Revised (NART-R), an accepted measure of premorbid IQ. Age and NART-R only were used to select this group and to prevent bias, selection was performed blind to cognitive, clinical and MRI data. Participants older than 65 years (the MCI group were all over 65) and with a verbal IQ not exceeding 2 SDs above the average patient IQ in the NART-R provided a matched sample of 20 healthy control participants.

Exclusion criteria for both groups were: a history of neurological disease or mental disorders (clinical disorders or acute medical conditions/physical disorders, as defined by DSM-IV-TR), including past history of moderate to severe head injury, prior or current drug or alcohol abuse, previous large-artery stroke or cerebral hemorrhage, known cervical, peripheral or coronary artery disease, structural heart disease or heart failure, and contraindications to MRI. Anxiety or antidepressant use was not an exclusion criterion, unless an individual met criteria for major depression. No patient with MCI met diagnostic criteria or had characteristic clinical features to suggest other degenerative disorders. An additional exclusion criterion for healthy participants was the past or current presence of subjective memory symptoms.

Ethical approval for the study was provided by the South East Wales Research Ethics Committee. All participants provided informed consent in accordance with the Declaration of Helsinki.

#### Cognitive Assessment

Neuropsychological assessment was performed over two 1.5 h testing sessions. Cognitive control was assessed with tasks that required the maintenance of a task set under speeded response conditions: attention switching was examined using alternation between letters and digits with a Verbal Trails Test. The Stroop Color-Word test was used to assess the suppression of response incongruent information (Trenerry et al., 1989). Verbal generation and fluency were measured with the verbal fluency tests from the D-KEFS for letters F, A, and S and for the categories of animals and boys' names. Motor planning skills based on spatial rules were assessed with the Tower of London test from the Delis and Kaplan Executive Function System battery (D-KEFS). The Digit Symbol Substitution test from the WAIS-III provided a measure of focused attention and psychomotor performance.

Free recall was assessed with the Free and Cued Selective Reminding Test (Grober et al., 1997). Additionally, the face recognition test from the Camden Recognition Memory Test (CRMT) was performed.

#### MRI Acquisition

Diffusion-weighted MRI data were acquired using a 3T GE HDx MRI system (General Electric) with a twice-refocused spinecho echo planar imaging sequence, providing whole oblique axial (parallel to the commissural plane) brain coverage (60 slices, 2.4 mm thickness, field of view 23 cm, acquisition matrix 96 × 96). Acquisition was peripherally gated to the cardiac cycle. TE (echo delay time) was 87 ms and parallel imaging (array spatial sensitivity encoding (ASSET) factor 2) was used. The b-value was 1,200 s/mm<sup>2</sup> . Data were acquired with diffusion encoded along 30 isotropically distributed directions and 3 non-diffusion-weighted scans, according to an optimized gradient vector scheme (Jones et al., 1999). Acquisition time was approximately 13 min.

T1-weighted structural MRI data were acquired using a 3D fast spoiled gradient recalled (FSPGR) echo sequence (matrix of 256 × 256 × 176, field of view of 256 mm × 256 mm × 176 mm, resulting in isotropic 1 mm resolution). The timing parameters were TR/TE/TI = 7.9/3.0/450 ms, and the flip angle was 20◦ .

#### Image Processing and Tractography

The acquired diffusion-weighted images were corrected for distortion and motion artifacts with reorientation of encoding vectors (Leemans and Jones, 2009) and modulation of the signal intensity by the Jacobian determinant of the transformation (Jones and Cercignani, 2010). The free-water elimination approach was used to correct for atrophy-related partial volume effects due to CSF contamination (Pasternak et al., 2009; Berlot et al., 2014).

Whole-brain tractography was performed using ExploreDTI<sup>1</sup> and a diffusion tensor model using every voxel as a seed point. A deterministic tracking algorithm estimated the principal diffusion orientation at each seed point and propagated in 0.5 mm steps along this direction. The fiber orientation was then estimated at the new location and tracking moved a further 0.5 mm along the direction that subtended the minimum change of principal direction. A streamline was traced until fractional anisotropy fell below 0.15 or the change in direction exceeded 60◦ .

Three-dimensional reconstructions of the cingulum and of temporal association tracts were derived. Detailed reconstruction algorithms and linked reproducibility data, showing good reproducibility, have been described previously (Metzler-Baddeley et al., 2011, 2012a,b).

<sup>1</sup>http://www.exploreDTI.com

Whole brain volume, normalized for head size, was estimated with SIENAX (Smith et al., 2002), part of FSL<sup>2</sup> (FMRIB Software Library, Version 5.0). White matter lesions were segmented and their total volume quantified using a multispectral imageprocessing tool, MCMxxxVI (Hernandez et al., 2010).

### Network Construction and Graph Theory-Based Analysis

Whole-brain tract reconstructions were transformed into Montreal Neurological Institute (MNI) space within ExploreDTI, using a non-rigid transformation utilizing B-splines. Gray matter was then parcellated into 90 cortical and subcortical regions, 45 for each hemisphere, using the automated anatomical labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002) (**Figure 1**). Each region was used to define a node of a network graph. Edges were defined by tractography streamlines connecting any pair of nodes. An edge was defined as present between two nodes if a streamline was reconstructed with start and end points in each. Networks were weighted by the number of reconstructed streamlines.

Network metrics were computed using Brain Connectivity Toolbox<sup>3</sup> (Rubinov and Sporns, 2010). We investigated measures of global and local network architecture: global efficiency, mean clustering coefficient and small-worldness.

#### Statistical Analysis

Global efficiency, clustering coefficient and small-worldness were compared between MCI and control groups using unpaired t-tests. Associations with cognitive scores were computed in each group separately using Pearson's product-moment correlation coefficients. Bonferroni correction for multiple comparisons was applied based on the number of network measures. Cognitive measures tend to be strongly correlated with each other and in these circumstances Bonferroni correction is vastly overconservative, so correction was not applied for the number of cognitive measures. Partial correlation coefficients were calculated accounting for potential confounding variables: age, gender, education (in years), total brain volume, and total white matter lesion volume.

Linear regression models were constructed for Category Fluency and Digit Symbol Substitution task performance to investigate mediation effects. Measures of tract microstructure that were used were based on previously determined associations between Category Fluency and Digit Symbol Substitution, and the microstructure of cingulum segments: left anterior fractional anisotropy in controls, and left posterior mean diffusivity in MCI. These associations were identified in a previous analysis of the same dataset (based on diffusion MRI but not including network graph or graph theory measures), detailed in Metzler-Baddeley et al. (2012a). Separate models were constructed that included: (i) tract microstructure alone; or (ii) both tract microstructure and a single network measure. Thus, the relationships between tract microstructure and cognition, and network topology and cognition were established, and the influence of tract

<sup>2</sup>http://www.fmrib.ox.ac.uk/fsl/

<sup>3</sup>https://sites.google.com/site/bctnet

microstructure on cognition while controlling for network topology was assessed. The mediation effect was assessed as a decrease in the value of the standardized regression coefficients (β) for the association between cingulum microstructure and cognition after inclusion of a network measure in the model. Estimates of direct and indirect causal effects were obtained from the models using the non-parametric bootstrapping approach, and the proportion mediated by the network measure was estimated (Imai et al., 2010). This approach allowed measurement of a partial mediation effect and was not aimed at showing full mediation (where inclusion of a mediator leads to a measured association between two factors falling to zero). To test specificity of the investigated relationships for cognitive control, a similar analysis was performed for episodic memory: parallel regression models were constructed with free recall as the dependent variable and fornix tissue volume fraction as the relevant singletract measure (Metzler-Baddeley et al., 2012b).

Structural equation modeling was performed within the statistical software package R<sup>4</sup> , using an approach analogous to previous studies (Lawrence et al., 2014; Knopman et al., 2015). Tract and network measures were tested for interaction in each model. No significant interaction was found; therefore interaction terms were not included in final models. For terms in all models, variance inflation factors indicated no significant multicollinearity (variance inflation factors <3).

### RESULTS

### Group Comparisons

Demographic, cognitive and general MRI measures for the groups are provided in **Table 1**. Structural networks of both healthy older adults and patients with MCI exhibited smallworld topology. There was no difference in small-worldness between groups. In contrast, both global efficiency and mean clustering coefficient were reduced in MCI. On the basis of group differences, global efficiency and mean clustering coefficient were taken forward to analysis of relationships with cognition (leading to Bonferroni-corrected significance equivalent to uncorrected p < 0.025).

### Relationship between Network Metrics and Cognitive Scores

In MCI, both global efficiency and mean clustering coefficient were associated with cognitive control (**Tables 2** and **3**). In contrast, there were no relationships between global network measures and episodic memory performance. Measures of network topology were not correlated with cognitive scores in control participants.

### Cognitive Control, Global Network Properties, and Individual Tract Structure

In MCI, the inclusion of global network properties led to an attenuation of the relationship between single tract

<sup>4</sup>http://www.r-project.org

performance (G). Age, gender, educational attainment, brain volume and volume of white matter hyperintensities were used as covariates.



Data are shown as mean (SD). A cube root transform was applied to white matter lesion volume. Significant differences (p < 0.05) are highlighted in bold. MCI – Mild Cognitive Impairment; NART-R – National Adult Reading Test-Revised; FCSRT – Free and Cued Selective Reminding Test; CRMT – Camden Recognition Memory Test; NBV – normalized brain volume; WML – white matter lesion

microstructure and cognition (**Tables 4** and **5**). For Category Fluency, both left posterior cingulum microstructure and mean clustering coefficient were significant independent predictors (**Table 5**).

**Figure 2** displays path diagrams of the mediation analysis. The magnitudes of mediation effects are summarized in **Figure 3**. The proportion of the effect of cingulum microstructure on cognitive scores, mediated by global efficiency, varied from

#### TABLE 2 | Univariate relationship between network topology and cognition in patients with MCI and healthy elderly.


Pearson product-moment correlations (r) of cognitive scores with global efficiency (Eglob) and mean clustering coefficient (C), with parenthetical p-values. Coefficients shown in bold reach significance after Bonferroni correction for number of network measures (uncorrected p < 0.025), but not number of cognitive tests.

TABLE 3 | Relationship between network topology and cognition in patients with MCI and healthy elderly, adjusting for covariates.


Partial correlation coefficients (r) of cognitive scores with global efficiency (Eglob) and mean clustering coefficient (C), covarying for age, gender, education, normalized brain volume, and total volume of white matter hyperintensities, with parenthetical p-values. Coefficients shown in bold reach significance after Bonferroni correction for number of network measures (uncorrected p < 0.025), but not number of cognitive tests.

22 to 35% (**Figure 3**). In patients, the mediation effect was strongest for the relationship between left posterior cingulum and Category Fluency, 31% of which was explained by global efficiency (p = 0.12) and 36% by mean clustering coefficient (p = 0.02). Mean clustering coefficient was also a significant partial mediator of the link between left anterior cingulum and Category Fluency in controls (19% of variance due to mediation effect, p = 0.04). Mediation effects of network topology were not demonstrated for episodic memory and the association between fornix structure and free recall, in either of the two groups (**Table 6**; **Figure 3**).

#### DISCUSSION

Mild Cognitive Impairment is often considered a prodrome of dementia. We showed previously that microstructure is altered in white matter tracts in MCI and that alterations in specific tracts relate to specific aspects of the cognitive deficit. The present analysis demonstrates that global properties of the structural connectome are also altered. Patients with MCI had reduced global efficiency and mean clustering coefficient, in comparison with matched controls. While whole-brain network measures were not related to episodic memory, measures of network efficiency and clustering were related to cognitive control in MCI. This was the case despite the fact that episodic memory deficits were the most consistent, indeed defining feature of the MCI group. Episodic memory impairment was a prerequisite for the diagnosis while only seven patients with MCI displayed additional executive deficits. This result suggests that global networks are perturbed in MCI, but are not critical to the core deficit in episodic memory, which relates to damage within the relatively narrow and circumscribed extended hippocampal network.

A relationship between network efficiency and executive function has been described in Alzheimer's disease (Reijmer et al.,

#### TABLE 4 | Regression models for measures of cognitive control in healthy elderly.


Models with fractional anisotropy of the left anterior cingulum (1), and additionally a network measure (2) as predictors. Displayed are standardized regression coefficients (β) with parenthetical p-values.

Eglob – global efficiency; C – mean clustering coefficient.


Models with mean diffusivity of the left posterior cingulum (1), and additionally a network measure (2) as predictors. Displayed are standardized regression coefficients (β) with parenthetical p-values.

Eglob – global efficiency; C – mean clustering coefficient.

2013), but also in other brain disorders such as traumatic brain injury (Caeyenberghs et al., 2012), small-vessel disease (Lawrence et al., 2014), and cerebral amyloid angiopathy (Reijmer et al., 2015). In patients with small-vessel disease and cerebral amyloid angiopathy, network measures were related only to executive function, but not memory performance. However, in these diseases episodic memory deficits are mild or absent, so this dissociation might have been explained by a lack of variance in memory scores. In the present study, conversely, episodic memory was impaired to a greater extent, and more consistently, than cognitive control. This dissociation therefore is more likely to reflect the functional anatomy of cognitive control and episodic memory in the brain and the dependence of cognitive control on a more diffuse network. Further, when correlations were controlled for the volume of white matter lesions, as well as other potential confounders, the pattern of associations remained consistent, and in some cases became stronger, indicating that small vessel disease did not account for the associations observed in this study. Mediation analyses suggested that the relationship between cingulum microstructure and cognitive control was partly mediated by global network topology, while no such link was observed for the relationship between fornix structure and episodic memory. These findings further underline a qualitatively different relationship between tracts and cognitive function for cognitive control and episodic memory.

One intriguing parallel to the pattern of results is that pathological processes also vary in whether they target local structures or more global infrastructure. For example, amyloid and tau pathologies have strong local predilections, at least early in the course of disease. Microvascular disease, on the other hand, leads to diffuse alterations in white matter microstructure so, potentially, it could have a general effect on network efficiency (Lawrence et al., 2014). One strength of the approach taken is that it provides a way to understand how coexistent pathologies could interact. For example, localized neurodegeneration and network-wide effects of diffuse microvascular disease could act synergistically to impair cognitive or executive control.

However, the contrasting relationships of network topology to episodic memory and cognitive control might also be related to methodology used. One possibility is that episodic memory depends on a network that more heavily involves subcortical structures and connections, particularly in the diencephalon, and that in turn topology of these networks is not strongly represented in whole-brain network metrics, constructed using current methods. Parcellation of nodes might be more effective for networks that involve multiple neocortical regions, such as those involved in cognitive control, than for networks with fine-grained subcortical anatomy. The AAL atlas used, as well as alternative parcellation techniques, do not include the mammillary bodies, for example, which are crucial structures within the extended hippocampal network involved in episodic memory.

The pattern of results suggests that damage to a tract such as the cingulum can degrade cognitive performance through two distinct roles of this tract – as a conduit for communication of specific information within a dedicated network for cognitive control, and as a more generic 'backbone' for communication across global brain networks. Previous work has shown that hub regions such as the anterior and posterior cingulate cortices, and their connections, might be important not only because they harbor critical functional specializations but also because they mediate connectivity across the structural network more broadly including, for example, in the case of the posterior cingulate cortex, tuning network metastability (Leech and Sharp, 2014).

A limitation of this study, common to studies based on tractography, is the risk of false positive and false negative connections. Weighting of network edges by the total number of reconstructed streamlines should reduce the impact of anatomically spurious edges as, in general, only a few outlier streamlines will run between regions that do not have a true connection. The choice of method for weighting edges is a controversial aspect of the application of graph theory to structural networks. Number of streamlines was used to offer consistency with previous studies and to avoid using microstructural measures known to be abnormal in MCI, but the effect of different weighting approaches has not been investigated in detail. Cognitive control is multifaceted and a number of

measures provide overlapping insights into these processes. The Bonferroni method is highly over-conservative in the presence of multiple inter-correlated measures. Correction was therefore applied for number of network measures but not for number of cognitive measures, so that the risk of false positive correlations may not be completely eliminated in the regression analyses. Similarly, a large number of mediation models could have been constructed based on different measures. To minimize

TABLE 6 | Regression models for free recall in healthy elderly and MCI.


Models with fornix volume fraction (1), and fornix volume fraction and a network measure (2) as predictors. Displayed are standardized regression coefficients (β) with parenthetical p-values.

Eglob – global efficiency; C – mean clustering coefficient.

the risk of mediation emerging by chance, we selected the two measures most consistently associated with cognition in regression analysis (**Tables 2** and **3**). In addition, a limitation of the mediation analysis performed is that we cannot make definite conclusions on the direction of the effect. Even though it seems less biologically plausible, our results do not exclude the possibility of cingulum microstructure mediating the effect of network topology on cognition.

Further insight into the dynamics of the relationship between 'local' and 'global' disease-related alterations could be gained by observing our population in a longitudinal setting, or additionally including a group of patients with more severe cognitive impairment. The current study does not extend to brain function, inferred from functional MRI data. It is possible that the topology of structural networks will not be entirely reflected by functional networks, which differ in being dynamic over short time scales. Finally, the interplay between 'local' and 'global' structural and functional changes might be of interest beyond cognitive function. Functional variation within the cingulate cortex and the large-scale networks might be related to the expression of specific clinical phenotypes, rather than disease-related alterations, such as the occurrence of hyperarousal, anxiety or hallucinations in neurodegenerative disorders (Franciotti et al., 2015). A similar approach could be used to test this hypothesis in Alzheimer's disease and other neurodegenerative disorders.

Potential treatments such as transcranial magnetic stimulation or direct current stimulation have largely been thought of in terms of localized effects on function. However, a number of studies show that treatment delivered locally can have effects on global network topology and dynamics (Polanía et al., 2011; Shafi et al., 2014). In principle, these wider effects could also be harnessed to restore network function. Our results suggest that for some functions – such as cognitive control – the ideal strategy may involve targeting both local and global alterations in brain structure and function.

#### AUTHOR CONTRIBUTIONS

RB contributed to study conception, data analysis, statistical analysis, writing and editing the manuscript. CM-B contributed to study conception, data collection, writing and editing the manuscript. MAI contributed to data analysis, statistical analysis, writing and editing the manuscript. DKJ contributed to data collection, data analysis, writing and editing the manuscript. MJO contributed to study conception and design, data collection, data analysis, writing, and editing the manuscript.

#### ACKNOWLEDGMENTS

This work was supported by the Medical Research Council, UK (MJO, grant refs G0701912 and MR/K022113/1). RB was

supported by the Slovenian Research Agency. CM-B is supported by an Alzheimer's Society and BRACE Alzheimer's charity research fellowship. DKJ is supported by the Wellcome Trust through a New Investigator Award. This study represents independent research part funded by the National Institute for

#### REFERENCES


Health Research (NIHR) Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King's College London. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health.

in communities – neurocognitive study. Stroke 46, 433–440. doi: 10.1161/ STROKEAHA.114.007847



Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., et al. (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 15, 273–289. doi: 10.1006/nimg.2001. 0978

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Berlot, Metzler-Baddeley, Ikram, Jones and O'Sullivan. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Spectral Variability in the Aged Brain during Fine Motor Control

Fanny Quandt <sup>1</sup> , Marlene Bönstrup<sup>1</sup> , Robert Schulz <sup>1</sup> , Jan E. Timmermann<sup>1</sup> , Maximo Zimerman2, 3, Guido Nolte<sup>4</sup> and Friedhelm C. Hummel 1, 2, 5, 6 \*

<sup>1</sup> BrainImaging and NeuroStimulation Laboratory, Department of Neurology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, <sup>2</sup> Institute of Neuroscience, Favaloro University, Buenos Aires, Argentina, <sup>3</sup> Institute of Cognitive Neurology, Buenos Aires, Argentina, <sup>4</sup> Department of Neurophysiology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, <sup>5</sup> Clinical Neuroengineering, Brain Mind Institute and Centre of Neuroprosthetics (CNP), Swiss Federal Institute of Technology (EPFL), Geneva, Switzerland, <sup>6</sup> Clinique Romande de Réadaptation, Swiss Federal Institute of Technology (EPFL Valais), Sion, Switzerland

Physiological aging is paralleled by a decline of fine motor skills accompanied by structural and functional alterations of the underlying brain network. Here, we aim to investigate age-related changes in the spectral distribution of neuronal oscillations during fine skilled motor function. We employ the concept of spectral entropy in order to describe the flatness and peaked-ness of a frequency spectrum to quantify changes in the spectral distribution of the oscillatory motor response in the aged brain. Electroencephalogram was recorded in elderly (n = 32) and young (n = 34) participants who performed either a cued finger movement or a pinch or a whole hand grip task with their dominant right hand. Whereas young participant showed distinct, well-defined movement-related power decreases in the alpha and upper beta band, elderly participants exhibited a flat broadband, frequency-unspecific power desynchronization. This broadband response was reflected by an increase of spectral entropy over sensorimotor and frontal areas in the aged brain. Neuronal activation patterns differed between motor tasks in the young brain, while the aged brain showed a similar activation pattern in all tasks. Moreover, we found a wider recruitment of the cortical motor network in the aged brain. The present study adds to the understanding of age-related changes of neural coding during skilled motor behavior, revealing a less predictable signal with great variability across frequencies in a wide cortical motor network in the aged brain. The increase in entropy in the aged brain could be a reflection of random noise-like activity or could represent a compensatory mechanism that serves a functional role.

#### Keywords: aging, motor control, entropy, oscillations, EEG

### INTRODUCTION

Physiological aging is paralleled by a decline of motor performance, most pronounced in demanding fine motor skills. At the higher age (Smith et al., 1999), elderly show a decrease of movement coordination (Stelmach et al., 1988; Wishart et al., 2000) with increasing variability of motor output (Cooke et al., 1989; Darling et al., 1989), along with a general movement slowing (Buckles, 1993). These behavioral changes are accompanied by alterations of the underlying brain network. During movements a more widespread neuronal network is recruited in the aged brain (Sailer et al., 2000; Ward and Frackowiak, 2003; Wu and Hallett, 2005; Naccarato et al., 2006; Rowe et al., 2006; Vallesi et al., 2010; Deiber et al., 2013). Additionally, elderly show higher magnitudes of movement-related desynchronization of oscillatory activity in frequency bands associated

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Yury (Juri) Kropotov, The Institute of the Human Brain of Russian Academy of Sciences, Russia Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece Franka Thurm, Dresden University of Technology, Germany

#### \*Correspondence:

Friedhelm C. Hummel friedhelm.hummel@epfl.ch

Received: 04 May 2016 Accepted: 30 November 2016 Published: 21 December 2016

#### Citation:

Quandt F, Bönstrup M, Schulz R, Timmermann JE, Zimerman M, Nolte G and Hummel FC (2016) Spectral Variability in the Aged Brain during Fine Motor Control. Front. Aging Neurosci. 8:305. doi: 10.3389/fnagi.2016.00305 with motor control (Sailer et al., 2000). Apart from the larger movement-related power decrease, few studies so far have observed differences in the frequency patterns with age. When reviewing lifespan changes in the alpha peak, Klimesch reported a drop of peak frequency with age (Klimesch, 1999). Moreover, previous studies found shifts of the most reactive peak frequency during rest (Gaál et al., 2010) and attention (Deiber et al., 2013). Hong and Rebec hypothesize that in order to compensate for a non-uniform decrease of nerve conduction in the aged brain, individual neurons increase their firing rate with a non-uniform pattern, leading to an irregular firing pattern which subsequently might lead to an unspecific broadband large-scale oscillatory response (Hong and Rebec, 2012). Such a broadening of the neuronal response has, to our knowledge, not been analytically addressed before and is rather difficult to detect when analyzing the data using movement specific narrow frequency bands as previously suggested in healthy young, such as the alpha or upper beta band (Pfurtscheller, 1989; Crone et al., 1998). In the healthy young, spectral changes in the mu (∼8–13 Hz) and beta band (∼14–30 Hz) are associated with voluntary movements (Pfurtscheller and Lopes da Silva, 1999). Their definite functional role still remains under debate, however, mu and beta rhythms are thought to represent separate functional processes with different time courses and distributions over the scalp. While the mu rhythm dominantly localizes to the post-central hand area, the beta rhythm localizes to pre-central areas (for a review please refer to (Cheyne, 2013; Brittain and Brown, 2014). Here, we investigate movement-related power changes over a broad frequency band from 8 to 25 Hz and aim to determine whether the implementation of motor tasks in the aged brain is in similar frequency bands compared to the young brain. One measure to characterize differences in the distribution of the spectral content is the spectral entropy H. Spectral entropy is an uncertainty measure borrowed from information theory. We solely employ the mathematical concept of entropy by treating the frequency spectrum as a probability density in order to describe the flatness and peaked-ness of a frequency spectrum (Inouye et al., 1991). An oscillatory activity with a flat frequency distribution and large variability results in a high spectral entropy, whereas a peaked signal such as a confined alpha or upper beta movementrelated desynchronization would result in a lower spectral entropy. It was our primary objective to mathematically quantify changes of the spectral content of oscillatory movement-related patterns in the aged brain by employing the spectral entropy. Electrophysiological data were recorded from elderly over the age of 60 and young participants while performing different skilled fine motor tasks. So far, most studies investigating the aged motor system have focused on one specific motor task. Here, we assessed different motor tasks, which allows us to make inferences on more generalized age-dependent motor network changes. The magnitude and spatial extent of movement-related power decrease was analyzed. Importantly, we characterized the distribution of the spectral content by spectral entropy. We hypothesized that the aged brain shows larger variability of oscillatory activation patterns with a broadening of the movement-related frequency band and higher spectral entropy.

### MATERIALS AND METHODS

#### Participants

Sixty-six healthy volunteers participated in different EEG experiments, consisting of an elderly group with 32 participants over the age of 60 (mean age 72.2 y/o ± 5.2 SD, range 61– 81 y/o, 19 females) as well as a young control group with 34 participants (mean age 25.5 y/o ± 3.3 SD, range 19–34, y/o, 14 females). A subgroup of 15 elderly and 16 young participants performed two different motor tasks. All participants were righthanded as confirmed by the Edinburgh handedness inventory (Oldfield, 1971), did not have a history of neurologic disorder and gave written informed consent. All elderly subjects were seen by a neurologist and did not show any cognitive impairment. Elderly subjects participating in the Finger Sequence Task, requiring learning of a digit sequence, all presented with a mini-mental state examination ≥28. The study conforms to "The Code of Ethics" of the World Medical Association (Declaration of Helsinki) and was approved by the local ethics committee of the Medical Association of Hamburg.

### Motor Tasks

Participants performed a motor task during EEG recording. All motor tasks required a movement in response to an external visual cue. In every participant, a 2–5 min pre- and postexperimental baseline was recorded at unconstrained rest with eyes open.

#### Finger Sequence Task

Seventeen elderly (range 61–81 y/o) and 18 young (range 19– 33 y/o) participants trained a sequence of 10 consecutive button presses [2 4 3 2 5 4 5 2 5 3, with (2) = index finger, (3) = middle, (4) = ring, (5) = little]. The sequence was trained until performance reached a stable level and participants were able to play the sequence at least ten times in a row without any mistakes at a pace of 1 Hz (Gerloff et al., 1997). Hence, the sequence was considered overlearned, ensuring constant baseline performance during the EEG session. Training was conducted either on the day prior or on the day of the experiment. During the following EEG experiment participants sat in front of a computer screen with the right arm positioned on a keyboard. Visual cues without any relation to the learned sequence ("#," "&," "+," "\$") were presented on a computer screen. The symbols paced the execution of the memorized, well-trained sequence at a frequency of 1 Hz and participants were asked to enter the finger sequence at the pace of the visual cues. The sequence was played with the right hand, using the index-, middle-, ring,- and little finger. Each participant performed 40 repetitions of the 10-digit sequence.

#### Pinch Grip and Whole Hand Grip Task

Fifteen elderly (range 67–79 y/o) and 16 young (range 20–34 y/o) participants performed repetitive pinch as well as whole hand grips lifting a weight positioned on a table in front of them. Participants were seated in front of a monitor with their arms placed on a custom-made platform. The right hand was placed on a socket installed on the platform with the elbow 90◦Flexed. The 200 g weight was lifted 10–20 cm of the table using the right thumb and index finger and reset immediately after. Instructions were visually presented on a screen, consisting of the word "pinch grip" or "whole hand grip," followed by a "GO!" cue 2–3 s later. Condition "pinch" and condition "whole" were presented in a random, counterbalanced order. The next trial was initiated 8– 10 s later. Each participant performed 80 pinch and 80 whole hand grips.

#### Recording Systems and Preprocessing

Data were sampled at 1000 Hz using a 63-channel EEG system positioned according to the 10–10 System of the American Electroencephalographic Society (using actiCAP <sup>R</sup> , Brain Products GmbH, Germany, Gilching; Electro-Cap International, Inc., Eaton, OH, USA) and referenced to the Cz electrode. The impedance of the EEG electrodes was kept below 25 k. Data were filtered from 0.2 to 256 Hz with a bandpass-filter of third order. Datasets were segmented into one second epochs for further analysis. Specifically, finger sequence data were segmented ± 500 ms around the visual cue and lifting task data were segmented from 300 to 1300 ms after the "GO" cue. Eye-movement artifacts were removed employing an independent component analysis (Makeig et al., 1996). Epochs containing electrode artifacts, muscle artifacts, head movements, or incompletely rejected blink artifacts were removed manually by visual inspection. In participants with great muscle artifacts a blind source separation-canonical correlation analysis was applied in order to correct these artifacts (De Clercq et al., 2006) as implemented in the eeglab-plugin meegpipe (https:// github.com/meegpipe/meegpipe/). Subsequently, data were re-referenced to a common average reference. Artifact rejection resulted in an overall number of µ = 117/260, SD = 63/44 trials (elderly/young; Finger Sequence Task), µ = 64/63, SD = 5/6 trials (elderly/young; Pinch Grip Task), and µ = 74/73, SD = 4/8 trials (elderly/young; Whole Hand Grip Task).

Pre- and post-experimental baselines were pooled and subsequently divided into 2000 ms segments and preprocessed jointly as described above. The Fieldtrip toolbox (Oostenveld et al., 2011) as well as custom written software using MATLAB Version 8.2.0 (R2013b, Mathworks Inc. Massachusetts) were used for EEG data analysis.

## EEG Data Analysis

#### Frequency Analysis

Power spectra were calculated from 8 to 25 Hz in steps of 1 Hz applying a fast Fourier transformation using one Hanning taper for each electrode and trial. In order to account for intersubject variability and decreasing power in higher frequencies, spectral power was expressed as the relative power (Powrel) defined by the percentage of power change during movement (Powmove) compared to baseline (Powbaseline; Gerloff et al., 1998; Pfurtscheller et al., 2003). Powbaseline, was obtained from the preprocessed baseline data and averaged across segments afterwards. Subsequently, Powrel was computed by:

$$Power\_{rel} = 100 \times \frac{Power\_{move} - Power\_{baseline}}{Power\_{baseline}} \tag{1}$$

Afterwards trials were averaged for each participant.

#### Spectral Entropy

Spectral entropy is an uncertainty measure borrowed from information theory. Here, we apply the entropy as a mathematical concept to describe the flatness of the frequency spectrum, which is treated as a probability density after appropriate normalization. A uniform flat signal with a high variability and a broad spectral content results in a high spectral entropy (H∼1), whereas a more predictable signal with a narrow, peaked power spectrum in a limited number of frequency bins yields a low spectral entropy (H∼0). The spectral entropy is calculated by:

$$H = \frac{-1}{\ln(N)} \sum p\_i \ln(p\_i) \tag{2}$$

with

$$p\_i = \frac{|Power\_{rel}\ (i)|}{\sum\_i Power\_{rel}\ (i)}\tag{3}$$

and with Powrel(i) being the relative power of frequency bin i and N being equal to the number of frequency bins (Inouye et al., 1991). In order to quantify the distribution of spectral power, we estimated the spectral entropy H in the broad frequency band between 8 and 25 Hz as well as in the frequency band showing greatest differences between the aged and young brain (13–19 Hz). The spatial extent of differences H was evaluated by calculating H for each electrode separately.

#### Source Analysis

Sensor data in the frequency band from 13 to 19 Hz were projected to source level in each sensor of each participant. The forward solution is constructed with a segmented template MRI brain (Holmes et al., 1998) using the boundary element method and a template grid of 8 mm spacing (Oostenveld et al., 2011). Individual electrode positions were determined using the Zebris localization system (CMS20, Zebris Medical GmbH, Isny, Germany) and realigned to the template MRI brain. A common filter for the frequency range from 13 to 19 Hz of movement period and baseline period was calculated based on the average real part of the cross-spectrum in that range using dynamic imaging of coherent sources (DICS; Gross et al., 2001) with source orientation chosen to maximize power using the Fieldtrip Toolbox. The DICS beamformer uses a frequency domain implementation of a spatial filter. Subsequently, the contrast was computed expressing a relative change of power as described in Equation (1).

#### Statistics

Firstly, it was the objective to analyze topographic age-group differences of the mean broadband power changes (8–25 Hz) for each motor task separately. Topographic age-group differences were statistically tested using an unpaired student's t-test (relative power, normally distributed) or a Wilcoxon rank sum test (entropy, non-normally distributed) corrected for multiple comparisons (63 channels) controlling the false discovery rate (FDR; Benjamini et al., 2001). Secondly, we tested age-group differences of single frequency bins over left sensorimotor cortex for each motor task separately, in order to demonstrate differences of the power distribution. Power distribution agegroup differences were statistically tested using an unpaired student's t-test corrected for multiple comparisons (18 frequency bins) controlling the FDR (Benjamini et al., 2001).

In addition, we estimated topographic age-group differences combining all participants of all three motor tasks in linear mixed effects models using R (CDT, 2008) and lme4 (Bates et al., 2015). The linear mixed effects models were calculated for entropy and relative power respectively. In order to correct for a potential influence of task, we entered task as a fixed effect. Participants were entered as a random intercept in order to correct for repeated testing. This model was calculated for each channel separately. We then extracted the p-value from our main effect of interest "group" and obtained 63 p-values (one per channel), which were then FDR corrected. Moreover, for post-hoc testing of task differences in young and elderly participants separately, we modeled the interaction of group and task to perform post-hoc testing using a pairwise comparison of least-square means.

#### RESULTS

### Power Amplitude Differences between Elderly and Young Participants

The topology of movement-related broadband power changes (8–25 Hz) revealed a more widespread spatial distribution of desynchronization in the elderly compared to young participants in all three tasks (**Figure 1A**). This difference of power was significant in electrodes covering sensorimotor cortex as well as in more frontal electrodes (electrodes as marked in **Figure 1A**, FDR corr., p < 0.05). Further probing the distribution in single frequency bins in electrodes covering the left sensorimotor cortex (mean of electrodes: FC3, C3, CP3), elderly participants showed a greater movement-related power decrease in all frequency bins (**Figure 1B**). This difference was significant from 13 to 20 Hz for Task 1, from 15 to 17 Hz in Task 2, and from 13 to 22 Hz in Task 3. Please refer to Supplementary Tables 1, 2 for t-test results and Supplementary Figure 1 for data distribution.

In order to identify responsible sources of oscillatory activity in the significant frequency band, we applied a beamforming technique. **Figure 2** displays the difference of movement-related power in the power band from 13 to 19 Hz between the aged and young brain, revealing that the aged brain recruits a more extended motor network of contralateral but also ipsilateral primary sensorimotor and secondary premotor areas including dorsal and ventral premotor cortex as well as the supplementary motor area (**Figure 2**).

### Differences in Spectral Entropy of Oscillations in Elderly and Young Participants

The shape of the distribution of the power spectrum in all three motor tasks differed with age (**Figure 1B**). Whereas, in young participants, a clear and peaked modulation of movement-related power decrease in the alpha and upper beta band was evident

FIGURE 1 | Power amplitude differences between elderly and young participants. (A) Topology of movement-related broadband power (8–25 Hz) for elderly and young participants for each task separately, averaged over participants. Stars mark the significant electrodes (unpaired t-test, FDR corr., p < 0.05). (B) Power in each frequency bin for elderly (red) and young (blue) averaged over electrodes covering the contralateral sensorimotor cortex (as framed by the rectangle in A). Black dots mark a significant difference between both groups in the corresponding frequency bin (unpaired t-test, FDR corr., p < 0.05). The x-axis shows the frequency in Hz, the y-axis displays the relative power (%) to baseline for elderly (red) and young participants (blue) separately.

(**Figure 1B**, blue bars), the aged brain displayed a more uniform flat curve of power decrease (**Figure 1B**, red bars). In order to mathematically quantify this disparity of spectral distribution, we calculated the spectral entropy H for elderly and young participants in each channel.

elderly and young of the Finger Sequence Task and Pinch Grip Task is rendered on the cortical surface, displayed from a top and left-side view, masked by power.

The group difference correcting for tasks and repeated testing of the same participants was assessed in linear mixed effect model for each channel separately. **Figure 3A** displays the estimated mean of the mixed model for each channel over the broadband spectrum from 8 to 25 Hz. Asterisks mark significant models with p < 0.05 (FDR corr.; for model results, please refer to Supplementary Tables 3, 4). The aged brain showed a higher spectral entropy H in electrodes covering frontal as well as sensorimotor areas. We further probed the spectral entropy in a more restricted frequency band (13–19 Hz, **Figure 3B**) in which relative power showed greatest differences between groups. Hence, we confirmed a flatter more uniform frequency spectrum with a broader spectral content in the aged population compared to younger people during different fine skilled motor tasks.

When assessing entropy differences in each task separately (mean value of relative power, **Figure 4**), we find a similar pattern in each task, with greater spectral entropy in the aged compared to the young brain in electrodes covering frontal as well as contra- and ipsilateral sensorimotor areas. **Figure 4** revealed that differences between motor tasks were mainly driven by the young participants. Elderly participants showed a similar activation

FIGURE 3 | Mixed model results (A) estimated mean power and entropy of the mixed effects model for each channel (8–25 Hz). Stars mark the significant group effect of the mixed model (FDR corr., p < 0.05). (B) Estimated mean power and entropy of the mixed model for each channel (13–19 Hz). Stars mark the significant group effect of the mixed model (FDR corr., p < 0.05).

pattern in all three tasks, whereas in young participants the neuronal activation pattern differs depending on motor tasks' complexity.

Post-hoc testing of task differences in the mixed effects model for the elderly and young group pointed toward differences in entropy between the Finger Sequence Task and the Pinch Grip and Whole Hand Grip Task in the young but not in the elderly participants (**Figure 5**).

### DISCUSSION

This study characterizes differences in the spectral content of motor control in healthy aging. By using the concept of entropy, we quantified differences in the spectral distribution between the aged and young brain and found a higher spectral entropy with a flat, uniform distribution of power in the aged brain in various fine skilled motor tasks. Whereas, the young brain showed lower entropy and a distinct peaked movement-related power decrease in the alpha and upper beta band, the aged brain exhibited a larger movement-related decrease of power most pronounced in the low beta frequency band along with a wider recruitment of the cortical motor network involving premotor areas.

### Reduced Frequency Specificity of the Aged Brain

Movement execution leads to distinct event-related desynchronization in the alpha and upper beta band over contra- and ipsilateral sensorimotor areas (Pfurtscheller, 1989; Crone et al., 1998). In task-related studies, these frequency bands have been often used for analysis of movement specific oscillatory changes. These motor-task-related frequency bands, however, have been determined based on data in the young brain. In contrast, these bands might not correspond to movement-related changes in the aged brain, as supported by previous studies, indicating changes in oscillatory activity patterns across the lifespan. Apart from a larger magnitude of the movement-related desynchronization within sensorimotor areas (Sailer et al., 2000; Mattay et al., 2002), age-related changes comprise shifts in resting state peak alpha frequency (Klimesch, 1999; Cottone et al., 2013), a decrease of alpha reactivity (Gaál et al., 2010), as well as changes of the dominant oscillator with age (Deiber et al., 2013). Hence, a priori knowledge on frequency specific bands as determined in the young brain might be arbitrary in the aged brain and might hinder the detection of distinct age-related oscillatory changes. For this reason, we characterized oscillatory changes by using the concept of spectral entropy allowing us to

circumvent the analysis of restricted frequency bands predefined from the young brain. Thereby, we observed differences in the spectral distribution of movement-related desynchronization over a broad frequency band (8–25 Hz) and found a widespread increase of spectral entropy during movement in the elderly compared to young, indicating that the movement-related broadband changes in the elderly are more variable and less predictable. This phenomenon was observed in three different fine motor skill tasks. The present finding underlines the notion that the increase of spectral entropy is a rather task unspecific phenomenon occurring during fine skilled motor control in the aged brain. Even though spectral entropy topographies showed differences in-between the Finger Sequence compared to the Pinch Grip and Whole Hand Grip task, these differences were solely derived from the young group. The aged brain on the other hand showed a uniform increase of entropy in all three tasks that did not statistically differ between tasks. One has to keep in mind, however, that in order to explicitly test, whether this difference of entropy over tasks was solely derived from the young group, one would have to conduct a crossover study, which includes execution of all three tasks in each participant.

Computational models of neuromodulation postulate that the aged brain exhibits deficient neuromodulatory mechanisms and consequently less distinctive neural pattern representations (Li and Sikström, 2002). Therefore, the increased spectral entropy could be a result of reduced coordination of enhanced synaptic activity of neuronal assemblies leading to greater variability of the neuronal responses. Correspondingly, increasing variability resulting in less consistent motor actions has been observed during healthy aging (Cooke et al., 1989; Darling et al., 1989; Contreras-Vidal et al., 1998; Sosnoff and Newell, 2011). On the one hand, the signal could be a result of inaccurate interregional neuronal communication leading to a breakdown of the frequency specificity, where the aged brain is not capable of keeping a certain frequency. Hence, the increased variability could be in line with the dedifferentiation of the aging brain (Deiber et al., 2013). On the other hand, the broadening of the frequency band could ensure to preserve the balance between energy consumption and entropy of the neural signal. Tsubo et al. have postulated that with higher uncertainty of the neural responses, the brain reduces the amount of energy necessary (Tsubo et al., 2012). Moreover, high variability of a signal has been suggested to result in an increase of performance and might be beneficial (Garrett et al., 2013). Hence, the increase of entropy could be a compensating mechanism to account for a decline of motor performance. In line Hanslmayr et al. speculated that higher neuronal desynchronization presents a greater richness of information, measured by the entropy and consider it as one mechanism serving memory encoding (Hanslmayr et al., 2012). We cannot, however, definitely infer whether a higher entropy is a reflection of random noise-like activity or if the irregular pattern is a compensatory mechanism that serves a functional role. Hence, higher entropy could mean that either more information is sent, or that more noise-like random activity is produced and sent. Further research will have to address this important issue, especially to determine its functional implication on behavior.

### Enhanced Spatial Recruitment in the Aged Brain

Several functional imaging and EEG studies have reported a more extended recruitment of brain areas during movement in the aged brain (Sailer et al., 2000; Mattay et al., 2002; Wu and Hallett, 2005; Naccarato et al., 2006; Rowe et al., 2006; Vallesi et al., 2010; Deiber et al., 2013). In line, we found activations in an extended motor network including bilateral primary motor and sensory areas as well as ipsilateral premotor areas, namely, dorsal and ventral premotor cortex, pre- and supplementary motor areas (**Figure 2**), most pronounced in the lower beta frequency band. The over-recruitment of brain areas might lead to more potential network configurations with higher noise interferences and hence greater variability of states giving rise to the unspecific frequency distribution of movement-related power changes determined here. The cause of this over-recruitment could be either compensation with greater recruitment of secondary motor areas, because of the subjective increase of task-related complexity, in order to achieve the same motor output (Zimerman et al., 2014), or an increase of the attentional load (Reuter-Lorenz and Cappell,

2008), or due to inefficient activations with reduced selectivity of neuronal networks and less distinct activation patterns (Li and Lindenberger, 1999; Riecker et al., 2006). However, this question cannot be definitely answered by this study, mostly because it lacks a functional outcome parameter (for a review see Grady, 2012).

### Possible Mechanisms of Dynamical Changes during Healthy Aging

The high variability of the spectral content along with the overrecruitment of secondary motor areas might be, on the one hand, a result of a decrease in selective local inhibition with greater background activity. On the other hand, a reduced selectivity of the network could be the consequence of a more general inhibition deficiency due to age-related structural and functional changes of the frontal cortex (Tisserand and Jolles, 2003; Rajah and D'Esposito, 2005). Moreover, a reduction of specific regulatory thalamic input could result in less distinct cortical activations. In Parkinson patients research has demonstrated the modulating influence of the basal ganglia-thalamocortical network on cortical oscillation patterns and motor control (de Hemptinne et al., 2013, 2015). These influences can be either of structural nature or can be evoked by intrinsic changes of synaptic properties. Furthermore, disrupted network dynamics might be a result of more subtle changes (McCarthy et al., 2012; Kopell et al., 2014; Voytek and Knight, 2015), such as neurochemical shifts and changes in synaptic binding potentials and receptor density. Future studies will have to further determine the underlying cause of oscillatory alterations in the aged brain.

#### REFERENCES


In summary, the aged brain exhibits a broadband, frequencyunspecific power desynchronization during movement as reflected by an increase of spectral entropy, revealing a less predictable signal with great variability across frequencies in a wide cortical motor network.

#### AUTHOR CONTRIBUTIONS

FQ conducted the research, analyzed the data, and drafted the manuscript. MB conducted the research, was involved in data analysis, revised the manuscript. RS, JT, MZ, GN were involved in data acquisition and analysis, revision of the manuscript. FH developed the experimental idea, involved in drafting and revising of the manuscript.

#### FUNDING

This research was supported by the German Research Foundation (DFG, SFB 936-C4 to FH and Z3 to GN) and the German Ministry of Science (BMBF, 01GQ1424B to FH).

#### ACKNOWLEDGMENTS

We thank Meike Mund for her help on data collection.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2016.00305/full#supplementary-material

and beta event-related desynchronization. Brain 121(Pt 12), 2271–2299. doi: 10.1093/brain/121.12.2271


of future elements in complex motor sequences. Brain 120(Pt 9), 1587–1602. doi: 10.1093/brain/120.9.1587


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Quandt, Bönstrup, Schulz, Timmermann, Zimerman, Nolte and Hummel. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Functional Integration in the Sensory-Motor System Predicts Aging in Healthy Older Adults

Hui He<sup>1</sup> , Cheng Luo<sup>1</sup> \*, Xin Chang<sup>1</sup> , Yan Shan<sup>1</sup> , Weifang Cao<sup>1</sup> , Jinnan Gong<sup>1</sup> , Benjamin Klugah-Brown<sup>1</sup> , Maria A. Bobes<sup>2</sup> , Bharat Biswal<sup>3</sup> and Dezhong Yao<sup>1</sup> \*

<sup>1</sup> The Key Laboratory for NeuroInformation of Ministry of Education, Center for Information in BioMedicine, High-Field Magnetic Resonance Brain Imaging Key Laboratory of Sichuan Province, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China, <sup>2</sup> Department of Biological Psychiatry, Cuban Neuroscience Center, La Habana, Cuba, <sup>3</sup> Department of Biomedical Engineering, New Jersey Institute of Technology, University Heights, Newark, NJ, USA

Healthy aging is typically accompanied by a decrease in the motor capacity. Although the disrupted neural representations and performance of movement have been observed in older age in previous studies, the relationship between the functional integration of sensory-motor (SM) system and aging could be further investigated. In this study, we examine the impact of healthy aging on the resting-state functional connectivity (rsFC) of the SM system, and investigate as to how aging is affecting the rsFC in SM network. The SM network was identified and evaluated in 52 healthy older adults and 51 younger adults using two common data analytic approaches: independent component analysis and seed-based functional connectivity (seed at bilateral M1 and S1). We then evaluated whether the altered rsFC of the SM network could delineate trajectories of the age of older adults using a machine learning methodology. Compared with the younger adults, the older demonstrated reduced functional integration with increasing age in the mid-posterior insula of SM network and increased rsFC among the sensorimotor cortex. Moreover, the reduction in the rsFC of mid-posterior insula is associated with the age of older adults. Critically, the analysis based on two-aspect connectivity-based prediction frameworks revealed that the age of older adults could be reliably predicted by this reduced rsFC. These findings further indicated that healthy aging has a marked influence on the SM system that would be associated with a reorganization of SM system with aging. Our findings provide further insight into changes in sensorimotor function in the aging brain.

Keywords: aging, resting state fMRI, functional connectivity, sensory-motor system, machine learning

### INTRODUCTION

Healthy aging is typically accompanied by functional and structural changes in the brain. Decrease in motor performance and movement coordination is one of the most consistent findings in older adults (Seidler et al., 2010; Allen et al., 2011; Hoffstaedter et al., 2014), and is an important aspect of physiological aging. The general slowing of movements accompanied with aging has been observed in previous studies (Birren and Fisher, 1995). The primary sensory-motor (SM) system plays a critical role for somesthesia and movement generation. Accumulating evidence suggests that age

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Roma Siugzdaite, Ghent University, Belgium Foteini Protopapa, International School for Advanced Studies, Italy Dina R. Dajani, University of Miami, USA

\*Correspondence:

Cheng Luo chengluo@uestc.edu.cn Dezhong Yao dyao@uestc.edu.cn

Received: 21 September 2016 Accepted: 02 December 2016 Published: 05 January 2017

#### Citation:

He H, Luo C, Chang X, Shan Y, Cao W, Gong J, Klugah-Brown B, Bobes MA, Biswal B and Yao D (2017) The Functional Integration in the Sensory-Motor System Predicts Aging in Healthy Older Adults. Front. Aging Neurosci. 8:306. doi: 10.3389/fnagi.2016.00306

**49**

related resting-state functional connectivity (rsFC) decreases in the SM network (Tomasi and Volkow, 2012). Furthermore, several prior studies have found both motor performance and age to be associated with connectivity strength in older adults, suggesting that it may serve as a biomarker of brain health and functional performance (Langan et al., 2010; Seidler et al., 2015).

Age-related functional and structural declines in the SM system and their possible impact on sensorimotor performance are quite well documented (Seidler et al., 2010). During an isometric handgrip task, previous study demonstrated that activity in the contralateral primary motor cortex, cingulate sulcus and a premotor cortex co-varied positively with increasing force output in younger adults, but was less prominent in older adults (Ward et al., 2008). These findings possibly indicate a reduced ability to modulate activity in appropriate motor networks in older adults (Seidler et al., 2010). In recent studies, with increasing age, the reduced rsFC between the mid-posterior insula and subthalamic nucleus (Mathys et al., 2014), as well as SMA and central insula (Hoffstaedter et al., 2014) that plays an important role in sensorimotor integration processing (Deen et al., 2011; Chang et al., 2013; Uddin, 2015), have been thought to be associated with the age of older adults. Furthermore, Seidler et al. (2015) found that greater rsFC of SM system was linked to better motor performance in healthy older adults. It was thus concluded that changes in the resting-state of the SM system might contribute to the sensorimotor performance observed in older adults (Seidler et al., 2015).

Several previous structural and functional studies using magnetic resonance imaging (MRI) scans have also shown developmental trajectories in brain maturation and aging (Dosenbach et al., 2010; Rodrigue and Kennedy, 2011; Cao et al., 2014, 2016; Khundrakpam et al., 2015). Khundrakpam et al. (2015) found that the top predictors of brain maturity were found in highly localized sensorimotor (precentral and postcentral gyrus, insula) and association areas (including middle and superior frontal gyrus) in normally growing children and adolescents. Similarly, Dosenbach et al. (2010) reported that rsFC of SM network contributed in estimating chronological age in the typically developing volunteers. However, fewer studies have examined the predictive model of chronological age in healthy older adults (Qiu et al., 2015). Aging of the brain's structure over the course of the adult lifespan has been characterized by decreased gray matter volume (GMV) in prefrontal cortex and primary sensory cortices (Rodrigue and Kennedy, 2011). Changes in the resting-state of the SM system might contribute to estimate the age of older adults (Qiu et al., 2015). Based on the existing literature, it is important to ascertain the intrinsic rsFC patterns of the SM system in older adults. Thus, we hypothesized that participants with advanced age would demonstrate abnormal SM system connectivity; moreover, we further speculated that the age of older individuals would be predicted by decreased intrinsic functional connectivity of the SM system.

In the present study, to validate our hypothesis, a cohort of healthy aging subjects was recruited in resting state fMRI test. First, we analyzed resting state fMRI data to evaluate the impact of healthy aging on the primary sensorimotor system from global (independent component analysis, ICA) and local (seed-based functional connection analysis) aspects. In addition, we used machine learning approaches from two-aspect connectome-based prediction frameworks contain multivariate pattern analysis (MVPA) and univariate pattern analysis (UVPA) tools to examine brainbased predictors of individual differences in the age of older adults.

#### MATERIALS AND METHODS

#### Subjects

Two groups of test subjects were recruited for this study, including 68 healthy right-handed older adults [age (mean ± SD): 63.5 ± 6.5 years (51–78 years); the years of education: 9.9 ± 3.2 years (6–14 years); n = 37 females] and fifty-seven healthy right-handed younger adults [age: 20.5 ± 2.2 years (18–26 years); the years of education: 13.9 ± 1.2 years (13– 16 years); n = 28 females]. None of the participants had a history of substance abuse, neurological or psychiatric disorders. All older subjects were assessed using neuropsychological and health test batteries including the health scale named Chinese 36-item short-form health survey (SF-36), which consisted of 36 items and tapped eight health concepts (Li et al., 2003), and the neuropsychological test named Montreal Cognitive Assessment (MoCA), which was specifically developed to screen for mild cognitive impairment (Nasreddine et al., 2005). Subjects with poor performance on the SF-36 and low MoCA score (<25) were excluded from this study. All the participants gave informed consent and the research protocol was approved by the Ethics Committee of the University of Electronic Science and Technology of China. All subjects were financially compensated for their time.

#### Imaging Data Acquisition

Images were acquired on a 3T MRI scanner (GE DISCOVERY MR750) at the MRI Research Center of University of Electronic Science and Technology of China. During scanning, foam padding and ear plugs were used to reduce head motion and scanning noise, respectively. Resting state functional MRI data were acquired using gradientecho echo planar imaging sequences (repetition time [TR] = 2000 ms, echo time [TE] = 30 ms, flip angle [FA] = 90◦ , matrix = 64 × 64, field of view [FOV] = 24 cm × 24 cm, slice thickness/gap = 4 mm/0.4 mm), with an eight channel-phased array head coil. A 510-second resting state scan (yielding 255 volumes) was collected from each of the subjects. Subsequently, high-resolution T1-weighted images were acquired using a 3- dimensional fast spoiled gradient echo (T1-3D FSPGR) sequence (TR = 6.008 msec, FA = 9 ◦ , matrix = 256 × 256, FOV = 25.6 cm × 25.6 cm, slice thickness = 1 mm, no gap, 152 slices). During resting-state fMRI, all subjects were instructed to have their eyes-closed and to move as little as possible without falling asleep.

#### fMRI Preprocessing

fnagi-08-00306 December 28, 2016 Time: 15:57 # 3

Data preprocessing was performed using SPM8<sup>1</sup> (Statistical Parametric Mapping). The first five volumes were discarded for the magnetization equilibrium from all fMRI scans. A series of preprocessing steps was performed for each subject: (1) slice timing correction; (2) head motion correction; (3) normalization: in detail, the mean images resulted from the motion correction step were segmented into gray matter, white matter, and cerebrospinal fluid using the "unified segmentation" (Ashburner and Friston, 2005). Then, we could get the resulting parameters of a discrete cosine transformation, which defines the deformation field to move subject data into Montreal Neurological Institute (MNI) space. The deformation was subsequently applied to transform each echo planar imaging volume into the MNI singlesubject space. The resulted images were resampled at 3 mm isotropic voxel size; (4) images were smoothed by an 8-mm full width at half maximum Gaussian; (5) temporal filtering was performed in band-pass 0.01–0.08 Hz (Fox et al., 2005); (6) nuisance signals were regressed out, including white matter, cerebrospinal fluid and global signal, and six motion parameters. Subjects who had a maximum translation in any of the cardinal directions larger than 1 mm or a maximum rotation larger than 1 ◦ were excluded from subsequent analysis. In addition, we also assessed framewise displacement translation (FDtranslation) and framewise displacement rotation (FDrotation) in both groups using the following formula:

$$\begin{split} FD\_{\text{translation}/\text{rotation}} &= \\ &\frac{1}{M-1} \sum\_{i=2}^{M} \sqrt{|\Delta \, d\_{\text{x}}|^2 + |\Delta \, d\_{\text{y}}|^2 + |\Delta \, d\_{\text{z}}|^2} \end{split}$$

where M is the length of the time courses (M = 250 in this study), xi , y<sup>i</sup> , and z<sup>i</sup> are translations/rotations at the ith time point in the x, y, and z directions, respectively, 1 DX<sup>i</sup> = X<sup>i</sup> − X<sup>i</sup> <sup>−</sup> <sup>1</sup>, and similar for Dyi and Dzi.

#### GMV Calculation

Controlling functional connectivity maps by adding the GMV as a covariate in the rsFC analysis (Damoiseaux et al., 2008) could increase the reliability of resting state fMRI studies and indicate whether changes in rsFC maps are associated with brain atrophy. To obtain the GMV, T1 weighted images were processed using SPM8 toolbox with spatial normalization to MNI-space using a diffeomorphic anatomical registration through exponentiated lie algebra (DARTEL), and segmentation into gray matter, white matter and cerebrospinal fluid. The segmented gray matter images were modulated using nonlinear deformation. Individual GMV of the whole brain was calculated by setting a threshold at a probability of 80%.

#### Sensory-Motor Connectivity Analysis

The SM system is a common resting state network reported in previous studies. In general, there are two common approaches to identify this system: a data-driven method and a hypothesisdriven method. The ICA is selected for the former; the typical choice for the latter is the seed-based rsFC analysis with seed at the motor and somatosensory cortex. These two methods were adopted in this study to evaluate the rsFC of the sensorimotor system in younger and older adult subjects.

First, the data-driven method, ICA, was performed in two groups. Group spatial ICA (Calhoun et al., 2001b) was conducted using GIFT software<sup>2</sup> (Version 2.0). We used minimum description length (MDL) (Li et al., 2007) to validate the number of ICA components. For computational feasibility, principal component analysis was used to reduce data dimensionality. The infomax algorithm was repeated 30 times in ICASSO<sup>3</sup> and the resulting components were clustered to estimate the reliability of the decomposition. Finally, spatial maps and time courses were reconstructed for each subject using the group ICA (GICA) backreconstruction method based on principal component analysis compression and projection (Calhoun et al., 2001a). The sensorimotor network component were visually inspected and selected.

The resting state networks comprising the primary motor somatosensory cortices were estimated using a seed-based analysis. Based on our previous research work (Luo et al., 2012), four nearly spherical regions (radius 6 mm) were selected from the bilateral primary motor cortex (right M1, MNI coordinates [47–15 57]; left M1, MNI coordinates [–44–15 58]) for the motor network, and the bilateral primary somatosensory cortex (right S1, MNI coordinates [53–26 59]; left S1, MNI coordinates [–49–26 60]) in the somatosensory network. The mean BOLD time series were extracted from these seeds. Subsequently, rsFC analysis was performed between the seed and every voxels in the brain. The resulting correlation coefficients were transformed to approximate a Gaussian distribution using Fisher's r-to-z transformation.

#### Statistical Analysis

Statistical analysis of the rsFC was performed in SPM8 for both seed-based rsFC and ICA. First, the whole brain GMV, years of education and gender were regressed as the potential confounding covariates in the general linear model for each group to correct for the effects of atrophy, education and gender on subsequent rsFC analysis. Then, the within-group Z-values maps were analyzed with a random effect one-sample t-test. Statistical maps of significant connections with each seed were created for each group. A threshold of P < 0.05 (false discovery rate corrected, cluster size >23 adjacent voxels (621 mm<sup>3</sup> ) was set to identify the significant level. Second, a two-sample t-test was performed with an explicit mask from the union set of the onesample t-test results of the two groups. The significance threshold of group differences was set to P < 0.05 (false discovery rate corrected) and cluster size >23 adjacent voxels (621 mm<sup>3</sup> ) in the mask.

<sup>1</sup>http://www.fil.ion.ucl.ac.uk/spm/

<sup>2</sup>http://mialab.mrn.org/software/gift/

<sup>3</sup>http://research.ics.tkk.fi./ica/icasso

### Connectome-Based Prediction Framework

fnagi-08-00306 December 28, 2016 Time: 15:57 # 4

To investigate the underlying relationship between altered functional properties in the SM system and age in older adults, we used machine learning algorithms in this study. The leave-one-out-cross-validation (LOOCV) strategy was used to estimate prediction accuracies (Lachenbruch and Mickey, 1968). Prediction process consisted of two steps: training and testing. In the training step, each older adult were designated the test sample in turns while the remaining samples were used to train the predictor model. The altered rsFCs (false discovery rate corrected P < 0.05, cluster size >23), which resulted from ICA and seed-based analysis in the training step, were used as features. Then, in the testing step, we predicted the ages of remaining older adult using the same feature. To predict the age of older adults from the local and global change in the SM system, we conducted two-aspect connectome-based prediction frameworks using MVPA on all altered rsFCs and UVPA on single altered rsFCs, respectively. Specific, the MVPA is based on support vector machine (SVM), and UVPA is a machine learning approach combines LOOCV with linear regression.

#### Multivariate Variable Pattern Prediction Analysis

In this study, a support vector regression (SVR) procedure (Smola and Scholkopf, 2004) was used to derive a brain aging of older adults from multivariate pattern. SVR is a supervised learning technique based on the concept of SVM in order to make realvalued predictions. We used the ε–SVR algorithm implemented in LIBSVM (Chang and Lin, 2011) to calculate the regression model used for estimating the brain aging of older adults. To achieve generalized performance, SVR attempts to minimize the training error within the ε tolerance and the complexity of the regressor (Smola and Scholkopf, 2004). A linear kernel SVR was used in this study. The epsilon parameter was set to its default value, ε = 0.001. During LOOCV, each older adult was designated the test sample in turns while the remaining samples were used to train the SVR predictor; the trained regression model is used to predict the testing example.

In detail, to improve the performance of the predictor, we first selected the features and then evaluated the age predictions using two nested stratified LOOCV loops (Ambroise and McLachlan, 2002; Huttunen et al., 2012). The features were selected in the inner LOOCV loop and the age predictions were evaluated in the outer LOOCV loop thus avoiding the problem of training on testing data. For each inner LOOCV loop, the correlation coefficient of each feature with the chronological age was computed on the data that is the training set of the outer LOOCV loop. The features were then separately ranked by the absolute value of the correlation coefficients in descending order. The model goodness criterion, which was the number of the ranked features that used in the outer LOOCV loop, was the correlation coefficient (r) between the chronological and estimated age. The ranked features, which could get the highest r between the chronological and estimated age, were retained, while the rest were eliminated. Since features ranking was based on a different subset of data for each of inner LOOCV, the selected features was slightly different among results of each inner

#### Univariate Pattern Prediction Analysis

We also conducted connectome-based prediction frameworks from univariate pattern based on single feature that was the altered rsFC in the older compared with younger adults in the training step. Here, a machine learning approach with LOOCV was used together with linear regression (Cohen et al., 2010). The age variable for older adults was referred to as "label". LOOCV was performed with this label. The dependent variable (age of older adults) and the independent variable (averaged value) were inputted into a linear regression algorithm. A linear regression model was established using altered rsFC chosen from the training step. Predicted values were obtained for the remaining older adult. This procedure was repeated to obtain a final result. The technical details are provided in Supplemental Information (see Results).

#### Model Prediction Evaluation

The two-aspect (MVPA and UVPA) models' accuracy in predicting older adults' age according to altered rsFCs were evaluated using two statistical measures. First, Pearson correlation coefficient [r(predicted, observed) ] was computed between chronological and estimated age. A nonparametric testing approach was used to test the null hypothesis of no significant correlation. The chronological ages were randomly permuted 1000 times, and the entire prediction process was carried out with each one of the randomized prediction labels. The statistical significance (p-values) of the permutation test represent the probability of observing the reported accuracy by chance [(number of permutation r(predicted, observed) < observed r(predicted, observed) ) + 1)/(number of permutations + 1)]. Only an extent threshold p < 0.05 is reported. Second, the mean absolute error (MAE) which measures the average magnitude of errors between chronological age and model predicted age was calculated. Low MAE value means better prediction than high MAE value.

#### Validation: Reproducibility

There is currently no consensus over whether the whole brain signal should be removed in the preprocessing of the resting-state fMRI data. The global signal is confounded with physiological noise, which has been reported by several studies (Birn et al., 2006), and should be removed (Fox et al., 2009). On the other hand, other studies have suggested that global signal regression (GSR) could introduce negative rsFC (Murphy et al., 2009; Weissenbacher et al., 2009), and is associated with the neuronal signal (Schölvinck et al., 2010). To ensure that the results were not outcome of GSR, we constructed the resting state networks of the primary motor somatosensory cortices using a seedbased rsFC analysis without GSR. Then, we also recomputed the prediction analysis, which included two-aspect connectomebased frameworks from MVPA and UVPA tools, based on the altered rsFCs resulted from GSR.

We further added the GMV as a control feature, combined with all altered rsFC features, decreased rsFC features of insula,


TABLE 1 | Significantly decreased functional connections among the SM network in older adults compared with younger adults.

BA, Brodmann area.

as well as increased rsFC features, respectively, in the UVPA and MVPA tools to compare the prediction contribution of the increased and decreased features in the altered SM system of older adults.

### RESULTS

#### Participant Fundamental Information

Sixteen older adults were excluded because of low MoCA score (five subjects), poor performance on the SF-36 (three subjects), and excessive head motion (eight subjects). Six younger adults were also excluded because of excessive head motion. Thus, 52 older subjects [age (mean ± SD): 63.2 ± 5.8 years (51–76 years), n = 30 females] and 51 younger subjects [age: 18–26 years (21.5 ± 1.9 years), n = 26 females] were included in further rsFC analysis. In addition, we compared the FDtranslation and FDrotation values between the remained subjects of two groups to evaluate the homogeneity of head motion between two groups. There were no significant differences between the two groups concerning FD values (two-sample two-tailed t-tests, T = 1.02, P = 0.31 for FDtranslation, and T = 1.20, P = 0.23 for FDrotation). There also were no significant differences between the two groups in gender (Chi square test, P = 0.49). Younger adults had more years of education compared with older adults (two-sample two-tailed t-tests, T = 9.35, P < 0.001). Compared with younger adults, significantly decreased whole GMV was found in older adults (two-tailed t-test, T = 4.32, P < 0.001).

### Analysis of Sensorimotor Network from ICA Analysis

Using GICA, 36 components were estimated by MDL criterion (Li et al., 2007), which include default mode network, auditory network, sensorimotor network, visual network, cerebellum network, and frontal-parietal network, for both groups. Because this study focused on SM system, the independent component (IC 15) including the supplementary motor area, sensorimotor cortex, and secondary somatosensory cortex, was selected as SM network, which is consistent with previous results (Smith et al., 2009). Compared with the younger adults, the older

group showed the significantly decreased functional connections among the main regions in the SM network, including SMA, pre/postcentral, superior parietal lobule, mid-posterior insula, and rolandic operculum (**Table 1**; **Figure 1**).

### Seed Based Functional Connectivity Analysis

The within-group rsFC maps were generated for each group. In the younger adults, the bilateral M1 was positively correlated with

the pre- and postcentral gyrus, middle occipital gyrus, superior temporal gyrus, SMA, putamen, and insula (**Figures 2A,B**). In the older adults, the bilateral M1 was positively correlated with similar brain regions such as in younger adults (**Figures 2A,B**). Relative to the younger adults, the older adults showed significantly increased rsFC seeded at bilateral M1 to pre- and postcentral gyrus and superior parietal lobule, while decreased rsFC was detected in the bilateral insula and rolandic regions (**Table 2**; **Figures 2A,B**). In the younger adults, the signals from pre- and postcentral gyrus, superior frontal gyrus, SMA, and insula were positively correlated with the signals from bilateral S1 (**Figures 2C,D**). In the older adults, the bilateral S1 were positively correlated with similar brain regions to those of the younger adults (**Figures 2C,D**). Compared to the younger adults, significantly increased connections were observed among the primary sensorimotor cortex and superior parietal lobule, while decreased connections were detected in the bilateral insula and rolandic regions (**Table 2**; **Figures 2C,D**). These results were largely preserved after accounting for the effects of global signal removal (**Figure 3**; Supplementary Table S1). Other details are provided in Supplemental Information (see Intrinsic Functional Connectivity Without Global Signal Regression Analysis).

To compare the contribution of the significantly increased and decreased rsFC in the altered SM system of older adults, MVPA was used in this study, since the contribution would be positively related with the performance of classifier (the detailed processing see Section "Comparison between Increased rsFC and Decreased rsFC through Multivariate Classification" in Supplemental Information). SVM classifiers were adopted here to classify older adults from younger adults using increased functional connections and decreased

#### TABLE 2 | Significant differences for resting-state functional connections with bilateral M1 and S1 in older adults compared with younger adults.


BA, Brodmann area

functional connections as features, respectively. Results show that linear SVM classifier with decreased rsFC score feature performs better than linear SVM classifier with increased rsFC score feature in terms of accuracy, sensitivity, specificity, and AUC value (**Table 3**; **Figure 4**). Other details are provided in Supplemental Information (see Comparison between Increased rsFC and Decreased rsFC through Multivariate Classification).

### Prediction of Older Adult's Chronological Age

We further examined the intrinsic functional connectivity of the sensorimotor system in relation to age in the older adults. According to the differences between groups in the training step, regions with significantly altered rsFC were chosen for the following machine learning prediction analysis: bilateral mid-posterior insula, superior parietal lobule, SMA, and pre/postcentral resulted from seed-based rsFC comparison, superior parietal lobule, mid-posterior insula, SMA and postcentral resulting from ICA comparison.

The result of MVPA [r(predicted, observed) = 0.463, p < 0.001, MAE = 3.993, **Figure 5**] represents that the age of older adults could be predicted through the features which come from fifty altered rsFC features. Five consensus features (left insula and left M1, left insula and right M1, left insula and right S1, right insula and left M1, right insula and left S1), which were used in the outer LOOCV loop, were observed. Furthermore, the univariate pattern connectome-based prediction analysis also revealed that, in older adults, age could be reliably predicted by the decreased

rsFC value in the right mid-posterior insula resulting from ICA analysis [r(predicted, observed) = 0.237, p = 0.026, MAE = 4.698, **Figure 5**), as well as through decreased rsFC values between sensorimotor cortex and bilateral mid-posterior insula (**Table 4**; **Figure 5**).

In addition, these results were also largely preserved after accounting for the effects of global signal removal (Supplementary Table S2; Supplementary Figure S1). More details are provided in Supplemental Information (see Detailed UVPA Prediction Steps and Results). The UVPA results resulted from features, which are not significant through permutation test, are provided in Supplemental Information (see Materials; Supplementary Tables S3.1–S3.3). Finally, the prediction analyses, which are based on different sets of features, show that the prediction results with insular features performs better than other sets of features. Other details are provided in Supplemental Information (see The Prediction Results Based on Different Sets of Feature; Supplementary Tables S4.1,S4.2).

#### DISCUSSION

Our results demonstrated that normal aging is associated with declining functional integration in the primary SM system using resting-state fMRI, and the individual age of older adults can



TABLE 4 | Resting-state functional connectivity (rsFC) predicts the age of older adults.

the univariate feature resulted from ICA analysis. The red lines denote the mean ± SD values (4.471 ± 0.25) of MAE.


UVPA, univariate pattern analysis; rsFC, resting-state functional connectivity; Ins, insula; MAE, mean absolute error

be reliably predicted by the intrinsic functional connectivity of mid-posterior insula through both MVPA and UVPA. The primary SM system was identified and evaluated in terms of two common approaches: ICA and seed-based rsFC analysis. The findings resulting from these two methods revealed robust age effects, indicating that decreases in primary SM system integration correspond with increasing age. In contrast to the declining function of the primary SM system, increased rsFC among primary sensorimotor regions were also found through seed-based rsFC analysis, which revealed that older adults might need a higher degree of anticipatory preparation for the declining sensorimotor function (Mathys et al., 2014; Song et al., 2014). These changes in rsFC might reflect a remodeling of function of the SM system with aging. These findings suggest that the functional connectivity of mid-posterior insula is modified with aging. These findings might provide further insight into changes in primary sensorimotor function underlying rest activity with aging.

The altered functional property of mid-posterior insula in primary SM system observed here is strikingly similar to previous findings. With increasing age, the reduced rsFC between SMA and central insula (Hoffstaedter et al., 2014), cerebellar seed and insula (Seidler et al., 2015), as well as posterior insula and SMA and other sensorimotor regions (Roski et al., 2013) may be associated with general impairments in somatosensory processing in older adults. In the present study, the significantly decreased rsFC of the mid-posterior insula was observed in both ICA and seed-based rsFC analysis in older adults relative to younger adults. The human insula cortex forms a distinct lobe and involves three major functionally distinct sub-regions (Chang et al., 2013). As one of the three sub-regions, the mid-posterior insula region is associated with sensorimotor processing (Stephani et al., 2011). The mid-posterior insula, a more high-level region in sensorimotor processing than the primary sensorimotor cortex, plays an important role in sensorimotor integration processing (Kurth et al., 2010; Nieuwenhuys, 2012; Chang et al., 2013). The mid-posterior insula has also been ascribed an integrative role, linking information from diverse sensorimotor functional regions and playing an important role in sensorimotor processing (Nieuwenhuys, 2012; Chang et al., 2013). Altogether, the key nodes of the SM network, the bilateral mid-posterior insula, showed strongly reduced rsFC in older adults. These findings might reflect that older adults loosened the integration of sensorimotor processing and indicate a reduced ability to modulate activity in the appropriate region of the sensorimotor system. In addition, functional differentiation of the insula cortex was already indicated by recent excellent studies (Nieuwenhuys, 2012; Chang et al., 2013). It is thought to play a role in functional integration between different functional systems by integrating information from diverse functional systems (Nieuwenhuys, 2012). It was reported to be involved in not only processing of the reciprocal influence of emotion and interoception, but also integrating between cognitive tasks

and emotion as well as sensation (Critchley, 2005). The decreased rsFC of the insula observed in the current study may influence the interregional integration among attention, emotion or other functional system in the older adults. This speculation was also validated in children. For example, the mid-posterior insula could mediate empathy when children observed a signal indicating others were receiving a pain stimulus by associating it with fronto-parietal attention network (Decety et al., 2008). These findings might be important for the future studies in cognitive disorders and healthy aging.

Furthermore, several previous studies which were either cortical thickness (Rodrigue and Kennedy, 2011; Khundrakpam et al., 2015) or activation fMRI studies (Dosenbach et al., 2010; Qiu et al., 2015) have reported that the SM system contributes to estimate the age of young adults and aging subjects. In this study, we found that the age of older adults could be predicted by decreased rsFC value between the mid-posterior insula and primary sensorimotor cortex, as well as decreased rsFC value of mid-posterior insula resulted from ICA analysis. The functional property of bilateral midposterior insula is the exclusive consensus features in the stage of selecting features in MVPA. These analyses revealed that weakening connections of mid-posterior insula contributed more to predicting the age of older adults than other features in SM system. Our findings provide new evidence that functional connectivity of mid-posterior insula in SM system is associated with the individual age of older adults. Interestingly, machine learning approaches revealed that the rsFC of mid-posterior insula in the SM system could also predict the age in older adults.

The primary somatosensory cortex is considered to be the main area of the SM system (Allen et al., 2011). Some previous studies based on ICA have demonstrated that the decreased integration of the SM network may be associated with perceptual impairments in patients with neurological disease (Luo et al., 2011; Li et al., 2015). Our findings from ICA also reflect the declining functional integration in sensorimotor areas in aging. Moreover, the decreased rsFC between the somatosensory cortex and mid-posterior insula was observed through seed-based rsFC analysis. Several recent studies have reported increasing rsFC in SM system with age (Langan et al., 2010; Song et al., 2014; Zhang et al., 2015). Hoffstaedter et al. (2014) reported that each S1/M1 showed age-related decrease of resting state rsFC with primary sensorimotor regions, while right S1/M1 featured agedependent increase of rsFC with SMA, superior parietal lobule. In several studies, increasing sensorimotor connectivity with age has been suggested to be compensatory (Mathys et al., 2014; Song et al., 2014). We have also documented that increased rsFC was found in some primary SM regions through the seed-based rsFC analysis. Although these results were different with the findings from ICA, both methods could evaluate the impact of healthy aging on the SM system from different aspects (global and local aspect). The observed results from ICA reveal that the declining functional integration (global aspect) was observed in SM system in aging. The findings from seedbased analysis might indicate that the increased rsFC (local aspect) in aging responds to the declining sensorimotor function. Some researchers also found the relationship between increased interhemispheric motor rsFC and reductions in interhemispheric inhibition with age (Fling and Seidler, 2012), suggesting that the increased rsFC may derive from age-related declines in inhibitory neurotransmitters. In addition, the linear SVM classifier with decreased rsFC score feature performs better than linear SVM classifier with increased rsFC score feature. The contribution of significantly changed SM system with decrease functional connectivity is stronger than that with increased functional connectivity.

Noteworthy, rsFC was related with behavior performance outside the MRI scanner (Seidler et al., 2015). Resting state connectivity could be regarded as offering a potential prediction indicator for task performance. Actually, some studies have reported that rsFC provided pre-task brain activation level, which was partly consistent with subsequent task results (Langan et al., 2010; Wang et al., 2010). Specifically, stronger resting state rsFC in hippocampal network might predict better memory task performance (Wang et al., 2010). Our findings of altered rsFC in primary SM system of older adults may be associated with the common decline of motor performance in aging.

### LIMITATIONS

While we believe our findings provide a further insight into changes in SM system in the healthy aging brain, there are a number of important caveats in interpreting these results. First, physiological noise should be considered in the rsFC analysis. In the present study we cannot eliminate cardiac and respiratory fluctuations completely through temporal filtering (band-pass 0.01–0.08 Hz). Second, we only delineated trajectories of the aging based on altered rsFC within primary SM system. We could not conclude that the top predictors were highly localized in primary SM system in healthy aging elder adults. The important regions or networks will also be examined in aging through the machine learning framework in future. Third, the current approach investigates the age in a cross-sectional rather than longitudinal fashion. However, we are following this cohort of older adults and will acquire data each year. The progressive effect of aging in the remodeling of rsFC in the SM system will be considered in future studies through a longitudinal analysis. Finally, testing for motor-related function was not included in the current study. Though no significant relationships were observed after we measured the association between the behavior features (the scores of SF-36 and MoCA) and age of older adults and altered rsFCs. Our findings may involve a confusion, in which the declined motor performance in older adults would affect the associations observed here. However, the physical functioning (PF) scores, which were extracted from the SF-36 test, may reflect a health scale about motor-related function to some degree. Thus, the partial correlations between the age of older adults and change rsFCs value were calculated, accounting for the effects of gender, years of education, whole brain GMV, and PF. Likewise, the relationship between the rsFCs of insula and the age of older adults were also found (Supplementary Table S5). More details are provided in Supplemental Information (see Correlations between Functional Properties and the Age of Older Adults Controlling for the Physical Functioning Related with Motor). The PF value was not the comprehensive behavior performance outside the MRI scanner. This defect would be investigated in the future study.

#### CONCLUSION

fnagi-08-00306 December 28, 2016 Time: 15:57 # 11

We analyzed the rsFC changes in the SM system in older adults compared to younger adults, which demonstrated significant remodeling of resting state primary sensorimotor system. The altered rsFC may be suggestive of the loosened integration of sensorimotor processing and might also imply the compensation in the primary sensorimotor network in older adults. Furthermore, we demonstrated that the MVPA and UVPA extract sufficient information from these decreased rsFC to make reliable predictions about individuals' chronological age across healthy aging. This study may help to investigate the potential reorganization of the SM system in the brain of older adults.

#### ETHICS STATEMENT

The study was approved by the Ethics Committee of University of Electronic Science and Technology of China in accordance with the Helsinki Declaration. Written informed consent was obtained

#### REFERENCES


from each patient and control subject. All the participants were volunteers. They were recruited from the local communities.

### AUTHOR CONTRIBUTIONS

Conceived and designed the work: HH, CL, BB, MB, DY. Acquired the data: XC, WC, JG, BB. Analyzed the data: HH, CL. Wrote the paper: HH, CL. All authors revised the work for important intellectual content. All of the authors have read and approved the manuscript.

### FUNDING

This work was supported by grants from the National Nature Science Foundation of China (grant number 81330032, 81271547, 91232725); Special-Funded Program on National Key Scientific Instruments and Equipment Development of China (grant number 2013YQ490859); the '111' project of China (grant number B12027) and the Program for Changjiang Scholars and Innovative Research Team (grant number IRT0910).

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2016.00306/full#supplementary-material

normal aging. Front. Aging Neurosci. 6:280. doi: 10.3389/fnagi.2014. 00280


anticorrelated functional networks. Proc. Natl. Acad. Sci. U.S.A. 102, 9673–9678. doi: 10.1073/pnas.0504136102


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 He, Luo, Chang, Shan, Cao, Gong, Klugah-Brown, Bobes, Biswal and Yao. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnagi-08-00306 December 28, 2016 Time: 15:57 # 12

# Neuroanatomical and Neuropsychological Markers of Amnestic MCI: A Three-Year Longitudinal Study in Individuals Unaware of Cognitive Decline

Katharina S. Goerlich<sup>1</sup> \* † , Mikhail Votinov1,2,3† , Ellen Dicks1,4, Sinika Ellendt<sup>1</sup> , Gábor Csukly<sup>5</sup> and Ute Habel1,2

<sup>1</sup> Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany, <sup>2</sup> Jülich Aachen Research Alliance (JARA) – Translational Brain Medicine, Aachen, Germany, <sup>3</sup> Institute of Neuroscience and Medicine (INM-10), Research Centre Jülich, Jülich, Germany, <sup>4</sup> Alzheimer Center and Department of Neurology, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, Netherlands, <sup>5</sup> Department of Psychiatry and Psychotherapy, Semmelweis University, Budapest, Hungary

Structural brain changes underlying mild cognitive impairment (MCI) have been wellresearched, but most previous studies required subjective cognitive complaints (SCC) as a diagnostic criterion, diagnosed MCI based on a single screening test or lacked analyses in relation to neuropsychological impairment. This longitudinal voxel-based morphometry study aimed to overcome these limitations: The relationship between regional gray matter (GM) atrophy and behavioral performance was investigated over the course of 3 years in individuals unaware of cognitive decline, identified as amnestic MCI based on an extensive neuropsychological test battery. Region of interest analyses revealed GM atrophy in the left amygdala, hippocampus, and parahippocampus in MCI individuals compared to normally aging participants, which was specifically related to verbal memory impairment and evident already at the first measurement point. These findings demonstrate that GM atrophy is detectable in individuals with amnestic MCI despite unawareness of beginning cognitive decline. Thus, individuals with GM atrophy in regions associated with verbal memory impairment do not necessarily need to experience SCC before meeting neuropsychological criteria for MCI. These results have important implications for future research and diagnostic procedures of MCI.

Keywords: mild cognitive impairment, voxel-based morphometry, subjective cognitive complaints, gray matter atrophy, amygdala, hippocampus, longitudinal

### INTRODUCTION

Age-related neurodegenerative diseases such as Alzheimer's disease (AD) impose a high social and financial burden for society that will increase in the following decades, given predictions of a 9% increase of people above 60% in 2050 (United Nations, Department of Economic and Social Affairs, 2013). Along with population aging, the prevalence of AD (currently 4.7%) is predicted to increase by approximately 225% by 2050 (Prince et al., 2013).

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Kiyotaka Nemoto, University of Tsukuba, Japan Mario Clerici, University of Milan, Italy

> \*Correspondence: Katharina S. Goerlich kgoerlich@ukaachen.de †Shared first authorship

Received: 05 December 2016 Accepted: 08 February 2017 Published: 22 February 2017

#### Citation:

Goerlich KS, Votinov M, Dicks E, Ellendt S, Csukly G and Habel U (2017) Neuroanatomical and Neuropsychological Markers of Amnestic MCI: A Three-Year Longitudinal Study in Individuals Unaware of Cognitive Decline. Front. Aging Neurosci. 9:34. doi: 10.3389/fnagi.2017.00034

Memory disturbances constitute early symptoms of AD, and with progressing impairment, other domains become affected, including language, problem-solving, and visuospatial perception (for a review, see Dubois et al., 2016). The neuropathological hallmarks of AD are brain atrophy, extracellular amyloid plaques, and neurofibrillary tangles (Kandel et al., 2013), yet its underlying cause remains unidentified (Zetterberg and Mattsson, 2014). Early diagnosis is in dire need for disease prevention and the development of new treatment strategies. Detecting AD at an early time point would enable early intervention and a timely start of treatment, possibly preventing disease progression.

A promising endeavor to provide such early diagnosis lies in the identification of the transitional state between normal aging and pathological cognitive decline. Since 1997, the term mild cognitive impairment (MCI) is used to mark the difference between continuous levels of cognitive impairment in normally aging controls compared to AD patients (Petersen et al., 1997). The concept of MCI is now widely accepted and continues to receive great attention in the literature as it represents a possible treatment target for AD. Its predictive power is reflected in annual conversion rates of up to 15% from MCI to AD (DeCarli, 2003) compared to conversion rates of 1–2% from normal aging to AD (Petersen et al., 1999).

Currently, MCI is diagnosed by: (a) abnormal cognitive function adjusted for age and education level, (b) self-reported cognitive complaints, (c) normal activities of daily living, and (d) absence of a dementia diagnosis (Winblad et al., 2004). Impairment severity is commonly assessed on the basis of general cognitive screening tests, such as the Mini-Mental State Examination (MMSE). However, such tests have been criticized for being not sufficiently specific regarding the subtle nature of MCI (Nickl-Jockschat et al., 2012; Forlenza et al., 2013). Even if specific neuropsychological tests for MCI diagnosis are used, studies often rely on only one diagnostic test (Bondi et al., 2014). Moreover, there is great variation in how the criteria for aberrant cognitive function are implemented. Thus, there is no standardized approach for diagnosing MCI, and prevalence rates are therefore highly dependent on the classification scheme used (Chapman et al., 2016; Clark et al., 2016). Furthermore, MCI was originally conceptualized as a prodromal stage for AD and thus focused on memory impairment (Petersen et al., 1999). However, MCI has been revealed to be a much broader construct affecting several cognitive domains (Winblad et al., 2004). Thus, extensive neuropsychological testing is required to identify early-stage MCI.

Critically, doubts have arisen regarding subjective cognitive complaints (SCC), which are currently implemented in the diagnosis of MCI. While these may be a marker of MCI (Clark et al., 2016), including SCC may lead to false omission of possible MCI candidates who are misdiagnosed as healthily aging individuals because they are not yet aware of beginning cognitive impairment (Edmonds et al., 2014). Such a lack of awareness has indeed been identified by several studies (e.g., Purser et al., 2006; Lenehan et al., 2012; Fragkiadaki et al., 2016; for a review, see Roberts et al., 2009), and the inclusion of SCC resulted in increased rates of false negative and false positive diagnoses (Lenehan et al., 2012). Moreover, subjective complaints are strongly related to individual differences in depression and neuroticism, casting further doubt on their reliability as a diagnostic marker of MCI (Reid and MacLullich, 2006). Taken together, the drawbacks of including SCC in the diagnosis of MCI seem to outweigh its benefits. Thus, the present study took a novel approach by investigating MCI in elderly individuals unaware of any cognitive impairment.

In addition to more thorough neuropsychological testing, the need for neuroanatomical biomarkers of MCI has been recognized (Ruan et al., 2016). For AD, the most widely studied biomarkers are: decreased cerebrospinal fluid amyloid beta (CSF Aβ), increased CSF tau, decreased fluorodeoxyglucose uptake on positron-emission tomography (FDG-PET), PET amyloid imaging, and structural MRI measures of cerebral atrophy. Importantly, strong evidence suggests that MRI, FDG-PET, and CSF tau biomarkers are already abnormal in the MCI phase of AD (Ewers et al., 2007; Sluimer et al., 2008; Shaw et al., 2009), and while both CSF tau and MRI are predictive of conversion from MCI to AD, the predictive power of structural MRI is greater. Moreover, cognitive measures correlate strongly with structural MRI, but not with CSF tau in patients with MCI (Vemuri et al., 2009). Structural MRI studies using longitudinal voxel-based morphometry (VBM) identified the medial temporal lobe (MTL) as a core region for progression of MCI to AD (Ferreira et al., 2011; Nickl-Jockschat et al., 2012; Yang et al., 2012). This is not surprising given that the MTL contains essential structures for memory consolidation (Kandel et al., 2013). The identification of progressing MCI patients can be made apparent already 4 years prior to conversion to AD on the basis of hippocampal atrophy patterns (Adaszewski et al., 2013). Comparing rates of hippocampal atrophy between healthy controls, MCI individuals and AD patients, annual atrophy rates of 1.9–2.8% were observed for the controls, 2.6–3.7% for MCI individuals, and 3.5% for AD patients (Jack et al., 2000). Moreover, meta-analyses indicate that a consistent atrophy pattern underlies MCI, comprising the amygdala, hippocampus, precuneus, and posterior cingulate gyrus (Nickl-Jockschat et al., 2012; Yang et al., 2012).

Although gray matter (GM) changes accompanying MCI have been well-researched, most of the previous studies identified MCI on the basis of general cognitive screening tests such as the MMSE, which may be insufficient to capture the subtle nature of MCI and presumably identifies MCI at a rather late stage, missing the beginning of cognitive decline. Even if the MMSE was not used for MCI diagnosis, previous VBM studies correlated GM atrophy with cognitive decline based on the MMSE and thus lack analyses in relation to specific behavioral impairment (e.g., Jack et al., 2000; Whitwell et al., 2007; Sluimer et al., 2008; Weiner et al., 2013).

The present study aimed to overcome these limitations by taking a novel, more sensitive approach: Elderly volunteers (minimum age: 50 years) from a community sample were included who had no subjective complaints, i.e., were not aware of potential cognitive impairment. MCI was then identified on the basis of a thorough neuropsychological test battery assessing memory, intelligence, executive functions, psychomotor speed, visuo-construction and visuo-spatial skills, attention, and language. To our knowledge, this is the

first VBM study investigating the longitudinal trajectory of morphological changes underlying early-stage MCI in individuals still unaware of beginning cognitive impairment. We predicted more pronounced region of interest (ROI) GM atrophy in individuals classified as MCI compared to HCs even without subjective awareness of cognitive deficits, and hypothesized faster atrophy rates in the MCI compared to the HC group. Moreover, we hypothesized ROI GM atrophy to be related to specific behavioral impairment in the MCI group only.

#### MATERIALS AND METHODS

#### Participants

Participants for the Helmholtz Alliance for Mental Health in an Aging Society (HelMA) study (Drexler et al., 2013; Chechko et al., 2014; Ellendt et al., 2016) were recruited through visitations to social facilities for elderly people (charity organizations and citizen centers) and advertisements in local newspapers. From 81 volunteers initially participating in the first measurement time point (T1), 43 participants (27 women, aged 50–83 years) eventually completed all three measurement time points (T1, T2, and T3) with a mean follow-up interval of 1.12 years, SD 0.38 years. Reasons for dropping out varied from refusing to further participate for time reasons or a lack of interest (13 participants), moving to another city (four participants), newly acquired MR contraindications following surgery during the course of the study (six participants), to illness (11 participants) and death (four participants).

Participants were included if they were 50 years or older, had sufficient German language and adequate visual performance abilities. Exclusion criteria comprised a diagnosis of dementia, neurological or psychiatric disorders according to DSM-IV axis I as assessed by the German version of Structured Clinical Interview (SKID; Wittchen et al., 1997), physical disease that could interfere with cognitive performance, lifetime or current drug addiction, seriously reduced vision, inability to follow the protocol, and medication use with possible cognitive side effects. The study was approved by the ethics committee of the medical faculty, RWTH Aachen University. All participants gave written informed consent and were paid for participation.

#### Neuropsychological Data Analysis

In addition to several dementia screening tests applied with the objective of eliminating dementia rather than diagnosing MCI, all participants underwent an extensive neuropsychological test battery evaluating memory, intelligence, executive functions, psychomotor speed, visuo-construction and visuo-spatial skills, attention, and language at each time point. **Table 1** provides an overview of all neuropsychological tests used for MCI assessment and diagnosis (for further details, see Drexler et al., 2013; Ellendt et al., 2016). The testing procedure was accomplished by trained psychologists. To counteract learning effects, tests were presented in counterbalanced order. Sessions took place in the mornings and lasted approximately 3 h. Short breaks were offered and if necessary, a second appointment was arranged.

TABLE 1 | Neuropsychological test battery used for the assessment and diagnosis of mild cognitive impairment (MCI).


The classification criteria for MCI were based on those described by Winblad et al. (2004). That is, (1) greater memory impairment than expected for age, (2) preserved activities of the daily living, and (3) absence of dementia. The criterion of memory impairment greater than expected for age was identified by an impaired score of at least 1.5 standard deviations (SD) below the mean according to normative datasets in at least one test assessing memory functioning (i.e., VLMT, Benton, WMS, memory assessing subtests of the CERAD-Plus battery). If the neuropsychological test comprised multiple subtests (as in VLMT and CERAD), at least two subtests had to indicate impairment, i.e., 1.5 SD below normative data, which is considered a conservative method for identifying MCI (Jak et al., 2009). The decision to classify subjects solely on their performance in tests regarding memory functioning can be accounted on the fact that amnestic MCI patients show a higher prevalence to develop AD compared to non-amnestic MCI patients, who have an increased risk of developing other types of dementia (Petersen and Morris, 2005). Note that impairment in cognitive domains other than memory was not observed in this sample. Thus, all MCI participants identified here can be classified as belonging to the single-domain amnestic MCI subtype. Correspondingly, the present results are of relevance specifically to amnestic MCI.

The original classification criteria proposed by Winblad et al. (2004) also include SCC but because the requirement of SCCs could lead to omitting subjects which would otherwise be classified as MCI (Edmonds et al., 2014), this criterion was not applied. In fact, all participants reported not to have experienced any difficulties regarding memory or any other cognitive domain beyond expected for their respective age. Thus, none of the participants were aware of any cognitive impairment.

The neuropsychological data were analyzed in SPSS 20 by means of multivariate analyses of covariance (MANCOVA) comparing performance between the MCI group and the HC group at each time point, corrected for age, gender, and education. The initial significance threshold was p < 0.05, and all results were Bonferroni corrected for multiple comparisons.

#### VBM Data Analysis

fnagi-09-00034 February 22, 2017 Time: 10:31 # 4

T1 anatomical images (TE: 3.03; TR: 2300 ms; FOV = 256 mm × 256 mm; number of sagittal slices = 176; voxel size: 1 mm × 1 mm × 1 mm) from the 43 study participants were acquired on a 3 Tesla Siemens <sup>R</sup> Trio MR scanner. Differences in GM volume between the MCI and HC group were assessed using longitudinal VBM (Ashburner and Friston, 2000; Ashburner and Ridgway, 2013) implemented in SPM12 (Wellcome Trust Centre for Neuroimaging, University College London, London, UK). The following preprocessing steps were applied: First, serial longitudinal registration, which produces one midpoint file and Jacobian determinants for each subject at each of the three time points. Next, each subject's midpoint average was segmented into GM, white matter (WM), and CSF. Then, GM images were computed for each subject at each time point using the respective Jacobian determinants. These images were spatially normalized by creating a customized group-specific template using the DARTEL approach and warping each of the individual GM segmentations onto this template. The warped GM segmentations were modulated to reflect the volume and smoothed using a Gaussian kernel of 8 mm at full width at half-maximum (FWHM). Data quality was ensured by visually checking each T1 image for abnormalities before preprocessing, checking volume orientation before smoothing during preprocessing, and by means of a sample homogeneity check after preprocessing, revealing no outliers.

Region of interest analyses were performed on GM volumes of eight anatomically defined a priori ROIs based on MCI meta-analyses (Nickl-Jockschat et al., 2012; Weiner et al., 2013): Left and right amygdala, hippocampus, parahippocampus, and precuneus. Anatomical ROIs for these regions were created using the automatic anatomic labeling (AAL) atlas templates (Tzourio-Mazoyer et al., 2002) provided by the WFU Pickatlas toolbox (Wake Forest School of Medicine, Winston Salem, NC, USA). Mean parameter estimates from these ROIs were extracted using MarsBaR<sup>1</sup> and analyzed by means of a Linear Mixed Model (LMM) Analysis in SAS 9.2. The factor group (MCI versus HC) was included in the LMM as a between-subjects fixed effect, and the model controlled for age, gender, education, and mean total intracranial volume (TIV, i.e., GM + WM + CSF). An unstructured covariance matrix was used, which was estimated by means of the null model estimation method.

Relationships between mean neuropsychological performance and mean GM volumes of the eight ROIs across the whole sample and within the HC and the MCI group were examined by means of partial correlations controlling for age, gender, education, and TIV. To this end, the residuals of each neuropsychological test as well as the residuals of the GM volumes of each ROI at each measurement point were calculated, controlling for age, gender, education, and TIV, and partial correlations were then performed between the neuropsychological test residuals and the ROI GM volume residuals. The Holm–Bonferroni correction was applied to control for multiple comparisons, resulting in an initial significance threshold of p < 0.006 (p = 0.05/n = 8 ROIs). Moreover, whole-brain analyses comparing GM volumes between the MCI and HC group at each time point were conducted at an uncorrected threshold p < 0.001.

## RESULTS

### Neuropsychological Performance

Based on the neuropsychological test battery, 16 participants were classified as MCI (6 male, mean age 66.13, SD 8.46 years, mean years of education 9.43, SD 1.59), and 27 participants were classified as HC (10 male, mean age 66.15, SD 6.15 years, mean years of education 10.5, SD 1.88). Groups did not differ in age (p > 0.05), but the HC group had more years of education (p < 0.05). The results on the neuropsychological tests for the HC and the MCI group are presented in **Table 2**, revealing significantly worse performance of the MCI group compared to the HC group on the total immediate recall and delayed recall subtests of the VLMT and the CERAD. At a lower significance threshold (p < 0.05 uncorrected), worse performance was additionally observed in the Benton test (correct drawings and number of mistakes) and the WMS-R subtests digit span forward and backward. The MMSE could not distinguish between the MCI and the HC group, as opposed to the TFDD, which indicated lower performance in MCI compared to HC participants (see **Table 2**). Note that none of the study participants converted to dementia during the course of the investigation.

### ROI GM Volumes

The LMM analyses on each ROI demonstrated a linear decrease in all ROI GM volumes in both groups over the course of the 3 years of observation. A group analysis between the MCI and HC group revealed significantly lower GM volume in the left amygdala, left hippocampus, and left parahippocampus in MCI participants compared to HCs at all three time points (see **Table 3**). Thus, MCI participants had significantly lower GM volume than controls in these areas already at the first time point, which then continued to decrease over the course of the following 2 years (effect of time in these ROIs: p < 0.001; see **Figure 1A**). However, there was no significant interaction between group and time point (p > 0.05), suggesting that atrophy rates were comparable between groups. Wholebrain analyses additionally indicated smaller left amygdala and left hippocampal volumes in the MCI compared to the HC group at each time point (see **Figure 1B**). No further activation outside of the ROIs was observed at p < 0.001. For visualization, **Figure 2** shows the decrease of ROI GM volumes across the age range of 50–83 years at T<sup>1</sup> in MCI and HC participants.

### Associations between ROI GM Volumes and Neuropsychological Performance

Relating mean neuropsychological performance with mean ROI GM volumes across all 3 years within the whole sample showed

<sup>1</sup>http://marsbar.sourceforge.net/

#### TABLE 2 | Neuropsychological test results for the MCI group compared to the HC group.


∗∗p < 0.006 Bonferroni-corrected, <sup>∗</sup>p < 0.05 uncorrected.

that the association between CERAD total immediate recall performance and larger GM volume in the left amygdala (partial r = 0.49, p < 0.002), left hippocampus (partial r = 0.51, p < 0.001), and left parahippocampus (partialr = 0.52, p < 0.001) remained, indicating a robust relationship between CERAD total immediate recall performance and GM volume in these regions. In addition, WMS-R digit span forward performance was linked to larger left parahippocampal volume (partial r = 0.44, p < 0.006).

Partial correlations between neuropsychological performance and ROI GM volumes within each group revealed that the abovementioned associations were solely driven by the MCI group as there were no significant correlations between neuropsychological performance and ROI GM volume in the HC group (all p > 0.05). In the MCI group, CERAD total immediate recall performance was highly correlated with GM volume in the left amygdala (partial r = 0.83, p < 0.001), left hippocampus (partial r = 0.72, p < 0.006), and left parahippocampus (partial


∗∗p < 0.006 Bonferroni-corrected, <sup>∗</sup>p < 0.05 uncorrected.

r = 0.80, p < 0.002), implying a robust relationship between verbal memory impairment and GM volume reduction in these regions. MMSE scores showed no correlation with ROI GM volumes (all p > 0.05). **Figure 3** visualizes the relationship between mean CERAD total immediate recall performance and mean GM volumes of the left amygdala, hippocampus, and parahippocampus in the MCI and HC group across the three time points.

To find out to what extent GM atrophy in these ROIs significantly predicted verbal memory impairment in MCI participants already at the first measurement point, a stepwise linear regression was performed with the residuals (i.e., controlling for age, gender, education, and TIV) of CERAD total immediate recall performance at T<sup>1</sup> as dependent variable and the residuals of GM volumes of the left amygdala, hippocampus, and parahippocampus at T<sup>1</sup> as predictors. From this analysis, a significant model emerged [F(1,24) = 5.79, p < 0.05, R <sup>2</sup> = 0.29], revealing that GM atrophy specifically in the left parahippocampus accounted for 29% of the variance in CERAD total immediate recall performance (β = 0.54, p < 0.05). Thus, left parahippocampal atrophy significantly predicted verbal memory impairment in the MCI group already at the first measurement point.

#### DISCUSSION

The aim of this longitudinal imaging study was to identify the relationship between neuropsychological and neuroanatomical changes associated with early stages of MCI identified by means of a thorough neuropsychological test battery in individuals unaware of cognitive impairment. Although several previous studies investigated the longitudinal trajectory of neuroanatomical alterations underlying MCI, these studies diagnosed MCI on the basis of a single general screening tests, lacked an analysis of these changes in relation to neuropsychological performance (e.g., Besson et al., 2015; Callahan et al., 2015; Dugger et al., 2015; Fellhauer et al., 2015; Granziera et al., 2015), or included subjective complaints as a diagnostic criterion. The present study aimed to overcome these limitations. Individuals classified as MCI who were unaware of beginning cognitive impairment exhibited verbal memory deficits (indicative of amnestic MCI) and GM atrophy in the left amygdala, hippocampus, and parahippocampus, compared to controls. Atrophy rates were comparable between groups, in contrast to our hypothesis of faster atrophy rates in MCI individuals than in controls. In line with our prediction of a specific relationship between GM atrophy and behavioral impairment in MCI individuals only, GM atrophy in the MCI group, but not the control group, was highly correlated with impaired verbal memory (CERAD total immediate recall). Moreover, GM atrophy in the left parahippocampus significantly predicted verbal memory impairment in MCI individuals already at T1, even without subjective awareness of cognitive impairment. This confirms our prediction of GM atrophy being evident in MCI individuals before they become aware of cognitive decline.

Lower GM volumes in the left amygdala, hippocampus, and parahippocampus are in accordance with previous studies reporting GM atrophy in the left MTL in MCI patients compared to controls (Ferreira et al., 2011; Nickl-Jockschat et al., 2012; Weiner et al., 2013; Csukly et al., 2016). Here, GM volume reduction in these regions was significantly linked to impaired total immediate recall performance (CERAD) in the MCI group, suggesting that GM atrophy in the MTL is specifically associated

with verbal memory impairment. This finding complements previous observations that cerebral atrophy correlates with measures of general cognition in MCI (Jack et al., 2010). Atrophy rates did not significantly differ between MCI participants and controls, although this might be expected with progressing MCI (e.g., Trivedi et al., 2006; for a review, see Chetelat and Baron, 2003). However, this absence of differences in atrophy rates may be explained by the restricted time frame of the present study, comprising 3 years, which was probably not long enough to capture significant differences in atrophy rates. Moreover, none of the study participants converted to dementia during the course of the investigation, indicating that indeed early-stage MCI was captured here, thus further explaining the absence of group differences in GM atrophy rates. This could also explain why no GM atrophy in the precuneus was identified here, suggesting that atrophy in this region may occur at later stages of MCI. Note

that MCI participants exhibited lower GM volumes compared to controls already at the first time point. This suggests that GM atrophy in the MCI group had already begun prior to enrolment in the study, warranting the inclusion of younger participants (<50 years) in future studies to be able to identify the point of divergence in medial temporal GM atrophy between MCI and normal aging.

According to the model of dynamic biomarkers of AD proposed by Jack et al. (2010), structural MRI is the last biomarker in the staging of the disease to become abnormal, preceded by Aβ-plaque biomarkers that are dynamic in early stages before the appearance of clinical symptoms, and by biomarkers of neurodegeneration that occur at later stages and correlate with symptom severity. MRI, FDG-PET, and CSF tau biomarkers are already abnormal in the MCI phase preceding AD. Findings of specific relationships between GM atrophy and neuropsychological impairment as observed in this study may be used in the future to increase prediction accuracy for conversion from MCI to AD.

The MMSE, a general cognitive screening test that has been commonly used to assess MCI in previous research did not show any correlations with GM volumes in the present MCI group and could not distinguish between the MCI and HC group: Both groups had an average score of 29 on the MMSE and MMSE scores remained intact over the course of the 3 years, with none of the participants having less than 27 points on the MMSE at any time point. Thus, the present MCI sample would not have been identified as such based on the results of this general screening test, despite evident cognitive impairment as demonstrated by significantly worse performance on the VLMT and CERAD (and tendencies toward impairment in the Benton and WMS-R). This is in line with previous findings showing that despite good sensitivity and specificity for diagnosing dementia, the commonly used cutoff scores of the MMSE do not show good accuracy for discrimination of MCI, misidentifying most of these subjects as having normal cognitive function (Kaufer et al., 2008). Since the MMSE is widely used in the MCI literature – and considering that individuals are classified as MCI with scores ranging from 23.1 to 28.7 (Nickl-Jockschat et al., 2012) – this not only entails the risk to include already more severely impaired individuals in MCI studies, but also causes problems for the comparison and evaluation of the results due to high variability in MCI diagnosis. Future studies should thus avoid relying on the MMSE alone and instead use a broader, more sensitive neuropsychological test battery to diagnose MCI and assess its severity. According to the present results, the total immediate and delayed recall scales of the VLMT and the CERAD seem especially sensitive to identifying early stage MCI in the absence of subjective awareness of cognitive decline. In contrast to the MMSE, the TFDD, a screening test aiming to detect early signs of dementia while differentiating cognitive problems due to depression, could distinguish between the MCI and the HC group in this study, although no between group differences on depression were detected. Thus, it may be worthwhile using the TFDD as an additional screening tool in future studies.

Importantly, SCCs are currently a diagnostic criterion for MCI. None of the individuals classified as MCI in this study were aware of any cognitive impairment (i.e., had no subjective complaints). Yet, the MCI group showed significantly worse performance on more than one test of the neuropsychological test battery. Confirming previous findings (Edmonds et al., 2014), this supports the idea to not include SCC as a diagnostic criterion for MCI in future studies in order to identify GM atrophy underlying early-stage MCI even without subjective awareness of beginning cognitive impairment.

#### Limitations

The sample size of the present study was small due to a dropout rate of almost 50%, resulting in only 43 participants (16 MCI) who completed all three measurement time points. High dropout rates are an inherent risk in longitudinal studies, particularly those including elderly individuals. Future studies should thus aim to include more as well as younger (<50 years) participants in order to pinpoint at what age structural brain changes in relation to beginning cognitive impairment become evident and divergent from normal aging.

### CONCLUSION

The present findings indicate that GM atrophy in the left MTL underlying MCI is specifically associated with verbal memory impairment. This underlines the importance of combining neuroanatomical markers of MCI with specific neuropsychological tests as it implies that age-related GM atrophy is only predictive of MCI if accompanied by specific cognitive deficits. Moreover, our findings show that reductions in GM volume are evident even if individuals are not yet aware of cognitive impairment, demonstrating that individuals with neuroanatomical evidence of atrophy in regions associated with verbal memory impairments do not necessarily need to experience subjective cognitive concerns before meeting neuropsychological criteria for MCI. Further, general cognitive screening test such as the MMSE may not be sensitive enough to identify early-stage MCI. These findings have important clinical implications as they highlight the need to discard SCCs from MCI diagnosis. Moreover, it would be useful to apply comprehensive neuropsychological batteries, possibly by means of computerized cognitive assessments rather than relying on a single cognitive screening test to identify MCI. Lastly, establishing regular neuropsychological testing for MCI already in individuals below the age of 50 years will help identify MCI at early stages, enabling early intervention and a timely start of treatment.

### AUTHOR CONTRIBUTIONS

Conception: UH. Organization: UH, KG, and MV. Execution: KG, MV, ED, and SE. Statistical Analysis: KG, MV, and GC. Manuscript: Writing of the first draft: KG and MV. Review and Critique: KG, MV, ED, SE, GC, and UH. All authors approved the final version of the manuscript.

### ACKNOWLEDGMENT

fnagi-09-00034 February 22, 2017 Time: 10:31 # 11

This work was supported by the Helmholtz Alliance "Mental Health in an Ageing Society," funded by the Initiative and

### REFERENCES


Networking Fund of the Helmholtz Association, and by the Brain Imaging Facility, a core facility of the Interdisciplinary Center for Clinical Research (IZKF) Aachen within the Faculty of Medicine at RWTH Aachen University.

findings of a community based sample. Curr. Alzheimer Res. doi: 10.2174/ 1567205014666161213120807 [Epub ahead of print].



predicting future clinical change. Neurology 73, 294–301. doi: 10.1212/WNL. 0b013e3181af79fb


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Goerlich, Votinov, Dicks, Ellendt, Csukly and Habel. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# White Matter Deterioration May Foreshadow Impairment of Emotional Valence Determination in Early-Stage Dementia of the Alzheimer Type

Ravi Rajmohan<sup>1</sup> \*, Ronald C. Anderson<sup>2</sup> , Dan Fang<sup>3</sup> , Austin G. Meyer <sup>4</sup> , Pavis Laengvejkal <sup>5</sup> , Parunyou Julayanont <sup>5</sup> , Greg Hannabas <sup>6</sup> , Kitten Linton<sup>7</sup> , John Culberson<sup>7</sup> , Hafiz M. R. Khan<sup>6</sup> , John De Toledo<sup>5</sup> , P. Hemachandra Reddy 1,8,9,10 and Michael O'Boyle1,3

<sup>1</sup>Department of Pharmacology and Neuroscience, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>2</sup>Department of Electrical and Computer Engineering, Texas Tech University, Lubbock, TX, USA, <sup>3</sup>Department of Human Development and Family Studies, Texas Tech University, Lubbock, TX, USA, <sup>4</sup>School of Medicine, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>5</sup>Department of Neurology, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>6</sup>Department of Public Health, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>7</sup>Department of Family Medicine, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>8</sup>Garrison Institute on Aging, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>9</sup>Cell Biology and Biochemistry, Texas Tech University Health Sciences Center, Lubbock, TX, USA, <sup>10</sup>Speech, Language and Hearing Sciences, Texas Tech University Health Sciences Center, Lubbock, TX, USA

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Michael Noll-Hussong, University of Ulm, Germany Jidan Zhong, Krembil Research Institute, Canada

> \*Correspondence: Ravi Rajmohan ravi.rajmohan@ttuhsc.edu

Received: 19 December 2016 Accepted: 10 February 2017 Published: 01 March 2017

#### Citation:

Rajmohan R, Anderson RC, Fang D, Meyer AG, Laengvejkal P, Julayanont P, Hannabas G, Linton K, Culberson J, Khan HMR, De Toledo J, Reddy PH and O'Boyle M (2017) White Matter Deterioration May Foreshadow Impairment of Emotional Valence Determination in Early-Stage Dementia of the Alzheimer Type. Front. Aging Neurosci. 9:37. doi: 10.3389/fnagi.2017.00037 In Alzheimer Disease (AD), non-verbal skills often remain intact for far longer than verbally mediated processes. Four (1 female, 3 males) participants with early-stage Clinically Diagnosed Dementia of the Alzheimer Type (CDDAT) and eight neurotypicals (NTs; 4 females, 4 males) completed the emotional valence determination test (EVDT) while undergoing BOLD functional magnetic resonance imaging (fMRI). We expected CDDAT participants to perform just as well as NTs on the EVDT, and to display increased activity within the bilateral amygdala and right anterior cingulate cortex (r-ACC). We hypothesized that such activity would reflect an increased reliance on these structures to compensate for on-going neuronal loss in frontoparietal regions due to the disease. We used diffusion tensor imaging (DTI) to determine if white matter (WM) damage had occurred in frontoparietal regions as well. CDDAT participants had similar behavioral performance and no differences were observed in brain activity or connectivity patterns within the amygdalae or r-ACC. Decreased fractional anisotropy (FA) values were noted, however, for the bilateral superior longitudinal fasciculi and posterior cingulate cortex (PCC). We interpret these findings to suggest that emotional valence determination and non-verbal skill sets are largely intact at this stage of the disease, but signs foreshadowing future decline were revealed by possible WM deterioration. Understanding how non-verbal skill sets are altered, while remaining largely intact, offers new insights into how non-verbal communication may be more successfully implemented in the care of AD patients and highlights the potential role of DTI as a presymptomatic biomarker.

Keywords: Alzheimer, chimeric faces, brain networks, DTI, fMRI, emotional valence

## INTRODUCTION

Alzheimer disease (AD) is marked by a progressive decline in cognitive functions. As this occurs, the affected individual becomes less capable of understanding the world around them. This occurs in part due to a loss of ability to interpret the body language and facial expressions of others. Face-processing represents a group of complex cognitive operations in which information is extracted from facial features in such a way that the observer is able to perceive pertinent information about the person they are viewing. Previous studies of moderate-stage AD patients showed that their ability to interpret basic emotional cues from faces is largely intact (Roudier et al., 1998; Luzzi et al., 2007). However, dysfunctions are detectable (Albert et al., 1991; Cadieux and Greve, 1997; Hargrave et al., 2002). Luzzi et al. (2007)showed that the ability to interpret emotional cues directly correlated to participants' performance on constructional praxis and visuospatial memory tasks, which are examples of nonverbally-mediated skills, in early-moderate-stage AD.

These findings, combined with the observation that stroke participants' performance on certain visuospatial tests correlates with the degree of localization of damage to the right parietal lobe (Luzzi et al., 1998), suggest that activity of the right parietal lobe is directly related to the interpretation of emotional cues. This concurs with both the current clinical and pathological picture of AD, which shows that non-verbal memory loss, as well as damage to the right parietal lobe, is a late-stage finding (Braak et al., 2006; Ally et al., 2009).

We attempted to correlate the non-verbal skills of participants with early-stage Clinically Diagnosed Dementia of the Alzheimer Type (CDDAT) with findings from both functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) measures during a non-verbally mediated face-processing task. Participants underwent a well-established, standardized, non-verbal skill test (the Rey-Osterrieth Complex figure B, visuospatial memory test (ROCFB-VMT); Luzzi et al., 2007) before entering the scanner to establish their visuospatial memory baseline. Participants then performed a variation of the Chimeric Face Test (CFT; Levy et al., 1983) that we created called the ''emotional valence determination task (EVDT)'' while in the scanner.

CDDAT patients were expected to perform as well as or slightly worse than neurotypicals (NTs) on the EVDT. Neuroimaging data was expected to show functional underactivation within the frontal (inferior frontal; left middle frontal) and parietal lobes (bilateral supramarginal gyri and left precuneus), given their association with non-verbal emotional processing, as well as increased activation of the bilateral amygdala and right anterior cingulate cortex (r-ACC), reflecting a greater reliance on these structures, during the EVDT (Nakamura et al., 1999; Fusar-Poli et al., 2009). Likewise, we hypothesized that DTI may show some deterioration of white matter (WM) tracts within frontoparietal connections, but not within amygdaloidal or cingulate tracts. Performance on the EVDT was expected to correlate to ROCFB-VMT performance. We hypothesized that this represents a relationship between emotional valence determination and visuospatial memory. By investigating the potential association of early-stage AD neuroimaging findings and non-verbal skill performance, we assessed the usefulness of these tests as cost effective supportive diagnostic markers of AD.

As may be expected, most task-related studies of AD focused on memory and displayed a pattern of early compensation followed by decreased activity. Pariente et al. (2005) noted increased activation during episodic memory encoding and recall in the posterior cingulate cortex (PCC), precuneus, parietal lobe and frontal lobe in early-stage AD patients with a mean Mini Mental State Exam (MMSE) score of 25 ± 1.8. Grady et al. (2003), Dickerson et al. (2005), Celone et al. (2006) and Zhou et al. (2008), on the other hand, observed decreased or even absent activation relative to NTs in the same regions in earlystage AD patients with mean MMSE scores of 21.1 ± 3.1, 22 ± 5, 21.3 ± 2.7, respectively. These findings support the assertion of Risacher and Saykin (2013) that increased activation occurs early on in an attempt to compensate, but ultimately leads to decreased activation once the disease burden is too great. We reasoned that this pattern of initially increased activation leading to decreased activation could be demonstrated for other cognitive domains as well because it represents a fundamental mechanism of the energetics of neurodegeneration. For a comprehensive review of AD discoveries through neuroimaging see Risacher and Saykin (2013).

Of equal importance is the understanding of how WM tracts are affected by the disease. WM alterations appear to parallel gray matter (GM) changes in that cortical abnormalities are greater in posterior brain regions relative to anterior regions at the earlystages of AD (Arnold et al., 1991; Braak and Braak, 1995). When the disease progresses, the neurofibrillary pathology advances from limbic to frontal structures, into higher-order association cortices, and finally into primary sensorimotor areas, which correlates with the clinical manifestations of AD (Braak and Braak, 1995). DTI-based tractography studies (Fellgiebel et al., 2005) and whole-brain DTI studies (Medina et al., 2006; Rose et al., 2006; Zhang et al., 2007) have consistently shown that fibers located deep in the posterior WM (e.g., the superior longitudinal fasciculus (SLF) and the posterior cingulum bundle (PCB)) are affected in patients with AD and mild cognitive impairment (MCI), a common precursor to AD. Bartzokis et al. (2003, 2004) have proposed that this may occur because as brain development takes place, later myelinated regions (cortical association areas) have fewer oligodendrocytes supporting greater numbers of axons (Bartzokis, 2004). DTI findings of decreased WM integrity in later myelinated regions at the onset of AD support this ''reversed demyelination'' construct (Medina and Gaviria, 2008). Furthermore, Huang et al. (2007) delineated a neuroanatomical pattern of functional alterations showing that changes in WM diffusion in parietal lobes correlated with scores of visuospatial skills.

Research on amnestic type MCI (aMCI) populations using neuropsychological tests of declarative memory extend this trend as they have demonstrated significant correlations between declining performance and decreases in posterior WM fractional anisotropy (FA), particularly in the PCB (Fellgiebel et al., 2005, 2008, Rose et al., 2006). Recalling the proposal of Bartzokis et al. (2004), disruptions in transcortical connectivity may serve as early contributors to the pathophysiology of dementia, as the observed WM deteriorations were embedded beneath cortical GM that is often affected early within the disease course. For a more in-depth review of DTI findings in AD, see Medina and Gaviria (2008).

Although memory deficits are a hallmark characteristic of AD, there may be little to gain from testing face-processing related to familiarity that isn't already known (Sperling et al., 2003; Golby et al., 2005; Winchester, 2009; Donix et al., 2013). On the other hand, the fundamentals of face-processing (e.g., emotional valence) have received little attention (Job, 2012) and warrant further investigation. The use of neuropsychological tests in combination with neuroimaging techniques stresses the importance of attempting to integrate pathological observations with clinical symptoms. In doing so, we are able to reinforce findings from either end of the spectrum to more efficiently develop our understanding of the disease. In order to investigate a cognitive operation as complex as face-processing, it will be necessary to use appropriate tests that can isolate its specific subdivisions. This is particularly important when performing an fMRI investigation, as some studies suggest that face-processing and recognition occurs in fractions of a second (Vuilleumier and Schwartz, 2001; Batty and Taylor, 2003). If such factors are not properly accounted for, it would be very difficult to remove these confounds from the fMRI data as the canonical hemodynamic response curve peaks between 4–6 s after the presented stimuli, making it far too slow to tease out these processes (Poldrack et al., 2011).

To this end, the EVDT was of particular value since it is a variation of the CFT. The CFT was originally developed by Levy et al. (1983) ''to index functional cerebral asymmetry for processing facial characteristics.'' It was shown to be highly reliable in detecting differences between right- and left-handers as well as being stable with regard to individual differences in perceptual asymmetries (Levy et al., 1983).

CFTs have previously been used to investigate asymmetries in the processing of emotions from facial expressions. Early works by Albert et al. (1991) and Cadieux and Greve (1997) suggested that by the moderate-stage, AD patients were impaired in recognizing emotions, but the authors noted that they could not rule out confounds ''due to the deficits in recognizing non-emotional facial features and in verbal processing''. Roudier et al. (1998) demonstrated that moderate-stage AD patients could accurately recognize when different emotions were displayed using the same human face. Hargrave et al. (2002) reconciled this difference through the use of a ''same-different'' emotion differentiation task in which participants were presented with a pair of photographs of different people and were asked to state if the two photographs in the pair were depicting the same or different emotions. From this, it was determined that moderatestage AD patients do, in fact, have difficulty differentiating emotions independent of verbally-mediated and non-emotional facial features. Indersmitten and Gur (2003) then determined that two separate circuits likely underlie emotional processing in facial asymmetries and, while the left hemiface/right cerebral hemisphere circuit was dominant for expressions of sad, fearful and happy, the right hemiface/left cerebral hemisphere proved to be more efficient on task performance in NTs.

Then, Luzzi et al. (2007) used a variation of the CFT called the Mona Lisa test (MLT) wherein they determined that when using a cartoon face, emotional valence determination was impaired in some participants with moderate-stage AD, but not significantly across the entire cohort. Luzzi et al. (2007) found that in those with impaired recognition of facial emotions, the impairment correlated to poor performance on constructional praxis and non-verbal memory test, but not to the verbal fluency test (VFT) or MMSE score. Therefore, the ability to recognize facial emotions correlated to non-verbal performance. Finally, a fMRI meta-analysis by Fusar-Poli et al. (2009) discovered that the processing of facial expressions for emotional valence was associated with neural activation in the parietal and frontal cortices in NTs, thereby making these ideal regions of interest for investigation of the emotional subdivision of face-processing.

All things considered, we chose to use a ''same-different'' task similar to Hargrave et al. (2002) to observe changes in brain activity and connectivity in early-stage CDDAT patients as it represents the most clinically relevant task given its translatability to real life scenarios. Their work did not incorporate neuroimaging techniques, however, representing a gap in knowledge which we intended to fill with the current study. The works of Roudier et al. (1998), Indersmitten and Gur (2003), Luzzi et al. (2007) and Fusar-Poli et al. (2009) established crucial observations that were necessary to proceed in such a manner as without them there would be too many potential confounds to consider using depictions of actual human faces, let alone a series of different faces, to assess emotional processing through neuroimaging. Additionally, by using non-chimeric variations of faces based on the Mattingley et al. (1993) chimeric faces, we more directly assessed the emotional valence components of face-processing than we could through the original CFT.

The ROCFB served as an ideal cognitive test for investigating the non-verbal underpinnings of emotional valence determination because it had been described by Luzzi et al. (2011) to be ''a valid instrument to assess non-verbal memory in adults and in the elderly''. Building upon the findings of Luzzi et al. (2007), we investigated the relationship between the ROCFB-VMT and the EVDT. We assessed the potential relationship between visuospatial memory (as measured by the ROCFB-VMT) and the EVDT to determine if the ability to discern emotional states from facial expressions is rooted in the brain's ability to recognize and retain visuospatial relations between shapes.

#### MATERIALS AND METHODS

#### Participant Identification and Selection

The University Medical Center Departments of Neurology, Family Medicine, and Geriatrics saw 171 patients for complaints of ''memory problems'' or ''dementia'' spanning a 6-month period (November 2015–April 2016). Potential participants were identified and classified into their respective categories by physician assessment. Of that, 100 were determined to have AD or its precursor, aMCI, in accordance with the guidelines outlined by McKhann et al. (2011) for ''dementia due to AD'' or ''MCI due to AD'' (Albert et al., 2011). Those in the possible early-stage AD category were selected based on an MMSE score (Folstein et al., 1975) of 26–21 in accordance with National Institute for health Care and Excellence guidelines (National Institute for Health and Care Excellence, 2011). Twenty of these patients were determined to fit our inclusion/exclusion criteria, four of whom agreed to participate. All CDDAT participants had an existing medical MRI scan interpreted by a radiologist within the last 5 years prior to the study. These individuals were grouped into the CDDAT group. Detailed inclusion/exclusion criteria for the CDDAT group are listed below:

Inclusion criteria


Exclusion criteria


Ten age-matched cognitively normal participants (5 males, 5 females) were recruited from a nearby senior living center and were grouped into the NT group. NTs were required to have an MMSE score >27 and meet all inclusion and exclusion criteria stated above, except for that concerning the existence of Alzheimer-related pathology. One male participant was retroactively removed after receiving a diagnosis of normal pressure hydrocephalus. One female participant from the NT group was removed due to a scanner script runtime error that prevented the EVDT from being run. Therefore, there were 4 CDDAT (1 female, 3 males) and 8 NT (4 females, 4 males) participants in this study. All participants were right-handed, except for 1 non-right-hand-dominant NT female, as confirmed by the Edinburgh Handedness Inventory-revised (EHI-r; Williams, 1986). The non-righthand-dominant NT female that was kept in the study was kept because the participant was shown to have no significant difference in either their accuracy, reaction time (RT), or within-group contrast mapping for either fMRI or DTI measures when compared to the rest of the control group.

Participants were informed that their participation was voluntary and they may withdraw from the study at any time and that their refusal to participate would have no impact on their level of care. This study was carried out in accordance with the recommendations of Texas Tech University Human Protections Internal Review Board with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol study was approved by the Texas Tech University Human Protections Internal Review Board.

#### Sample Size

A 2004 meta-analysis by Henry et al. (2004) demonstrated an effect size of r = 0.73 for a test of semantic fluency in patients with Dementia of the Alzheimer Type (DAT). From this it was determined that an assumed effect size of r = 0.8 would achieve 80% power with a minimum sample size of 4. Our behavioral analyses, therefore, seek to determine if an impairment on either the EVDT or ROCFB-VMT is greater than or equal to the known deficit of verbal fluency, as this currently represents the most consistent and pronounced clinical manifestation of the disease at this stage. Effect size for each test was calculated using Cohen's d and converted to its r-equivalent (Cohen, 1992).

The number of necessary participants was based on the Desmond and Glover (2002) report that single voxel-level power of 80%, corrected for multiple comparisons can be achieved with a sample size of 24 per group. According to Desmond and Glover (2002), ''Sample size increases power because the standard error of the mean decreases by the square root of N'' (N = the number of subjects). In our study, we had a minimum of four subjects within each group. Therefore, we may expect to reach of 32.7% power at the single voxel level (i.e., since we have 1/6 of the number of subjects we will reach 41% of the power achieved by having 24 per group or 33% overall; the square root of 4/24 = 0.41<sup>∗</sup> 0.8 = 0.328). As such, even by the most conservative of estimates (assuming a power of 10 exponential) our minimum detection for 80% power is 316 voxels (80%/32%= 2.5; 102.5= 316). Given that this is a region-of-interest-based investigation (ROI) >1000 voxels, however, we may reasonably expect to surpass 80% power for each ROI with a sample size of ≥4 per group.

#### Participant Demographics

Participant demographics are summarized in **Table 1**. We recruited four right-handed Caucasian patients (3 males, 1 female) with a diagnosis of aMCI or early-stage AD between the ages of 73–93 (median age 83.5 ± 8.4) with MMSE scores between 26–23 (median of 24.5 ± 1.3). Two of the males (MMSE scores of 26 and 25) had an existing diagnosis of ''MCI due to AD'' (Albert et al., 2011). The remaining male and female had an existing diagnosis of ''dementia due to AD'' (MMSE scores of 24 and 23, respectively; McKhann et al., 2011). Eight (7 right-handed and 1 non-right-hand-dominant) cognitively



CDDAT, clinically diagnosed dementia of the Alzheimer type; NT, neurotypical; F, female; M, male; R, right-handed; nR, non-right-handed; CA, Caucasian.

normal Caucasian participants (4 males, 4 females; all with MMSE scores of 30) between the ages of 79–91 (median age 80 ± 4.0) also participated. Handedness was determined using the EHI-r (Williams, 1986).

There was no significant difference in median age (p = 0.561), but there was an observable difference in median MMSE score for the two groups (p = 0.014). While there may be some concern given our lack of equal sex-distribution (1 female and 3 males) within the CDDAT group, no differences in performance were discernable based on sex within our NT group (4 females, 4 males) across either cognitive test (ROCFB-VMT or EVDT) for either accuracy or RT (data not shown) nor for withingroup sex-based contrast mapping for fMRI or DTI (data not shown).

#### Administration of Rey Osterrieth Complex Figure B Visuospatial Memory Test

Participants were first asked to copy the ROCFB as accurately as possible, without tracing it, on a different sheet of paper that had the same dimensions as the paper containing the figure. The size of the model was 5.5 cm × 8 cm, and the sheet of paper on which it was printed was half the size of an 8.5<sup>00</sup> × 11<sup>00</sup> sheet. Participants were then asked to reproduce the ROCFB from memory on a different sheet of paper after an interval of 5 min, during which they were engaged with the VFT (Luzzi et al., 2007). The rate of retention was calculated as a percentage of the score obtained on reproduction from memory compared to the original figurecopying score. Digital photographs were taken of the drawings produced by participants to ensure accuracy of scoring. The test was administered and scored following the protocol detailed in Luzzi et al. (2011).

### Imaging Methodology and Analyses

#### Stimulus Presentation and Participant Response

After completing the ROCFB-VMT, all participants were briefly trained on the EVDT before entering the scanner to ensure task comprehension by allowing them to practice on a single example trial of each condition type. Pairs of faces based on the chimeric faces of Mattingley et al. (1993) were presented in an eventrelated design for the EVDT using Eprime 2.0. Presentation order was counterbalanced across participants using a Latin square design. The question ''Are the emotions the same?'' was displayed upon the screen for 4 s inside the scanner. A set of 40 trails were randomized and displayed for 4 s (s) each with a jittered interstimulus interval (ISI) randomized for five time points between 800–1200 ms, by 100 ms apiece. Participants responded using a fiber optic controller held in the right hand where button 1 was pressed by the index finger and button 2 was pressed by the middle finger. Participants pressed one of two buttons to indicate Yes (button 1) or No (button 2) in response to the question. An example of this test is given in **Figure 1**.

#### Scanning Parameters

All images were acquired with a 3T Siemens MR system (Skyra, Germany) at the Texas Tech Neuroimaging Institute. The T1 anatomic scan parameters were: TR: 1900 ms; TE: 2.49 ms; FOV: 240; Flip angle: 9, Voxel size = 0.9 × 0.9 × 0.9 mm; slice number: 192. The fMRI parameters were: TR: 2500 ms; TE: 20.0 ms; FOV: 231; Flip angle: 75, Voxel size = 2.5 × 2.5 × 3.0 mm; slice number: 41. In each fMRI dataset, there are 80 volumes. The DTI parameters were: TR: 5000 ms; TE: 95 ms; FOV: 220; B/W: 1562, Voxel size = 1.7 × 1.7 × 4.0 mm; slice number: 32. There were 64 directions for DTI.

#### Image Preprocessing **fMRI**

Image preprocessing steps included removing non-brain structures by Brain Extraction Tool (BET), motion correction by using Motion Correction for FSL Linear Registration Tool (MCFLIRT), temporal high-pass filtering with a cutoff period of 24 s, spatial smoothing with a 5 mm Gaussian full width, half maximum (FWHM) algorithm, and co-registering of the functional images to the high resolution T1 structure images in their native space using boundary border registration (BBR) and FSL Linear Registration Tool (FLIRT; Jenkinson and Smith, 2001; Jenkinson et al., 2002) at 12 degrees of freedom to the Korean Normal Elderly (KNE96; Lee et al., 2016) standard brain space.

#### **DTI**

FSL Diffusion Toolbox 3.0 (FDT) from FMRIB Software Library (FSL 5.0.5) was used to complete the construction of and preprocessing for all anatomical brain networks for all subjects. It processed DICOM/NIfTI files into diffusion metrics (e.g., FA) that were ready for statistical analysis at the voxel-level after performing corrections for image alignment and artifact clean-up (Top-Up) and local field distortions (eddy current correction; Smith et al., 2004).

#### Image Processing **fMRI**

fMRI data processing was carried out using FMRI Expert Analysis Tool (FEAT) Version 6.00, part of FSL<sup>1</sup> (Worsley, 2001). The time series for the behavioral events of the EVDT were analyzed for the following conditions:

<sup>1</sup>www.fmrib.ox.ac.uk/fsl

FIGURE 1 | Examples of the emotional valence determination test (EVDT). EVDT examples in response to the question: "Are the emotions the same?" (left) "yes"; two neutral faces. (center) "yes"; two happy faces. (right) "no"; happy (top) and neutral (bottom).

For ''presentation of stimuli scenarios'' (i.e., when the participant was shown two happy faces, two neutral faces, or one of each)—brain activity from the first 2 s following the presentation of a stimulus was recorded for each trial. In the event a participant responded within <2 s, the interval between the presentation of the stimulus and 250 ms before the response was used for the recording interval. This was done to avoid brain activity artifacts related to the button press. Instances where a participant made no response before the presentation of the next stimulus were discarded.

For ''participant response'' scenarios (i.e., when the participant chose ''yes'' (y) or ''no'' (n) in response to the question: ''Are the emotions the same?'')—brain activity from the first 2 s following the press of a button was recorded for each trial. In the event a participant responded within <2 s before the presentation of the next stimulus, that interval was used for the recoding time. Instances where a participant made no response before the presentation of the next stimulus were discarded.

A summary of fMRI contrasts is given in **Table 2**. A total of five subject-level contrast maps were created for all subjects using threshold free cluster enhancement (TFCE) of z > 2.3 and

TABLE 2 | Summary of functional magnetic resonance imaging (fMRI) contrasts.


A total of five subject-level contrast maps were created for all subjects. EVDT, emotional valence determination test.

a cluster corrected significance threshold of p < 0.05. These modeled time series were convolved with the double gamma hemodynamic response function (dg-HRF), which was modeled from a combination of Gaussian functions. Group-level contrast maps were created using FMRIB's Local Analysis of Mixed Effects (FLAME1). The thresholds for group level activation maps were created using TFCE of z > 1.5 and a clustercorrected significance threshold of p < 0.05. The exact regions of brain activity were determined using the KNE96 coordinate space and Harvard-Oxford cortical and subcortical structural atlases.

#### **DTI**

Voxel-based group differences were calculated for the FA images using Tract-Based Spatial Statistics (TBSS; Smith et al., 2006; Smith and Nichols, 2009; Cheon et al., 2011). TBSS linearly registered individual FA images in native space and then to the FA template via the FLIRT command of FSL. The resultant warping transformations were then used to convert images of diffusion (i.e., FA) to Montreal Neuroimaging Institute (MNI152) space with a spatial resolution of 1 × 1 × 1 mm. For statistical inference, including correction for multiple comparisons, permutation testing was used (Nichols and Holmes, 2002; Cheon et al., 2011) as implemented by RANDOMISE of the FSL software package. Five hundred permutations were performed for significant group differences at a threshold of 0.2; corresponding to p < 0.05, corrected for multiple comparisons using TFCE (Smith and Nichols, 2009; Cheon et al., 2011). WM tracts were identified using the Johns-Hopkins University ICBM-DTI-81 WM labels and probabilities of tract accuracy were assessed using the Johns-Hopkins University WM Tractography atlas.

#### Behavioral Methodology and Analyses

All statistical calculations for behavioral analyses were performed using Rstudio Desktop (version 0.99.896, Rstudio, Inc., Boston, MA, USA).

#### Comparison of Test Performance Amongst Study Groups

Mann-Whitney U-tests with false discovery rate corrections of α = 0.05 for multiple comparisons to reach significance at p < 0.05 were used to evaluate differences in test performance between the study groups. A one-sample t-test with mu = 30 and two-sample U-test were used to calculate the group difference in MMSE score given that the variance for the NT group was zero. Both tests revealed a significant result. Additionally, the zero variance for the NT group on the MMSE is within reason for cognitively normal elderly individuals, as persons without cognitive deficits may be expected to receive a perfect score (30/30) as was seen here (Folstein et al., 1975).

#### Correlation of CFTs to Cognitive Test Performance

Pearson correlation coefficients, with false discovery rate corrections of α = 0.05 for multiple comparisons to reach significance at p < 0.05, were used to correlate CFT performance to performance on their respective cognitive tests.

#### Assessing Tests as Classifiers

Fischer Exact test, with corrections for multiple comparisons of α = 0.05 to reach significance at p < 0.05, were used to determine strength of each test as a classifier between CDDAT and NT.

### RESULTS

#### Behavioral Data

We found no significant difference in median ROCFB-VMT or EVDT score for the two groups (NT = 58 ± 26.2

vs. CDDAT = 41 ± 23.6; Cohen's d = −0.196, r-equivalent effect size = −0.0975, p = 0.555, NT = 85 ± 8.8 vs. CDDAT = 87 ± 13; Cohen's d = 0.671, r-equivalent effect size = 0.318, p = 0.969, respectively). There was also no significant different RT for the EVDT (CDDAT = 2166 ± 461 ms vs. NT = 2084 ± 493 ms, not shown). Therefore, neither ROCFB-VMT nor EVDT function as a potential classifier between NT and CDDAT given their nearly identical overlap. All of these findings suggest that no observable difference in either emotional valence determination or visuospatial ability exists between these two groups at this stage of the disease. Additionally, we found no correlation between the ROCFB-VMT and EVDT performance (R <sup>2</sup> = 0.022, p = 0.648). Thus, at this stage of the disease there appeared to be no connection between emotional valence determination ability and visuospatial memory decline; this finding, however, may be the result of a ceiling effect in participant performance. Behavioral results are summarized in **Figure 2**.

#### Imaging Results

#### fMRI Results

There were no significant differences in areas of activation between groups upon stimulus presentation regardless of the emotional pairing displayed (data not shown). Furthermore, no differences in areas of activation were observed between groups with relation to when they chose ''yes'' or ''no'' in response to the question: ''Are the emotions the same?'', thus indicating no fundamental difference in perception or processing of the faces (data not shown).

#### DTI Results

CDDAT did not show higher FA values in any areas. NTs showed higher FA values than CDDATS in the right inferior longitudinal fasciculus (r-ILF), right posterior thalamic radiations (r-PTR) and the bilateral PCC (b-PCC) and superior longitudinal fasciculi (b-SLF). For the SLF, the differences between NTs and CDDATs were greater in the left than the right hemisphere, suggesting WM deterioration has occurred in the CDDAT group. No significant differences were observed in any of the three major amygdala connection pathways: the amygdalofugal, the stria terminalis, or the anterior commissure. Imaging results are summarized in **Figure 3**.

### DISCUSSION

#### Significance of Behavioral Findings

Consistent with our hypothesis and the previous findings of Albert et al. (1991) and Luzzi et al. (2007), CDDAT participants displayed no difference in accuracy or RT for either the ROCFB-VMT or the EVDT. These findings support the notion that non-verbal abilities, as measured by the ROCFB-VMT, and emotional valence determination, as measured by the EVDT, are largely intact in early-stage AD patients.

All of the subjects were combined together when analyzing the correlation of CFTs to cognitive test performance. Although it would be possible to add a categorical variable representing disease/no disease to the model to generate a multivariate linear model that controls for disease status, that would be ideal if the goal was to produce the most predictive model or if we were trying to demonstrate the causal dependence of the dependent variable on the independent variable.

However, there are several reasons we chose not to control for disease status. First, in principle, people who might apply these tests would likely not know disease status prior to administration. Thus, a model that controls for a variable that they don't have available might confuse its application. Second, we are not proposing that CFT is causing cognitive test performance so much as they are simply correlated; that is, our goal is not necessarily to construct the most explanatory possible model. Controlling for disease status makes that impossible.

FIGURE 3 | Pertinent diffusion tensor imaging (DTI) results for the EVDT. Significant differences were seen in the posterior cingulate cortex (PCC; light blue crescent, top) and superior longitudinal fasciculi (blue crescent, bottom), but not anterior cingulate (red crescent, top) or amygdaloidal connections (pink circle, bottom). Areas where NT > CDDAT for fractional anisotropy (FA) measures are shown in red. Green lines represent the FA skeleton. NT, Neurotypicals; CDDAT, Clinically Diagnosed Dementia of the Alzheimer Type.

### Significance of fMRI Findings

Further support for our hypothesis comes from the fMRI results. We hypothesized that, in the absence of differences in behavioral performance, there would likely be no difference in brain activity. We also hypothesized that, if such a difference were seen, it would occur within the parietal lobes, amygdalae, or r-ACC based on fMRI findings in NTs reported by Fusar-Poli et al. (2009). The lack of differences between the brain activity patterns of the two groups upon stimulus presentation and during the selection of a response suggests that early-stage AD patients perceive the stimuli and choose a response in a fundamentally similar way to NTs, thereby strengthening the notion that these processes are not yet demonstrably affected by the disease.

### Significance of DTI Findings

Although our behavioral and functional imaging results seem promising with regard to the retention of non-verbal skill sets, a more concerning observation was made through DTI. Consistent with our hypothesis, we saw lower FA values in the b-SLF, especially within the left hemisphere, and b-PCC, but no differences in the amygdalae or r-ACC. While there were no significant differences in the major amygdala or anterior cingulate connection pathways, the changes in the SLF and the PCC are consistent with both the pathological progression of WM deterioration in AD (Braak and Braak, 1997; Bartzokis, 2004; Leech and Sharp, 2014) and proposed mechanisms of emotional valance (Maddock et al., 2003; Fusar-Poli et al., 2009).

In congruence with Indersmitten and Gur (2003), we propose the greater deterioration of the left-SLF tracts observed herein suggests a decline in processing efficiency for emotion, but a retention of the dominant circuit, which is served by tracts of the right-SLF. Finally, Kosaka et al. (2003) determined that the PCC plays an important role in face-processing for the transition of a face from being unrecognized to being acknowledged as familiar in NTs. Therefore, the early deterioration of PCC connecting fibers may play a significant role in AD patients' impaired ability to recognize loved ones independent of semantic memory loss. Taken together, our findings further support the reverse demyelination hypothesis, proposed by Medina and Gaviria (2008), and Bartzokis's (2004) assertion of the role of myelin damage predating neuronal loss and its possible role in exacerbating the progression of AD.

### Limitations

We acknowledge that the instances in which our behavioral analyses did not find a significant difference between our NT and CDDAT groups does not necessarily mean that impairments are not present at this time. Instead, it indicates that such deficits, if they exist, are less severe than those of semantic verbal fluency, which currently represents the most consistent and pronounced clinical manifestation of this disease at this stage (Henry et al., 2004). Additionally, the perfect score with zero variance obtained on the MMSE by our control group makes them an idealized group for comparison, but it also makes them less representative of the NT geriatric population as a whole. It is important to note, however, that they were not intentionally selected to obtain a perfect score and that this is an incidental result.

As previously mentioned, the inability to directly correlate ROCFB-VMT: EVDT performance may be due to a ceiling effect. Replicating this experiment with a more difficult visuospatial test, such as the Rey Osterrieth Complex figure A (ROCFA) may result in an observable difference between NTs and early-stage CDDAT, but given that a similar correlation has been established for the MLT and ROCFB in moderate-stage AD by Luzzi et al. (2007) such an observation may be more academic than clinically relevant.

It is important to note, however, that since the etiology of AD is unknown and requires post mortem confirmation it is possible that our findings may not extend to all individuals affected with the ailment. Other potential cofounders could be differences in the demographics and comorbidities in the two groups. Due to restrictions related to the approved Human Protections Board protocol, we do not have any additional medical or demographic information (e.g., marital status, list of medications or medical comorbidities that are not of a neurologic or psychiatric nature) on our participants than what is currently described. An alternative interpretation is that although there is significant structural damage, the lack of functional differences between the groups may represent a type of structural compensation independent of reverse demyelination. Conversely, these observations may be rooted in fundamental issues of neurodegeneration and may therefore extend to other neurodegenerative diseases making these findings simultaneously of larger interest to the field as a whole, but less exclusive to AD. Additional investigations will be necessary to access the reproducibility of these findings in other populations as well as their specificity to AD.

## CONCLUSION

While we observed that both groups performed similarly on both skill sets, we found no correlation between the two, due in part to the massive variations in ROCFB-VMT scores. Yet, since both groups scored within normative ranges after adjusting for age and education (Becker et al., 1987), we conclude that both skill sets are intact within these participants at this stage of the disease. Furthermore, there were no significant differences in BOLD fMRI activation with regard to either stimulus presentation or participant response, strengthening our assertion that these skill sets are, as of yet, unaffected. Lastly, these findings are consistent with the previous work by Luzzi et al. (2007). The deterioration of WM tracts within the b-SLF and b-PCC may foreshadow the impending decline of these functions, as would be consistent with clinical observations (Moore and Wyke, 1984; Haupt et al., 1991; Liu et al., 1991) and autopsy findings (Braak and Braak, 1997) for the progression of the disease.

In conclusion, our discoveries highlight the potential of WM tractography as a presymptomatic biomarker for AD. Based on our observations we suggest that when caring for patients with suspected early-stage AD, use direct/pronounced body language and facial expressions over verbal commands whenever possible. Emphasizing non-verbally mediated social cues has the added benefit of strengthening social interaction, which may help slow the progression of symptoms (Bennett et al., 2006) and reduce the risk of depression (Yaffe et al., 1999).

### AUTHOR CONTRIBUTIONS

RR created and executed the project, analyzed and interpreted the data, and wrote the manuscript. RCA and DF assisted in the neuroimaging experimental design and analyses. AGM and HMRK assisted in experimental design and statistical analyses and interpretation. PL, PJ, GH, KL, JC and JT determined participant eligibility and consulted on medical interpretation and significance. PHR and MB provided consultation on experimental design and interpretation.

### REFERENCES


#### ACKNOWLEDGMENTS

Funding for this research was provided by the I. Wylie Briscoe College of Human Sciences Endowment for Alzheimer Research. PHR is supported by NIH grants (AG042178, AG047812) and the Garrison Family Foundation. This study was completed in agreement with the Alzheimer Disease Neuroimaging Initiative Image and Data Archive (ADNI LONI IDA) data sharing policy. Raw and preprocessed images are available at http://adni.loni.usc.edu/data-samples/access-data/. Portions of this article are taken from the dissertation (Rajmohan, 2016) in compliance with Frontiers in Aging Neuroscience's rules and regulations.


impairment and AD: a diffusion tensor imaging study. Neurobiol. Aging 27, 663–672. doi: 10.1016/j.neurobiolaging.2005.03.026


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Rajmohan, Anderson, Fang, Meyer, Laengvejkal, Julayanont, Hannabas, Linton, Culberson, Khan, De Toledo, Reddy and O'Boyle. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution and reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# A Neurochemical Basis for Phenotypic Differentiation in Alzheimer's Disease? Turing's Morphogens Revisited

Heather T. Whittaker <sup>1</sup> and Jason D. Warren1, 2 \*

*<sup>1</sup> Department of Neurodegenerative Disease, UCL Institute of Neurology, University College London, London, UK, <sup>2</sup> Dementia Research Centre, UCL Institute of Neurology, University College London, London, UK*

Keywords: Alzheimer's disease, dementia, phenotype, network, Turing, protein

### INTRODUCTION

The pathological process underpinning Alzheimer's disease (AD) can manifest as any of several distinctive clinico-anatomical syndromes (Warren et al., 2012). The factors that drive this phenotypic variation remain unclear but are likely to hold important insights into the mechanisms whereby local neurotoxic effects of pathogenic proteins are scaled to distributed brain networks. A theoretical framework for understanding morphological differentiation in biological systems was first outlined in now classic work by Turing (1952), who showed computationally that diffusion of two or more tissue chemicals or "morphogens" reacting across an embryonic cellular network is sufficient to scale initial random fluctuations into stable, often strikingly asymmetric patterns. Here we propose that Turing's theory predicts the phenotypic diversity of AD, as a fundamental consequence of two interacting pathogenic proteins (phosphorylated tau and beta-amyloid) that spread diffusively through a common, distributed neural network.

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Rik Vandenberghe, KU Leuven, Belgium Moira Steyn-Ross, University of Waikato, New Zealand*

\*Correspondence:

*Jason D. Warren jason.warren@ucl.ac.uk*

Received: *10 January 2017* Accepted: *13 March 2017* Published: *29 March 2017*

#### Citation:

*Whittaker HT and Warren JD (2017) A Neurochemical Basis for Phenotypic Differentiation in Alzheimer's Disease? Turing's Morphogens Revisited. Front. Aging Neurosci. 9:76. doi: 10.3389/fnagi.2017.00076*

### THE PROBLEM OF ALZHEIMER'S DISEASE: CLINICAL DIVERSITY ON A COMMON PATHOLOGICAL SUBSTRATE

Canonically, AD presents with episodic memory impairment attributable to dysfunction of hippocampi and connected circuitry traversing the mesial temporal lobes. However, several other variant clinical presentations of AD are well-recognized: these include a "visual" variant led by visuoperceptual and visuospatial deficits; "logopenic aphasia" led by language impairment; and a "frontal" variant led by executive and behavioral decline (Warren et al., 2012). These clinical syndromes have associated profiles of regional brain dysfunction and atrophy which are likely to reflect differential involvement of a core, distributed temporo-cingulo-parietal network and its projections by the diffusive spread of pathogenic proteins (Seeley et al., 2009; Pievani et al., 2011; Warren et al., 2012). In the healthy brain, the core network mediates stimulus-independent thought (hence its designation as the so-called "default-mode network") and it is targeted early and relatively selectively by the pathological process in AD (Buckner et al., 2009; Simic et al., 2014).

All AD phenotypes are characterized by pathological tissue accumulation of neurofibrillary tangles containing abnormally-phosphorylated tau and extracellular plaques containing betaamyloid. Though their precise relation remains contentious, beta-amyloid and phosphorylated tau are central to current concepts of AD biology, and the action of toxic oligomers on synapses may instigate a cascade of intra- and extra-cellular events leading ultimately to the tissue expression of AD (Ittner and Götz, 2011; Morris et al., 2014). While the regional tissue distribution of pathology may vary between AD syndromes (Murray et al., 2011; Mesulam et al., 2014; Martersteck et al., 2016; Ossenkoppele et al., 2016), any mapping between histology and phenotype is likely to be complex. This raises an apparent paradox: why should the pathological process in AD manifest as a handful of diverse but consistent patterns, rather than as a single uniform signature or a stochastic spectrum of random tissue damage?

### TURING'S THEORY OF MORPHOGENESIS AND ITS LEGACY

Turing's reaction–diffusion theory of morphogenesis posits that an initially stable, homogeneous cell array containing "formproducing" chemicals (morphogens) may depart from stability due to stochastic fluctuations in the array. It is assumed that morphogens diffuse and react, such that a morphogen may excite its own formation and diffusion by autocatalysis, or inhibit these processes in another morphogen; it is further required that morphogens have different rates of diffusion. Excitationinhibition coupling between the morphogens tends to focus autocatalysis locally into zones separated by intervening regions where inhibitory effects predominate (if this is not the case, then catastrophic instability occurs and growth of the reaction will halt). Over time, a "wave-like" pattern of inhomogeneous morphogen concentrations develops across the cell array and transmits a corresponding pattern of cellular effects. For the case of two interacting morphogens, the resultant patterns resemble standing waves and become more salient over time, while for three or more morphogens, more complex behaviors emerge.

Turing showed that it is relatively straightforward mathematically to extend the reaction–diffusion framework of "homogeneity breakdown" from a ring to a sphere (or shell) of cells. Since Turing's original formulation, his theory has been shown to hold for an extraordinary variety of applications, ranging from coat pigmentation patterns in animals to predator-prey relationships in ecosystems, crime hotspots in communities, sand ripples, and galaxy formation (Murray, 1990; Ball, 2015). Turing effects have also been shown to operate on electrophysiological neural network parameters that do not require physical transfer of "morphogens" (Jirsa and Kelso, 2000; Hutt and Atay, 2005; Steyn-Ross et al., 2009, 2013).

### TRANSLATING TURING: FROM MORPHOGENS TO NEURODEGENERATIVE PATHOGENS

Our proposal to extend Turing's theory to AD pathogenesis was motivated by the empirical resemblance of AD neuroanatomical phenotypes to Turing reaction-diffusion patterns in other biological and physical systems. The two pathogenic proteins integral to the development of AD are clearly dissimilar to the morphogens of developmental biology (Tiberi et al., 2012): the pathogenic proteins of AD are "form-destroyers" rather than form-producers and any analogy must be qualified. Nevertheless, these AD proteins are likely to possess the key Turing morphogen attributes of diffusive spread and mutual reaction (Ittner and Götz, 2011; Warren et al., 2013): the "inhomogeneities" they produce are departures from brain network health, expressed as regional neural dysfunction and damage. Our idea is sketched in **Figure 1**.

Pathogenic protein effects on synaptic function and intercellular connectivity determine the final common pathway of protein diffusion and reaction at network level. Computational modeling of neural network behavior has established that certain synaptic connectivity properties can generate spatial Turing instabilities over macroscopic scales via long-range electrophysiological field effects (Jirsa and Kelso, 2000; Hutt and Atay, 2005; Steyn-Ross et al., 2009, 2013). In particular, Turing activation patterns emerge where the firing rates of connected neurons are governed by disproportionate excitatory vs. inhibitory inputs acting over different spatial ranges. Besides physical diffusion between neurons (Warren et al., 2013), tau and beta-amyloid have complex effects on synaptic and neurotransmitter physiology that might establish such Turing field effects. These proteins react extensively in an intricate "pas de deux" that is likely to produce net toxic gain-offunction as well as loss-of-function effects at synaptic (and by extension, network) level (Winklhofer et al., 2008; Ittner and Götz, 2011; Leighton and Allison, 2016; Ovsepian et al., 2016). While ultimately the interaction of tau and beta-amyloid is additive in promoting the spread of AD pathology, at a given stage during evolution of the disease the proteins might plausibly have mutually reciprocal effects on synaptic function and network connectivity: for example, prior to undergoing pathogenic misfolding tau protein protects against beta-amyloidinduced neuronal dysfunction (Dawson et al., 2010), while tau and beta-amyloid associate with distinct network profiles in the aging brain (Sepulcre et al., 2016).

Remarkably, connectivity properties of the default-mode network may make it intrinsically more susceptible to Turing effects than other large-scale brain networks (Steyn-Ross et al., 2009, 2013). This might explain why the network is selectively targeted by the dual-protein pathological process in AD and why this process is phenotypically differentiated. Electrophysiological Turing patterns of neuronal dysfunction developing within the network would be translated into neuronal damage and death, thereby fixing the electrophysiological patterns into the structural atrophy patterns that constitute AD phenotypes (**Figure 1**).

### TRANSLATING TURING: SOME KEY PROBLEMS

We now consider certain important challenges in extending the Turing framework to AD.

One immediate consideration is the geometry (and relatedly, the finite number) of diffusive AD patterns. Turing structures are highly dependent on system boundary and scaling constraints as well as specific diffusion characteristics of the relevant morphogens, which are generally not known a priori (Murray, 1990). The intrinsic "wavelength" of the putative reaction– diffusion process in AD is likely to be substantially larger than an individual cell or cortical column and may be amplified by involvement of longer-range network projections, which have been shown to support macroanatomical Turing structures

(Nakamasu et al., 2009; Steyn-Ross et al., 2009, 2013; Kondo, 2016).

While Turing's concept of autocatalysis is broadly supported by empirical evidence for auto-propagation of pathogenic proteins in AD and other proteinopathies (Hardy and Revesz, 2012; Warren et al., 2013), the relation between phosphorylated tau and beta-amyloid remains a key unresolved issue. A Turing model would require them to fill the roles of "activator" and "inhibitor:" this need not of course imply that either protein has a protective role, but would predict a reciprocal relationship over some spatial scale or temporal interval. At present this is difficult to assess directly in human disease, however it may be pertinent that phenotype and tissue damage have been found generally to correlate with the regional distribution of phosphorylated tau but not beta-amyloid (Morris et al., 2014; Ossenkoppele et al., 2016).

A further key issue is that the clinico-anatomical patterns constituting AD phenotypes are eventually unstable, evolving and converging to a global distribution of tissue damage with disintegration of the network that instantiates the reaction– diffusion process. As originally proposed, Turing's theory eschewed scenarios far from the onset of inhomogeneity and did not explicitly model changing temporal dynamics: these scenarios are clearly apposite to AD and may be more effectively addressed using more recent extensions of the theory, including the incorporation of temporal Hopf instabilities (Kondo et al., 2009; Steyn-Ross et al., 2009, 2013; Kondo, 2016).

### TESTING THE IDEA AND FUTURE DIRECTIONS

In its simplest form, Turing's model depends on four parameters for each pathogenic protein: the rate of production; the rate of diffusion; the rate of degradation; and the magnitude of their interaction. This should make the model relatively amenable to experimental evaluation in artificial neural networks using computational techniques or indeed, in vitro neural circuits or transgenic animals. Computational models incorporating biologically-realistic neuronal and circuit parameters have been shown to generate complex Turing behavior for both the healthy brain and selected disease states such as epilepsy and schizophrenia (Jirsa and Kelso, 2000; Steyn-Ross et al., 2009, 2013). These models should be extended to simulate the effects of pathogenic protein properties on synaptic function and tissue spread. The advent of tau-PET neuroimaging (in conjunction with well-established amyloid imaging) opens an avenue to directly compare tau and beta-amyloid tissue deposition profiles in patients (Ossenkoppele et al., 2016). Ultimately there is a need for direct histopathological examination of human brain tissue: while inevitably subject to ascertainment bias (toward more advanced disease), this could be somewhat offset by improved definition of the culprit molecular species and their sites of action within local tissue circuits (Ittner and Götz, 2011; Morris et al., 2014). Although it is unlikely that a Turing reaction-diffusion process is the sole influence governing phenotypic differentiation in AD, it might act as an essential driver that is modulated by other endogenous and environmental factors (Murray, 1990; Ball, 2015; Kondo, 2016).

The molecular nexopathies paradigm of neurodegeneration rests on a coherent conjunction of pathogenic protein and network characteristics (Warren et al., 2013). Turing effects might underpin the peculiar vulnerability of the brain's defaultmode network to AD nexopathy (Steyn-Ross et al., 2009). At the same time, a Turing model of AD pathogenesis might suggest that the neurodegenerative process "unravels" the events of normal neural network ontogeny (Tiberi et al., 2012), implying that embryological differentiation and disease-related de-differentiation exploit intrinsically similar mechanisms. An important motivation for examining models such as Turing's in this context is to deconstruct the apparent complexity of

#### REFERENCES

Ball, P. (2015). Forging patterns and making waves from biology to geology: a commentary on Turing (1952) "The chemical basis of morphogenesis." Philos. Trans. R. Soc. Lond. B Biol. Sci. 370:20140218. doi: 10.1098/rstb.2014.0218

neurodegenerative disease phenomenology to more tractable building blocks. Phenotypic heterogeneity in AD is often ascribed to the operation of still unidentified genetic and epigenetic modifiers of disease expression in particular neural systems (Murray et al., 2011; Warren et al., 2012; Mesulam et al., 2014; Martersteck et al., 2016; Ossenkoppele et al., 2016): if valid, a Turing process would provide a parsimonious mechanism encompassing all variant phenotypes and inherent to the primary disease. This in turn might have implications for development of novel biomarkers and therapeutic interventions targeting the factors that scale reaction–diffusion processes dynamically across the compromised network. At least in principle, macroanatomical confirmation of a Turing signature could help discriminate between candidate molecular mechanisms that drive the observed patterns of neural damage (Kondo et al., 2009; Kondo, 2016).

Finally, Turing's theory may be broadly applicable to other forms of pathological aging and a range of neurodegenerative proteinopathies besides AD. The interaction of C9orf72 products and TDP-43 in frontotemporal dementia is one recent candidate (Vatsavayai et al., 2016) but the application need not be restricted to diseases with two protein pathogens; the role of the second morphogen in the Turing model might be taken by a nonpathogenic tissue factor. Human neurodegenerative diseases may further vindicate the far-reaching potency of Turing's original idea.

### AUTHOR CONTRIBUTIONS

Both HW and JW contributed substantially to the conception of the work, drafted, and revised the work critically for important intellectual content and gave final approval of the version to be published. Both agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

### ACKNOWLEDGMENTS

We thank our reviewers for constructive suggestions and drawing our attention to relevant previous work. HW was supported by Wellcome Trust Strategic Award 098330 for the London Down Syndrome Consortium (LonDowns) "An integrated system to study the development and therapeutic amelioration of cognition and dementia." The Dementia Research Centre is supported by Alzheimer's Research UK, the Brain Research Trust and the Wolfson Foundation. This work was funded by the Alzheimer's Society and the NIHR Queen Square Dementia Biomedical Research Unit. JW was supported by a Wellcome Trust Senior Clinical Fellowship (Grant No. 091673/Z/10/Z).

Buckner, R. L., Sepulcre, J., Talukdar, T., Krienen, F. M., Liu, H., Hedden, T., et al. (2009). Cortical hubs revealed by intrinsic functional connectivity: mapping, assessment of stability, and relation to Alzheimer's disease. J. Neurosci. 29, 1860–1873 doi: 10.1523/JNEUROSCI.5062- 08.2009


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Whittaker and Warren. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Resting State fMRI Reveals Increased Subthalamic Nucleus and Sensorimotor Cortex Connectivity in Patients with Parkinson's Disease under Medication

Bo Shen<sup>1</sup> , Yang Gao<sup>2</sup> , Wenbin Zhang<sup>3</sup> , Liyu Lu<sup>1</sup> , Jun Zhu<sup>1</sup> , Yang Pan<sup>1</sup> , Wenya Lan<sup>1</sup> , Chaoyong Xiao<sup>4</sup> and Li Zhang<sup>1</sup> \*

<sup>1</sup> Department of Geriatrics, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China, <sup>2</sup> Department of Computer Science and Technology, Nanjing University, Nanjing, China, <sup>3</sup> Department of Neurosurgery, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China, <sup>4</sup> Department of Radiology, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Florian Beissner, Hannover Medical School, Germany Huiling Tan, University of Oxford, UK Roxana Gabriela Burciu, University of Florida, USA

> \*Correspondence: Li Zhang neuro\_zhangli@163.com

Received: 15 November 2016 Accepted: 10 March 2017 Published: 04 April 2017

#### Citation:

Shen B, Gao Y, Zhang W, Lu L, Zhu J, Pan Y, Lan W, Xiao C and Zhang L (2017) Resting State fMRI Reveals Increased Subthalamic Nucleus and Sensorimotor Cortex Connectivity in Patients with Parkinson's Disease under Medication. Front. Aging Neurosci. 9:74. doi: 10.3389/fnagi.2017.00074 Functional connectivity (FC) between the subthalamic nucleus (STN) and the sensorimotor cortex is increased in off-medication patients with Parkinson's disease (PD). However, the status of FC between STN and sensorimotor cortex in on-medication PD patients remains unclear. In this study, resting state functional magnetic resonance imaging was employed on 31 patients with PD under medication and 31 healthy controls. Two-sample t-test was used to study the change in FC pattern of the STN, the FC strength of the bilateral STN was correlated with overall motor symptoms, while unilateral STN was correlated with offside motor symptoms. Both bilateral and right STN showed increased FC with the right sensorimotor cortex, whereas only right STN FC was correlated with left-body rigidity scores in all PD patients. An additional subgroup analysis was performed according to the ratio of mean tremor scores and mean postural instability and gait difficulty (PIGD) scores, only the PIGD subgroup showed the increased FC between right STN and sensorimotor cortex under medication. Increased FC between the STN and the sensorimotor cortex was found, which was related to motor symptom severity in on-medication PD patients. Anti-PD drugs may influence the hyperdirect pathway to alleviate motor symptoms with the more effect on the tremor subtype.

Keywords: Parkinson's disease, subthalamic nucleus, sensorimotor cortex, functional connectivity, hyperdirect pathway, on-medication

### INTRODUCTION

Parkinson's disease (PD) is the second most common progressive neurological degenerative disorder caused by dopamine deficits in the substantia nigra pars compacta (Lees et al., 2009). Impairment of the respective functions of parallel cortico-basal ganglia-thalamo-cortical circuits causes various symptoms (DeLong, 1990; Helmich et al., 2010). The subthalamic nucleus (STN) is one of the preferred targets in deep brain stimulation (DBS) treatment of PD patients, with greater

clinical benefits in motor symptom improvement than those obtained by stimulating other sites (Volkmann et al., 2004; Odekerken et al., 2013). However, the exact mechanism of this stimulation remains unknown. Therefore, STN may play a role in the motor control in PD patients (Chiken and Nambu, 2014).

In a healthy brain, the STN stimulates the internal segment of the globus pallidus, leading to increased inhibition of the ventrolateral thalamus. Consequently, the motor activity is increased within the primary somatosensory cortex (S1), primary motor cortex (M1), and premotor cortical area (Weintraub and Zaghloul, 2013). This phenomenon is an indirect pathway that is depressed by dopamine. The indirect pathway is overactive in PD patients, leading to hyperactivity of the STN (Alexander and Crutcher, 1990). Furthermore, the fast hyperdirect feedback loop from supplementary motor area and M1 cortical projections to the STN via glutamatergic neurons needs further investigation (Tewari et al., 2016).

Resting state functional MRI (rs-fMRI) is a relatively novel technique which is easily carried out in large populations. However, the biological origin and relevance of these slow neuronal activity components are still poorly understood, the latter observation that spontaneous BOLD activity is specifically organized in the resting human brain, which has generated a new avenue of neuroimaging research (Deco et al., 2011). Subsequently, rs-fMRI is also a well-accepted tool in the noninvasive study of neurological and psychiatric disorders at a network level in vivo (Zhang and Raichle, 2010). In offmedication PD patients, an increased functional connectivity (FC) between the STN and hand M1S1 areas was found in the non-tremor subgroup with the FC strength correlating with rigor scores (Baudrexel et al., 2011), while increased FC between these two areas was also discovered in early drug-naïve PD patients (Kurani et al., 2015). Primary data in the α- and β-frequency EEG bands showed a burst oscillatory local field activity in the STN and an increased FC between STN and motor cortical in PD patients (Hammond et al., 2007; Lalo et al., 2008). Therefore, increased oscillations in the STN may be a factual reason for the abnormal activity of the M1S1 cortex. A consistent conclusion was the increased FC of STN at different stages in off-medication PD patients.

Only two articles reported the FC of STN and motor area in normal PD patients while in the on-medication. Fernández-Seara et al. (2015) showed an increased FC between the STN and the motor cortex just like in off-medication PD patients using arterial spinlabeled (ASL) perfusion fMRI, whereas Mathys et al. (2016) did not find a change in the FC between the two areas. Aside from the different methods in these two articles, we speculate that choosing patients from a broad severity range may benefit FC change analysis, as previous research shows that a broad range of severity is needed when combining the de novo and moderate PD groups into the correlation analysis (Kurani et al., 2015), while Fernández-Seara et al. (2015) preferred early-state PD patients (mean HY = 1.83) and Mathys selected patients with a mean duration of 6 years. Litvak et al. (2011) found that dopaminergic medication modulated the resting beta network by combining magnetoencephalographic and subthalamic local field potential recordings. However, the correlation between decreased FC strength and decreased motor symptoms in on-medication PD patients using fMRI technology was still unknown.

Hence, in this work, we selected PD patients with different severities (HY from 1 to 4 on-medication based, duration from 1 to 18 years) to assess the change in FC of STN. We tested whether changes in FC between STN and whole brain may exist, as well as the correlation with the motor symptom, because motor symptoms exist after drug administration.

### MATERIALS AND METHODS

#### Participants

We conducted a prospective case – control study of 36 PD patients and 31 healthy controls in the Department of Geriatrics, Nanjing Brain Hospital between July 2015 and March 2016. Patients were included in the study if they were aged 18 years or older, satisfied the standard UK Brain Bank criteria for PD (Hughes et al., 1992), and experienced at least one of the following symptoms: severe response fluctuations, dyskinesias, painful dystonias, or bradykinesia. Exclusion criteria included history of other neurological or psychiatric diseases, and cognitive impairment based on the PDD criterion in 2007 (Dubois et al., 2007). We defined anti-Parkinsonian medication to include any drug designed to alter symptoms of PD or any drug that slows the progression of PD, levodopa equivalent daily dose (LEDD) was calculated with previous research (Tomlinson et al., 2010). All PD patients were scanned twice in off-medication and in 60–90 min after taking anti-Parkinsonian medication. Only the on-medication measurements were analyzed. All participants had written informed consent and the study was approved by the Medical Research Ethical Committee of Nanjing Brain Hospital, Nanjing, China.

### Assessment of PD Motor and Cognition Symptoms

Motor impairment in patients with PD was assessed by items of Part III (motor part) of the Unified Parkinson's disease Rating Scale (UPDRS) and H&Y staging scale for both the "on" and "off " states. Unilateral limb tremor scores are the sum of hand tremor scores and lower limb scores from UPDRSIII. The mean tremor score was derived from the sum of items 16 and 20–26 on the UPDRS, while a mean score was derived from five postural instability and gait difficulty (PIGD) items (Stebbins et al., 2013). Patients were classified as having tremor-domain Parkinson's disease (TD-PD) when the ratio of the mean tremor score to the mean PIGD score was ≥ 1.5 and as having PIGD-PD when this ratio was ≤ 1, others were included as having mixed subtype PD. Overall cognition condition was assessed using the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) (Chen et al., 2016).

### MRI Data Acquisition Protocol

MR-imaging was carried out on a 3.0-T MR scanner system (Siemens, Verio, Germany) with an 8-channel phased-array head coil for signal reception and a whole body coil for radio frequency transmission. Subjects were instructed to lie still, relax, and not think of anything in particular, while being required to keep their eyes open to avoid falling asleep. No subject reported to have fallen asleep when routinely asked immediately after examination.

#### Image Acquisition

fnagi-09-00074 April 1, 2017 Time: 17:14 # 3

Functional scans of the brain were acquired using a gradient echo EPI sequence with the following parameters: repetition time (TR) = 2000 ms, echo time (TE) = 25 ms, matrix size = 64 × 64, field of view (FoV) = 240 mm × 240 mm, 33 slices with 4 mm slice thickness and 0 mm inter-slice gap, and scan duration of 8 min and 6 s. Axial anatomical images were acquired using a 3D-MPRAGE sequence (TR = 1900 ms; TE = 2.48 ms; flip angle [FA] = 9 ◦ ; matrix = 256 × 256; FoV = 250 mm × 250 mm; slice thickness = 1 mm; and gap = 0 mm; slices covered the whole brain, with registration and functional localization). Patients were scanned during on-state medication, resulting in a 4D data set consisting of 240 volumes of functional data for subsequent FC analysis.

### Data Preprocessing

Preprocessing was carried out using Data Processing Assistant for Resting-State fMRI (DPARSF; Chao-Gan and Yu-Feng, 2010<sup>1</sup> ) which is based on Statistical Parametric Mapping (SPM8)<sup>2</sup> .

The first 10 volumes of the BOLD data for each subject were discarded, while the remaining images were corrected by realignment, accounting for head motion. Three patients with head motions exceeding 3 mm of translation, or a rotation of 3◦ , throughout the course of the scan were excluded from the study. The remaining functional images were coregistered to the individual T1-weighted images and were then segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) tissue maps using a unified segmentation algorithm followed by non-linear normalization into the Montreal Neurological Institute space. Resultant functional images were re-sampled to 3-mm isotropic voxels, and spatially smoothed with a Gaussian kernel full width at half maximum = 4 mm × 4 mm × 4 mm. The resulting fMRI data were band-pass filtered (0.01 < f < 0.08 Hz) to reduce low-frequency drift and high-frequency physiological, respiratory, and cardiac noise. Any linear trend was then removed. Subsequently, six head motion parameters and the mean time series of global activities, WM, and CSF signals were introduced as covariates into a random effects model to remove possible effects of head motion, global activities, WM, and CSF signals on the results. For each patient, the global signal is obtained by averaging the time series of voxels in all brain tissues, everyone has an unique global signal, global signals was thought as including the breath, heart rate, and other noise. So in the data preprocessing step, global signals are regressed out at the single-subject level.

The bilateral STNs were defined as regions of interest from the WFU\_pickatla (Talairach brain atlas theory), which automatically generated segmented atlas region of interest templates in the MNI space (Sakurai et al., 2017). Each STN was about 81 mm<sup>3</sup> , the centers of the left and right STN were [−9, −12, −8] and [9, −14, −8], the location and extent of STN are displayed in **Figure 1**. The mean time series of bilateral STN were extracted. Furthermore, a voxel-wise FC analysis was performed by computing the temporal correlation between the mean time series of the combined left and right STN and the time series of each voxel within the whole brain. The correlation coefficients of each voxel were normalized to z-scores with Fisher's r-to-z transformation. Therefore, a z-score map for the entire brain was created for the STN of each subject. Finally, the region where the significant difference between PD and HC, was used as the mask to extract the mean Z value for each PD patient. The correlation of Z-values and motor symptoms was measured with SPSS20, and motor symptoms included UPDRSIII, TD scores, PIGD scores, duration, bilateral tremor and rigidly scores, as the tremor and rigidly are the common symptoms of PD. The TD score was derived from the sum of items 16 and 20–26 on the UPDRS, while PIGD score was derived from five PIGD items.

### Statistical Analysis

One-sample t-tests were conducted on the z-score maps of the two groups. Then, between-group two-sample t-tests were performed within the whole brain mask, with age, gender, education, and GM volumes as covariates, to detect significant differences between the two groups, GM volumes were calculated by the SPM in the latter voxel based morphometry step. With in-group multiple comparisons, all T maps had a threshold of p < 0.005, while the cluster extent was calculated according to alphasim correction based on REST software (voxel-level p value < 0.005; cluster size: >69 voxels; determined by a Monte-Carlo simulation resulted in a cluster-level significance threshold of P < 0.005). Correlation between the FC strength of the bilateral STN and overall motor symptoms have been evaluated, while unilateral STN FC was correlated with contralateral motor symptoms, and motor symptoms included UPDRSIII, TD scores,

FIGURE 1 | ROI presentation of subthalamic nucleus (STN). ROI definition of the STN, bilateral STNs were defined as regions of interest from the WFU\_pickatla, the centers of the left and right STN were [−9, −12, −8] and [9, −14, −8].

Seed Region and FC Analysis

<sup>1</sup>http://rfmri.org/DPARSF

<sup>2</sup>http://www.fil.ion.ucl.ac.uk/spm

PIGD scores, duration, bilateral tremor and rigidly scores, as the tremor and rigidly are the common symptoms of PD. All correlation analyses were performed using the SPSS20.0 software package.

#### Voxel Based Morphometry (VBM)

To test whether the change in FC pattern was associated with the structural atrophy, gray and white matter volumes were measured based on the SPM8. VBM procedure involved the segmentation of the original structural MRI images in native space into GM, WM, and CSF tissues, then GM and WM images were normalized to templates in stereotactic space to acquire optimized normalization parameters, which were applied to the raw images. GM images were smoothed using an 8-mm full width at half maximum isotropic Gaussian kernel. The last, we employed a general linear model, using age and sex as covariates. Comparison between PD patients and healthy controls was again carried out with the two-sample t-test option provided in the SPSS software, with significant difference set at p < 0.05.

### RESULTS

### Clinical and Neuropsychological Evaluations

A total of 31 PD patients were included in our study (excluding three patients whose head motions exceeded 3 mm of translation and another two patients who could not bear the noise of the MRI). The 31 PD patients contained varying motor symptom severity and durations, with 18 of the 31 patients being more affected at the left side of the body in terms of UPDRS III. For subsequent analyses, due to the smaller size of TD patients, patients with tremor-dominant (n = 5) and mixed type (n = 9) were pooled and referred to as the tremor subgroup (n = 14) similar to the previous study (Baudrexel et al., 2011). No significant differences in age, sex, education, MMSE, and MOCA were found for the three groups. **Table 1** summarizes the detailed demographic and clinical characteristics of the three groups (patients with PD and healthy controls).

### STN FC in Healthy Controls

Mean resting state FC z-score maps of STN in healthy controls are displayed in **Figure 2**. The correlation of left and right STN with whole brain was measured with DPARSF and REST software, as a result, the left and right STN FC pattern were compared with the two-sample t-test to seek the differences, and there is no difference. The result showed that most of the positive z-score values were found bilaterally in the brainstem, caudate nucleus, putamen, thalamus, and the cerebellum. Moreover, relatively small positive z-score values were found in the Frontal Lobe, which included middle frontal gyrus, superior frontal gyrus, frontal eyes field, pre-motor, and supplementary motor cortex in the right cerebrum. In contrast, negative z-score values were found in bilateral precuneus, which is the core area of the DMN, cuneus, lingual gyrus, middle occipital gyrus, left superior occipital gyrus, right middle temporal gyrus, left primary visual cortex, and calcarine.

### Between-Group Differences of STN FC

Compared with healthy controls, the PD patients exhibited increased right STN FC with the right M1S1, which contains the precentral and postcentral gyrus (p < 0.005, cluster size: >69 voxels, multiple-comparison correction using AlphaSim in REST), while no decreased areas were found as shown in **Figure 3** and **Table 2**. Increased FC patterns of bilateral STN with the right M1S1 were also found as shown in **Figure 3** and **Table 2**. The left STN FC pattern had no change in area, while all T maps showed no decrease in area. The PIGD showed the increased FC between

TABLE 1 | Demographic and neuropsychological characteristics of all subjects (on-medication).


MMSE, Mini-Mental State Examination; TD, tremor-dominant; PIGD, postural instability and gait difficulty; UPDRS, unified Parkinson's disease rating scale; LEDD, levodopa equivalent daily dose; p<sup>a</sup> -value for the gender difference was obtained by chi-square test, others were obtained using one-way analyses of variance.

the right STN and bilateral M1S1 compared with the HC with on changed pattern in TD subgroup as shown in **Figure 4** and **Table 3**. Z-values of right STN and right M1S1 in the different groups as shown in **Figure 5**.

## STN–M1S1 FC and Motor Symptom Correlation Analysis

Finally, the region where the significant M1S1 clustered, as a result of the between-group analysis, was used as the mask to extract the mean Z value for each PD patient. The Z-values of right STN showed no correlation with tremor and PIGD scores. By contrast, the Z-values of right STN and right M1S1 showed a correlation with left lumbar rigidity scores from UPDRS (r = 0.414, p = 0.021) and LEDD (r = 0.435, p = 0.014) in 31 PD patients (**Figure 6**). What's more, there were no correlation of Z-values and motor symptoms in the PD subgroup.

### ROI Analysis of Right STN–Right M1S1 FC

A special ROI centered at the t maximum from the previous experiments was used as functional representation of right M1S1 for further research. This special ROI was defined by the changed

area of STN FC pattern between PD and HC. It was in the right hemisphere and it's peak coordinate was (45, –11, 36), the detailed information of M1S1 were seen in **Table 2**. Compared with the HC, both the PD subgroup showed the increased FC (TD: p = 0.002, PIGD: p < 0.001), A direct statistical comparison of the two PD subgroups under medication again yielded no significant results similar to the previous study (Baudrexel et al., 2011) (**Figure 5**).

#### VBM

Voxel Based Morphometry did not reveal significant differences between patients and healthy controls for gray matter volume and WMV, with detailed information in **Table 4**.

#### DISCUSSION

The STN is the preferred target in DBS surgery to normalize aberrant patterns of STN and to improve cardinal motor symptoms. However, the mechanism is still unclear. The current study found that the combined bilateral STN and right STN showed increased FC with M1S1 in PD patients under medication. These findings provide new evidence of increased FC between STN–M1S1. Furthermore, severe motor symptoms related to the changes in STN–M1S1 FC may also exist in PD patients despite the effects of medication.

Resting state fMRI provided a new way of viewing the hyperdirect pathway, although its mechanism remains unknown.

#### TABLE 2 | REST group comparison results indicating increased STN FC in PD patients as compared to healthy controls.


Localization, voxels sizes, T-values, and MNI coordinates of bilateral (A) and right (B) STN FC pattern as compared with healthy controls. Spatial distribution of significant voxels with respect to their locations according to the automated anatomical labeling AAL template.

FIGURE 4 | Between postural instability and gait difficulty (PIGD) subgroup and HC of difference in the right STN resting state FC. Between PIGD subgroup and HC of difference in the STN resting state FC, results are in MNI space, red color represents the increased correlation while the blue color represents the decreased correlation.

#### TABLE 3 | REST group comparison results of right STN in PIGD patients.


Localization, voxels sizes, T-values, and MNI coordinates of right STN FC pattern as compared with healthy controls. Spatial distribution of significant voxels with respect to their locations according to the AAL template. PIGD, postural instability gait difficulty.

Moreover, the overactivity of the hyperdirect pathway was proven in the off-medication PD patients and animal PD models (Dejean et al., 2008). Kahan et al. (2014) found a decrease in the effective FC of the hyperdirect pathway with STN stimulation, with the relationship between decreased hyperdirect coupling strength and improved clinical severity being particularly interesting. Local field potential recordings from the STN of patients undergoing surgery for DBS revealed strong oscillatory activity, particularly in the beta band (13–35 Hz) (Kuhn et al., 2004). STN beta oscillations were reduced by the application of levodopa and DBS (Priori et al., 2004; Kuhn et al., 2008). Furthermore, STN beta power reduction correlated with clinical improvement (Kuhn et al., 2008). The STN has also been involved in reactive global inhibition through the hyperdirect path (Vink et al., 2005; Zandbelt et al., 2013), with the increased 2.5–5 Hz phase activity of STN leading to increased response thresholds and slower responses (Tewari et al., 2016). Previous studies showed increased FC of STN with M1S1 in off- and onmedication patients. Adding our conclusions, the overactivity of the so-called hyperdirect loop and the occurrence of motor symptoms could be simultaneously depressed with the effects of medication.

Compared with the ASL research, the same conclusions regarding the increased correlation of bilateral STN and M1S1 and the similar correlation between FC and LEDD were seen, which may show the modulation of medication on brain activity (Litvak et al., 2011). Multimodel research correlating ASL, fMRI, and EEG research outputs would help us know more about the mechanism of this disease. The difference in the FC pattern was that changed areas in our research were all in right hemisphere, whereas they found the left part to be mainly

FIGURE 6 | Results of the correlation of FC values with levodopa equivalent daily dose (LEDD) and motor symptom. Left graph showed the correlation diagram of right STN–M1S1 area FC values with corresponding UPDRS III right lumbar rigor scores (hand scores plus leg scores) in patients, right graph showed the correlation diagram of same FC values with LEDD.



GMV, Gray matter volume; WMV, white matter volume; volumes are represented as the mean ± standard deviation. For comparisons of demographics, P-values were obtained using one-way analyses of variance. TD, tremor-dominant; PIGD, postural instability gait difficulty.

altered. This may be because our patients were more onset in the left side of body since the change in FC was only based on the seed of the right STN, while no significant difference was found in the seed of the left STN. Furthermore, previous research performed right-sided surgery faster but with higher errors (Obeso et al., 2013), which means that the right STN is more likely to be involved in action inhibition (Tewari et al., 2016). This abnormal FC in the unilateral cerebral hemisphere is consistent with previous studies wherein the unilateral hemispheric basal ganglia-thalamo-cortical circuit modulated the contralateral movement (Kahan et al., 2014).

We selected unilateral STN as the seed to find its relationship to the offside motor symptom because of the unsymmetrical severity of motor symptoms. Our findings were similar to that of Baudrexel's research wherein FC strength with right STN and right M1S1 were correlated with left-leg rigor scores. Previous studies showed an association between rigor strength and increased oscillatory activity within the STN (Kuhn et al., 2009). We found that the relationship between FC strength and rigidity scores may be because of the stability and objectivity of the rigidity symptoms, whereas tremors were always affected by mood. A little difference between our research with previous offmedication study is that the PIGD showed the most significant enhanced FC pattern. In our study, some people presented serious tremors in the off-state, which disappeared after taking anti-PD drugs. This finding is consistent with previous studies showing that tremors are more affected by drugs (Connolly and Lang, 2014), along with increased FC strength. Katz et al. (2015) found TD patients had greater mean overall motor improvement than PIGD patients after STN DBS, measured by UPDRS-III. Anti-PD drugs or DBS may have a more effect on the hyperdirect pathway of TD patients to resolve the abnormal activity of

#### REFERENCES


STN. Therefore, in studies involving on-state PD patients, the burst FC pattern is also related with the motor symptoms and PD subgroup showed the different hyperdirect pathway under medication.

### CONCLUSION

We are the first to prove the increased FC patterns and the correlations between rigidity symptoms of varying severity in PD patients under medication. Our findings further suggest that PIGD and tremor symptoms might be linked to an different coupling of these areas in on-medication PD patients. Moreover, anti-PD drugs may changed the hyperdirect pathway, thereby altering the motor symptoms.

### AUTHOR CONTRIBUTIONS

LZ designed the study and revised it critically for important intellectual content. BS performed the research and drafted the manuscript, YG and YP helped in data analyses, WZ, CX, LL, and JZ help in clinical data collection and analyses, and made patient follow-ups, and WL edited the paper.

### FUNDING

This study was supported by the Nanjing Science and Technology Development Program (201503039) and special funds of the Jiangsu Provincial Key Research and Development Projects (BE2016614).



Parkinson's disease. Exp. Neurol. 189, 369–379. doi: 10.1016/j.expneurol.2004. 06.001


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Shen, Gao, Zhang, Lu, Zhu, Pan, Lan, Xiao and Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Age-Related Differences in Reorganization of Functional Connectivity for a Dual Task with Increasing Postural Destabilization

Cheng-Ya Huang1, 2, Linda L. Lin<sup>3</sup> and Ing-Shiou Hwang4, 5 \*

<sup>1</sup> School and Graduate Institute of Physical Therapy, College of Medicine, National Taiwan University, Taipei, Taiwan, <sup>2</sup> Physical Therapy Center, National Taiwan University Hospital, Taipei, Taiwan, <sup>3</sup> Institute of Physical Education, Health and Leisure Studies, National Cheng Kung University, Tainan, Taiwan, <sup>4</sup> Institute of Allied Health Sciences, College of Medicine, National Cheng Kung University, Tainan, Taiwan, <sup>5</sup> Department of Physical Therapy, College of Medicine, National Cheng Kung University, Tainan, Taiwan

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Roumen Kirov, Institute of Neurobiology (BAS), Bulgaria Thomas Stoffregen, University of Minnesota, USA Bahar Güntekin, Istanbul Medipol University, Turkey Foteini Protopapa, Scuola Internazionale di Studi Superiori Avanzati, Italy

#### \*Correspondence:

Ing-Shiou Hwang ishwang@mail.ncku.edu.tw

Received: 15 September 2016 Accepted: 28 March 2017 Published: 12 April 2017

#### Citation:

Huang C-Y, Lin LL and Hwang I-S (2017) Age-Related Differences in Reorganization of Functional Connectivity for a Dual Task with Increasing Postural Destabilization. Front. Aging Neurosci. 9:96. doi: 10.3389/fnagi.2017.00096 The aged brain may not make good use of central resources, so dual task performance may be degraded. From the brain connectome perspective, this study investigated dual task deficits of older adults that lead to task failure of a suprapostural motor task with increasing postural destabilization. Twelve younger (mean age: 25.3 years) and 12 older (mean age: 65.8 years) adults executed a designated force-matching task from a level-surface or a stabilometer board. Force-matching error, stance sway, and event-related potential (ERP) in the preparatory period were measured. The forcematching accuracy and the size of postural sway of the older adults tended to be more vulnerable to stance configuration than that of the young adults, although both groups consistently showed greater attentional investment on the postural task as sway regularity increased in the stabilometer condition. In terms of the synchronization likelihood (SL) of the ERP, both younger and older adults had net increases in the strengths of the functional connectivity in the whole brain and in the fronto-sensorimotor network in the stabilometer condition. Also, the SL in the fronto-sensorimotor network of the older adults was greater than that of the young adults for both stance conditions. However, unlike the young adults, the older adults did not exhibit concurrent deactivation of the functional connectivity of the left temporal-parietal-occipital network for posturalsuprapostural task with increasing postural load. In addition, the older adults potentiated functional connectivity of the right prefrontal area to cope with concurrent force-matching with increasing postural load. In conclusion, despite a universal negative effect on brain volume conduction, our preliminary results showed that the older adults were still capable of increasing allocation of neural sources, particularly via compensatory recruitment of the right prefrontal loop, for concurrent force-matching under the challenging postural condition. Nevertheless, dual-task performance of the older adults tended to be more vulnerable to postural load than that of the younger adults, in relation to inferior neural economy or a slow adaptation process to stance destabilization for scant dissociation of control hubs in the temporal-parietal-occipital cortex.

Keywords: aging, EEG, dual task, functional connectivity, balance control

### INTRODUCTION

Maintenance of postural balance requires attentional resources corresponding to the degree of postural threat (Remaud et al., 2012). Postural response requires complexity and numerous micro-adjustments, and stable bilateral stance is principally regulated by an automatic process using brainstem synergy (Honeycutt et al., 2009). Increasing postural destabilization shifts postural control to a more controlled process involving the frontal and cortical-basal ganglia loop (Jacobs and Horak, 2007; Boisgontier et al., 2013). Due to the additional attentional investment, postural response becomes more regular in the controlled process (Donker et al., 2007; Stins et al., 2009). Central resource allocation of a postural-suprapostural dualtask is an elaborate trade-off, flexibly depending on response compatibility of the two subtasks. Addition of a secondary task (or suprapostural task; Mitra and Fraizer, 2004) to a postural task does not necessarily result in dual-task degradation due to resource competition (Chen and Stoffregen, 2012; Stoffregen, 2016); instead, postural response can be integrated with suprapostural activity to facilitate suprapostural performance (Stoffregen et al., 1999; Prado et al., 2007). On the other hand, although a great number of the studies conducted on dual tasks have employed two cognitive tasks, very few neuroimaging studies have focused on postural-suprapostural dual tasks because of methodological constraints. With the eventrelated potential (ERP) of scalp electroencephalogram (EEG), Huang and Hwang (2013) reported that the amplitudes of the N1 and P2 waves in the preparation period prior to executing a secondary motor task varied with task loads of the postural and suprapostural tasks, respectively. The N1 amplitude reflected anticipatory arousal and postural response preceding the forcematching (Adkin et al., 2008; Mochizuki et al., 2008; Sibley et al., 2010; Huang et al., 2014). An increasing N1 amplitude in the sensorimotor and parietal areas implies more attentive control required for postural destabilization (Huang and Hwang, 2013; Little and Woollacott, 2015). On the other hand, P2 amplitude related to neural resource for visuomotor processing of the subsequent force-matching event, with a greater P2 amplitude associated with a less task load of force-matching (Huang and Hwang, 2013; Hwang and Huang, 2016).

In older adults, the shrinkage of a wide range of cortical areas causes evolving dysfunction of a dual task (Fernandes et al., 2006; Hartley et al., 2011). Degeneration of the frontal-parietal network specifically impairs executive processes keyed to dual tasking, such as response inhibition, task switching (Cole et al., 2013), and selective attention to relevant information (Mozolic et al., 2011). Age-related dual task deficits also manifest with a resource ceiling (Geerligs et al., 2014) and compensatory recruitment of additional brain resources (Hartley et al., 2011; Boisgontier et al., 2013), especially when the dual task places high demands on those resources. Behavioral studies have shown that postural destabilization can multiply the dual task cost for the elderly of a postural-suprapostural task. Past researches employing a choice reaction time task (Shumway-Cook and Woollacott, 2000) and digit 2-back task (Rapp et al., 2006; Doumas et al., 2008) highlighted age-related differences in suprapostural performance by increasing stance instability rather than degrading visual stimulation of the suprapostural task. In effect, the brain's residual capacity wanes with aging, which causes negative postural penetrability to suprapostural processing. Hence, older adults often adopt a postural prioritization strategy to keep attentional resources on the postural task (Lacour et al., 2008; Liston et al., 2014). However, behavioral data that elucidate the neural correlates of resource allocation in older adults during a postural-suprapostural task are very limited. Although, neural evidence of age-related deficits has been extensively sought using classic dual tasks, the findings up to date cannot be directly applied to postural-suprapostural dual tasks on account of issues of response compatibility (Salo et al., 2015). For instance, the task quality of a suprapostural motor task such as juggling or traycarrying takes advantage of stance stability (Balasubramaniam et al., 2000; Wulf et al., 2004; McNevin et al., 2013), whereas parallel loading of two cognitive tasks always causes mutual interference. To our knowledge, no studies have investigated alterations in information transfer for a postural-suprapostural task performed by older adults, despite the degeneration of the white matter integrity of the brain with aging (Furst and Fellgiebel, 2011; de Groot et al., 2016). Hence, it is worthwhile to characterize the differences in the functional connectivity of the frontal/prefrontal areas to other cortical regions [such as the parietal (Gontier et al., 2007) and premotor areas (Marois et al., 2006)] of young and older adults during a postural-suprapostural task.

A challenging postural set-up is a sensitive way to highlight age-related differences in a postural-suprapostural task. To explore the underlying neural mechanisms of dual-task interference of a postural-suprapostural task, we investigated age effect on ERP dynamics for force-matching from the level-surface to stabilometer stances during the preparatory period of the particular dual task. Defined as the time window between the execution beep and onset of the force-matching act, the preparatory period consists of posture-dependent N1 and supraposture-dependent P2 waves that encrypt cognitive processing of pre-movement stance regulation, task switching from the posture subtask to supraposture subtask, and planning of the subsequent force-matching act (Huang and Hwang, 2013; Huang et al., 2014; Hwang and Huang, 2016). For the young healthy adults, a postural-suprapostural task with increasing postural instability caused reorganization of functional connectivity in the preparatory period with anterior shift of processing resources and dissociation of control hubs in the parietal-occipital cortex (Huang et al., 2016). Within the brain connectome context, this study aimed to extend on previous work by exploring differences in the component amplitudes (N1 and P2) and functional connectivity of the ERP in young and older adults during the performance of a

**Abbreviations:** NFE, normalized force error; RT, reaction time; SampEn, sample entropy; AMF\_RMS, root mean square value of ankle fluctuation movements; AMF\_SampEn, sample entropy of ankle fluctuation movements; GF, global frontal; SM, sensorimotor; PO, parietal-occipital; SL, synchronization likelihood; NBS, network-based statistics; FSM, fronto-sensorimotor; TPO, temporal-parietaloccipital.

suprapostural motor task with increasing postural challenge. This study hypothesized that, with increasing postural load, young adults would exhibit smaller changes than older adults in the component amplitudes of ERP (N1 and P2) and functional connectivity, especially those for the fronto-parietal network in the preparatory period. We also hypothesized that topological reorganization of functional connectivity due to increasing postural load would differ in the two populations.

#### MATERIALS AND METHODS

#### Subjects

Twelve young healthy adults (5 female and 7 male, age: 25.25 ± 1.25 years, range 21–33 years) and 12 older healthy adults (5 female and 7 male, age: 65.83 ± 1.01 years, range: 61–73 years) participated in this study. Subjects were volunteers from the local community and university campus who responded to a poster or a network advertisement. All of the participants were righthanded and had no history of neurological or musculoskeletal diagnoses. The older adults in this study, who had regular exercise habits, had experienced no falls in the previous 6 months. They participated in the postural-suprapostural experiment after signing personal consent forms approved by the local ethics committee (University Hospital, National Cheng Kung University, Taiwan).

#### Procedures

Before the main experiment, each participant was instructed to stand on a stabilometer in a shoulder-width stance with their arms hanging by their sides. The stabilometer was a wooden platform (50 × 69 cm) with a curved base (height: 18.5 cm). When the platform of the stabilometer was in the horizontal position, the midline of the platform (34.5 cm from the front/rear edge) passed through the anterior aspect of the participant's bilateral lateral malleolus. The positions of the participant's feet were used in the following experiment. Then the maximal angle of anterior tilt was determined from the readings of an electrogoniometer (Model SG110, Biometrics Ltd., UK; output accuracy: 1 mv = 0.09 degrees) on the ankle joint as the participants tilted the stabilometer with maximum plantarflexion of the ankle joint. In addition, we determined the force of each participant's maximum voluntary contraction (MVC) from three attempts of the right thumb-index precision grip during quiet upright stance. The stabilometer is commonly used to train balance in clinics and provides postural challenge for single postural task (Wulf et al., 2001; McNevin et al., 2003; Chiviacowsky et al., 2010) and postural-suprapostural dual-task in the laboratories (Wulf et al., 2003; Huang et al., 2014, 2016; Hwang and Huang, 2016). Therefore, we used the stabilometer to produce postural destabilization in this study.

The formal experiment required the participants to conduct a dual task (suprapostural force-matching and postural tasks) with on-line visual feedback under two different randomized stance conditions (level-surface vs. stabilometer). A monitor that displayed force output, ankle movement, and the target signals was placed 60 cm in front of the subject at eyelevel. The subject conducted a thumb-index precision grip to couple a target line of 50% MVC force (pre-determined in the experiment) and concurrently maintained a stable upright stance with minimal ankle movement on a wooden level surface or a tilted stabilometer. Participants were not told to prioritize either task, and they were instructed to perform both postural and force-matching tasks as well as possible. The stabilometer produced less postural disturbance than was used in our previous studies (Hwang and Huang, 2016) because the balance capacity of the elderly participants was poorer than that of the young adults. The postural task in the level-surface and stabilometer conditions required the participants to couple the ankle joint angle derived from the readings of the electrogoniometer to the target line, based on visual feedback. The target lines for the postural task in the level-surface and stabilometer conditions were set at the horizontal surface and 50% of the maximal anterior tilt, respectively (**Figure 1A**). The postural tasks are known as postural tasks of visual internal focus (Huang et al., 2014), with which the participants should control upright stance with ankle angular displacement (or an internal aspect of body movement). Utilization of an internal focus for a postural task will interfere with postural automatic processes, especially when difficulty is added to stance control in a dual task condition for the elderly (Chiviacowsky et al., 2010). To minimize the potential visual load during the concurrent tasking, the target signals for posture and force-matching were carefully scaled at the same vertical position of the monitor for each participant (**Figure 1A**). We fully understood that the relative task difficulty of the postural and suprapostural tasks was a critical determinant of the reciprocal effect of the posturalsuprapostural task. An earlier pilot experiment had shown that the present dual task setup would not significantly degrade the force-matching accuracy of the young adults between the levelsurface and stabilometer conditions (Hung et al., 2016). In this particular dual task design, stance destabilization was expected to produce a decline in force-matching performance due to increasing postural threat (stabilometer vs. level-surface) in the older participants (Boisgontier et al., 2013). With this design, we were able to examine the age effect on the compensatory mechanisms underlying perseverance of quality of the secondary motor task when balance contexts varied.

Execution of the suprapostural force-matching in an experimental trial was first cued by a warning signal (an 800 Hz tone lasting for 100 ms). Upon hearing an executive tone (a 500 Hz tone lasting for 100 ms), the participants then started a quick thumb-index precision grip (force impulse duration <0.5 s) to couple instantaneously the peak precision-grip force with the force target on the monitor. The warning-executive signal pairs were randomly presented at different intervals of 1.5, 1.75, 2, 2.25, 2.5, 2.75, or 3 s (**Figure 1B**). The interval between the end of the executive tone and the beginning of the next warning tone was 3.5 s. There were a total of 14 warning-executive signal pairs in an experimental trial (80 s per trial) and six experimental trials of the postural-motor dual task for each stance condition. Both young and older subjects were allowed for a fixed rest duration between trials (1 min) to minimize fatigue effect.

#### Experimental Setting

A 40-channel NuAmps amplifier (NeuroScan Inc., EI Paso, TX, USA) with Ag-AgCl scalp electrodes was used to record scalp voltage fluctuations from different 30 EEG channels (Fp1/2, Fz, F3/4, F7/8, FT7/8, FCz, FC3/4, Cz, C3/4, CPz, CP3/4, Pz, P3/4, T3/4, T5/6, TP7/8, Oz, and O1/2). The ground electrode was placed along the midline ahead of Fz. Electrodes placed above the arch of the left eyebrow and below the eye were used to monitor eye movements and blinks. The impedances of all the electrodes were below 5 k and were referenced to linked mastoids of both sides. The EEG data was recorded with a band-pass filter set at 0.1–100 Hz and with a sampling rate of 1 kHz. The electrogoniometer was attached to the dominant ankle joint to record the angular motion of the ankle joint. The electrogoniometer consisted of two sensors. One sensor was placed at the dorsum of the right foot between the second and third metatarsal heads, and the other sensor was fastened along the midline of the middle third of the anterior aspect of lower leg. A load cell (15-mm diameter × 10 mm thickness, net weight = 7 g; Model: LCS, Nippon Tokushu Sokki Co., Japan) on the right thumb was used to record the level of force-matching. All physiological data were synchronized and digitized at a sampling rate of 1 kHz in LabVIEW software (National Instruments, Austin, TX, USA).

## Data Analyses

#### Behavior Data

Normalized force error (NFE) of force-matching was used to represent suprapostural performance in the present study. Force-matching error was represented in terms of NFE, or <sup>|</sup>TF−PGF<sup>|</sup> TF × 100% (PGF: peak grip force; TF: target force; **Figure 2**). The NFEs of all force-matching events were averaged across trials for each subject in the level-surface and stabilometer conditions. The reaction time (RT) of force-matching was denoted as the timing interval between the executive tone and the onset of grip force. Postural performance was characterized with the fluctuation properties of ankle movement during the interval between the warning signal and the onset of force pulse. We applied root mean square (RMS) and sample entropy (SampEn) to assess the amplitude and complexity of the ankle movement fluctuations (AMF\_RMS and AMF\_SampEn) after resampling the kinematic data to 125 Hz. SampEn is an appropriate entropy measure for reliably quantifying the variability structure of biological data with a short

length (Yentes et al., 2013). The mathematical formula for SampEn was

$$\text{SampEn}(m, r, N) = -\log(\frac{\sum\_{i=1}^{N-m} A\_i}{\sum\_{i=1}^{N-m} B\_i})$$

where r = 15% of the standard deviation of the ankle movement fluctuations, m is the length of the template (m = 3), and N is the number of data points in the time series. Ai is the number of matches of the ith template of length m + 1 data points, and Bi is the number of matches of the ith template of length m data points. A SampEn close to 0 represents greater periodicity (or regularity), while a value near 2 represents higher complexity (or irregularity). Higher regularity (or lower SampEn value) of postural sway represents the more attentional focus being paid to postural control, and vice versa (Donker et al., 2007; Borg and Laxåback, 2010; Kuczynski et al., ´ 2011).

#### Component Amplitudes and Functional Connectivity of Multi-Channel ERP

ERP data was analyzed off-line with the NeuroScan 4.3 software program (NeuroScan Inc., EI Paso, TX, USA). Prior to ERP quantitative analysis, third-order trend correction and eye movement correction protocols were applied to the entire set of recorded data to remove the DC shift and eye movement artifacts. The eye movement artifacts were removed from the EEG using regression analysis (Semlitsch et al., 1986), and the number of eye blinks in each trial was roughly 10–15 across subjects. After eye movement was removed, the EEG data were conditioned with a low-pass filter (40 Hz/48 dB roll-off), and then the conditioned EEG data were segmented into epochs of 700 ms, including 100 ms before the onset of each execution signal. Epochs were all baseline-corrected at the pre-stimulus interval. Poor epochs, such as those affected by excessive drift or eye blinks, were discarded by visual inspection (rejection rate of inappropriate trials: <8%). The remaining artifact-free epochs were averaged for an experimental trial in the level-surface and stabilometer conditions, and then the ERP data were also grouped according to a two-factor design (population: the young and older adults; postural task: level-surface and stabilometer stances).

As postural-suprapostural behaviors involve information mastery dependent upon the fronto-motor-parietal network (Huang and Hwang, 2013), we expected age-related differences in regional activity of ERP due to increasing difficulty of the postural subtask of the dual task in the global frontal (GF: Fp1, Fp2, F3, Fz, F4, F7, and F8), sensorimotor (SM: C3, Cz, C4, CP3, CPz, and CP4), and parietal-occipital (PO: P3, Pz, P4, O1, Oz, and O2) areas for the level-surface and stabilometer conditions. The N1 and P2 amplitudes were quantified as the peak amplitude in two separate time windows (80–150 ms, 150–240 ms after executive signal onset). The ERP of each electrode contained N1 and P2 components, which were selectively averaged to obtain amplitudes of the N1 and N2 of the above-mentioned areas. For instance, amplitudes of N1 and P2 recorded from the electrodes of the Fp1, Fp2, F3, Fz, F4, F7, and F<sup>8</sup> were averaged to represent the size of N1 and P2 of the global frontal area.

Based on multi-channel ERP signal, we also quantified statistical interdependencies of non-stationary ERP in the preparatory period with one of the most popular approaches, synchronization likelihood (SL). The SL measures the degrees of linear and non-linear dimensions of EEG/MEG coupling within cortical networks (Leistedt et al., 2009; Boersma et al., 2011). Theoretically, SL takes into account the recurrences of state space vectors occurring at the same moment that are converted from two time-series of interest (Stam et al., 2005). SL can sensitively detect slight variations in the coupling strength for a fine time scale (Stam and van Dijk, 2002), which is appropriate for resolve ERP synchronization patterns in a short period. An SL close to 0 indicates no coupling; an SL of 1 indicates complete coupling. For brevity, detailed descriptions of SL calculation (Stam and van Dijk, 2002; Stam et al., 2003) and parameter settings (Montez et al., 2006) can be found in previous works. Computation of the SL across all pairs of ERP data of the channels in the preparatory phase (the time interval between the executive tone and the force-matching onset) produced a square 30 × 30 SL adjacent matrix. Each entry in the SL adjacent matrix represented the connectivity strength within the functional networks. For each participant, the overall SL adjacent matrix from all experimental trials in the level-surface or stabilometer condition was averaged. SL thresholds from 0.1 to 0.9 were selected to build functional connectomes of different connection strengths. The SL adjacent matrix was rescaled with the proportion of strongest weights, such that all other weights below a given threshold (including SL on the main diagonal) were set to 0. Namely, the selection of the SL threshold of 0.1 merely accounted for the strongest 10% of the weights in the SL adjacent matrix (or functional connectivity in the functional connectome). SL was calculated with the functions of HERMES for Matlab (Niso et al., 2013). The mean value of SL for all the electrode pairs was defined as SL\_All. The mean values of SL that connected to the specified areas, the fronto-sensorimotor (SL\_FSM), and parietal-occipital (SL\_PO) areas, were determined for the level-surface and stabilometer conditions.

#### Statistical Analyses

The purpose of this study was to examine the neural mechanisms underlying age and stance effects on postural-suprapostural performance. The current experimental design focused on the neural mechanisms responsible for differential stance effects on force-matching accuracy between young and older adults. Two way repeated measures ANOVA with population (young and older) and postural load (level-surface and stabilometer) were used to examine the significance of differences in behavior parameters (NFE, RT, AMF\_RMS, and AMF\_SampEn), and the mean SL of the areas of interest (SL\_All, SL\_FSM, and SL\_PO) across different threshold values. The level of significance of the above-mentioned statistical analyses was set at p = 0.05. The significance of the post-hoc test for stance and age effects was p = 0.0125 using the Bonferroni correction. Moreover, networkbased statistics (NBS) were performed to vigorously identify stance-related changes in the functional connectivity of all the node pairs for the young and older groups. For each group, paired t-tests were independently performed at each synchronization value, and t-statistics larger than an uncorrected threshold of t(13) = 3.012 (p = 0.005) were extracted into a set of supra-threshold connections. Then we identified all connected components in the adjacency matrix of supra-threshold links and saved the number of links. A permutation test was performed 5,000 times to estimate the null distribution of the maximal component size, and the corrected p-value was calculated as the proportion of permutations for which the most connected components consisted of two or more links. Methodological details of NBS are documented in Zalesky et al. (2010). The age effect on the topological distribution of significant stance-related differences in synchronization value were examined with visual inspection. Statistical analyses were performed in Matlab (Mathworks Inc. Natick, MA, USA) and SPSS v.19.0 (SPSS Inc. Chicago, IL, USA). All data are presented as mean ± standard error.

### RESULTS

#### Behavior Performance

**Figure 3** shows means and standard errors of task performance of force-matching and postural response for young and older groups under the level-surface and stabilometer conditions. The ANOVA results revealed that NFE was subject to both stance and age effects [stance: F(1, 22) = 10.36, p = 0.004; age: F(1, 22) = 4.60, p = 0.043; stance × age: F(1, 22) = 3.04, p = 0.095]. On account of a marginal interaction effect, we continued the post-hoc analysis which indicated that NFE of the older group was more susceptible to stance configuration and the older adults performed worse force-matching in the stabilometer condition than in the levelsurface condition (p = 0.002). In contrast, NFE of the young group was not affected by stance configuration (p = 0.307). The

fluctuation movements; AMF\_SampEn, sample entropy of ankle fluctuation movements.

RT of the force-matching was not age dependent [F(1, 22) = 1.57, p = 0.223], but varied with stance pattern [F(1, 22) = 4.55, p = 0.044] without a significant interaction [F(1, 22) = 1.06, p = 0.315; Young: level-surface = 307.2 ± 7.8 ms, stabilometer: 326.4 ± 5.6 ms; Older: level-surface = 303.0 ± 8.3 ms, stabilometer: 308.9 ± 4.5 ms]. In terms of RMS, the magnitude of ankle movement fluctuations was also a function of age and stance configuration [stance: F(1, 22) = 67.22, p < 0.001; age: F(1, 22) = 8.50, p = 0.008; stance × age: F(1, 22) = 7.63, p = 0.011]. Post-hoc analysis revealed that both the young and the older adults exhibited greater ankle movement fluctuations during the stabilometer stance than during surface stance (p < 0.001). In particular, the ankle movement fluctuations of the older adults were greater than those of the young adults in the stabilometer condition (p < 0.001). In addition, irregularity of the ankle movement fluctuations was subject to stance configuration rather than to age effect [stance: F(1, 22) = 73.46, p < 0.001; age: F(1, 22) = 2.62, p = 0.120; stance × age: F(1, 22) = 1.79, p = 0.125]. Increases in stance difficulty resulted in a consistently lower AMF\_SampEn (more regularity) of the ankle movement fluctuations in the young and older groups (p < 0.001).

#### ERP Component Amplitude

**Figure 4** show pooled ERP profiles of the each electrode of the young and older adults in the level-surface and stabilometer conditions. Stance-related differences in the ERP profiles were evident in the anterior portions of the cortex, irrespective of the populations. The N1 amplitude in the GF and SM areas varied significantly with age [GF: F(1, 22) = 8.14, p = 0.009; SM: F(1, 22) = 5.54, p = 0.028], but not with stance [GF: F(1, 22) = 2.74, p = 0.112; SM: F(1, 22) = 0.31, p = 0.582] or interaction effects [GF: F(1, 22) = 0.01, p = 0.919; SM: F(1, 22) = 0.15, p = 0.699]. However, N1 amplitude of the PO areas did not significantly vary with stance and age effects [stance: F(1, 22) = 0.03, p = 0.860; age: F(1, 22) = 1.66, p = 0.211; stance × age: F(1, 22) = 0.44, p = 0.513]. In contrast, the P2 amplitudes in the GF, SM, and PO areas were all dependent on stance configuration [GF: F(1, 22) = 12.32, p = 0.002; SM: F(1, 22) = 13.37, p = 0.001; PO: F(1, 22) = 6.03, p = 0.022], rather than on age [GF: F(1, 22) = 0.17, p = 0.683; SM: F(1, 22) = 0.71, p = 0.408; PO: F(1, 22) = 2.30, p = 0.143] or interaction effects [GF: F(1, 22) = 0.79, p = 0.382; SM: F(1, 22) = 0.34, p = 0.566; PO: F(1, 22) = 0.05, p = 0.818; **Figure 5**].

#### Functional Connectivity of ERP in the Preparatory Phase

**Figure 6** presents the mean SL (SL\_All) of all electrode pairs in the level-surface and stabilometer conditions and stance-related change in SL (1SL\_All) as a function of threshold value. For the both groups, SL\_All tended to be larger in the stabilometer condition than in the level-surface condition. **Table 1** shows the detailed results of ANOVA for age and stance effects on SL\_All across different thresholds. For thresholds of 0.1 and 0.2, main effects of stance and age on SL\_All were not significant (p > 0.05). For thresholds of 0.3–0.9, SL\_All was subject to a main effect of stance, and SL was significantly larger in stabilometer condition than in the level-surface condition (p < 0.05). **Figure 7A** presents the mean SL of the electrode pairs in the fronto-sensorimotor network (SL\_FSM) for the levelsurface and stabilometer conditions and stance-related change in SL (1SL\_FSM) as a function of threshold value. **Table 2** summarizes the ANOVA results for age and stance effects on SL\_FSM across different thresholds. For all threshold values, SL\_FSM varied with age and stance configuration (p < 0.05), except for a marginal effect of age for a threshold setting of 0.2. That was, the SL\_FSM of the young and older adults increased in the stabilometer condition for all threshold values (p < 0.006), and the older adults exhibited a larger SL\_FSM than the young adults in the level-surface and stabilometer conditions (p < 0.05). The most remarkable difference in SL modulation for stance difficulty increment between the young and older groups was in the PO area (**Figure 7B**). **Table 3** summarizes the ANOVA results for age and stance effects on SL\_PO across different thresholds. For threshold values of 0.2–0.4, SL\_PO was significantly subject to the interaction effect of age and stance configuration (p < 0.05). For the young adults, post-hoc analysis further showed that SL\_PO in the stabilometer condition was smaller than that in the level-surface condition (p < 0.0125). Notably, such a stancedependent decline in SL\_PO at lower threshold value was not present in the older group (p > 0.0125). Interaction effect of age and stance configuration on SL\_PO for the threshold values of 0.8 and 0.9 was also significant (p < 0.05). Particularly at the threshold value of 0.9, post-hoc analysis revealed that the SL\_PO for the older adults potentiated with increasing postural load (p = 0.009), but not the SL\_PO of the young adults (p > 0.05). The stance-related modulations of the SL\_PO between the young and older adults were opposite for the selection of threshold value (**Figure 7B**).

The significance of spatial distribution change in SL (threshold value = 0.3) with respect to stance configuration was examined with NBS. The threshold was selected to contrast the alterations in the brain wiring diagram at relatively stronger functional connectivity. For the strongest SL with thresholds set at 0.1 and 0.2, the stance-dependent difference in SL variables was not always evident between the young and older adults (**Tables 1**–**3**). **Figure 8** presents the pooled adjacent matrix of SL of preparatory ERP in the level-surface and stabilometer conditions for the young and older groups (threshold value = 0.3). The SL difference of all electrode pairs between the level-surface and stabilometer conditions was labeled with the adjacent matrix of t-values (t > 1.771: stabilometer SL > level-surface SL, p < 0.05; t < −1.771: level-surface SL > stabilometer SL, p < 0.05; **Figure 9**, upper row). The results of NBS indicated that changes in stance configuration significantly altered the brain functional connectivity in both groups (p = 0.0002, corrected; **Figure 9**, lower row). In addition, there were notable topological differences in dual task organization of supra-threshold connectivity for the young and older adults to cope with increasing postural load. The young adults in the stabilometer condition exhibited a global potentiation of supra-threshold connectivity in the fronto-sensorimotor cortex and reduction in supra-threshold connectivity between the left temporal area and the parietal-occipital cortex, as compared with the level-surface condition. In contrast, when postural load increased, the older adults enhanced supra-threshold

connectivity in the fronto-sensorimotor cortex of the bilateral hemispheres and between the frontal and right prefrontal cortex. No significant suppression of supra-threshold connectivity was noted for conducting force-matching with increasing postural load in the older group.

### DISCUSSION

The present postural-supraposatural task produced an expected outcome: the suprapostural performance and the size of postural sway of the older adults were more vulnerable to increasing

postural load than those of the young adults (**Figure 3**). Contrary to the idea of attention withdrawal from the postural task to facilitate supraposture performance (Donker et al., 2007; Kuczynski et ´ al., 2011), additional investment of neural resource on the postural task was necessary to prepare for concurrent force-matching with increasing postural load, in light of increased AMF\_RMS and decreased AMF\_SampEn for the young and older adults in the stabilometer condition. Our behavior results imply distinct brain mechanisms to cope with posture destabilization in young and older adults during a postural-suprapostural dual task. In terms of ERP connectivity analysis, the aged brain exhibited compensatory recruitment of the right prefrontal network and lack of sufficient neural economy for task switching from a postural task to a secondary force-matching act, when postural load multiplied during the stabilometer stance.

#### Increase in the Strength of Functional Connectivity for Postural Destabilization

The primary finding of this study was that the increase in postural load from level-surface to stabilometer stances is associated with a marked increase in global functional connectivity (SL\_All) for the young and older adults (**Figure 6**, **Table 1**). Generally speaking, our data revealed the feasibility of utilizing central residual capacity for the healthy elderly to deal with stance instability during a postural-suprapostural task.



N.S., non-significance.

−, post-hoc analysis should not be processed due to non-significant interaction effect.

The whole fronto-parietal network is integrated to coordinate a postural-suprapostural task (Karim et al., 2013; Ferraye et al., 2014). On account of a stance-related increase in functional connectivity of the fronto-sensorimotor network (SL\_FSM) for both the young and older adults (**Figure 7A**, **Table 2**), it was plausible that force-matching from stabilometer stance caused a shift to a state in which the frontal control predominated, linking to increase in attentional demand to posture stabilization. In fact, several previous studies reported a parallel enhancement of cortical recording from the frontal cortex and supplementary motor area following posture perturbation (Mihara et al., 2008; Fujita et al., 2016). Also, the stabilometer fluctuation movements aggravated externally-induced retinal image motion, so as to hampered precise visual target location (Sipp et al., 2013; Hülsdünker et al., 2015) and then enhance mid-frontal activity for action monitoring and error processing prior to forcematching (Mihara et al., 2008). The stance-dependent increases in SL\_FSM also accounted for the unexpected lack of an increase in N1 amplitude in the stabilometer condition. Our previous work on a posture-motor task revealed that force-matching from unipedal stance led to a greater N1 amplitude than force-matching from bipedal stance (Huang and Hwang, 2013). Originated in the fronto-central region (Adkin et al., 2008), N1 amplitude reflects monitoring of the attentional states (Huang and Hwang, 2013; Huang et al., 2014) and sensory processing (Sibley et al., 2010) of postural perturbation in a posturalsuprapostural task. The insignificant variation in N1 amplitude in this study may partly due to the use of a stabilometer of low curvature that did not produce as much stance instability as unipedal stance would have. Moreover, the most appealing explanation to reconcile the paradoxical finding is that a dual task may not necessarily alter regional activation, instead altering the interactions of the frontal/prefrontal areas with other cortical regions [such as parietal (Gontier et al., 2007) and premotor areas (Marois et al., 2006)]. Of note, the older adults exhibited a stronger SL\_FSM than the young adults (**Figure 7A**, **Table 2**). Although, the older adults seemingly recruited more central resource in the fronto-sensorimotor network in the both stance conditions, yet it could not nicely explain why dual-task performance of the older adults tended to be more vulnerable to higher postural load.

#### Lack of Neural Economy in the Elderly

The interaction effect of age and stance configuration of SL\_PO plays a critical role in age-dependent differences in dual-task performance with increasing postural load. The young adults showed surprising desynchronization of the PO network, in view of the decline in SL\_PO with increasing postural load at the threshold values of 0.2, 0.3, and 0.4 (**Figure 7B**, **Table 3**). During concurrent execution of force-matching in the stabilometer condition, the young adults appeared to avoid division of attentional resources toward multisensory information by dissociating the neuro-anatomical implementation in the PO network. As stabilometer stance did not cause inferior forcematching performance in the young adults (**Figure 3**), the scenario suggests neural economy (Schubert, 2008) or adaptive resource sharing (Mitra and Fraizer, 2004), with which the young adults could minimize the dual-task cost to facilitate task switching for the subsequent force-matching event (Huang et al., 2016). In contrast, the older adults increased the weak functional connectivity (threshold value = 0.9) of the PO network in the stabilometer condition (**Figure 7B**, **Table 3**). The genesis of the relatively weak connectivity simply taxed a limited central resource from the aged brain, because there was no significant performance benefits associated with increasing postural load for the older adults (**Figure 3**).

Further supporting the notion of age-related deficits in neural economy for a postural-suprapostural task is the topology of the wiring diagram (**Figure 9**). The left temporal lobe is known to handle the timing of complex movements with auditory cues (Nakai et al., 2005). Previous fMRI studies revealed that the superior temporal sulcus and posterior middle temporal gyrus of the left hemisphere are more selective to body actions and actions performed on other objects, respectively (Jellema and Perrett, 2006; Assmus et al., 2007). Studies of the macaque monkey (Perrett et al., 1989; Jellema and Perrett, 2003) have shown the superior temporal sulcus to be modulated by body posture during target reaching. Hence, the functional connectivity of the left temporal-parietal-occipital network (TPO network) in a postural-suprapostural task might serve to identify the execution beep (distinguishing the tone from the warning signal) and integration of sensory information from the parietal cortex regarding body schema representation (Pellijeff et al.,

2006) as well as stance fluctuations (Noppeney et al., 2005; Vangeneugden et al., 2011). For the young adults, multimodal sensory integration to detect postural instability using the TPO network was conditionally disengaged before force-matching. In that brief moment, postural control could be temporarily regulated by automatic responses using postural synergy in the midbrain (Jacobs and Horak, 2007). The suppression of information transfer in the left TPO network would augment resource availability upon retrieval of spatial information of the visual target for force-matching (Chelazzi et al., 1993; Kastner and Ungerleider, 2001), such as occurs in solving task conflicts (Schall et al., 2002; Schulz et al., 2011) and facilitating task-switching from stabilometer stance to forcematching (Hwang and Huang, 2016) using frontal executive function. In fact, changes in P2 amplitude with respect to stance configuration also support the argument that the elderly paid less attention to force-matching preparation. In experiments of forcematching to couple a static or a dynamic target, P2 amplitude is inversely related to attentional focus on the visual target of the force-matching task (Huang and Hwang, 2013; Huang et al., 2014). Hence, greater P2 amplitude in the stabilometer condition (**Figure 5**, the right column) might indicate that the participants did not focus well on the force-matching event during posture challenge, especially in older adults.


TABLE 2 | Summary of ANOVA results for age and stance effects on synchronization likelihood of the electrode pairs in the fronto-sensorimotor area (SL\_FSM).

−, post-hoc analysis should not be processed due to non-significant interaction effect.

TABLE 3 | Summary of ANOVA results for age and stance effects on synchronization likelihood of the electrode pairs in the parietal-occipital area (SL\_PO).


−, post-hoc analysis should not be processed due to non-significant interaction effect.

A slower adaptive process for the aged brain is an alternative explanation for why control hubs in the TPO network of the older adults was not dissociated with high postural load. According to the free energy principle (Friston et al., 2006; Friston, 2010), the predictive coding is generated by communication with actual sensory feedback connections to update cortical representations on a trial-by-trial basis (Panichello et al., 2013). When the incoming sensory information coincides with the predictive coding, free energy is minimized. Instead, the brain keeps estimating most-likely likelihood from the information changes in the sensory feedback with environment contexts. For those young adults who could more quickly adapt to stabilometer stance before force-matching, the left TPO network was less activated for a small prediction error when descending prediction efficiency interpreted the actual sensory input. A natural consequence of aging causes slow adaptation and deviance detection of environmental changes. For our older cohort, the TPO network in the stabilometer condition was not suppressed, because they still kept reinforcing internal generative model by comparing of predictions and actual sensory inputs (primarily the ventral visual sources) till free energy was optimally minimized (Panichello et al., 2013).

#### Compensatory Recruitment of the Right Prefrontal Network in the Elderly

Unlike the young adults, the older adults revealed stancerelated enhancements of functional connectivity between the right pre-frontal and frontal areas (**Figure 9**, right). In contrast to the level-surface condition, the associated force-matching with the stabilometer stance recruited more attentional resources to deal with increases in postural sway and stance-induced difficulty in target detection prior to force-matching. Recently, the prefrontal area has been linked to balance control, especially when unexpected external postural perturbation is provided (Maki and McIlroy, 2007; Mihara et al., 2008). In healthy adults and stroke patients, the right prefrontal lobe plays a greater role than the left prefrontal lobe in stance control during postural perturbation (Ugur et al., 2000; Fujita et al., 2016). Prefrontal lateralization is related to resetting eye positions in accordance with using spatial working memory processes

Frontiers in Aging Neuroscience | www.frontiersin.org

(Mihara et al., 2008; Fujita et al., 2016). The dorsolateral prefrontal cortex was also found to play a critical role in goaldirected behavior (de Wit et al., 2009), integrating environmental contexts with information on body positioning of the elderly (Wang et al., 2016). The prefrontal area, which receives cerebellum influences via the thalamus (Middleton and Strick, 2001), has dense projections to the pontine nuclei (Ramnani et al., 2006) for reflexive control of postural balance following external stance perturbation (Hartmann-von Monakow et al., 1981; Mihara et al., 2008). Previous behavioral studies have shown that the elderly often prioritize attentional allocation to posture maintenance above supraposture performance with increasing stance destabilization, known as a "posture-first strategy" (Shumway-Cook and Woollacott, 2000; Doumas et al., 2008). Hence, the age-dependent compensatory recruitment of the right prefrontal lobe is in a good agreement with the greater need for spatial attention of the elderly adults in the posturalsuprapostural task with increasing postural load. The prevailing attentional focus on the posture subtask of the elderly added to the difficulty in task-switching at the cost of inferior forcematching performance. In fact, for healthy cognitive aging, additional recruitment of cognitive processes mediated by the prefrontal cortex and its vast interconnections were found to increase with task demand (Allali et al., 2014; Toepper et al., 2014), in accordance with the compensation-related utilization of neural circuits hypothesis (Reuter-Lorenz and Cappell, 2008).

groups, (t > 1.771: stabilometer SL > level-surface SL, p < 0.05; t < −1.771: level-surface SL > stabilometer SL, p < 0.05). The lower plots display the results of connectivity analysis with network-based statistics (threshold value = 0.3). A contrasting wiring diagram shows topological distributions of the suprathreshold connectivity that vary with stance difficulty increase for the young and older groups. Red line: stabilometer connectivity of supra-threshold > level-surface connectivity of supra-threshold, p < 0.005; blue line: level-surface connectivity of supra-threshold > stabilometer connectivity of supra-threshold, p < 0.005.

#### Methodology Issues

To date, SL has most commonly been used to characterize multiple synchronized neural sources in the brain (Stam and van Dijk, 2002) using low-density (Smit et al., 2012; Herrera-Díaz et al., 2016) or high-density EEG (Polanía et al., 2011). The major methodological advantage of using SL is that it can sensibly detect slight and intricate variations in the coupling strength (Koenis et al., 2013), resolving rapid synchronization patterns in the non-stationary ERP profile at a short time scale (Stam and van Dijk, 2002; Betzel et al., 2012). Some simulation studies have argued that SL could bring about spurious coupling due to a volume conduction effect (Stam et al., 2005; Tognoli and Kelso, 2009). The authors agree that an increase in postural challenge probably led to enhanced volume conduction for the recruitment of more neurons in the stabilometer condition, especially for those functional connectivity grouped with neighboring regions. If the physical synchronization did exist, we could not completely deny overestimation of the functional connectivity within neighboring electrodes. However, physical synchronization does not rationally explain our major finding of age-related differences in connectivity reorganization with increasing postural load. In contrast to the young adults, the older adults did not exhibited a polarization modulation of spatiallydistributed communities for the FSM and TPO networks with increasing postural load (**Figure 9**). Particularly for the stronger functional connectivity, the stance-related modulations of the two distinct networks in young adults and lack of the paralleling connectivity change in the elderly were hard to reconcile with the global rise or fall of state transition to intermittent activity of a single common-source volume conduction. Moreover, we did not intend to specify any age-related differences in functional connectivity of neighboring electrodes or small-range excitability of a recording electrode, as we did accentuate agedependent parametric changes on a network basis (**Figures 5**– **7**). Although, some measures of functional connectivity, such as phase lag index, have been proposed to minimize common sources (Stam et al., 2007), phase-based approaches that can be more sensitive to noise measure temporal-spatial properties of functional connectivity that are quite different from SL (Vinck et al., 2011). Phase-based approaches are not appropriate for accessing the inter-dependences of two short-length ERP profiles in the presence of non-stationarities (Cohen, 2015). As no quantification of functional connectivity is perfect, future work may consider a combination of local cerebral hemodynamic properties and spontaneous neural activity. However, it is not

#### REFERENCES


feasible to fully rule out potential common sources with the present setup.

### CONCLUSION

At the neural level, the present work first reveals neural underpinnings which are associated with (but not necessarily causal to) why dual-task performance of the older adults is tended to be more easily affected by postural load increment. In comparison with the young adults, high postural load produced more dual-task cost of the older adults, pertaining to lack of neural economy to timely deactivate the left TPO network. In addition, the age-dependent compensatory recruitment of the right prefrontal lobe indicates that the older adults may need a greater spatial attention for dual-task control against fluctuating stabilometer movements.

#### ETHICS STATEMENT

National Cheng Kung University Hospital Research Ethics Committee. A physical therapist explained the study purpose and experiment procedure for each participant. All participants gave informed consent to participate according to a protocol approved by the local ethics committee (University Hospital, National Cheng Kung University, Taiwan). There is no vulnerable populations in the present study.

### AUTHOR CONTRIBUTIONS

Substantial contributions to the conception or design of the work or the acquisition: CH, LL, and IH. Analysis or interpretation of data for the work: CH and IH. Drafting the work or revising it critically for important intellectual content: CH and IH. Final approval of the version to be published and agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved: CH, LL, and IH.

#### FUNDING

This research was supported by a grant from the Ministry of Science and Technology, Taiwan, ROC, under grant no. MOST 103-2314-B-002-007-MY3 and MOST 104-2314-B-006- 016-MY3.

Assmus, A., Giessing, C., Weiss, P. H., and Fink, G. R. (2007). Functional interactions during the retrieval of conceptual action knowledge: an fMRI study. J. Cogn. Neurosci. 19, 1004–1012. doi: 10.1162/jocn.2007.19.6.1004


is preserved in Alzheimer disease. Am. J. Geriatr. Psychiatry 14, 52–61. doi: 10.1097/01.JGP.0000192490.43179.e7


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Huang, Lin and Hwang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# MEG Beamformer-Based Reconstructions of Functional Networks in Mild Cognitive Impairment

Maria E. López1,2 \* † , Marjolein M. A. Engels<sup>3</sup>† , Elisabeth C. W. van Straaten4,5 , Ricardo Bajo<sup>6</sup> , María L. Delgado7,8, Philip Scheltens<sup>3</sup> , Arjan Hillebrand<sup>4</sup> , Cornelis J. Stam<sup>4</sup> and Fernando Maestú2,6,8

<sup>1</sup> Laboratory of Neuropsychology, Universitat de les Illes Balears, Palma de Mallorca, Spain, <sup>2</sup> Networking Research Center on Bioengineering, Biomaterials and Nanomedicine, Madrid, Spain, <sup>3</sup> Alzheimer Center and Department of Neurology, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, Netherlands, <sup>4</sup> Department of Clinical Neurophysiology and MEG Center, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, Netherlands, <sup>5</sup> Nutricia Advanced Medical Nutrition, Nutricia Research, Utrecht, Netherlands, <sup>6</sup> Laboratory of Cognitive and Computational Neuroscience, Center for Biomedical Technology, Complutense University of Madrid and Technical University of Madrid, Madrid, Spain, <sup>7</sup> Seniors Center of the District of Chamartín, Madrid, Spain, <sup>8</sup> Department of Basic Psychology II, Complutense University of Madrid, Madrid, Spain

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Chunbo Li, Shanghai Jiao Tong University, China Manousos A. Klados, Max Planck Institute for Human Cognitive and Brain Sciences (MPG), Germany Xu Lei, Southwest University, China

> \*Correspondence: Maria E. López

meugenia.lopez@uib.es †These authors have contributed equally to this work.

Received: 15 December 2016 Accepted: 04 April 2017 Published: 25 April 2017

#### Citation:

López ME, Engels MMA, van Straaten ECW, Bajo R, Delgado ML, Scheltens P, Hillebrand A, Stam CJ and Maestú F (2017) MEG Beamformer-Based Reconstructions of Functional Networks in Mild Cognitive Impairment. Front. Aging Neurosci. 9:107. doi: 10.3389/fnagi.2017.00107 Subjects with mild cognitive impairment (MCI) have an increased risk of developing Alzheimer's disease (AD), and their functional brain networks are presumably already altered. To test this hypothesis, we compared magnetoencephalography (MEG) eyes-closed resting-state recordings from 29 MCI subjects and 29 healthy elderly subjects in the present exploratory study. Functional connectivity in different frequency bands was assessed with the phase lag index (PLI) in source space. Normalized weighted clustering coefficient (normalized Cw) and path length (normalized Lw), as well as network measures derived from the minimum spanning tree [MST; i.e., betweenness centrality (BC) and node degree], were calculated. First, we found altered PLI values in the lower and upper alpha bands in MCI patients compared to controls. Thereafter, we explored network differences in these frequency bands. Normalized Cw and Lw did not differ between the groups, whereas BC and node degree of the MST differed, although these differences did not survive correction for multiple testing using the False Discovery Rate (FDR). As an exploratory study, we may conclude that: (1) the increases and decreases observed in PLI values in lower and upper alpha bands in MCI patients may be interpreted as a dual pattern of disconnection and aberrant functioning; (2) network measures are in line with connectivity findings, indicating a lower efficiency of the brain networks in MCI patients; (3) the MST centrality measures are more sensitive to detect subtle differences in the functional brain networks in MCI than traditional graph theoretical metrics.

Keywords: mild cognitive impairment, magnetoencephalography, phase lag index, brain networks, minimum spanning tree

### INTRODUCTION

fnagi-09-00107 April 22, 2017 Time: 15:37 # 2

### Mild Cognitive Impairment

Mild cognitive impairment (MCI) is an intermediate state between normal aging and dementia. Patients suffering from MCI have an increased risk of developing dementia, in particular Alzheimer's disease (AD; Shah et al., 2000; Farias et al., 2005). Those MCI subjects with positive biomarkers for AD (i.e., amyloid deposition and neural injury markers such as accumulations of intracellular tau or medial temporal lobe atrophy [MTA]) are regarded to be at the symptomatic pre-dementia phase of AD, and are often referred to as "MCI due to AD" (Albert et al., 2011). The pathological processes that these biomarkers indicate are well described in AD (Braak and Braak, 1991) and are known to produce synaptic disruptions. AD has been described as a "disconnection syndrome," not only at the cellular level, but also at the macroscale, since the connections in the brain networks also seem to be disrupted (Blennow et al., 1996; Selkoe, 2002; Delbeuck et al., 2003; Arendt, 2009; Takahashi et al., 2010). In fact, structural and functional changes have been described in MCI subjects, suggesting that this disconnection of brain networks already begins during the MCI-stage of AD (Pijnenburg et al., 2004; Koenig et al., 2005; Buldú et al., 2011; Wang et al., 2013).

### Functional Connectivity

One of the main concepts used to understand how the different brain areas interact is functional connectivity (Friston, 1994), which reflects the statistical interdependencies between twotime series of physiological activity. Several resting-state electroencephalography (EEG) and magnetoencephalography (MEG) studies have found decreases in functional connectivity, especially in higher frequency bands (i.e., alpha and beta bands), in MCI patients compared to healthy controls (Moretti et al., 2008; Gómez et al., 2009; López et al., 2014b; Cuesta et al., 2015). This change in synchronization pattern is quite similar to that described for AD patients (Stam and van Dijk, 2002; Jeong, 2004; Stam et al., 2006), although increased connectivity has also been described for AD involving posterior brain regions (Stam et al., 2006; Alonso et al., 2011). Reductions in functional connectivity in MCI have also been observed between regions of the default mode network (DMN), with a parallel disruption of the anatomical connections (Garcés et al., 2014; Pineda-Pardo et al., 2014). However, some resting-state studies comparing different MCI groups have also detected a specific hypersynchronization pattern in high frequency bands (alpha and beta) in those MCI subjects who finally developed AD (López et al., 2014a), or those who presented abnormal concentration of phospho-tau (p-tau) protein in the cerebrospinal fluid (CSF; Canuet et al., 2015). In a recent multicenter study, this profile of hypersynchronization was used to obtain a high percentage of correct classification of MCI and healthy controls (Maestú et al., 2015).

### Brain Networks

Based on the estimated functional connectivity between time series, a weighted network can be reconstructed using graph theory (Bullmore and Sporns, 2009). For this purpose, brain systems are described as sets of nodes (i.e., brain regions or sensors) and links (i.e., functional connections between nodes). The topology of these networks can then be characterized, for example, providing information about the local integration of the network (e.g., the "clustering coefficient") as well as the global integration (e.g., the "path length"; see Rubinov and Sporns, 2010 for a review). A small-world network has a high local connectedness (quantified by a large clustering coefficient) and a high global integration (quantified by a short path length) and has been regarded as a network with an optimal topology for the transfer of information. AD patients exhibit brain networks that appear to have a sub-optimal topology in which the networks have shifted toward a more random configuration. This was mainly characterized by a loss of small-worldness (see Tijms et al., 2013 for a review), supporting the hypothesis of the disconnection syndrome. However, Tijms et al. (2013) also show that the results differ drastically between studies. Only few studies, using different approaches and modalities, have explored the network topology in MCI, reporting a disturbed balance between local and global integration [functional Magnetic Resonance imaging (fMRI); Wang et al., 2013; MEG; Buldú et al., 2011; Pineda-Pardo et al., 2014]. Methodological difficulties make the comparison between networks of different sizes and different edge densities challenging, if not impossible (Lee et al., 2006; van Wijk et al., 2010; Stam, 2014; Tewarie et al., 2015). This might lead to contradicting results that could be due to differences in modalities (Tijms et al., 2013), but also due to methodological biases (van Wijk et al., 2010). One solution is to reconstruct the minimum spanning tree (MST; Stam, 2014). The MST is a sub-graph of the complete network, which forms a backbone of the original graph. It is uniquely defined whilst avoiding the arbitrary choices of traditional approaches, therefore solving the limitations of previous graph studies (van Wijk et al., 2010; Tewarie et al., 2015). Despite its advantages and application in other neurological disorders (Ortega et al., 2008; Fraschini et al., 2014; Olde Dubbelink et al., 2014; Otte et al., 2015; Tewarie et al., 2015), there is only one fMRI study and one EEG study that has used the MST in comparing healthy elders and AD patients (Çiftçi, 2011; Engels et al., 2015), and none that studied the MST in patients with MCI.

For this reason, we performed an MEG study with MCIs and healthy controls with the aim to characterize how the functional network organization in the MCI stage differs from that of controls. To this end, we estimated MEG resting-state functional connectivity between cortical regions and characterized the topology of the reconstructed MST. We expected to find subtle differences in functional connectivity between MCI patients and controls while the graph theoretical measures will show a clear disrupted topological pattern in MCI. We expected that the novel MST measures give more insight in the network changes of MCI than the traditional network measures (normalized Cw and normalized Lw).

### MATERIALS AND METHODS

fnagi-09-00107 April 22, 2017 Time: 15:37 # 3

### Subjects

Magnetoencephalography recordings were obtained from 58 subjects (29 MCI patients and 29 healthy elderly subjects). The MCI group was recruited from the Hospital Universitario San Carlos (Madrid), and the control group from the Seniors Center of the District of Chamartin (Madrid). All subjects were right handed (Oldfield, 1971) and native Spanish speakers.

All participants were screened by means of standardized diagnostic instruments and also received an exhaustive neuropsychological assessment. To evaluate their global and cognitive functional status there were used: the Spanish version of the Mini-Mental State Examination (MMSE; Lobo et al., 1979), the Global Deterioration Scale (GDS; Reisberg et al., 1982), the Functional Assessment Questionnaire (FAQ; Pfeffer et al., 1982), the Geriatric Depression Scale-Short Form (GDS-SF; Yesavage et al., 1982), the Hachinski Ischemic Score (Rosen et al., 1980), the questionnaire for Instrumental Activities of Daily Living (Lawton and Brody, 1969), and the Functional Assessment Staging (FAST; Auer and Reisberg, 1997).

Additionally, all subjects underwent an extensive neuropsychological assessment to explore their cognitive functioning by using the following tests: direct and inverse digit span test (DDS and IDS, Wechsler Memory Scale III, WMS-III; Wechsler, 1997), immediate and delayed recall (IR and DR, WMS-III; Wechsler, 1997), phonemic and semantic fluency (PhF and SF, controlled oral word association test; Benton and Hamsher, 1989), ideomotor praxis of Barcelona test (IP; Peña-Casanova, 1990), Visual Object and Space Perception Test (VOSP; Warrington and James, 1991), Boston Naming Test (BNT; Kaplan et al., 1983), and Trail-Making Test (TMT), parts A and B (TMT-A and TMT-B; Reitan, 1958).

The MCI diagnosis was established according to the National Institute on Aging- Alzheimer Association (NIA-AA) criteria (Albert et al., 2011), which includes: (i) self- or informant-reported cognitive complaints; (ii) objective evidence of impairment in one or more cognitive domains; (iii) preserved independence in functional abilities; and (iv) not demented (McKhann et al., 2011). Besides meeting the clinical criteria, MCI participants had signs of neuronal injury (hippocampal volume measured by magnetic resonance imaging (MRI). So, they might be considered as "MCI due to AD" with an intermediate likelihood (Albert et al., 2011).

None of the participants had a history of psychiatric or neurological disorders (other than MCI). General inclusion criteria were: an age between 65 and 80, a modified Hachinski score ≤ 4, a short-form Geriatric Depression Scale score ≤ 5, and T1/T2-weighted MRI within 12 months and 2 weeks before MEG screening without indication of infection, infarction, or focal lesions (rated by two independent experienced radiologists; Bai et al., 2012). Patients were off those medications that could affect MEG activity, such as cholinesterase inhibitors, 48 h before recordings.

### Ethics Statement

Methods were carried out in accordance with the approved guidelines. The study was approved by the Hospital Universitario San Carlos Ethics Committee (Madrid), and all participants signed a written informed consent prior to participation.

### MEG Acquisition

Magnetoencephalography signals were measured by a 306 channel Vectorview system (Elekta Neuromag Oy) at the Center for Biomedical Technology (Madrid, Spain), inside a magnetically shielded room (VacuumSchmelze GmbH, Hanau, Germany) within several days after the neuropsychological assessment. Brain magnetic fields were recorded in the morning during a task-free 3 min eyes-closed resting-state condition, while subjects sat comfortably. Participants were instructed to move as little as possible, and were monitored during the recording to ensure that they did not fall asleep.

Sampling frequency was 1000 Hz with an online filter with bandwidth 0.1–300 Hz. The position of the head inside the sensor array was determined using a head-position indicator (HPI) with four coils attached to the scalp (two on the mastoids and two on the forehead). These four coils along with the head shape (∼500 points) of each subject (referenced to three anatomical fiducials: nasion, and left-right preauricular points) were acquired using a three-dimensional Fastrak Polhemus system (manufacturer: Polhemus, Inc., USA). Vertical ocular movements were measured by two bipolar electrodes attached above and below the left eye, and a third one to the earlobe, for electrical grounding.

Maxfilter software (version 2.2, Elekta Neuromag Oy) was used to remove noise from the MEG data using the temporal extension of signal space separation (tSSS) with movement compensation (Taulu and Simola, 2006). Flat channels, or those that contained excessive artifacts, were manually discarded after visual inspection of the data by one of the authors (M. E. López) before estimation of the SSS coefficients. The tSSS filter was subsequently used to remove noise signals that SSS failed to discard, typically from noise sources near the head, using a subspace correlation limit of 0.9 (Medvedovsky et al., 2009) and a sliding window of 10 s.

#### MRI Acquisition

3D T1 weighted anatomical brain MRI scans were collected with a General Electric 1.5T MRI scanner, using a high resolution antenna and a homogenization PURE filter [Fast Spoiled Gradient Echo (FSPGR) sequence with parameters: TR/TE/TI = 11.2/4.2/450 ms; flip angle 12◦ ; 1 mm slice thickness, a 256 × 256 matrix and FOV 25 cm].

FreeSurfer software (version 5.1.0; Fischl et al., 2002) was used to obtain the hippocampal volumes, which were normalized with the overall intracranial volume (ICV) of each subject.

### Co-registration and Beamforming

The outline of the scalp, as obtained from the subject's structural MRI, was used for co-registration with the MEG data using the VUmc Amsterdam co-registration surface matching software, resulting in an estimated co-registration accuracy of

approximately 4 mm (Whalen et al., 2008). A single sphere fitted to the scalp surface was used as a volume conductor model for the beamformer analysis. An atlas-based beamformer was used to project the MEG sensor signals to 78 cortical regions-of-interest (ROIs) from the Automatic Anatomical Labeling (AAL) atlas (see Supplementary Material) (Tzourio-Mazoyer et al., 2002; Gong et al., 2009). Based on the broad-band (0.5–48 Hz) beamformer weights, time series of neuronal activity were reconstructed for the voxel with the maximum power within a ROI for each frequency band separately, i.e., a virtual electrode that was representative for that specific ROI was reconstructed. A detailed description of this procedure is given in Hillebrand et al. (2012).

#### MEG Analysis

Per subject, five artifact free trials of approximately 16.384 s (four times 4096 samples) were selected after careful visual inspection, giving a total of 20 epochs of 4096 samples for further analysis. Time-series of neuronal activation were computed for the six frequency bands: delta (0.5–4 Hz), theta (4–8 Hz), lower alpha (8–10 Hz), upper alpha (10–13 Hz), beta (13–30 Hz), and gamma (30–48 Hz). Selected epochs were converted to ASCIIfiles and imported into an in-house developed software package BrainWave version 0.9.125, developed by one of the authors (C. J. Stam) and available at: http://home.kpn.nl/stam7883/brainwave. html.

#### Functional Connectivity

Functional connectivity was assessed with the PLI, which quantifies the consistency of a phase relationship between two signals while zero-lag (mod π) phase differences are ignored (Stam et al., 2007b). Therefore, the PLI is insensitive to spurious interactions caused by the effects of volume conduction and/or field spread (Stam et al., 2007b; Hillebrand et al., 2012). The PLI ranges between 0 and 1 in which 0 represents no consistent coupling or coupling with zero-lag and one represents consistent phase-lagged coupling. First, the instantaneous phase for each time series is computed by taking the argument of the analytic signal (Stam et al., 2007b) as computed using the Hilbert transform. Second, we calculate the asymmetry of the distribution of instantaneous phase differences between two time series:

$$\text{PLI} = |\lnot \text{sign}[\sin(\Delta \Phi\_t)]| > |\lnot \text{sign}[\}|$$

where the phase difference 18<sup>t</sup> is defined in the interval [-π, π], <> denotes the mean value, sign stands for signum function, | | indicates the absolute value, and t corresponds to time samples 1, . . ., Ns, where Ns is the number of samples. By calculating the PLI values between all pairs of ROIs, we obtained a 78 × 78 adjacency matrix, which we used for the network analyses (see below).

#### Small-Worldness

A low characteristic path length (L) and a high clustering coefficient (C) characterize a small-world network (Watts and Strogatz, 1998). In an unweighted network, the C represents the probability that two nodes are connected when they share a neighboring node and the L represents the average of the shortest distance between pairs of nodes, with distance defined by the number of links between nodes. From the weighted graph, the weighted clustering coefficient (Cw) and weighted characteristic path length (Lw) were calculated as described in Stam et al. (2007a). Fifty random control networks were created by randomly shuffling the PLI values in each adjacency matrix while keeping the matrix symmetry intact. For each ensemble of 50 random networks, the average Cw (random) and Lw (random) were computed. The observed network values were divided by the average values obtained for the random networks in order to create normalized values. The resulting normalized clustering coefficient (normalized Cw) and normalized path length (normalized Lw) were used for further analyses. These measures were computed for each epoch, and then averaged over the epochs for each subject.

#### Minimum Spanning Tree

We constructed the MST from the weighted adjacency matrix containing the PLI values. The MST is a unique subgraph that connects all nodes in the network by the strongest connections (defined as the network links with the highest PLI values) without forming cycles (Stam et al., 2014), and was reconstructed using Kruskal's algorithm (Kruskal, 1956). By using 1/PLI as input to the algorithm, strongest connections are likely to be included in the MST, as long as no cycles are formed. The MST was characterized by the following measures: degree, betweenness centrality (BC), eccentricity, degree distribution (κ), the number of leafs, degree correlation (R), tree hierarchy and diameter. MST measures were computed for each epoch, and then averaged over the epochs for each subject. The degree describes how many links each node has. BC is a measure of the importance of a node within the network. The BC of node i is defined as the number of shortest paths in the network that run through a specific node, divided by the total number of shortest paths from any node to all other nodes in the MST. In our calculations, we used the maximum BC across all nodes as well as per node for further testing. The eccentricity of a node is defined as the longest distance between that node and any other node in the network. The degree distribution is formed by the likelihood (P) that a randomly chosen node of the network will have degree κ (the number of connections of a specific node); it is a plot of P(κ) as a function of κ (Stam and Reijneveld, 2007). The degree correlation is an index of how much the degree of a node is correlated to the degree of nodes it is connected to. The leaf number is the number of nodes that have a degree of 1 representing the "leafs" or "extremities" of the network. Tree hierarchy quantifies the tradeoff between large scale integration in the MST and the overload of central nodes. The diameter represents the longest path in the MST. For more information about these measures, we refer to Boersma et al. (2013), Stam et al. (2014), Tewarie et al. (2015).

#### Statistical Analysis

Subject's characteristics were tested using independent samples t-tests or chi-square tests where appropriate using SPSS (20.0 for windows). For the PLI, permutation testing, based upon t-statistics, was used for each pair of regions [among the 78 studied, (78 × 77)/2 in total], and in each frequency band, with the aim of comparing both groups (Maris and Oostenveld, 2007).

To this end, participants were randomly divided into two sets with the same size as the original groups (29 vs. 29 subjects). This procedure was repeated 2000 times (2000 permutations). A new t-test between each pair of regions [(78 × 77)/2 couples] in each frequency band, was then carried out using these two newly created groups, getting a t-value for each pair of regions and each frequency band. After sorting these 2001 t-test results [2000 corresponding to "randomly divided" groups and another one for the original (MCI vs. Control) subject's distribution], only p-values within the 5% of lower values were considered statistically significant [note that this is for each pair of regions and each frequency band studied: hence it was repeated ((78 × 77)/2) × (number of frequency bands) times]. These analyses were uncorrected for the number of permutations performed and therefore serve as exploratory analyses (see **Figures 1, 2** and **Tables 2, 3**). Afterward, we performed an FDR correction on the data with the goal to examine the statistical significant results that survive a multiple comparison correction. For the network analysis, we focused on those frequency bands in which the connectivity analyses showed (uncorrected) significant differences between the groups. Again, corrected and uncorrected (exploratory) permutation testing was used to compare the

precentral gyrus (53) – left postcentral gyrus (16) – left anterior cingulate and paracingulate gyri (36) – right superior occipital gyrus (61) – right cuneus (65)

for the lower alpha band. AAL numbers appear in parentheses.

groups. All statistical analyses were performed using MATLAB (R2015b, Mathworks).

### RESULTS

### Demographics

Subject characteristics are shown in **Table 1**. Controls and MCI patients did not differ in age, gender, or educational level. As expected, the scores of MMSE and two measures of episodic memory (immediate and delayed recall) were both lower in MCI patients compared to controls. Additionally, hippocampal volumes were both lower in the MCI group.

#### Functional Connectivity

Significant differences between MCI patients and controls in PLI using the uncorrected permutation tests were obtained in the lower (8–10 Hz) and upper (10–13 Hz) alpha band (**Figure 1**). In the lower alpha band, compared to the control group, PLI in the MCI group was lower between left superior frontal gyrus -orbital part- and left gyrus rectus, and between left superior temporal gyrus and left insula (p < 0.01). Additionally, in the same frequency band, MCI subjects showed higher PLI values than controls between four regions, namely between right superior frontal gyrus (medial orbital) and right precentral gyrus, right superior frontal gyrus (dorsolateral) and left superior frontal gyrus (dorsolateral), left anterior cingulate and paracingulate gyri and left postcentral gyrus, and between right cuneus and right superior occipital gyrus.



Mean ± standard deviation (SD) and p-values. The p-values were obtained by twoindependent samples t-test (<sup>∗</sup> ) or chi-square test (∗∗) where relevant. M, male; F, female; MMSE, Mini Mental State Examination score; RH\_ICV and LH\_ICV, right and left hippocampal volume normalized with intracranial volume, respectively. Highest education completed, using five levels: (1) Illiterate, (2) Primary studies, (3) Elemental studies, (4) High school studies, and (5) University studies.

(17), right supramarginal gyrus (58) and right anterior cingulate and paracingulate gyri (75) for the upper alpha band. AAL numbers appear in parentheses.

Furthermore, in the upper alpha band, MCI subjects presented lower PLI values than controls between four regions, namely between right gyrus rectus and right olfactory cortex, left insula and left middle temporal gyrus, right parahippocampal gyrus and right fusiform gyrus, and right inferior temporal gyrus and left precuneus.

After FDR-correction for multiple comparisons, none of these significant differences survived.

#### Small-worldness

We focused the network analysis on those frequency bands in which there were differences in connectivity, namely in lower and upper alpha bands.

Using the uncorrected permutation tests, no differences between controls and MCI subjects were found for normalized Cw and normalized Lw.

#### Minimum Spanning Tree

We found group differences in BC and degree in lower and upper alpha bands using the uncorrected permutation tests (see below).

Betweenness centrality results are shown in **Figure 2** and **Table 2**. In the lower alpha band, there were no differences between controls and MCI patients in the maximum value of BC globally. However, compared to the healthy controls, the MCI group showed higher BC values in eight brain areas: the left preand post-central gyri, the left and right superior parietal gyri, the right middle frontal gyrus, the right superior and middle temporal gyrus (temporal pole) and the right insula; and lower values in one brain area: the left anterior cingulated/paracingulate gyri (see Supplementary Material).

In the upper alpha band, we did not find global differences between the groups, but there were differences in BC for specific brain areas. Compared to controls, MCIs exhibited higher values in three brain areas: the left superior frontal gyrus dorsolateral (lSFGdor), the right supramarginal gyrus (rSMG), and the right anterior cingulate/paracingulate gyri (rACG); and lower BC values in three brain areas: the left middle temporal gyrus (lMTG), the right fusiform gyrus (rFFG) and the right parahippocampal gyrus (rPHG).

Finally, the MCI group showed higher MST degree values than controls in the lower alpha band in five brain areas: the left superior frontal gyrus, dorsolateral, the left postcentral gyrus, the left Heschl gyrus, the right middle frontal gyrus and the right calcarine fissure and surrounding cortex; while they exhibited a lower degree value in one brain area: the right olfactory cortex. In the upper alpha band, MCI subjects exhibited higher degree values in four brain areas: the left superior frontal gyrus, dorsolateral, the left superior parietal gyrus, the rSMG and the


#### TABLE 2 | Mean ± standard deviation (SD) of the betweenness centrality (BC) in the lower and upper alpha band, p-values were obtained using permutation testing without correction for multiple comparisons across regions.

Arrows indicate lower or higher values in the MCI group.

right anterior cingulate and paracingulate gyri; and lower degree values in three brain areas: the lMTG, the right inferior frontal gyrus, orbital part and the rPHG. MST degree results are shown in **Figure 2** and **Table 3**.

After FDR-correction for multiple comparisons, none of these significant differences survived. Also, we did not find any differences between the controls and the MCI group for any of the other MST measures in these two frequency bands.

#### DISCUSSION

With the aim to corroborate our hypothesis about the differences in both functional connectivity and network organization between healthy aging and MCI, we performed a functional connectivity (PLI) and MST analyses in resting state MEG data. The main finding of this study was the detection of differences in both functional connectivity and brain network topology in a group of patients with MCI compared to controls. Note however, that these results are exploratory and the significance between the groups did not survive FDR correction for multiple comparisons. The uncorrected connectivity results showed that the MCI patients exhibited more increases than decreases in PLI values in the lower alpha band, and decreases in the upper alpha band. As differences in connectivity between both groups were found in the alpha band, we examined differences of network's topography in this frequency band by using concepts from graph theory. We did not find any group difference in weighted clustering and path length, but regionally we obtained higher BC and degree values when examining the MST in the MCI group in lower alpha band, and both increases and decreases in the upper alpha band.

Mild cognitive impairment patients demonstrated lower PLI values in the lower alpha band that affected frontal and temporal brain areas within the left hemisphere. Using EEG, an overall decrease in the lower alpha band has been observed in AD


TABLE 3 | Mean ± standard deviation (SD) and p-values after a t-test without correction for multiple comparisons for degree values in the control group compared with the MCI group in lower and upper alpha bands.

Arrows indicate lower and higher values in the MCI group.

patients (Stam et al., 2006, 2009), and also in MCI patients (Babiloni et al., 2006). In a recent study performed with EEG data in AD (Engels et al., 2015), the decrease of connectivity in the lower alpha band was related to the severity of the disease, mainly over posterior areas. However, in the present study, MCI patients also showed an increase in connectivity between intra- and interhemispheric frontal areas, and in right posterior regions. This intra- and inter-hemispheric increase in connectivity has been usually described in the MCI population while performing a cognitive task. Pijnenburg et al. (2004) found an increase in lower alpha band in MCIs compared to subjects with subjective memory complaints (SMC) during a visual working memory (WM) task. Jiang (2005) and Zheng et al. (2007) obtained higher coherence values in both lower and upper alpha bands during an arithmetic WM paradigm in MCIs compared to healthy controls. In addition, an MEG study performed in progressive MCI patients (pMCI) found a higher synchronization in those patients who finally developed AD, compared with those who remained stable over time (stable MCI, sMCI), in lower alpha and upper alpha bands while performing a memory task (Bajo et al., 2012). In the same vein, a recent resting-state MEG study which did not divide the alpha band into two sub-bands, found that patients with MCI that eventually converted to AD, exhibited a higher connectivity in this frequency range than those MCI patients that remained stable over time, between the right anterior cingulate and temporo-occipital brain regions (López et al., 2014a). Our findings add to the current knowledge that results of functional connectivity in MCI patients are dependent on the region and on the frequency band. However, there is no consistent increase or decrease in connectivity in patients with MCI compared to controls during resting state. Therefore, we conclude that, the increases and decreases of functional connections observed in the MCI population in the lower alpha band may reflect the aberrant functioning until the breakdown of the system, which characterizes AD.

The increase in PLI values found in the lower alpha band in patients with MCI has been commonly considered as a compensatory mechanism. This interpretation was related to the attempt of the brain to overcome the damage caused by the disease in the networks involved in cognitive functioning (see Grady, 2012; Scheller et al., 2014 for reviews). In the case of healthy controls, this mechanism would not be needed

López et al. MEG Functional Networks in MCI

while AD patients would not compensate any more due to the severity of the disease. Nonetheless, recent studies postulated that instead of being a compensatory mechanism, it would be a pathological characteristic of MCI patients (de Haan et al., 2012; López et al., 2014a,b). During the course of the disease, there is a loss of GABAergic synapsis caused by the accumulation of β-amyloid (Aβ) plaques (Garcia-Marin et al., 2009), producing an inhibitory deficit. The loss of inhibitory interneurons in the cortex would induce increasing brain activity/connectivity in MCI patients, leading to an aberrant functioning (disinhibition) until the breakdown of the system, which is what occurs in AD.

In agreement with what has previously been described in some AD studies (Stam et al., 2002; Pijnenburg et al., 2004), we obtained lower connectivity values in the MCI group in the upper alpha band mainly concerning temporal and parietal brain areas. As far as we know, no studies have been performed describing this finding in MCI. However, considering the alpha band as one (normally from 8 to 13 Hz), some authors have revealed this decrease in connectivity in MCI patients compared to controls (Koenig et al., 2005; Garcés et al., 2014; Cuesta et al., 2015). Our results point out that networks that are usually implicated in episodic memory, olfactory function, visuospatial processing or executive functioning (previously described in the Results section) are already impaired in MCI patients. These results may indicate that in MCI the disconnection that characterizes AD would have already started, probably contributing to the cognitive deficits observed in this population. According to the increases and decreases obtained in PLI values, which have been also described in previous studies, it might be considered that during the symptomatic pre-dementia phase of AD, two mechanisms could be coexisting in MCI: disconnection and aberrant functioning.

To elucidate about the meaning of this duality of hyper and hypo connectivity we decided to evaluate the functional network organization using the network theory approach. We started with two of the most basic network parameters: the characteristic path length and the clustering coefficient. As firstly described by Watts and Strogatz (1998), these two measures together form the concept of the small-world network topology whereas the network architecture combines an efficient balance between local (short range) and global (long range) connectivity. This small-world configuration is thought to be better suited for information transfer and thus presumably for cognitive processing rather than the topology of random or regular networks (Bassett and Bullmore, 2006; Stam et al., 2007a). We did not find differences in terms of clustering and path length in our MCI cohort. In other studies, however, an increased path length and decreased clustering coefficient in MCI was found (Xiang et al., 2013; Zhang et al., 2015) and therefore MCI mimics results of AD studies (Stam et al., 2007a). MCI has been referred to as an intermediate state between healthy aging and AD in terms of their network topology (Seo et al., 2013). Our cohort did not differ in terms of the small-world parameters clustering coefficient and characteristic path length and therefore the exploration of different network measures is interesting since they may be more sensitive for the subtle changes in MCI.

Studying brain networks using measures like the clustering coefficient and the characteristic path length give useful insights within datasets of similar network sizes and link densities, but cause a comparison problem when these requirements are not met. This problem is thoroughly explained in a paper by van Wijk et al. (2010). It stresses the comparison problem between networks, not only because of the differences in network sizes (number of nodes) and degree but also due to arbitrary choices that have to be made (i.e., the threshold for the link density within a weighted network). This was the main reason for the use of the MST (Stam, 2014). Using MST, no arbitrary choices have to be made in case of unique functional connectivity values: it does not require setting a threshold and the number of nodes and links is fixed. It can be regarded as the backbone of a network (Çiftçi, 2011; Yu et al., 2015). In the present study, we found differences in two measures of centrality when comparing the MST of MCI patients and healthy controls. The MST is regarded as the backbone of a functional network since it merely involves the strongest links of the network (Stam, 2014). The MST-BC as a measure for centrality has previously shown a shifted hub location in patients with AD in high frequency bands (Engels et al., 2015). In our study, we found increased BC values in MCI patients as well as some decreases in lower and upper alpha bands. The degree, also a measure of centrality, was also found to be reduced mainly in the temporal regions. As with the functional connectivity measures, we thus found a dual pattern in the MCI population. These findings may suggest that the loss of BC/degree, mainly in temporal areas, may reflect that these areas are weakened in the brain network while frontal and parietal compensate for this malfunction. It also may reflect that some brain areas lose control within the network while others are functioning in a more aberrant way. In conclusion, although after correcting for multiple comparisons the significant differences in MST disappeared, this study showed that the classical network measures (normalized Lw and Cw) did not distinguish between MCIs and controls during resting state, but MST analysis may be a new and useful procedure to characterize and differentiate both populations. Although the reduction in centrality in temporal regions was not reported in the one study evaluating the MST BC in AD (Engels et al., 2015), this finding can be understood in the light of the disease pathology, which involves the temporal lobe. The differences between these two studies may be explained by differences in age difference (patients in our cohort were older) and therefore parietal pathology may be relatively less present (Adriaanse et al., 2012).

This study has a number of strengths and limitations. A strong point is that we used the PLI as a measure of functional connectivity since it reduces the bias due to volume conduction and/or field spread (Stam et al., 2007b). Another strong point is the use of conventional network measures (i.e., the normalized clustering coefficient and the characteristic path length), which are well described in literature, and MST parameters, that offer an arbitrary-free method for comparing networks with different properties. Our source-space analyses included 78 regions of interest according to the AAL atlas. This is a commonly used atlas, but our approach could be applied to other atlases as well. Besides these advantages, this study has several limitations as

well which should be taken into account. Our results may have been influenced by methodological choices such as the selection of artifact-free epochs by one of the authors (M. M. Engels). Epochs with signs of artifacts, drowsiness were discarded. One of the other authors (M. E. López) checked the selected epochs and therefore, we expect that the epochs we have selected for our final analyses are artifact-free. An important consideration for this approach was that we did not want to apply data cleaning approaches (e.g., Delorme et al., 2006) that could modify the connectivity structure of the data, and thereby bias subsequent functional connectivity and network analyses. Consistent with our previous work, we therefore opted to rely on thorough visual inspection for the selection of artifact free data segments.

Finally, it should be pointed that our MCIs were recruited from a clinical context. Several studies have reported that it is easier to find more cases of MCIs within a clinical population and also that the rate of conversion to AD per year is higher in a clinical setting compared to the general population (Farias et al., 2005; Jelic et al., 2006). Although NIA-AA clinical criteria is standard for all subjects (Albert et al., 2011), our findings may be more representative of the clinical than the community population.

Please note that the significant differences described in this study were not corrected for multiple testing. The FDR-corrected results did not show any significant group differences. Therefore, these results are presented as an exploratory study that can be used as a guide for regions and measures that show a trend toward significance between MCI and controls.

Our results revealed differences between MCI patients and controls. These patients did not have dementia yet, although they have an increased risk of developing it. Although these patients only have minor cognitive deficits, the functional connectivity and network differences are striking, suggesting a possible

#### REFERENCES


causative role. Therefore, measures of functional connectivity, and the network parameters derived from these inter-areal functional connections, may help to characterize the very early stages of dementia.

#### AUTHOR CONTRIBUTIONS

ML performed the MEG recordings, wrote the main manuscript, and prepared the figures; ME pre-processed the MEG data, collaborated with the MEG data analysis, and wrote the main manuscript; EvS supervised the study; RB collaborated with the statistical analysis; MD collected the sample; PS supervised the study; AH collaborated with the statistical analysis and supervised the study; CS supervised the study; FM collaborated with the experimental design and supervised the study. All authors reviewed the manuscript.

#### ACKNOWLEDGMENTS

This study was supported by two projects from the Spanish Ministry of Economy and Competitiveness (PSI2009-14415-C03- 01 and PSI2012-38375-C03-01), and a postdoctoral fellowship to ML (FJCI-2014-22730) from the Spanish Ministry of Economy and Competitiveness.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2017.00107/full#supplementary-material

impairment. J. Neurosci. 32, 4307–4318. doi: 10.1523/JNEUROSCI.5061- 11.2012



cognitive impairment and Alzheimer's disease during a working memory task. Clin. Neurophysiol. 115, 1332–1339.


synapses. Neurobiol. Aging 31, 1145–1152. doi: 10.1016/j.neurobiolaging.2008. 07.021


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 López, Engels, van Straaten, Bajo, Delgado, Scheltens, Hillebrand, Stam and Maestú. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Altered Functional Connectivity of the Basal Nucleus of Meynert in Mild Cognitive Impairment: A Resting-State fMRI Study

Hui Li 1,2† , Xiuqin Jia1,2† , Zhigang Qi 1,2 , Xiang Fan<sup>1</sup> , Tian Ma<sup>1</sup> , Hong Ni <sup>1</sup> , Chiang-shan R. Li 3,4,5 and Kuncheng Li 1,2 \*

<sup>1</sup>Department of Radiology, Xuanwu Hospital, Capital Medical University, Beijing, China, <sup>2</sup>Beijing Key Lab of MRI and Brain Informatics, Beijing, China, <sup>3</sup>Department of Psychiatry, Yale University School of Medicine, New Haven, CT, USA, <sup>4</sup>Department of Neuroscience, Yale University School of Medicine, New Haven, CT, USA, <sup>5</sup>Beijing Huilongguan Hospital, Beijing, China

Background: Cholinergic dysfunction plays an important role in mild cognitive impairment (MCI). The basal nucleus of Meynert (BNM) provides the main source of cortical cholinergic innervation. Previous studies have characterized structural changes of the cholinergic basal forebrain in individuals at risk of developing Alzheimer's disease (AD). However, whether and how functional connectivity of the BNM (BNM-FC) is altered in MCI remains unknown.

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Stephen J. Gotts, National Institute of Mental Health (NIH), USA Veena A. Nair, University of Wisconsin-Madison, USA

#### \*Correspondence:

Kuncheng Li cjr.likuncheng@vip.163.com †These authors have contributed

equally to this work.

Received: 20 November 2016 Accepted: 18 April 2017 Published: 04 May 2017

#### Citation:

Li H, Jia X, Qi Z, Fan X, Ma T, Ni H, Li CR and Li K (2017) Altered Functional Connectivity of the Basal Nucleus of Meynert in Mild Cognitive Impairment: A Resting-State fMRI Study. Front. Aging Neurosci. 9:127. doi: 10.3389/fnagi.2017.00127 Objective: The aim of this study was to identify alterations in BNM-FC in individuals with MCI as compared to healthy controls (HCs), and to examine the relationship between these alterations with neuropsychological measures in individuals with MCI.

Method: One-hundred-and-one MCI patients and 103 HCs underwent resting-state functional magnetic resonance imaging (rs-fMRI). Imaging data were processed with SPM8 and CONN software. BNM-FC was examined via correlation in low frequency fMRI signal fluctuations between the BNM and all other brain voxels. Group differences were examined with a covariance analysis with age, gender, education level, mean framewise displacement (FD) and global correlation (GCOR) as nuisance covariates. Pearson's correlation was conducted to evaluate the relationship between the BNM-FC and clinical assessments.

Result: Compared with HCs, individuals with MCI showed significantly decreased BNM-FC in the left insula extending into claustrum (insula/claustrum). Furthermore, greater decrease in BNM-FC with insula/claustrum was associated with more severe impairment in immediate recall during Auditory Verbal Learning Test (AVLT) in MCI patients.

Conclusion: MCI is associated with changes in BNM-FC to the insula/claustrum in relation to cognitive impairments. These new findings may advance research of the cholinergic bases of cognitive dysfunction during healthy aging and in individuals at risk of developing AD.

Keywords: basal nucleus of Meynert, basal forebrain, cholinergic bases, mild cognitive impairment, functional connectivity

### INTRODUCTION

Mild cognitive impairment (MCI), as a syndrome of cognitive decline without demonstrable alteration in daily activities (Gauthier et al., 2006), frequently precedes the development of Alzheimer's disease (AD; Petersen, 2011). Epidemiological studies reveal that 3% to 19% of the adults older than 65 years of age suffer from MCI (Gauthier et al., 2006). Most of the patients with MCI experience poor memory as the primary symptom (Kawas, 2003). Memory formation and maintenance is supported by a wide swath of cerebral structures and endogenous acetylcholine plays a critical role in the modulation of information acquisition, encoding, consolidation and retrieval (Blokland et al., 1992; Boccia et al., 2003, 2004, 2009; Power et al., 2003; Winters and Bussey, 2005). As the most important component of the basal forebrain cholinergic system (BFCS; Liu et al., 2015), the basal nucleus of Meynert (BNM) provides main sources of cholinergic innervation to the cerebral cortex (Gratwicke et al., 2013). In MCI, β-amyloid (Aβ) deposition, neurofibrillary tangles and trophic support reduction impair cholinergic functions of the BNM (Mesulam et al., 1986; Ruberg et al., 1990; Vogels et al., 1990). It is plausible that dysfunctional BNM-FC may contribute to cognitive decline in patients with MCI (Whitehouse et al., 1981; Mesulam et al., 1983; Nordberg, 1993; Cullen et al., 1997; Grothe et al., 2010; Ferreira-Vieira et al., 2016).

In humans the BNM is located in the BFCS with an extent of approximately 13–14 mm anterio-posteriorly and 16–18 mm medio-laterally (Gratwicke et al., 2013). The small volume and irregular border has impeded research of the BNM. Extant imaging studies have focused on volume changes in the BNM in association with MCI (Hanyu et al., 2002; Zaborszky et al., 2008; Grothe et al., 2012) with reduced volume extending into the whole BFCS in patient with AD (George et al., 2011; Grothe et al., 2013, 2014), as compared with healthy aging (Kilimann et al., 2014). For instance, volume reduction of the BNM and temporal lobe structures was associated with impairment in delayed recall in MCI patients (Grothe et al., 2010). These findings are consistent with neuronal loss in the BNM during the prodromal stage of AD and greater neuronal loss in the BNM than other cortical and subcortical structures in advanced AD (Arendt et al., 2015).

Investigators have suggested that functional alterations likely precede structural atrophy and examination of cerebral functional connectivity may be critical to understanding the etiologies of many neuropsychiatric disease (Greicius et al., 2004; Fox and Raichle, 2007; Qi et al., 2010; Liang et al., 2011). Low-frequency fluctuations of blood oxygenation level–dependent (BOLD) signal that occur during rest reflect connectivity between functionally related brain regions (Biswal et al., 1995; Fair et al., 2007; Fox and Raichle, 2007). Studies of this ''spontaneous'' activity describe the intrinsic functional architecture of the sensory, motor, cognitive systems (Fox and Raichle, 2007; Manza et al., 2015; Kann et al., 2016; Zhang et al., 2016) and may provide useful information on network functional integrity. For instance, numerous studies have demonstrated altered resting state functional connectivity (rsFC) in patients with MCI and AD (Wang et al., 2006; Sorg et al., 2007; Liang et al., 2012). Compared with healthy controls (HCs), selected areas of the default mode network, including the posterior cingulate, and the executive attention network demonstrated reduced network-related activity in individuals with MCI (Sorg et al., 2007). Patients with mild AD demonstrated disruption in hippocampal connectivity to the medial prefrontal cortex, ventral anterior cingulate cortex, posterior cingulate cortex, right superior and middle temporal gyrus (Wang et al., 2006). The network dysconnectivity appears to aggravate as the illness progresses.

On the other hand, whether or how BNM-FC is altered in MCI or AD remains unknown. On the basis of a previous study that delineated the BNM of post-mortem human brains in a standard stereotaxic space (Zaborszky et al., 2008), we recently characterized whole-brain rsFC of the BNM in a large cohort of healthy adults (Li et al., 2014). Furthermore, we showed that BNM connectivity to the visual and somatomotor cortices decreases while connectivity to subcortical structures including the midbrain, thalamus and pallidum increases with age (Li et al., 2014). These findings of age-related changes of cerebral functional connectivity of the BNM may facilitate research of the neural bases of cognitive decline in health and illness. Here, we pursued this issue by investigating whole-brain rsFC of the BNM in individuals with MCI. We hypothesized that, in comparison with age-matched control participants, BNM connectivity with cortical-subcortical regions would be disrupted in MCI patients in association with cognitive dysfunction.

### MATERIALS AND METHODS

#### Subjects and Assessments

One-hundred and one MCI patients were recruited from the Xuanwu Hospital of Capital Medical University in Beijing. MCI patients met the core clinical criteria stipulated by the National Institute on Aging and the Alzheimer's Association (Albert et al., 2011) that include: (a) complaint of a change in cognition; (b) impairment in cognitive function, especially episodic memory; (c) ability to maintain independence in daily activities; (d) not demented; and (e) Clinical Dementia Rating (CDR) score = 0.5, with a score of at least 0.5 on the memory domain (Petersen et al., 2001).

One-hundred and three age- and gender-matched HCs were recruited from the community for comparison. The criteria for the HCs was as follows: (a) no current or previous diagnosis of any neurological or psychiatric disorders; (b) no neurological deficiencies in physical examinations; (c) absence of abnormal findings on brain MRI; (d) no complaints of cognitive changes; and (e) a CDR score of 0. Additional exclusion criteria for both MCI and HCs participants included contraindications for MRI such as use of cardiac pacemakers and claustrophobia.

All participants underwent history-taking, complete physical examination and neuropsychological evaluation with Mini-Mental State Examination (MMSE), CDR, Montreal Cognitive Assessment (MoCA) and Auditory Verbal Learning Test (AVLT). MMSE has been a widely used cognitive test (Nilsson, 2007) since its publication in 1975 (Folstein et al., 1975), with excellent sensitivity and moderate specificity in the diagnosis of dementia in specialist settings (Mitchell, 2009). MoCA is a brief, stand-alone cognitive screening tool with high sensitivity and specificity for the detection of MCI and excellent test-retest reliability (Nasreddine et al., 2005). CDR is used to stage the severity of AD in longitudinal studies and clinical trials covering six cognitive categories, including memory, orientation, judgment and problem solving, community affairs, home and hobbies and personal care (Morris, 1993). AVLT provides a standardized procedure to evaluate verbal learning and memory of supra-span lists of words (Volkmar, 2013) and has been proved to be useful in the diagnosis of MCI and could predict disease progression (Zhao Q. et al., 2015). Demographic and assessment data of the participants are shown in **Table 1**.

The study was conducted under a research protocol approved by the Ethics Committee of the Xuanwu Hospital, in accordance with the Declaration of Helsinki. All participants were given a detailed explanation of the study and signed an informed consent prior to the study.

#### MRI Data Acquisition

MRI data were acquired with a 3-Tesla Trio scanner (Siemens, Erlangen, Germany). All participants were asked to hold still, with their eyes closed. Foam padding was employed to limit head motion and headphones were used to reduce scanner noise. Resting-state functional magnetic resonance imaging (rs-fMRI) images were acquired using an echo-planar imaging (EPI) sequence with a repetition time (TR)/echo time (TE)/flip angle (FA) = 2000 ms/40 ms/90, field of view (FOV) = 256 mm, 28 axial slices, slice thickness/gap = 4/1 mm, bandwidth = 2232 Hz/pixel and number of repetitions = 239. The 3D T1-weighted anatomical image was acquired with a magnetization-prepared rapid gradient echo (MPRAGE) method with the following parameters: TR/TE/inversion time (TI)/FA = 1900 ms/2.2 ms/900 ms/9, bandwidth = 199 mm,



Note: Values represent means ± standard deviation; MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rate; MoCA, Montreal Cognitive Assessment; AVLT, Auditory Verbal Learning Test; P values were derived from the Student t-test comparing the 2 groups except for "gender" where the P value was obtained using the χ 2 test.

matrix = 256 × 224, 176 sagittal slices with 1 mm thickness.

### MRI Data Preprocessing and Functional Connectivity Analysis

Rs-fMRI data were preprocessed using the statistical parametric mapping software SPM8 (Wellcome Department of Imaging Neuroscience, London, UK) and seed-to-voxel correlation analysis was carried out by the functional connectivity (CONN) toolbox v17a (Whitfield-Gabrieli and Nieto-Castanon, 2012). The first 10 functional images were discarded to reduce the fluctuation of MRI signal in the initial stage of scanning. The remaining 229 images of each individual subject were first corrected for slice timing to reduce the within-scan acquisition time differences between slices and then realigned to eliminate the influence of head motion during the experiment. All subjects included in the present study exhibited head motion less than 1.5 mm in any of the x, y, or z directions and less than 1.5◦ of any angular dimension and volume-level mean framewise displacement (FD) less than 0.30 (with a mean FD across all subjects of 0.12 ± 0.06; Power et al., 2012). Next, the realigned images were spatially normalized to the Montreal Neurological Institute (MNI) space and resampled them to 2 × 2 × 2 mm<sup>3</sup> . Subsequently, the functional images were smoothed with a 4-mm full width at half maximum (FWHM) isotropic Gaussian kernel. After preprocessing, images were then band-pass filtered to 0.008 ∼0.09 Hz to reduce the influence of noise. Further denoising steps included regression of the six motion parameters and their first-order derivatives, regression of white matter and cerebrospinal fluid (CSF) signals following the implemented CompCor strategy (Behzadi et al., 2007) and a linear detrending. In second-level covariance analysis, the covariates included age, gender, education level, mean FD (Power et al., 2015), as well as global correlation (GCOR; Saad et al., 2013). Localization of brain region was conducted with xjView<sup>1</sup> .

The seed BNM was, as defined by Li et al. (2014), based on a stereotaxic probabilistic maps of the basal forebrain, that contains the magnocellular cholinergic corticopetal projection neurons (Zaborszky et al., 2008). In the latter study, after a T1-weighted MRI scan, 10 human postmortem brains were made into histological serial sections and stained by silver. The positions and the extent of each part of the basal forebrain were microscopically delineated, 3D reconstructed and warped to the reference space of the MNI brain. Magnocellular cell groups in the subcommissural-sublenticular region of the basal forebrain were defined as the BNM (Vogels et al., 1990; de Lacalle et al., 1991). To consider the influence of volumetric difference in BNM between patients with MCI and HCs, we compared gray matter volume of BNM between MCI and HCs groups.

The correlation coefficients between the seed voxels and all other brain voxels were computed to generate correlation maps.

<sup>1</sup>http://www.alivelearn.net/xjview

For group analyses the correlation coefficients were converted to z-value using Fisher's r-to-z transformation to improve normality (Lowe et al., 1998).

#### Statistical Analysis

Clinical data and neuropsychological measures were analyzed using SPSS 19, with the Student's t-test conducted for continuous variables and the chi-squared test for dichotomous variables. The BNM-FC maps were analyzed based on a voxel-wise comparison across the whole brain. The individual z value was entered into a random effect one sample t-test to determine brain regions showing significant connectivity to the BNM within each group. Results within-group were thresholded at voxellevel p < 0.05 (FWE corrected) and cluster size >100 voxels. Two-sample t-test was performed to compare BNM-FC between MCI and HCs with age, gender, education level, mean FD, as well as GCOR as nuisance covariables. For between-group comparisons, we used 3dClustSim program in AFNI (version: December 8, 2016<sup>2</sup> ) to conduct the multiple comparison corrections within a group GM mask (obtained by a threshold of 0.2 on the mean GM probability map of all subjects) using a voxel-wise threshold of p < 0.0005 (uncorrected) and cluster size >69 voxels, which corresponded to a corrected p < 0.05

<sup>2</sup>http://afni.nimh.nih.gov/pub/dist/doc/program\_help/3dClustSim.html

(using AFNI's updated autocorrelation function estimation; Cox et al., 2017).

### Correlation Analysis

Region-of-interest (ROI) analysis was performed on the regions showing significant BNM-FC changes in MCI as compared to HCs. For each subject the mean BNM-FC across all voxels of each ROI was extracted and computed. Partial correlation analysis was then conducted to evaluate the relationship between the BNM-FC to these ROIs and raw scores of clinical assessments, controlled for age, gender, education level, mean FD, as well as GCOR. Statistical significance was set at p < 0.05, Bonferroni corrected for multiple comparisons.

### RESULTS

#### Demography and Neuropsychological Assessment

As shown in **Table 1**, no significant difference in age, gender and FD was found between the patients with MCI and HCs. Compared to HCs, MCI patients showed significantly lower education level (p < 0.05) and cognitive decline in the MMSE (p < 0.001), MoCA (p < 0.001) and AVLT (p < 0.001) score. Meanwhile, MCI patients showed

FIGURE 1 | Whole brain functional connectivity of the basal nucleus of Meynert (BNM) in healthy controls (HCs) (A) and mild cognitive impairment (MCI) (B). Numbers in the figure indicate the Z coordinate in Montreal Neurological Institute (MNI); results within-group were thresholded at voxel-wise p < 0.05 (FWE corrected) and cluster size >100 voxels. BNM, basal nucleus of Meynert; colorbar indicates t-score.

significantly higher CDR score, as compared with HCs (p < 0.001).

### Whole Brain BNM-FC in HCs and MCI Patients

Within group analysis revealed that the positive connectivity between the BNM and many brain regions both in HCs and individuals with MCI. These regions included the bilateral frontal lobe extending into the orbital/rectal cortex, medial/inferior/middle gyrus, temporal lobe extending into (para) hippocampus/amygdala, basal ganglia including the caudate/putamen/claustrum, as well as thalamus (**Table 2** and **Figure 1**).

### Altered BNM-FC in MCI Patients Compared to HCs

Compared with HCs, significantly decreased BNM-FC was detected in the left insula/claustrum in MCI patients (**Figure 2A** and **Table 2**). No brain regions showed increased BNM-FC in MCI patients when compared to HCs. In addition, no significant difference was observed in the gray matter volume of BNM between MCI and HCs (p = 0.067).

### Correlation of BNM-FC with Neuropsychological Scores

Pearson's correlation showed that BNM-FC with the left insula/claustrum was positively associated with the memory performances measured by AVTL-immediate recall (r = 0.41, p < 0.0001) in MCI patients (**Figure 2B**). On the other hand, no significant correlation was detected between BNM-FC and neuropsychological measures in HCs.

### DISCUSSION

The present study examined the BNM-FC in MCI patients in comparison to HCs in a relatively large sample of participants. The findings showed decreased BNM-FC at

TABLE 2 | Regions of basal nucleus of Meynert-functional connectivity (BNM-FC) from one-sample t test of each group and regions showing decreased BNM-FC in MCI as compared to HCs.


Note: Results within-group were thresholded at voxel-wise p < 0.05 (FWE corrected) and cluster size >100 voxels. Results between-group were corrected for multiple comparisons using an uncorrected p < 0.0005 and cluster size >69 voxels to yield a corrected p < 0.05 (using AFNI's updated autocorrelation function estimation). BG, basal ganglia; Lt, left; Rt, right.

FIGURE 2 | (A) Decreased BNM-left insula/claustrum connectivity in MCI patients; and (B) linear correlation of BNM-left insula/claustrum connectivity and AVTL-immediate recall performance in MCI patients. BNM, basal nucleus of Meynert; Ins/CL, insula/claustrum; AVLT, Auditory Verbal Learning Test; FC, functional connectivity; ∗∗∗indicates p < 0.0001.

the left insula/claustrum in MCI patients, as compared to HCs. Furthermore, decreased BNM-insula/claustrum functional connectivity was positively associated with impaired performance as measured by the AVLT- immediate recall in MCI patients.

The insula plays a critical role in cognition, emotion, sensory processing and autonomic control (Naqvi et al., 2007; Allen et al., 2008; Kurth et al., 2010). It is also involved in integrating somatosensory, autonomic and cognitive-affective information to guide behavior (Christopher et al., 2014) and switching brain network activities to support various aspects of cognitive functions (Seeley et al., 2007). Studies in monkeys showed that the density of acetylcholinesterase containing fibers (Mesulam et al., 1984) and choline acetyltransferase activities are particularly rich in the insula (Mesulam et al., 1986). Previous studies revealed that the insula was affected in individuals with MCI (Xie et al., 2012) and at risk of developing AD (Foundas et al., 1997) and insula atrophy effectively discriminated AD patients from the healthy populations (Fan et al., 2008; Insel et al., 2015). The Aβ plaques, neurofibrillary tangles and significant volume atrophy have all been reported in the insula in MCI and AD patients (Karas et al., 2004; Braak et al., 2006; He et al., 2015; Ting et al., 2015; Zhao Z. L. et al., 2015). Furthermore, wholebrain correlational analyses revealed that cognitive performance was associated with the volume (Farrow et al., 2007; Lu et al., 2016; Tillema et al., 2016), functional activity and intrinsic connectivity of the insula in MCI patients (Xie et al., 2012; Nellessen et al., 2015).

The striatum, including claustrum, putamen/pallidum and caudate (Chikama et al., 1997) are of central importance to cognitive motor functions (Christopher et al., 2015). The claustrum is an important albeit less studied part of the striatum. Previous studies reported that the claustrum was activated during episodic memory retrieval in HCs (Schwindt and Black, 2009). Recent work demonstrated that the claustrum receives input from multiple brain regions such as the parietal and the medial temporal structures (Park et al., 2012; Torgerson et al., 2015), while projecting to the insula and frontal pole (Burman et al., 2011). Via this anatomical connectivity, the claustrum may play an important role in binding multiple sources of information to facilitate memory-guided behavior. The claustrum showed neurofibrillary, amyloid pathology (Morys et al., 1996) and reduced choline acetyltransferase (Ohara et al., 1996, 1999; Gill et al., 2007) in both MCI and AD. Wang et al. (2016) reported that disrupted amygdala connectivity with the claustrum in AD patients. Thus, in line with these studies, the present work reveals the existence of altered BNM-FC at the claustrum in MCI patients.

Greater decrease in BNM-FC with insula/claustrum was associated with lower verbal episodic memory scores in MCI patients. The claustrum and the insula are functionally and structurally connected (Chikama et al., 1997). The striatum receives insular projections primarily to support motivated behavior, including reward processing and approach and avoidance learning (Chikama et al., 1997). The perisylvian division of the cholinergic fiber bundles originating from BNM traveled within the claustrum and supplied innervations to the insula (Selden et al., 1998). Previous studies revealed that the insula/claustrum is a crucial hub of the brain network integrating information from multiple brain regions through its extensive reciprocal connections to neocortical, limbic and paralimbic structures (Crick and Koch, 2005; Fernández-Miranda et al., 2008; Park et al., 2012; Torgerson et al., 2015). More recently, Seo and Choo (2016) demonstrated positive correlations between memory performance and regional cerebral glucose metabolism in bilateral claustrum and the left insula in MCI. A single photon emission computed tomography study also reported significant correlation between verbal memory test performance and brain perfusion in the left insula and claustrum in AD patients (Nobili et al., 2007). Thus, the positive correlation between verbal episodic memory scores and the BNM-FC with the left claustrum/insula suggests decline in episodic memory in MCI patients in association with the inefficient information integration and decreased functional connectivity at the insula and claustrum.

#### REFERENCES

Albert, M. S., DeKosky, S. T., Dickson, D., Dubois, B., Feldman, H. H., Fox, N. C., et al. (2011). The diagnosis of mild cognitive impairment due to Alzheimer's disease: recommendations from the national institute on aging-Alzheimer's association workgroups on diagnostic guidelines for Alzheimer's disease. Alzheimers Dement. 7, 270–279. doi: 10.1016/j.jalz.2011.03.008

#### CONCLUSION

In conclusion, MCI patients present significant changes of BNM-FC at the left insula and the claustrum in association with immediate recall during AVLT. These new findings may advance research of the cholinergic bases of cognitive dysfunction during healthy aging and in individuals at risk of developing AD.

### LIMITATIONS

Several limitations need to be considered for this study. First, those are a cross-sectional study finding and a longitudinal study is needed to understand how BNM-FC may be related to disease progression. Longitudinal follow-up of the patients would allow us to examine whether these neural phenotypes could predict onset of AD in this MCI cohort. Second, the matching of MCI and HCs groups were marginally significant. Although these factors were included as covariates in data analyses, we could not rule out their potential impact on the current results. Finally, human brains are highly varied among different demographics (e.g., gender, age and race). A recent study has demonstrated that the Chinese brain atlas improved accuracy and reduced anatomical variability during registration (Liang et al., 2015). In the current study, the data of participants were normalized to MNI152. Future studies involving Chinese populations should be considered to normalize to Chinese 2020 (a typical statistical Chinese brain template).

#### AUTHOR CONTRIBUTIONS

HL and XJ carried out data collection and analysis and wrote the manuscript. ZQ helped with data interpretation. XF, TM and HN carried out data collection. CRL contributed to conceptualization of the study and revision of the manuscript. KL contributed to conceptualization and design of the study and revised the manuscript.

#### ACKNOWLEDGMENTS

This work was supported partly by grants from the National Natural Science Foundation of China (31400958, 81471649, 81571648, 61672065 and 81370037), Clinical Medicine Development Special Funding from the Beijing Municipal Administration of Hospital (ZYLX201609) and Key Projects in the National Science and Technology Pillar Program during the Twelfth Five-year Plan Period (2012BAI10B04).

Allen, J. S., Emmorey, K., Bruss, J., and Damasio, H. (2008). Morphology of the insula in relation to hearing status and sign language experience. J. Neurosci. 28, 11900–11905. doi: 10.1523/JNEUROSCI.3141 -08.2008

Arendt, T., Brückner, M. K., Morawski, M., Jäger, C., and Gertz, H. (2015). Early neurone loss in Alzheimer's disease: cortical or subcortical? Acta Neuropathol. Commun. 3:10. doi: 10.1186/s40478-015-0187-1


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Li, Jia, Qi, Fan, Ma, Ni, Li and Li. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# tDCS Over the Motor Cortex Shows Differential Effects on Action and Object Words in Associative Word Learning in Healthy Aging

Meret Branscheidt <sup>1</sup> \* † , Julia Hoppe2† , Nils Freundlieb2,3 , Pienie Zwitserlood<sup>4</sup> and Gianpiero Liuzzi 1,2

<sup>1</sup>Department of Neurology, University Hospital Zurich, Zurich, Switzerland, <sup>2</sup>Department of Neurology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, <sup>3</sup>Brain Stimulation, Department of Psychiatry and Psychotherapy, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, <sup>4</sup>Department of Psychology, University of Münster, Münster, Germany

Healthy aging is accompanied by a continuous decline in cognitive functions. For example, the ability to learn languages decreases with age, while the neurobiological underpinnings for the decline in learning abilities are not known exactly. Transcranial direct current stimulation (tDCS), in combination with appropriate experimental paradigms, is a well-established technique to investigate the mechanisms of learning. Based on previous results in young adults, we tested the suitability of an associative learning paradigm for the acquisition of action- and object-related words in a cohort of older participants. We applied tDCS to the motor cortex (MC) and hypothesized an involvement of the MC in learning action-related words. To test this, a cohort of 18 healthy, older participants (mean age 71) engaged in a computer-assisted associative word-learning paradigm, while tDCS stimulation (anodal, cathodal, sham) was applied to the left MC. Participants' task performance was quantified in a randomized, cross-over experimental design. Participants successfully learned novel words, correctly translating 39.22% of the words after 1 h of training under sham stimulation. Task performance correlated with scores for declarative verbal learning and logical reasoning. Overall, tDCS did not influence associative word learning, but a specific influence was observed of cathodal tDCS on learning of action-related words during the NMDA-dependent stimulation period. Successful learning of a novel lexicon with associative learning in older participants can only be achieved when the learning procedure is changed in several aspects, relative to young subjects. Learning success showed large interindividual variance which was dependent on non-linguistic as well as linguistic cognitive functions. Intriguingly, cathodal tDCS influenced the acquisition of action-related words in the NMDA-dependent stimulation period. However, the effect was not specific for the associative learning principle, suggesting more neurobiological fragility of learning in healthy aging compared with young persons.

#### Keywords: associative word learning, healthy aging, transcranial direct current stimulation, motor cortex, language functions

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

José M. Delgado-García, Pablo de Olavide University, Spain Lucía Amoruso, University of Udine, Italy

> \*Correspondence: Meret Branscheidt mbransc1@jhu.edu

†These authors have contributed equally to this work.

> Received: 23 January 2017 Accepted: 24 April 2017 Published: 15 May 2017

#### Citation:

Branscheidt M, Hoppe J, Freundlieb N, Zwitserlood P and Liuzzi G (2017) tDCS Over the Motor Cortex Shows Differential Effects on Action and Object Words in Associative Word Learning in Healthy Aging. Front. Aging Neurosci. 9:137. doi: 10.3389/fnagi.2017.00137

**Abbreviations:** ANOVA, analysis of variance; ISI, interstimulus interval; MC, motor cortex; MEP, motor evoked potential; MMST, mini-mental status-test; tDCS, transcranial direct current stimulation; TMS, transcranial magnetic stimulation; VAS, visual analog scales; VLMT, verbal learning and memory test.

## INTRODUCTION

With increasing life expectancy, quality of life and social participation in older people is more and more dependent on fluid cognitive functioning. However, the ability to acquire new skills, for example to learn a new language, decreases with age (Flöel et al., 2012; Zimerman et al., 2013). The neurobiological aspects underlying language learning in healthy aging constitute a novel research area and are still not well understood. It is known that acquired knowledge and well-trained skills are commonly preserved in healthy aging, while the formation of new memory contents becomes increasingly difficult (Cohen, 1979; Light and Burke, 2009). By consequence, everyday language comprehension shows little abnormalities, whereas linguistic information processing in more challenging situations, for example learning new words, deteriorates with age (Service and Craik, 1993; Light and Burke, 2009).

Associative learning of a new lexicon has been successfully implemented in experimental settings with precise control over stimulus frequency and exposure time (Breitenstein and Knecht, 2002; Dobel et al., 2009; Liuzzi et al., 2010). While these paradigms yielded robust results for different word classes in young adults, testing the suitability and determining learning success in an older population was the scope of the present work.

The framework of embodied semantics is based on the hypothesis that motor areas activated by execution and observation of actions are also involved in processing linguistic information related to these actions (Hauk and Pulvermüller, 2004; Gallese and Lakoff, 2005; Pulvermüller, 2005). Evidence to support this theory comes from fMRI, lesion and electrophysiological studies highlighting a tight functional link between the motor and language systems (Flöel et al., 2003; Kemmerer et al., 2012). Building on this, recent studies have explored the possibility of influencing one system to alter function in the other domain. For instance, it has been shown that listening to food action related sentences results in effector specific excitability changes in the food motor area but not the hand motor area and vice versa (Buccino et al., 2005). Also, activation of motor areas (e.g., by allowing manual gestures or suppressing them) can improve certain aspects of language (Rauscher et al., 1996; Pine et al., 2007). Additionally, Liuzzi et al. (2010) could demonstrate that left motor cortex (MC) stimulation was causally involved in learning a novel action word vocabulary. Given the findings for young adults with brain stimulation, we hypothesized an influence of the left MC on the acquisition of a novel action word lexicon. In particular, we explored whether the associative learning principle was specifically altered by brain stimulation in older adults.

Transcranial direct current stimulation (tDCS) is a non-invasive electrical brain stimulation technique, which has been successfully used to improve learning in non-linguistic domains in healthy aging (Hummel et al., 2010; Flöel et al., 2012; Zimerman et al., 2013; Park et al., 2014). While the efficacy of tDCS in improving language function at various levels has been demonstrated in young healthy adults (for an overview see Miniussi et al., 2008; Cotelli et al., 2008), the effect of tDCS on language acquisition in an older population remains an open question. In young people, Liuzzi et al. (2010) demonstrated that tDCS over the left MC affected associative learning of a novel action-word lexicon. tDCS has specific NMDA-dependent and plasticity-related effects that are necessary for the coupling of actions with novel words (Liebetanz et al., 2002; Liuzzi et al., 2010). This allows for a characterization of learning word-to-semantic couplings in a neurobiologically defined way. We here investigated associative word learning of action- and object-related vocabulary applying MC-tDCS in healthy older participants to investigate whether the associative learning principle described in the young is altered in older adults.

We hypothesized that tDCS over the left MC influences associative word learning in healthy, older participants. We were especially interested in whether tDCS has a specific effect on word classes (action- vs. object-related words) and on specific response styles corresponding to Hebbian assumptions.

### MATERIALS AND METHODS

#### Participants

A total of 18 participants were enrolled in the study protocol: 12 females, mean age: 70.6 ± 5.7 years, age range 61–82 years. According to the Edinburgh inventory of handedness, 17 participants were right-handed and one person was born left-handed but retrained to be right-handed during early childhood. All participants were native German speakers and spoke 2.0 ± 1.4 foreign languages. Formal years of education ranged from 8 to 13 years (10.2 ± 1.9).

Participants were not bilingual, had no history of neurological or psychiatric diseases, especially no severe head traumas, seizures, no metal implants in the head/neck region nor pacemaker implantation and did not use neuroactive (e.g., antidepressants, anticonvulsants etc.) or recreational drugs (>6 cups of coffee/day, >50 g of alcohol/day).

To characterize cognitive profiles, participants were screened with a comprehensive battery of neuropsychological tests. Current general cognitive status was assessed with the mini-mental status-test (MMST, Folstein et al., 1975) and the verbal learning and memory test (VLMT, Helmstaedter et al., 2001), verbal fluency with the Regensburg verbal fluency test (Aschenbrenner et al., 2000), visuo-spatial memory and executive abilities with the Rey-Osterrieth complex figure test (Rey, 1959), attention span with the d2-test, working memory with digit spans, and logical reasoning using the Horn Intelligence test (Brickenkamp, 2002). Participants whose performance was more than two standard deviations above or below the age-adjusted mean were to be excluded, but all screened participants were within these boundaries.

This study was carried out in accordance with the recommendations of the local ethics committee at the University of Hamburg and the Deutsche Forschungsgesellschaft (DFG). All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the ethics committee of the University of Hamburg.

#### Stimulus Material

#### Images

An associative word-learning paradigm, previously established and extensively pretested, was used (Breitenstein and Knecht, 2002; Liuzzi et al., 2010; Freundlieb et al., 2012). Participants were presented with spoken pseudowords (e.g., kage, gafo), together with different photographs of actions or objects. For details regarding the generation and compilation of the stimulus material see Liuzzi et al. (2010); Freundlieb et al. (2012).

Actions involving either hands and arms (e.g., knocking or eating), or the whole body (e.g., running or boxing) were taken from a set of photos of everyday actions. Images were previously evaluated regarding quality and suitability for the learning paradigm; assessing naming agreement, quality of depiction, motion association, involvement of a particular body part (head/face/mouth; arm/hand; leg/foot; whole body), daily-life frequency of execution, and frequency of personal everyday performance. Two different pictures were chosen for each action, illustrated by various actors and shot from different perspectives or in different locations (for further details on these images, see Freundlieb et al., 2012).

Object images (e.g., house, tree) were evaluated for recognizability, associations with body parts or motion. Two pictures were selected for each object, depicting the same object concept in two different ways (e.g., different houses). Images were taken from different angles, with different surroundings and without visible human body parts (for details see Freundlieb et al., 2012).

For both visual stimulus sets (action/object images), pictures were converted to grayscale, centered and adjusted for potential distracting features (e.g., text or background objects).

#### Pseudowords

Thirty-four 4-letter pseudowords were taken from an attested language-learning paradigm (Freundlieb et al., 2012). The spoken stimuli (e.g., binu, gafo) complied with the phonotactics of German, had neutral emotional valence and limited associations with existing words. The novel words had a stimulus duration of 970.3 ± 127.4 ms and the same maximum volume. All stimuli were spoken by the same female voice.

#### Word Learning Paradigm

During the associative learning task, correct (to be learned) and incorrect picture/pseudoword couplings were presented, with the proportion of correct pairings increasing over time. Participants had to decide intuitively whether word and picture matched or not. Spoken pseudowords were presented over earphones, while pictures were presented on a computer display. The onset of picture presentation was 200 ms after the onset of the spoken pseudoword. Participants answered by left- (correct) or right- (incorrect) clicks on a computer mouse with their right hand. Single-trial duration was 3000 ms, and only responses obtained within this time window were taken into account for analyses. The interstimulus interval (ISI) was 2000 ms. In contrast to the visual presentation duration used in young healthy participants (Freundlieb et al., 2012), we extended their duration (3000 ms) for our cohort of older participants. Pilot experiments showed that most healthy older participants did not learn with short picture presentation times.

Participants took part in three learning sessions (see description below and **Figure 1**). During each session, they had to learn 34 pseudowords (17 action- and 17 object-related words), with two different images for each concept. Each session was divided into five blocks, separated by 2 min. breaks. Each block consisted of 136 trials, with a total of 680 trials per learning session. Over the course of the five blocks, correct couplings appeared 10 times (five times for each image of the action/object), whilst each pseudoword was also incorrectly paired with 10 different actions/objects (resulting in a correct/incorrect ratio of 10:1). We ensured that the same auditory/visual pair, as well as the same type of coupling (correct/incorrect) did not occur more than two times in a row.

Every participant completed three sessions on three separate days, with anodal, cathodal or sham tDCS and three parallel lexicon versions (lexicon 1, 2 or 3). The order of interventions and lexicon versions was pseudo-randomized and counterbalanced across sessions. Training sessions were separated by at least 2 and maximum 3 weeks. One week prior to the first tDCS training session, participants were familiarized with the learning paradigm, using a small lexicon of five words. Dependent measures were collected within each session: (1) correct responses during each training session; and (b) translation of pseudowords into German after training. No feedback was given on translation performance (for details see **Figure 1**).

#### Transcranial Direct Current Stimulation

Transcranial magnetic stimulation (TMS) was used to determine the hand region in the left MC in each participant immediately prior to tDCS application. The so-called ''hotspot'' for the hand region was identified as the position where the highest MEP amplitudes could be consistently evoked in the right first dorsal interosseous muscle (Chen et al., 1998). TMS was delivered by a Magstim 200 stimulator connected to a figure-8 shaped coil (7 cm in diameter, Magstim Co.).

tDCS was administered via two sponge electrodes (Eldith; soaked in 0.9% saline solution) connected to a DC-stimulator (Eldith; serial no. 0006). Either the anode or cathode was placed as stimulating electrode over the left hemispheric ''hotspot'' of the hand motor area (surface area 25 cm<sup>2</sup> ). The reference electrode was positioned over the contralateral supraorbital region (surface area 35 cm<sup>2</sup> ). Stimulation started immediately at the beginning of the learning paradigm, and intensity was increased in a ramp-like fashion over 10 s until 1 mA for verum and sham stimulation. In case of anodal or cathodal tDCS stimulation, the current intensity remained constant for 20 min. In contrast, the

FIGURE 1 | Word learning paradigm. (A) Participants were shown a photographic illustration depicting an action/object paired with a spoken pseudoword. They had to decide intuitively if the presented coupling was a correct or incorrect match. Two-hundred milliseconds after onset of the sound, the picture appeared. Responses had to be given within a time window of 3000 ms. The interstimulus interval (ISI) was 2000 ms between two trials. (B) One learning session consisted of 680 single trials, subdivided into five blocks with a 2 min break. Each pseudoword was shown 10 times with the "correct" action/object (depicted by two different images) as well as 10 times with different "incorrect" images. (C) Study design: prior to the first learning session participants were screened with a neuropsychological test battery (NP) and were introduced to the learning paradigm with a small lexical pre-test (PT) consisting of five words. Each participant completed three learning sessions on three different days with three different lexicon sets. During the task, participants received anodal, cathodal or sham tDCS in a randomized order across learning sessions (1 mA for 20 min; double blind application). At the start of each session the stimulation side over the left motor cortex (MC) was localized using transcranial magnetic stimulation (TMS). tDCS stimulation and the learning paradigm started simultaneously, the latter exceeding the stimulation for approximately 20 min. After each session, patients were asked to translate the acquired pseudowords into notions of their native language (T). Abbreviations: NP, neuropsychological testing; PT, pre-test; T, translation test.

sham stimulation had a duration period of 30 s during which the intensity was ramped down to zero during the following 10 s. This ''fade-in/fade-out'' approach is standard best-practise procedure for sham stimulation in tDCS and mimics the cutaneous sensations experienced for verum stimulation (Gandiga et al., 2006). A person not involved in the experiment and data analysis entered the stimulation parameters. Experimenters and participants were blinded for stimulation type.

The neurophysiological effects of tDCS to the MC have been shown to outlast the stimulation period, with effects depending on current intensity and stimulation duration (Nitsche and Paulus, 2000; Nitsche et al., 2003). On this basis, a stimulation period of 20 min was regarded as sufficient for the learning paradigm lasting 40 min.

We evaluated participants' appraisal of attention, unpleasant sensations (i.e., discomfort/pain) and fatigue with questionnaires using visual analog scales (VAS) as control parameters.

### Data Analysis

Two outcome measures were selected to determine successful learning: (1) the percentage of correct responses in learning blocks over time; and (2) the translation rate for pseudowords into native language after each training session. In a subanalysis, we also calculated the learning success for different response types in each block over time.

To investigate the influence of tDCS, we performed a three-factorial rmANOVA on the dependent variable percentage of correct decisions, with the within-subject factors ''stimulationanodal/cathodal/sham'', ''word classobject/action'' and ''blocks1–5''. We further investigated how stimulation might change response behavior for the word classes differently by two separate three-factorial rmANOVAs for objects and actions, with the within-subject factors ''stimulationanodal/cathodal/sham'', ''blocks1–5'' and ''response typehit/miss/corr\_reject/false\_alarm''. For translation rates, differences between stimulation sessions were analyzed with a two-factorial rmANOVA, with the within-subject factors ''stimulationanodal/cathodal/sham'' and ''word classobject/action''.

Performance scores of the neuropsychological test battery were probed for linear association with the translation rate and percentage of correct decisions in block five, using Pearson correlation coefficients. A one-way ANOVA with the withinsubject factor ''stimulationanodal/cathodal/sham'' was calculated for the VAS outcomes for attention/fatigue/discomfort/pain.

Before application of parametric tests, normal distribution of the dependent variables was tested using Shapiro-Wilk tests and quantile-quantile plots). All ANOVA results were Greenhouse-Geisser corrected if assumptions of sphericity were violated. Paired two-tailed t-tests were used for the analysis of the predicted effects of stimulation on word class. Results were considered significant at p < 0.05 and Cohen's d is reported as a measure of effect size. All data are expressed as mean ± standard error unless stated otherwise. Statistical analyses were done using SPSS 22.0<sup>r</sup> and GraphPad Prism<sup>r</sup> Software.

#### RESULTS

#### Learning Success

#### Percentage of Correct Responses

First, we report the effect of tDCS on the performance of associative learning: The number of correct responses increased significantly over the five blocks (blocks1–5, F(4,68) = 30.44, p = 0.000). Averaged over all stimulation conditions, participants started at chance level of 49.33 ± 0.94% and reached 69.32 ± 3.69% in block 5 (see **Figure 2**). However, anodal and cathodal tDCS stimulation over the left MC did not result in significantly different percentages of correct responses (stimulationanodal/cathodal/sham F(2,34) = 1.51, p = 0.235; stimulationanodal/cathodal/sham <sup>∗</sup>blocks1–5: F(8,136) = 1.16, p = 0.337), overall mean accuracy in block 5: anodal: 69.61 ± 3.8%; cathodal: 66.54 ± 4.1%; sham: 71.81 ± 4.1%). For all stimulation conditions together, correct responding during learning was better for object- than for action-related words (word classobject/action, F(1,17) = 9.59, p = 0.007. There was no significant interaction between blocks and word class (blocks1–5 <sup>∗</sup> word classobject/action F(4,68) = 1.09, p = 0.368).

Next, we evaluated whether tDCS over the left MC affects word classes differently, as shown in a younger population for cathodal tDCS and action words (Liuzzi et al., 2010). Regarding correct responses, neither the interaction of word classobject/action ∗ stimulationanodal/cathodal/sham nor the three-way interaction

of blocks1–5 ∗ stimulationanodal/cathodal/sham <sup>∗</sup>word classobject/action reached significance in the rmANOVA (F(2,34) = 2.33, p = 0.112, respectively F(8,136) = 0.73, p = 0.661; see **Figure 2**). However, as predicted from earlier results in young participants, there was a significant effect of cathodal stimulation on reduction of correct responses of action-related, but not of object-related words (cathodal vs. sham stimulation for actions: 64.1 ± 4.2% vs. 71.2 ± 4.1%, t(17) = 2.21 p = 0.02, d = 0.41, cathodal vs. sham stimulation for objects: 69.0 ± 4.3% vs. 72.9 ± 4.2%, t(17) = 1.53, p = 0.144, d = 0.22).

We also looked at the distribution of correct and incorrect response types (see ''Materials and Methods'' Section) for the different word classes/stimulation types. There was no interaction of response typehit/miss/corr\_reject/false\_alarm ∗ stimulationanodal/cathodal/sham nor a three-way interaction of blocks1–5 ∗ response typehit/miss/corr\_reject/false\_alarm ∗ stimulationanodal/cathodal/sham in actions (F(6,102) = 479.74, p = 0.234, respectively F(24,408) = 56.10, p = 0.221) or objects (F(6,102) = 0.75, p = 0.507, respectively F(24,408) = 1.12, p = 0.358). Note that cathodal stimulation seemed to lead to lower correct rejection and higher false alarm rates for actions compared to objects.

#### Translation

We additionally tested whether participants were able to transfer the learnt association between pseudo-word and images to their native language. Across all stimulation conditions, participants were able to translate a mean of 34.32 ± 5.47% pseudowords into German. tDCS stimulation did not significantly influence overall translation rates (stimulationanodal/cathodal/sham: F(2,34) = 2.34, p = 0.107, overall mean translation: anodal: 32.68 ± 5.1%; cathodal: 31.06 ± 5.5%; sham: 39.22 ± 5.6%). Even though the percentage of correct responses differed for word class, there was no significant difference in overall translation rates (word classobject/action, F(1,17) = 1.15, p = 0.299; objects: 51.76 ± 8.1%; actions: 48.24 ± 8.9% of all correctly


TABLE 1 | Data for the two outcome measures, as a function of stimulation and word class.

translated object and action words, respectively). The two-way interaction stimulationanodal/cathodal/sham <sup>∗</sup>word classobject/action was also not significant (F(2,34) = 2.80, p = 0.075, see also **Table 1**). Our findings suggest that aged-populations might have a limited ability for transfer of associated learning content.

#### Neuropsychological Evaluation

Finally, we tried to identify cognitive and stimulationrelated factors that could potentially influence a participant's performance in the associative learning task. Learning success and translation rates correlated with performance scores in the VLMT (verbal learning and memory ability; r = 0.597, p = 0.009 and r = 0.604, p = 0.008, respectively) and logical reasoning (r = 0.610, p = 0.007 and r = 0.495, p = 0.037, respectively, see **Figure 3**), while the results for visuo-spatial memory and executive abilities, attention and working memory did not correlate with language learning performance.

Stimulation-related factors did not appear to have a noticeable effect on participant performance. Verum and sham tDCS evoked similar minimal painful sensations (F(2,34) = 1.972, p = 0.155; mean VAS 1-10 anodal: 1.56 ± 0.2; cathodal: 2.39 ± 0.5; sham: 1.72 ± 0.3). Ratings for discomfort, fatigue and attention were comparable in all groups (discomfort, F(2,34) = 0.081; mean VAS 1-10 anodal: 4.06 ± 0.6; cathodal: 3.89 ± 0.5; sham: 3.78 ± 0.5; p = 0.923; fatigue, F(2,34) = 1.125, p = 0.336; mean VAS 1-10 anodal: 5.28 ± 0.7; cathodal: 4.61 ± 0.6; sham: 5.22 ± 0.7; attention, F(2,34) = 0.412, p = 0.665; mean VAS 1-10 anodal: 3.78 ± 0.5; cathodal: 4.33 ± 0.6; sham: 3.89 ± 0.6).

#### DISCUSSION

The aim of this study was to investigate associative word learning and the effect of tDCS over the MC during acquisition of a novel vocabulary in healthy older participants. Using a previously established associative language-learning paradigm (Liuzzi et al., 2010; Freundlieb et al., 2012), older participants acquired a novel vocabulary for everyday objects or actions. No specific effect of anodal or cathodal tDCS application on overall learning success was found, although as predicted there was an impairment of action word acquisition after cathodal tDCS stimulation.

The learning paradigm used here has been applied with some variations in previous studies (Breitenstein and Knecht, 2002; Dobel et al., 2009). Freundlieb et al. (2012) tested the efficiency of the paradigm for learning of novel object and action words in young adults in a single session design, without tDCS stimulation. Paradigm B of their study was similar to our approach except for their shorter visual presentation duration of 1400 ms. Young adults were able to translate 66% of novel words and showed reliable acquisition of the new lexicon, with an overall accuracy rate of 86% in the last block. Performance during the training did not differ between action-related and objectrelated words, but the translation test revealed significantly better learning of object-related (79%) than of action-related words (53%).

In our study, older participants did not perform as well as the young subjects from Freundlieb et al. (2012) both with respect to translation (66% in young adults, 39% in our older cohort in the sham condition) and correct decisions during training (86% correct decisions in the last block in young adults, 71% in our older population in the sham condition). Additionally, pretesting showed that learning in the older participants was only possible when picture presentation time was more than doubled compared with the duration for young participants. We cannot draw any firm conclusions about which specific process is most compromised in the older participants. Aging may impair both non-linguistic (visuospatial and auditory processing, working memory) as well as linguistic functioning (phonological processes, binding phonological information with semantic context). It can only be concluded that associative word learning is relatively preserved, but is considerably slowed down by aging.

Liuzzi et al. (2010) evaluated the functional role of the MC in action-word acquisition in young healthy participants. Using the same associative learning task that coupled spoken words with action-related pictures only, participants learned one set of 76 pseudowords during four consecutive training sessions. Either anodal, cathodal or sham tDCS stimulation over the MC was administered for 20 min prior to each session. Differences between their and our associative-learning paradigm concerned the number of novel words (76 overall vs. 34 per session), the use of action pictures or action and object pictures, visual presentation time (1400 vs. 3000 ms), the ratio between correct and incorrect couplings (4:2 vs. 10:1), a longer training duration (with training sessions on four consecutive days, with learning success measured on day four vs. three different sets and learning assessments each day) and the design (anodal, cathodal, sham between vs. within subjects). Liuzzi et al. (2010) could demonstrate that cathodal tDCS to the left MC led to a significantly reduced translation of novel action words compared to sham stimulation, whereas no effect was seen for anodal stimulation. Control experiments with object-related pseudowords or stimulation over prefrontal areas indicated a

responses in block 5 and the performance scores of (A) VLMT (verbal learning and memory test) respectively (B) logical reasoning. Right side: correlations between number of correctly translated words and performance scores of (C) VLMT or (D) logical reasoning.

semantic and topographic specificity of the observed effect after cathodal tDCS over the MC (Liuzzi et al., 2010).

In older participants, we also observed that action-word acquisition after cathodal tDCS was significantly lower compared to the sham group. However, the effect was weaker compared with our previous study in young participants. This might be due to the single-session crossover design vs. repeated training over 4 days. Furthermore, while the study design of Liuzzi et al. (2010) allowed for consolidation effects overnight, the older participants were tested on the same day. It has been shown that the time for memory consolidation can have a crucial influence on novel word acquisition in an associative learning paradigm (Geukes et al., 2015). The effect of cathodal tDCS could only be shown in the NMDA-dependent stimulation period, supporting the idea that the MC is functionally connected during learning of novel action words. The pattern of results makes other effects like shifts of membrane polarization during the initial 20 min of stimulation less likely, given that inadvertent general effects on word learning could also not be shown in this study. As in young participants, we could not enhance learning with anodal tDCS over the MC.

In our participant group, verbal-learning ability showed good correlation with successful lexical acquisition. This finding is in line with previous findings in young healthy participants that showed a positive correlation of verbal-memory abilities with associative object word learning (Breitenstein and Knecht, 2002). More intriguing was the finding that logical reasoning and associative word learning were correlated. This is consistent with findings in aphasic patients, suggesting that preservation of executive functions may promote therapeutic outcome (Fillingham et al., 2005, 2006; Nicholas et al., 2005; Fridriksson et al., 2006). Different fMRI studies have shown that besides some task-specific regions decreasing in activity, especially prefrontal activation increases in older adults (Park and Reuter-Lorenz, 2009). The correlation of executive cognitive functions such as logical reasoning with lexical acquisition could hence suggest a compensatory strategy to maintain behavioral performance.

To summarize, associative learning imposes minimal demands on conscious effort compared with declarative vocabulary learning. This makes computer-based associative learning paradigms a promising tool for language learning in healthy aging. The ability for semantic transfer seems to be compromised in older people. Thus, training needs to be more intense in frequency and presentation times in order to build stable word representations.

Non-invasive brain stimulation techniques such as tDCS can be used to probe the interaction of specific brain areas with cognitive performance, to gain a better understanding of age-related changes of learning. Beyond this, the thorough knowledge of age-dependent cognitive skills in healthy older people might help to find predictive factors for language recovery in aphasic patients and to improve speech and language therapy.

#### AUTHOR CONTRIBUTIONS

GL: study conception and design. MB, JH, GL: acquisition, analysis and interpretation of data. MB, JH, GL: drafting

#### REFERENCES


the manuscript. GL, PZ, NF: critical revision. MB, JH, NF, PZ, GL: final approval of the version to be published and agreement to be accountable for all aspects of the work.

#### FUNDING

The present study was supported by the Deutsche Forschungsgemeinschaft (DFG LI 1892/1-1 to GL).

#### ACKNOWLEDGMENTS

We thank Annette Baumgärtner, Friedhelm Hummel and Christian Gerloff for their valuable comments on the present study and manuscript.


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Branscheidt, Hoppe, Freundlieb, Zwitserlood and Liuzzi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Lower Activation in Frontal Cortex and Posterior Cingulate Cortex Observed during Sex Determination Test in Early-Stage Dementia of the Alzheimer Type

Ravi Rajmohan<sup>1</sup> \*, Ronald C. Anderson<sup>2</sup> , Dan Fang<sup>3</sup> , Austin G. Meyer <sup>4</sup> , Pavis Laengvejkal <sup>5</sup> , Parunyou Julayanont <sup>5</sup> , Greg Hannabas <sup>6</sup> , Kitten Linton<sup>7</sup> , John Culberson<sup>7</sup> , Hafiz Khan<sup>6</sup> , John De Toledo<sup>5</sup> , P. Hemachandra Reddy 1,8,9,10 and Michael W. O'Boyle<sup>3</sup>

<sup>1</sup>Department of Pharmacology and Neuroscience, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>2</sup>Department of Electrical and Computer Engineering, Texas Tech University, Lubbock, TX, United States, <sup>3</sup>Department of Human Development and Family Studies, Texas Tech University, Lubbock, TX, United States, <sup>4</sup>School of Medicine, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>5</sup>Department of Neurology, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>6</sup>Department of Public Health, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>7</sup>Department of Family Medicine, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>8</sup>Garrison Institute on Aging, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>9</sup>Department of Cell Biology and Biochemistry, Texas Tech University Health Sciences Center, Lubbock, TX, United States, <sup>10</sup>Department of Speech, Language and Hearing Sciences, Texas Tech University Health Sciences Center, Lubbock, TX, United States

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Bahar Güntekin, Istanbul Medipol University, Turkey Vasileios Papaliagkas, Aristotle University of Thessaloniki, Greece

> \*Correspondence: Ravi Rajmohan ravi.rajmohan@ttuhsc.edu

Received: 17 December 2016 Accepted: 05 May 2017 Published: 22 May 2017

#### Citation:

Rajmohan R, Anderson RC, Fang D, Meyer AG, Laengvejkal P, Julayanont P, Hannabas G, Linton K, Culberson J, Khan H, De Toledo J, Reddy PH and O'Boyle MW (2017) Lower Activation in Frontal Cortex and Posterior Cingulate Cortex Observed during Sex Determination Test in Early-Stage Dementia of the Alzheimer Type. Front. Aging Neurosci. 9:156. doi: 10.3389/fnagi.2017.00156 Face-labeling refers to the ability to classify faces into social categories. This plays a critical role in human interaction as it serves to define concepts of socially acceptable interpersonal behavior. The purpose of the current study was to characterize, what, if any, impairments in face-labeling are detectable in participants with early-stage clinically diagnosed dementia of the Alzheimer type (CDDAT) through the use of the sex determination test (SDT). In the current study, four (1 female, 3 males) CDDAT and nine (4 females, 5 males) age-matched neurotypicals (NT) completed the SDT using chimeric faces while undergoing BOLD fMRI. It was expected that CDDAT participants would have poor verbal fluency, which would correspond to poor performance on the SDT. This could be explained by decreased activation and connectivity patterns within the fusiform face area (FFA) and anterior cingulate cortex (ACC). DTI was also performed to test the association of pathological deterioration of connectivity in the uncinate fasciculus (UF) and verbally-mediated performance. CDDAT showed lower verbal fluency test (VFT) performance, but VFT was not significantly correlated to SDT and no significant difference was seen between CDDAT and NT for SDT performance as half of the CDDAT performed substantially worse than NT while the other half performed similarly. BOLD fMRI of SDT displayed differences in the left superior frontal gyrus and posterior cingulate cortex (PCC), but not the FFA or ACC. Furthermore, although DTI showed deterioration of the right inferior and superior longitudinal fasciculi, as well as the PCC, it did not demonstrate significant deterioration of UF tracts. Taken together, early-stage CDDAT may represent a common emerging point for the loss of face labeling ability.

Keywords: Alzheimer, face-processing, neuroimaging, chimeric faces, brain networks, neurodegeneration

## INTRODUCTION

Alzheimer disease (AD) is a debilitating disorder marked by a progressive decline in cognitive functions in which the affected individual becomes less capable of understanding the world around them often leading to fear, isolation and depression. Part of this isolation stems from a loss of the ability to classify faces into social categories; a process known as facelabeling. This plays a critical role in human interaction as it serves to define concepts of socially acceptable interpersonal behavior.

In this study, we attempted to determine how verbal impairments may relate to face-labeling. The current study attempted to characterize, what, if any, impairments in face-labeling are detectable in participants with earlystage clinically diagnosed dementia of the Alzheimer type (CDDAT) through the use of the sex determination test (SDT). Psychological research on categorical grouping of faces previously hypothesized that this was likely a verballymediated process (Chance and Goldstein, 1976; Perdue et al., 1990; Otten and Wentura, 1999). More recent work, however, has shifted emphasis away from verbal-labels in favor of visuospatial constructs such as ''face space'' upon which facial characteristics are imposed and the composite of which are then assigned to a category (Valentine, 1991). This view has since gained ground through findings from computer-based geometric modeling (Valentine, 2001) and neuroimaging (Oruç et al., 2011; Contreras et al., 2013). These observations suggest that social categorization is likely a two-step process: one in which the visuospatial components of a face must be decoded and grouped with similar faces, and then a verbal-label is generated to describe to which group the face belongs, thereby making the verbal component significantly more removed than previously thought.

Existing clinical work, however, has established that verbal impairments occur early within the course of AD (Chobor and Brown, 1990; Locascio et al., 1995; Henry et al., 2004) as do impairments in gender discrimination (Rombouts et al., 2005). These impairments are congruent with the pathological progression of the disease (Braak et al., 2006; Jack et al., 2008) and we therefore sought to identify what functional and connectivity-based alterations may underlie this pathology.

To this end, the ability to identify the sex and race of faces was specifically localized to activity within the fusiform face area (FFA) in neurotypicals (NT) by Contreras et al. (2013). Potential damage to this region is of particular interest in the context of AD, given the aforementioned pathology within the temporal lobes (Braak et al., 2006) and consistently observed verbal impairments (Henry et al., 2004).

We interpret these findings to suggest that during earlystage AD, FFA activity is still intact, but the decreased temporal derivative in the anterior cingulate cortex (ACC) observed by Rombouts et al. (2005) hints at a developing inability to successfully discriminate between genders. It is important to note, however, that neither of these studies used chimeric face tests (CFTs) and the gender discrimination task (GDT) is slightly, but fundamentally, different from our SDT.

Therefore, we hypothesized that the increased demand on the FFA caused by presenting CFTs, which are substantially harder to process than natural faces, would lead to an observable difference in FFA activity and task performance between our CDDATs and NTs if one were readily existent. In order to further stress this difference, the SDT asked participants to determine if the two faces were of the same sex, and not which sex they were (as the GDT did). By doing so, we placed greater emphasis on making an executive decision related to a specific verballabel (i.e., greater emphasis is placed on a participant's ability to recognize if two faces are either ''men, women, or one of each'') with less emphasis on their ability to generate verbally-mediated labels (e.g., ''men, women, or both''). This, in turn, would likely flush out differences in ACC activity (Bush et al., 2000), similar to Rombouts et al. (2005), thereby giving us the opportunity to assess the relative importance of FFA and ACC activity within a single task.

Lastly, verbal skill tests have been used to define the progression of cognitive decline in AD (Verma and Howard, 2012). Pineda-Pardo et al. (2014) correlated fractional anisotropy (FA) values of several known speech areas to performance on the verbal fluency test (VFT) in patients with mild cognitive impairment of the amnestic type (aMCI), a common precursor to AD, showing that there may be structural as well as functional alterations underlying this deterioration. Therefore, establishing what association may exist between face-processing and verbal skills, as well as their underlying circuits of activity and connectivity, is of great importance in order to assess their potential for assisting in the early detection of this disease.

fMRI studies on aMCI patients routinely focused on memory and showed increased activation in individuals with the mildest forms of impairment while those who showed decreased activity were more likely to be in the most severe end of the spectrum (Dickerson et al., 2005; Celone et al., 2006; Johnson et al., 2006; Trivedi et al., 2008). This led Risacher and Saykin (2013) to conclude that the initial increases observed in the earlystages of impairment represent attempted compensation while the decreased activity seen upon greater impairment is due to deterioration of the system after it is no longer capable of compensating.

This pattern of early compensation followed by decreased activity is repeated in AD patients. Pariente et al. (2005) noted increased activation during episodic memory encoding and recall in the posterior cingulate cortex (PCC), precuneus, parietal lobe and frontal lobe in early-stage AD patients with a mean Mini Mental State Exam (MMSE) score of 25 ± 1.8. Grady et al. (2003), Dickerson et al. (2005), Celone et al. (2006) and Zhou et al. (2008) on the other hand, observed decreased or even absent activation relative to NTs in the same regions in earlystage AD patients with mean MMSE scores of 21.1 ± 3.1, 22 ± 5, 21.3 ± 2.7, respectively, suggesting that the differences in observations are congruent with previous findings in aMCI patients with regard to progression of symptoms. In particular, the default mode network (DMN; which includes the PCC, medial parietal lobe and medial frontal lobe) shows decreased activity at rest, decreased connectivity and reduced deactivation upon task initiation in AD patients (Grady et al., 2003; Greicius et al., 2004; Buckner et al., 2005; Celone et al., 2006).

It is important to note that this significant difference in neuroimaging findings occurred in the same clinically specified stage (i.e., early-stage AD) as defined by MMSE scores (between 26–21, National Collaborating Centre for Mental Health (UK), 2007) demonstrating the importance of integrating neuroimaging results with clinical observations. For a comprehensive review of AD discoveries through neuroimaging, see Risacher and Saykin (2013).

Of equal importance is the understanding of how white matter (WM) tracts are affected by the disease. Previous DTI studies have shown reduced FA in the parietal and temporal WM, the corpus callosum and the posterior cingulum fibers (Medina and Gaviria, 2008). WM alterations appear to parallel the previously discussed gray matter (GM) changes in that cortical abnormalities are greater in posterior brain regions relative to anterior regions at the early-stages of AD (Arnold et al., 1991; Braak and Braak, 1995). When the disease progresses, the neurofibrillary pathology advances from limbic to frontal structures, into higher-order association cortices, and finally affects primary sensorimotor areas, which correlates with the clinical manifestations of AD. DTI-based tractography studies (Fellgiebel et al., 2005) and whole-brain DTI studies (Medina et al., 2006; Rose et al., 2006; Zhang et al., 2007) have consistently shown that fibers located deep in the posterior WM (e.g., the superior longitudinal fasciculus and the posterior cingulum bundle) are affected in patients with AD and aMCI. Bartzokis et al. (2003, 2004) and Bartzokis (2004) have proposed that this may occur because as brain development takes place, later myelinated regions (cortical association areas) have fewer oligodendrocytes supporting greater number of axons compared with earlier myelinated regions. The oligodendrocytes in the cortical association areas therefore have higher metabolic demands in order to maintain their widely distributed axons, making them more susceptible to pathological processes. The DTI findings of decreased WM integrity in later myelinated regions at the onset of AD support this ''reversed demyelination'' construct (Medina and Gaviria, 2008).

Following the abundance of evidence of WM compromise in AD, a pilot study using this approach found that both aMCI and AD groups showed significant reductions in FA within the temporal lobes (Huang et al., 2007) and regional estimations of WM integrity been linked to MMSE scores (Folstein et al., 1975), with lower scores being associated with decreased WM integrity in cerebral posterior regions of AD subjects (Rose et al., 2000; Bozzali et al., 2002; Yoshiura et al., 2002; Duan et al., 2006).

Research on aMCI populations using neuropsychological tests of declarative memory extend this trend as they have demonstrated significant correlations between declining performance and decreases in posterior WM FA, particularly in the posterior cingulum bundles (Fellgiebel et al., 2005, 2008; Rose et al., 2006). Harking back to the proposal of Bartzokis et al. (2004), disruptions in transcortical connectivity may serve as early contributors to the pathophysiology of dementia as the observed WM deteriorations were embedded beneath cortical GM that is often affected early within the disease course. For a more in-depth review of DTI findings in AD, see Medina and Gaviria (2008).

Although memory deficits are a hallmark characteristic of AD, there may be little to gain from testing face-processing related to familiarity that isn't already known (Sperling et al., 2003; Golby et al., 2005; Winchester, 2009; Donix et al., 2013). On the other hand, fundamentals of face-processing (e.g., categorical labeling), as well as its relation to verbal skill sets, are areas that have received little attention (Job, 2012) despite their crucial role in daily life and therefore warrant further investigation.

The use of neuropsychological tests in combination with neuroimaging techniques stresses the importance of attempting to integrate pathological observations with clinical symptoms. In doing so, we are able to reinforce findings from either end of the spectrum as well as more efficiently develop our understanding of the disease and what measures may be taken to help those afflicted. In order to investigate a cognitive operation as complex as face labeling, it will be necessary to use appropriate tests that can isolate it as a specific subdivision of face-processing. This is particularly important when performing an fMRI investigation, as some studies suggest that processing and recognition occur automatically (Vuilleumier and Schwartz, 2001) while others say this can occur within 100 ms of being presented with a face (Batty and Taylor, 2003). If such factors are not properly accounted for, it would be very difficult to remove these confounds from the data as the canonical hemodynamic response curve suggests that blood flow in response to neuronal activity peaks between 4 s–6 s after the presented stimuli, making it far too slow to tease out processes that are nearly instantaneous by comparison (Poldrack et al., 2011).

To this end, a modified version of a specific neuropsychological test will be of particular value when investigating face labeling in AD. The CFT was originally developed by Levy et al. (1983) ''to index functional cerebral asymmetry for processing facial characteristics''. Through neuroimaging and lesion studies the FFA has been identified as being highly involved in many face-processing tasks (Kanwisher et al., 1997). In 2008 an fMRI study by Yovel et al. (2008) on face-processing concluded that FFA asymmetry is a highly stable individual characteristic that underlies the well-established left-visual-field superiority for face recognition (Yovel et al., 2008), making it an ideal region of interest when investigating the effects of AD on face-processing.

CDDAT participants were expected to show poor performance on the SDT, which could be explained by decreased activation and connectivity patterns within the FFA (Contreras et al., 2013) and ACC (Rombouts et al., 2005). This was expected to correlate to poor performance on the VFT. This would have established that sex determination is affected in early-stage AD (as was previously demonstrated by Rombouts et al., 2005, but not using Chimeric Faces). Said decrease in performance would likely be due to altered brain activity/connectivity patterns within the FFA, which may correlate sex determination and verbal fluency performance, thereby strengthening the notion that sex determination is a verbally-mediated face labeling process.

### MATERIALS AND METHODS

### Participant Identification and Selection

The University Medical Center Departments of Neurology, Family medicine, and Geriatrics saw 171 patients for complaints of ''memory problems'' or ''dementia'' by during a 6-month period (November 2015–April 2016). Hundred of these patients were determined to have AD or its precursor, aMCI, in accordance with the guidelines outlined by McKhann et al. (2011) for ''dementia due to AD'' or ''mild cognitive impairment due to AD'' (Albert et al., 2011) by physician assessment. In accordance with National Institute for health Care and Excellence guidelines (National Collaborating Centre for Mental Health (UK), 2007), those in the possible early-stage AD category were selected based on an MMSE score of 26–21. Twenty of these patients were determined to fit our inclusion/exclusion criteria; four of whom agreed to participate. All CDDAT participants had an existing medical MRI scan interpreted by a radiologist within the last 5 years prior to the study. These individuals were grouped into the CDDAT category. Detailed inclusion/exclusion criteria for the CDDAT group are listed below:

Inclusion Criteria


Exclusion Criteria


Ten age-matched cognitively normal participants (5 males, 5 females), determined by physician assessment, were recruited from a nearby senior living center and were grouped into the NT category. NTs were required to have an MMSE score >27 and meet all inclusion and exclusion criteria stated above, except for that concerning the existence of Alzheimer-related pathology. One male participant was retroactively removed after receiving a diagnosis of normal pressure hydrocephalus.

All control participants are also patients followed by our physicians, not merely elderly individuals from the community who have been screened using an MMSE. Although our Human Protections protocol prevents us from accessing or further commenting on medical information on these patients other than their MMSE scores, our physicians selected patients for both the control and experimental group in accordance with the accepted international guidelines outlined by Albert et al. (2011) for the diagnosis of mild cognitive impairment. An example of the fact that the control individuals were appropriately screened for possible neurological deficits is evidenced by the retroactive removal of one of the NTs after a suspicion by clinical evaluation led to a clinical MRI scan which demonstrated normal pressure hydrocephalus. The physicians promptly informed us of the finding and the participant was removed from the study.

Participants were informed that their participation was voluntary and they may withdraw from the study at any time and that their refusal to participate would have no impact on their level of care. This study was approved by the Texas Tech University Human Protections Internal Review Board. The subject was given a copy of the informed consent form and each section of the informed consent form was verbally reviewed. Subjects were given time to ask any questions they may have. If the subject verbally agreed to participation, the subject was asked to briefly summarize their understanding of the research project and what they were agreeing to do. Any minor misconceptions were corrected and the subject was again allowed to verbally decide. After this process, the subject was allowed to sign the consent form and received a copy of the signed consent form to keep. Written informed Consent was obtained in the presence of the primary care giver.

#### Sample Size

Henry et al. (2004) calculated an effect size of r = 0.73 for a test of semantic fluency in patients with Dementia of the Alzheimer Type (DAT). That effect size r is equivalent to Cohen's d = 2.14. To achieve 80% power, we needed a minimum sample size of n = 5. Given our heterogeneous group sizes of four disease and nine control, we were able to achieve 90% power.

#### Participant Demographics

We recruited four right-handed Caucasian patients (3 males, 1 female) with a diagnosis of aMCI or early-stage AD between the ages of 73–93 (median age 83.5 ± 8.4) with MMSE scores between 26–23 (median of 24.5 ± 1.3). Two of the males (MMSE scores of 26 and 25) had an existing diagnosis of ''mild cognitive impairment due to AD'' (Albert et al., 2011). The remaining male and female had an existing diagnosis of ''dementia due to AD'' (MMSE scores of 24 and 23, respectively; McKhann et al., 2011). Nine (8 right-handed and 1 non-right-handdominant) cognitively normal Caucasian participants (4 males, 5 females; all with MMSE scores of 30) between the ages of

#### TABLE 1 | Participant demographics.


CDDAT, clinically diagnosed dementia of the Alzheimer type; NT, neurotypical; F, female; M, male; R, right-handed; nR, non-right-handed; CA, Caucasian.

79–91 (median age 80 ± 3.8) also participated. Handedness was determined using the Edinburgh Handedness Inventory-revised (EHI-r; Williams, 1986). There was no significant difference in median age (p = 0.582), but there was an observable difference in median MMSE score for the two groups (p = 0.011).

While there may be some concern given our lack of equal sex-distribution (1 female and 3 males) within the CDDAT group, no differences in performance were discernable based on sex within either our NT or CDDAT groups across any cognitive test (VFT or SDT) for either accuracy or reaction time (RT; data not shown) nor for within-group sex-based contrast mapping for fMRI or DTI (data not shown). A summary of participant demographics is given in **Table 1**.

#### Administration of Verbal Fluency Test

Participants were asked to list all the names of animals they could recite in 1 min. The same procedure was then repeated for names of fruit and finally for colors. Audio recordings were taken of the verbal tasks. Scores were expressed as the sum of the three tasks.

### Imaging Methodology and Analyses

#### Stimulus Presentation and Participant Response

After completing the VFT and Rey Osterrieth Complex Figure B (ROCFB), all participants were briefly trained on the SDT before entering the scanner to ensure task comprehension by allowing them to practice on a single example trial of each condition type (Rajmohan, 2016). Chimeric faces (Mattingley et al., 1993) were presented in an event-related design using Eprime 2.0. Presentation order was counterbalanced across participants using a Latin square design. The question ''Are the two people of the same sex?'' was displayed on the screen inside the scanner for 4 s. A set of 40 trials were randomized and displayed for 4 s each with a jittered inter-stimulus interval (ISI) randomized for five time points between 800 ms and 1200 ms, by 100 ms apiece. Participants responded using a fiber optic controller held in the right-hand where button 1 was pressed by the index finger and button 2 was pressed by the middle finger. Participants pressed one of two buttons to indicate Yes (button 1) or No (button 2). An example of the possible stimulus presentations is given in **Figure 1**.

#### Scanning Parameters

All images were acquired with a 3T Siemens MR system (Skyra, Germany) at the Texas Tech Neuroimaging Institute. The T1 anatomic scan parameters were: TR: 1900 ms; TE: 2.49 ms; FOV: 240; Flip angle: 9, Voxel size = 0.9 × 0.9 × 0.9 mm; slice number: 192. The fMRI parameters were: TR: 2500 ms; TE: 20.0 ms; FOV: 231; Flip angle: 75, Voxel size = 2.5 × 2.5 × 3.0 mm; slice number: 41. The DTI parameters were: TR: 5000 ms; TE: 95 ms; FOV: 220; B/W: 1562, Voxel size = 1.7 × 1.7 × 4.0 mm; slice number: 32.

#### Image Preprocessing

#### **fMRI**

Image preprocessing steps included removing non-brain structures by Brain Extraction Tool (BET), motion correction by using Motion Correction for FSL Linear Registration Tool (MCFLIRT), temporal high-pass filtering with a cutoff period of 24 s, spatial smoothing with a 5 mm Gaussian full width, half maximum (FWHM) algorithm, and co-registering of the functional images to the high resolution T1 structure images in their native space using boundary border registration(BBR) and FSL Linear Registration Tool (FLIRT; Jenkinson and Smith, 2001; Jenkinson et al., 2002) at 12 degrees of freedom to the Korean Normal Elderly (KNE96; Lee et al., 2016) standard brain space.

#### **DTI**

FSL Diffusion Toolbox 3.0 (FDT) from FMRIB Software Library (FSL 5.0.5) was used to complete the construction of and preprocessing for all anatomical brain networks for all subjects. It processed DICOM/NIfTI files into diffusion metrics (e.g., FA) that were ready for statistical analysis at the voxel-level after performing corrections for image alignment and artifact clean-up (Top-Up) and local field distortions (eddy current correction; Smith et al., 2004).

#### Image Processing **fMRI**

fMRI data processing was carried out using FMRI Expert Analysis Tool (FEAT) Version 6.00, part of FMRIB's Software Library<sup>1</sup> (FSL). The time series for the behavioral events were analyzed for the following conditions.

For ''presentation of stimuli scenarios'' (i.e., when the participant was shown two male faces, two female faces, or one of each)—Brain activity from the first 2 s following the presentation of a stimulus was recorded for each trial. In the event a participant responded within <2 s, the interval between the presentation of the stimulus and 250 ms before the response was used for the recording interval. This was done to avoid brain activity artifacts related to the button press. Instances where a participant made no response before the presentation of the next stimulus were discarded.

For ''participant response'' scenarios (when the participant chose ''yes'' or ''no'' in response to the question: ''Are the two people of the same sex?'')—Brain activity from the first 2 s following the press of a button was recorded for each trial. In the

<sup>1</sup>www.fmrib.ox.ac.uk/fsl

FIGURE 1 | Examples of the sex determination test (SDT). SDT examples in response to the question: "ARE THE TWO PEOPLE OF THE SAME SEX?" (left) "yes"; two males (center) "yes"; two females (right) "no"; male (top) and female (bottom).

event a participant responded within <2 s before the presentation of the next stimulus, that interval was used for the recording time. Instances where a participant made no response before the presentation of the next stimulus were discarded.

A summary of fMRI contrasts is given in **Table 2**. A total of five subject-level contrast maps were created for all subjects


A total of five subject-level contrast maps were created for all subjects. SDT, sex determination test.

using threshold free cluster enhancement (TFCE) of z > 2.3 and a cluster corrected significance threshold of p < 0.05. These modeled time series were convolved with the double gamma hemodynamic response function (dg-HRF), which was modeled from a combination of Gaussian functions. Group-level contrast maps were created using FMRIB's Local Analysis of Mixed Effects (FLAME1). The thresholds for group level activation maps were created using TFCE of z > 1.5 and a clustercorrected significance threshold of p < 0.05. The exact regions of brain activity were determined using the KNE96 coordinate space and Harvard-Oxford cortical and subcortical structural atlases.

#### **DTI**

Voxel-based group differences were calculated for the FA images using Tract-Based Spatial Statistics (TBSS; Smith et al., 2006; Smith and Nichols, 2009; Cheon et al., 2011). TBSS linearly registered individual FA images in native space and then to the FA template via the flirt command of FSL. The resultant warping transformations were then used to convert images of diffusion (i.e., FA) to MNI152 (Montreal Neuroimaging Institute) space with a spatial resolution of 1 × 1 × 1 mm. For statistical inference, including correction for multiple comparisons, permutation testing was used (Nichols and Holmes, 2002; Cheon et al., 2011) as implemented by RANDOMISE of the FSL software package. Five-hundred permutations were performed for significant group differences at a threshold of 0.2; corresponding to p < 0.05, corrected for multiple comparisons using TFCE (Smith and Nichols, 2009; Cheon et al., 2011). WM tracts were identified using the Johns-Hopkins University ICBM-DTI-81 WM labels and probabilities of tract accuracy were assessed using the Johns-Hopkins University WM Tractography atlas.

#### Behavioral Methodology and Analyses

All statistical calculations for behavioral analyses were performed using Rstudio Desktop (version 0.99.896, Rstudio, Inc., Boston, MA, USA).

#### Comparison of Test Performance Amongst Study Groups

Mann-Whitney U-tests with false discovery rate (FDR) corrections of α = 0.05 for multiple comparisons to reach significance at p < 0.05 were used to evaluate differences in test performance between the study groups. A one-sample t-test with mu = 30 and two-sample U-test were used to calculate the group difference in MMSE score given that the variance for the NT group was 0. Both tests revealed a significant result. Additionally, the zero variance for the NT group on the MMSE is within reason for cognitively normal elderly individuals, as persons without cognitive deficits may be expected to receive a perfect score (30/30) as was seen here (Folstein et al., 1975).

#### Correlation of CFTs to Cognitive Test Performance

Pearson correlation coefficients, with FDR corrections of α = 0.05 for multiple comparisons to reach significance at p < 0.05, were used to correlate SDT performance to VFT performance.

#### Assessing Tests as Classifiers

Fischer Exact test, with FDR corrections of α = 0.05 for multiple comparisons to reach significance at p < 0.05, were used to determine strength of each test as a classifier between CDDAT and NT.

## RESULTS

### Behavioral Data

Consistent with previous studies, the median VFT score was significantly different between NTs and CDDATs (54 ± 8 vs. 29 ± 9, respectively; p = 0.023, Cohen's d = 3.01, r-equivalent effect size = 0.833). VFT was a successful classifier at an optimal score range of 31–35 (p = 0.014). Additionally, while SDT performance was not significantly different between median NT and CDDAT scores (97.4 ± 5.0 vs. 81.0 ± 14.8%; p = 0.069, Cohen's d = 2.05, r-equivalent effect size = 0.715), this was due to half of the CDDAT participants performing substantially worse than NT (70.0 ± 0.3%), while the other half performed very similar (94.8 ± 3.5%). As such, SDT was not a successful classifier at any score (optimal score range 69–85). CDDAT RT, however, was not significantly different from NTs for SDT (1757 ms vs. 2007 ± 457 ms, not shown), suggesting intact comprehension of task instructions despite their varied performance. We found no connection between verbal ability and face-labeling (VFT ∼ SDT; R<sup>2</sup> = 0.218, p = 0.161). Behavioral results are summarized in **Figure 2** and **Table 3**.

### Imaging Data

#### fMRI Results

Imaging results are summarized in **Figure 3** and **Table 3**. There were no significant differences in areas of activation between the NT and CDDAT groups upon stimulus presentation regardless of the sex of the faces, indicating no fundamental difference in perception of stimuli. NTs, however, were shown to have greater activity in the left superior frontal gyrus (l-SFG) and left posterior cingulate cortex (l-PCC) when responding ''yes'' to the question ''Are the two people of the same sex?'' CDDAT did not show higher values in any areas for any contrasts.

#### DTI Results

NTs showed higher FA values than CDDATS in the right inferior longitudinal fasciculus (r-ILF), right posterior thalamic radiations (r-PTR) and the bilateral PCC (b-PCC) and superior longitudinal fasciculi (b-SLF). For the SLF, differences between NTs and CDDATs were greater in the left than the right hemisphere. No tracts showed higher FA values for the CDDAT group than NT group. FA measures did not demonstrate significant differences throughout the uncinate fasciculi (UF). However, CDDATs were noted to have less frontotemporal connectivity than NTs via lower FA values throughout the r-ILF and r-SLF as well as the b-PCC, but not ACC. Imaging results are summarized in **Figure 3**.

### DISCUSSION

### Significance of Behavioral Findings

In this study, we chose to use the VFT given its previous use by Luzzi et al. (2007) and Pineda-Pardo et al. (2014) and its validation by Robert et al. (2003), which gave us confidence in its suitability as a measure of verbal ability. Although the VFT was a successful classifier between CDDAT and NTs, we were unable to establish a direct correlation between VFT and SDT performance. Furthermore, SDT performance was not significantly lower in CDDATs compared to NTs. Given the dichotomous performance of our CDDAT group on the SDT, however, along with their significantly lower performance on the VFT, we reason that, while the SDT

CDDAT, Clinically diagnosed dementia of the Alzheimer type. Thickened line is VFT classifier score at 31–35; p = 0.014.

may not directly correlate with verbal proficiency, it may still represent an accurate assessment of face-labeling ability. The discrepancy between VFT and SDT performance can therefore be explained as the verbally-mediated component of the SDT (i.e., observing two faces and determining whether they are a ''man'' or ''woman'' and then determining if they are ''same'' or ''different'') being less demanding of verbal ability than the VFT, which requires spontaneous word generation within the confines of an abstract category. Therefore, it is reasonable to expect impairments in verbally-mediated processes to manifest sooner and more drastically through the VFT than the SDT. Support for this argument may be inferred from our behavioral results, which show that, while SDT score varies greatly between the highest and lowest (standard deviation of 14.8%) performance by CDDAT, there is less change in VFT (standard deviation of 9%).

In spite of this, the SDT still has clinical merit for individuals with suspected dementia as it mirrors a more realistic point of assessment (i.e., can patients determine the sex of a person they are talking to?) in relation to social interaction and activities of daily living than the word-finding nature of the VFT. Lastly, the split in CDDAT performance gives us confidence that this stage of the disease does, in fact, represent the initial stages of face-labeling deterioration given that Luzzi et al. (2007) observed significant impairment by the moderate-stage in AD patients.

#### TABLE 3 | Summary of Verbal Fluency Test (VFT) and Sex Determination Test (SDT) results.


NT, Neurotypical; CDDAT, Clinically Diagnosed Dementia of the Alzheimer Type; VFT scores are reported as total number of words.

#### Significance of fMRI Findings

Our fMRI findings further support this notion, in that, they did not show an observable difference in FFA or temporal lobe activity between CDDAT and NTs, even when specifically contrasting the subgroup of low-performing CDDAT individuals (data not shown). While it is possible that the lack of difference in findings within the FFA and temporal lobes for the subgroup contrast map may be attributable to a lack of power due to small sample size (n = 2; design efficiency of 0.447), their absence from the total group contrast map, in the presence of lower activity within the l-PCC and l-SFG, demonstrates that impairments in verbally-mediated label generation or fundamental aspects of face-processing were at the very least less pronounced than differences in attention and self-referential decision making processes. This is consistent with common associations of both the l-PCC and l-SFG that has been established by previous work (Maddock et al., 2003; Goldberg et al., 2006; Leech and Sharp, 2014).

Additionally, while decreases in PCC activity are commonly associated with AD (Leech and Sharp, 2014), the lack of this observation within the CFT and EVDT for our CDDAT participants (Rajmohan, 2016) highlights its exclusiveness to the SDT, and is likely a reflection of the intended demanding nature of the task as previously discussed. In spite of this, no difference was observed in either the ACC or the FFA. This leads us to conclude that although their verbal skills are declining, early-stage CDDATs are still capable of a similar level of verbal-labeling required for this task and have fundamentally

FIGURE 3 | (A) Significant fMRI activations for SDT when participants responded "yes" in left-Superior Frontal Gyrus and left-Posterior cingulate cortex (PCC; threshold free cluster enhancement, threshold free cluster enhancement, TFCE of z > 1.5 and a cluster-corrected significance threshold of p < 0.05). (B) Pertinent DTI results for the SDT. Significant differences were observed for the superior longitudinal fasciculi (number 1) and inferior longitudinal fasciculi (number 2) for the left (top left) and right (top right) hemispheres and for the PCC (blue crescent outline), but not the anterior cingulate cortex (ACC; red crescent outline) for the left (bottom left) and right (bottom right) hemispheres. For Figures: areas where NT brain activity > CDDAT brain activity are shown in red and CDDAT brain activity > NT brain activity are shown in blue. For fMRI figures: areas where NT brain activity > CDDAT brain activity are shown in red. For DTI figures: areas where NT > CDDAT for fractional anisotropy (FA) measures are shown in red. Green lines represent the FA skeleton. NT, Neurotypicals; CDDAT, Clinically Diagnosed Dementia of the Alzheimer Type. (C) Significant fMRI activations for SDT when participants responded "yes" are listed for the NT > CDDAT contrast (TFCE of z > 1.5 and a cluster-corrected significance threshold of p < 0.05). No significant fMRI activations for SDT were seen for the CDDAT > NT contrast.

intact face-processing abilities. Instead, impairments in executive decision-making likely explain the discrepancies in scores seen here, which is consistent with the findings of Rombouts et al. (2005).

### Significance of DTI Findings

The lack of difference in UF tract integrity, as assessed by FA, contradicts our initial hypothesis, but strengthens the notion that the connections most crucial to face labeling are largely intact. However, the lower FA values in the r-ILF, r-SLF and PCC observed in the CDDAT group suggest deterioration of right-sided frontotemporal connectivity and posterior cingulate fiber bundles, respectively. This deterioration in connectivity would likely explain the simultaneously observed decrease in activity in the l-PCC and l-SFG during CDDAT response as a loss in connectivity reduces system efficiency, which in turn manifests as decreased task-related signal activity. Support for this hypothesis also comes from our simultaneous investigations of CFT and EVDT performance in these participants, which showed that behavioral impairments correlated to deterioration of WM tracts, but no observable differences in brain activity (Rajmohan, 2016). These findings give insight into how the deterioration of WM tracts may predate decreases in brain activity in AD. It is therefore of great interest to see how decline in verbal fluency and face-labeling may occur in tandem with the deterioration of WM tracts and decreasing brain activity as the disease progresses.

### Limitations

Since the etiology of AD is unknown, it is possible that it is the culmination of many factors that may vary in both course and presentation. Therefore, the findings noted in this study population may not extend to all affected individuals due to issues of hereditary or environment. We also acknowledge that aMCI has an insidious course and often goes undetected for years in the absence of rigorous neuropsychiatric testing. It has been noted that in 6 years' time about 70% of aMCI patients develop AD (Mauri et al., 2012). Therefore, while we cannot rule out the possibility that participants in our control group may inevitably progress to aMCI, they did not meet the accepted international diagnostic criteria for this diagnosis at the time of the study.

Other potential cofounders could be differences in the demographics and comorbidities in the two groups. Due to restrictions related to the approved Human Protections Board protocol, we do not have any additional medical or demographic information (e.g., marital status, list of medications or medical comorbidities that are not of a neurologic or psychiatric nature) on our participants than what is currently described. Although we found statistically significant differences in both connectivity and activation patterns between our CDDAT and NT participants, it is important to acknowledge that this is a pilot study that is limited by its small sample size. As such, it is imperative to demonstrate its reproducibility on a larger scale.

While it may seem odd that the areas of WM deterioration do not directly overlap with those of task-related decrease in activity for all noted instances, this may be because the observable areas of decreased activity are those most crucial for performing the task at hand (in this case, left-hemispheric areas, given the indirect verbal component of this task) and are therefore the first to be detectable in a compromised system. We cannot rule out the possibility that differences in performance are based on dysfunctions in processing facial characteristics that predate verbal labeling based solely on the lack of differences in activity observed through the ''stimuluspresentation'' contrast map. Therefore, early-stage CDDATs may perceive the sex of faces in a different manner than NTs. A task that focused more specifically on the overlapping regions of interest may more directly demonstrate a correlation between tract integrity and decreased activity, but such a task would likely be of less clinical significance than the SDT.

Additionally, while performance on the VFT was consistent with previous reports (Henry et al., 2004), our SDT findings differed from those of Rombouts et al. (2005). This is attributable to several factors. First, our CDDAT population consisted of a milder stage of impairment (median MMSE score of 24.5 ± 1.3 vs. 22.5 ± 2.2). Second, the split in CDDAT performance previously mentioned makes our subgroups (n = 2) too small to achieve significant power in spite of the >20% difference in performance, even though this was over twice the effect size (9%) observed by Rombouts et al. (2005) given that their study had six times as many participants. Finally, the SDT is a similar, yet fundamentally different, task from the GDT used by Rombouts et al. (2005) in that the SDT places greater emphasis on decision-making related to the grouping of chimeric faces based on sex, whereas the GDT more directly asks a participant to assign a face to its appropriate sex. All of these factors combine to allow us to suggest that these findings are, in fact, in congruence with those of Rombouts et al. (2005), but merely represent a different circumstance.

Finally, the inability to find a significant difference between our NT and CDDAT groups for the SDT does not necessarily mean that impairments are not present at this time. Instead, it indicates that such a deficit, if it exists, is less severe than that of verbal fluency, which currently represents the most consistent and pronounced clinical manifestation of this disease (Henry et al., 2004) and was observed within this study.

### CONCLUSION

Given that declines in verbal fluency are already present by this stage, one way to reduce confusion is to use simple, concrete words. We therefore suggest the use of binary categorization (e.g., ''yes'' or ''no'') strategies in place of word labels when dealing with early-stage AD patients. Performance by our CDDAT participants on the SDT suggests that they are able to perform ''same/different'' classifications more readily than verbally-mediated tasks that emphasize word generation. Additionally, given that our fMRI data suggest greater strain upon verbally-mediated decision making, avoid excessive open-ended questioning in favor of directed questions with discrete responses. By following these simple strategies, we may ease the frustration experienced by both patients and their care-givers.

Although we highlighted several concerns related to how the findings of this investigation may fit into the larger literature, these seemingly limiting differences in many ways provide us with a great deal of new information. By noting the split in CDDAT performance in this study within the context of the performance for MCI (which were shown to perform as well as NTs) and AD patients (who were impaired on the task) from Rombouts et al. (2005), we can be quite confident that we have identified the initial stages of face-labeling impairment. From this, we can infer that: (1) verbal fluency decline predates loss of face-labeling ability; (2) differences in brain activity patterns, along with a lack of difference in UF tracts within the earlystages of face-labeling impairment suggest a more crucial role for executive function than verbal generation when performing the SDT; and (3) destruction of WM tracts may predate GM changes, as observed by the deterioration of frontotemporal WM connections; as such, greater emphasis should be placed on the anatomical and cellular pathology of WM in the study of AD pathogenesis as was suggested by Brun and Englund (1986) nearly 30 years ago.

#### REFERENCES


#### AUTHOR CONTRIBUTIONS

RR created the experimental design, ran the study, analyzed the data, interpreted the results and wrote the manuscript. RCA and DF consulted on neuroimaging design, analysis and interpretation. AGM and HK consulted on statistical analysis and interpretation. PL, PJ, GH, KL, JC and JT consulted on experimental design, patient recruitment and selection, and running of the study. PHR consulted on experimental design and pathological interpretation. MWB consulted on experimental design and cognitive psychological interpretation.

### ACKNOWLEDGMENTS

The funding for this research was provided by the I. Wylie Briscoe College of Human Sciences Endowment for Alzheimer Research. PHR is supported by National Institutes of Health (NIH) grants AG042178, AG047812 and the Garrison Family Foundation. This study was completed in agreement with the Alzheimer Disease Neuroimaging Initiative Image and Data Archive (ADNI LONI IDA) data sharing policy. Raw and preprocessed images are available at http://adni.loni.usc.edu/data-samples/access-data/. Portions of this article are taken from the dissertation of Rajmohan (2016) in compliance with Frontiers in Aging Neuroscience's rules and regulations.


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer VP and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.

Copyright © 2017 Rajmohan, Anderson, Fang, Meyer, Laengvejkal, Julayanont, Hannabas, Linton, Culberson, Khan, De Toledo, Reddy and O'Boyle. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Melatonin Supplementation, a Strategy to Prevent Neurological Diseases through Maintaining Integrity of Blood Brain Barrier in Old People

#### Wen-Cao Liu<sup>1</sup> , Xiaona Wang2,3 , Xinyu Zhang2,3 , Xi Chen<sup>4</sup> \* and Xinchun Jin2,3 \*

<sup>1</sup>Department of Emergency, Shanxi Provincial People's Hospital, Taiyuan, China, <sup>2</sup>Jiangsu Key Laboratory of Translational Research and Therapy for Neuro-Psycho-Diseases and Institute of Neuroscience, Department of Neurology, the Second Affiliated Hospital of Soochow University, Suzhou, China, <sup>3</sup>School of Pharmacy, Key Laboratory of Molecular Pharmacology and Drug Evaluation (Yantai University), Ministry of Education, Yantai University, Yantai, China, <sup>4</sup>Department of Core Facility, the People's Hospital of Baoan Shenzhen, Shenzhen, China

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Panteleimon Giannakopoulos, Université de Genève, Switzerland Wei Wang, Stowers Institute for Medical Research, United States

#### \*Correspondence:

Xi Chen beating\_u5@hotmail.com Xinchun Jin xinchunjin@gmail.com

Received: 08 March 2017 Accepted: 10 May 2017 Published: 24 May 2017

#### Citation:

Liu W-C, Wang X, Zhang X, Chen X and Jin X (2017) Melatonin Supplementation, a Strategy to Prevent Neurological Diseases Through Maintaining Integrity of Blood Brain Barrier in Old People. Front. Aging Neurosci. 9:165. doi: 10.3389/fnagi.2017.00165 Blood brain barrier (BBB) plays a crucial role in maintaining homeostasis of microenvironment that is essential to neural function of the central nervous system (CNS). When facing various extrinsic or intrinsic stimuli, BBB is damaged which is an early event in pathogenesis of a variety of neurological diseases in old patients including acute and chronic cerebral ischemia, Alzheimer's disease and etc. Treatments that could maintain the integrity of BBB may prevent neurological diseases following various stimuli. Old people often face a common stress of sepsis, during which lipopolysaccharide (LPS) is released into circulation and the integrity of BBB is damaged. Of note, there is a significant decrease of melatonin level in old people and animal. Melatonin has been shown to preserves BBB integrity and permeability via a variety of pathways: inhibition of matrix metalloproteinase-9 (MMP-9), inhibition of NADPH oxidase-2, and impact on silent information regulator 1 (SIRT1) and nucleotide-binding oligomerization domainlike receptor family pyrin domain-containing 3 (NLRP3) inflammasome. More important, a recent study showed that melatonin supplementation alleviates LPS-induced BBB damage in old mice through activating AMP-activated protein kinase (AMPK) and inhibiting gp91phox, suggesting that melatonin supplementation may help prevent neurological diseases through maintaining the integrity of BBB in old people.

#### Keywords: melatonin, blood brain barrier, neurological diseases, old people, lipopolysaccharide

**Abbreviations:** AMPK, AMP-activated protein kinase; BBB, blood brain barrier; CNS, central nervous system; iNOS, inducible nitric oxide synthase; LPS, lipopolysaccharide; MMP-9, matrix metalloproteinase-9; nNOS, neuronal nitric oxide synthase; NADPH, nicotinamide adenine dinucleotide phosphate; NF-κB, nuclear factor kappa-light-chain-enhancer of activated B cells; NLRP3, nucleotide-binding oligomerization domain-like receptor family pyrin domain-containing 3; ROS, reactive oxygen species; SIRT1, silent information regulator 1; TBI, traumatic brain injury; TJPs, tight junction proteins; TLR4, toll like receptor 4.

## INTRODUCTION

## The Blood Brain Barrier Damage and Neurological Diseases

The blood brain barrier (BBB) is a regulated interface between the peripheral circulation and the central nervous system (CNS; Jin et al., 2014). BBB, which is composed of cerebral microvascular endothelial cells, neurons, astrocytes, pericytes and the extracellular matrix, plays a key role in maintaining homeostasis of microenvironment that is essential to neural function of the CNS (Hawkins and Davis, 2005). When facing various extrinsic or intrinsic stimuli (Weiss et al., 2009), BBB is damaged and BBB dysfunction is an early event in pathogenesis of a variety of neurological diseases in old patients including vascular cognitive impairment, amyotrophic lateral sclerosis, Alzheimer's disease, neuropathic pain, brain trauma, acute and chronic cerebral ischemia, multiple sclerosis, and brain infections (Rosenberg, 2012; please see **Figure 1**). Treatments that could maintain the integrity of BBB will have important roles in preventing stimuli-produced neurological diseases.

The tight junction proteins (TJPs), composed of occludin, claudin and zo-1, are key components of the BBB (Hawkins and Davis, 2005). It seals the interendothelial cleft forming a continuous blood vessel, leads to high endothelial electrical resistance, and allows low paracellular permeability of watersoluble substances from the blood into brain parenchyma (Stamatovic et al., 2008). Free radicals of oxygen and nitrogen and the proteases, matrix metalloproteinases (MMPs) and cyclooxgyenases, play key roles in the early and delayed BBB disruption as the neuroinflammatory response progresses (Liu and Rosenberg, 2005). During an injury, free radicals and proteases attacked the cell membranes and degraded the TJPs between endothelial cells and the integrity of BBB is damaged

(Jin et al., 2013; Liu et al., 2016; Wang et al., 2016). It is worth of note, death of endothelial cells of microvessels is also a major contributor to the disruption of BBB integrity (Simard et al., 2007). Therefore, protective effect on intergrity of BBB should consider both death of endothelial cells of microvessels and degradation of TJPS.

## BBB Disruption Induced by Lipopolysaccharide (LPS)

Lipopolysaccharide (LPS) could produce neuroinflammation (Shi, 2015), promoting the generation of reactive oxygen species (ROS) in cerebral microvascular endothelial cells and BBB disruption (Seok et al., 2013). Worth of note, LPS has been shown to increase BBB permeability in vitro (Nonaka et al., 2004) and compromise BBB integrity in young (Ruiz-Valdepeñas et al., 2011; Zhou T. et al., 2014) and old mice (Wang et al., 2017). More interesting, LPS has been shown to induce BBB dysfunction via nicotinamide adenine dinucleotide phosphate (NADPH) oxidase-derived ROS (Liu et al., 2012; Zhao et al., 2014). NADPH oxidases, a major source of ROS generation in the brain, critically contributes to BBB disruption under various neurological disorders (Kahles et al., 2007). Of note, gp91phox is the catalytic subunit of NADPH oxidase and BBB disruption is significantly reduced in gp91phox knockout mice compared to wild-type mice after stroke (Kahles et al., 2007) and reduction of gp91phox expression has shown protective effect against ischemia-induced brain injury and BBB damage (Liu et al., 2008, 2011). More importantly, Wang et al. (2017) showed that LPS increased gp91phox expression in both endothelial cells and in old mice, suggesting that gp91phox up-regulation may be an important mechanism responsible for LPS-induced BBB permeability increase in old mice.

## Relationship between Melatonin and Aging

Melatonin, which is produced mainly in the pineal gland, retina and the gastrointestinal tract, plays important roles in many physiological and biochemical functions (Bubenik and Konturek, 2011), such as acting as an anti-inflammatory and immunoregulating molecule as well as a circadian rhythm regulator (Manchester et al., 2015). Melatonin is a potent free radical scavenger, lack of melatonin may result in decreased antioxidant function in the old people which have significant influence not only on aging per se but also on the incidence or severity of age-related diseases (Karasek, 2004). In addition, oxygen radical detoxification processes was significantly decreased during aging and there was a obvious downregualtion in pineal biosynthetic activity in aging hamster (Bubenik and Konturek, 2011). More interesting, melatonin levels in serum and brain decline as a result of aging (Bubenik and Konturek, 2011; Hill et al., 2013). In addition, melatonin has been reported to regulate aging and neurodegeneration through energy metabolism, epigenetics, autophagy and circadian rhythm pathways (Jenwitheesuk et al., 2014).

## Beneficial Role of Melatonin in Sepsis

Sepsis is a systemic inflammatory response to infection that causes severe neurological complications (Zhao et al., 2015) and it is a common stress that old people often face (Martin et al., 2006), in which LPS is released into circulation (Shukla et al., 2014).

Melatonin has been shown to restore the mitochondrial production of ATP in septic mice (López et al., 2006a), block the septic response by disrupting connection of the nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) with nucleotide-binding oligomerization domain-like receptor family pyrin domain-containing 3 (NLRP3) in mice (El Frargy et al., 2015) and improve survival in a zymosan A-induced rat model of sepsis/shock (Reynolds et al., 2003). In addition, melatonin has been shown to protect organs against sepisinuduced injury. For example, melatonin improved cardiac mitochondria and survival rate in rat septic heart injury (Zhang et al., 2013) through inhibition of inducible nitric oxide synthase (iNOS) and preservation of neuronal nitric oxide synthase (nNOS; Ortiz et al., 2014) and attenuated sepsis-induced cardiac dysfunction via a PI3K/Akt-dependent mechanism (An et al., 2016). Furthermore, melatonin protected liver bioenergetics from sepsis-induced damage (Basile et al., 2004), modified cellular stress in the liver of septic mice by reducing ROS and increasing the unfolded protein response (Kleber et al., 2014), protected against sepsis-induced functional and biochemical changes in rat ileum and urinary bladder (Paskaloglu et al., 2004), improved colonic anastomotic healing in a rat experimental sepsis model (Ersoy et al., 2016) and counteracted inducible mitochondrial nitric oxide synthase-dependent mitochondrial dysfunction in skeletal muscle (Escames et al., 2006) and diaphragm (López et al., 2006b) in septic mice.

### Melatonin's Effect on LPS-Induced Injury

Melatonin has been shown to ameliorate LPS-induced brain injury in neonatal rats (Wong et al., 2014), alleviate LPS-induced placental cellular stress response in mice (Wang et al., 2011) as well as LPS-induced hepatic SREBP-1c activation and lipid accumulation in mice (Chen et al., 2011). Of note, melatonin shown protective effect against BBB damage induced by various stimuli, including transient focal cerebral ischemia in mice (Chen et al., 2006), excitotoxic injury in neonatal rats (Moretti et al., 2015) and methamphetamine-induced inflammation (Jumnongprakhon et al., 2016). Therefore, decreased levels of melatonin in old mice may contribute to the BBB disruption when facing various extrinsic or intrinsic stimuli because melatonin has demonstrated its protective effects against LPS-induced injury to the heart (Lu et al., 2015), brain (Carloni et al., 2016), lung (Lee et al., 2009) and liver (Wang et al., 2007) by scavenging a variety of free radicals (Manchester et al., 2015). Interestingly, chronic melatonin treatment has also shown reduction of age-dependent inflammatory process in senescenceaccelerated mice (Rodríguez et al., 2007). In a recent study, Wang et al. (2017) showed that 1 week melatonin treatment significantly alleviated LPS-induced BBB damage accompanied by reduction of occludin and claudin-5 degradation, suggesting that melatonin supplementation is important in decreasing sepsis and neuroinflammation-induced TJPs degradation as well as BBB damage.

### Possible Molecular Mechanism Underlying Melatonin's Effect on LPS-Induced BBB Damage in Old Mice

Melatonin has shown protective effect on BBB integrity via a variety of pathways: inhibition of the toll like receptor 4 (TLR4)/NF-κB signaling pathway in neonatal rats (Hu et al., 2017), inhibition of NADPH oxidase-2 (Jumnongprakhon et al., 2016), inhibition of MMP-9 (Alluri et al., 2016), inhibiton of AMP-activated protein kinase (AMPK) activation (Wang et al., 2017) and impact on silent information regulator 1 (SIRT1; Zhao et al., 2015) and NLRP3 inflammasome (Rahim et al., 2017).

### AMPK Activation

AMPK activation has been shown to play important role in maintaining the integrity of BBB (Liu et al., 2012) and it is also reported that LPS inhibits the activation of AMPK, a serine/threonine protein kinase regulating cellular and organismal metabolism (Wang et al., 2017). Interestingly, AMPK activation has been shown to alleviate LPS-induced BBB disruption in both in vitro cell model (Zhao et al., 2014) and in vivo mice model (Zhou X. et al., 2014; Wang et al., 2017). Activation of AMPK also demonstrated protective effect against diabetes-induced BBB damage by inhibiting NADPH oxidase expression upregulation in brain capillary endothelial cells (Liu et al., 2012). In a recent study, Wang et al. (2017) showed that AMPK activation by melatonin reduced LPS-induced BBB damage in old mice and AMPK activation by metformin decreased LPS-induced gp91phox up-regulation in brain capillary endothelial cells (**Figure 2**). AMPK activation might be important in maintaining the integrity of BBB in old patients and AMPK dysfunction might play a key role in the initiation and progression of neurological disorders in old people. Therefore, activation of AMPK may be a strategy to reduce neurological disorders following sepsis and neuroinflamation-induced BBB damage in old people.

### Matrix Metalloproteinase-9 (MMP-9)

MMP-9 has been shown to play important role in BBB damage (Jin et al., 2013, 2015; Cai et al., 2015) and melatonin has been shown to bind to MMP-9 to act as its endogenous inhibitor. Melatonin treatment provided protection against traumatic brain injury (TBI)-induced BBB hyperpermeability via MMP-9 inhibition (Alluri et al., 2016), indicating its potential as a therapeutic agent for BBB damage.

#### Silent Information Regulator 1 (SIRT1)

SIRT1 was reported to be beneficial in sepsis. Using EX527, a SIRT1 inhibitor, the authors figured out that melatonin alleviated BBB damage in mice which subjected to cecal ligation and puncture via SIRT1 to inhibit inflammation, apoptosis and oxidative stress (Zhao et al., 2015).

#### NLRP3 Inflammasome

Aging and sepsis triggered NLRP3 inflammasome activation (Volt et al., 2016), which has been shown to be involved in the innate immune response during inflammation (Rahim et al., 2017). Furthermore, NLRP3 inflammasome activation was showed to be associated with the upregulation of apoptotic signaling pathway in various inflammatory diseases (Volt et al., 2016) and melatonin attenuated subarachnoid hemorrhage-induced BBB damage via attenuating the expressions of NLRP3 (Dong et al., 2016).

#### Dark Side/Downsides of Melatonin Supplementation

Although acute toxicity of melatonin is extremely low in both animal and human studies, melatonin may still cause minor adverse effects, such as headache, insomnia and nightmares (Malhotra et al., 2004). Based on previous studies, melatonin could be used as a daily supplement to delay or prevent changes associated with age. However, long-term side effect of melatonin has to be tested, because melatonin has been used as a contraceptive for women which could have reproduction alterations by consumption of melatonin (Voordouw et al., 1992). In addition, there was a decrease in sperm motility in male rats (Gwayi and Bernard, 2002), and long-term administration

#### REFERENCES


of melatonin inhibited testicular aromatase levels (Luboshitzky et al., 2002). It does not matter to provide old people with daily melatonin to prevent neurological diseases even if these two side effects may happen as they would not have reproduction anymore. Other side effects should be considered, for example, melatonin may accelerate the development of autoimmune conditions (Mattsson et al., 1994), increase atherosclerosis in the aorta in hypercholesterolemic rats (Tailleux et al., 2002) and produce opposite effects in cancer treatment with poorly timed administration (Bartsch and Bartsch, 1981).

#### Conclusion

In conclusion, decreased melatonin levels may account for the BBB damage in old people who often face the common stress of sepsis and neuroinflammation. Melation supplementation treatment significantly inhibits such events. Therefore, continuous daily melatonin supplementation may help prevent sepsis and neuroinflammation-related neurological diseases through maintaining the integrity of BBB in old people. Since melatonin has low toxicity profile and high efficacy in many pathophysiological states, it should be more commonly tested/used in the medical and veterinary arenas. Further studies are needed to verify the important significance of daily melatonin supplementation in old people.

#### AUTHOR CONTRIBUTIONS

W-CL, XW, XZ, XC and XJ wrote the manuscript and XC, XJ obtained the funding. XW drew the figures. All authors have approved the final version of this review article.

#### ACKNOWLEDGMENTS

This work was supported by National Natural Science Foundation of China (81671145), by Natural Science Foundation of Jiangsu Province of China (L221506415) and by grants from Shenzhen Science and Technology Innovation Commission (JCYJ20150402152005623). This work was also partly supported by Priority Academic Program Development of Jiangsu Higher Education Institutions of China.


inflammatory process in senescence-accelerated mice. J. Pineal Res. 42, 272–279. doi: 10.1111/j.1600-079x.2006.00416.x


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Liu, Wang, Zhang, Chen and Jin. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Dissecting the Molecular Mechanisms of Neurodegenerative Diseases through Network Biology

Jose A. Santiago, Virginie Bottero and Judith A. Potashkin\*

Department of Cellular and Molecular Pharmacology, The Chicago Medical School, Rosalind Franklin University of Medicine and Science, North Chicago, IL, United States

Neurodegenerative diseases are rarely caused by a mutation in a single gene but rather influenced by a combination of genetic, epigenetic and environmental factors. Emerging high-throughput technologies such as RNA sequencing have been instrumental in deciphering the molecular landscape of neurodegenerative diseases, however, the interpretation of such large amounts of data remains a challenge. Network biology has become a powerful platform to integrate multiple omics data to comprehensively explore the molecular networks in the context of health and disease. In this review article, we highlight recent advances in network biology approaches with an emphasis in brainnetworks that have provided insights into the molecular mechanisms leading to the most prevalent neurodegenerative diseases including Alzheimer's (AD), Parkinson's (PD) and Huntington's diseases (HD). We discuss how integrative approaches using multi-omics data from different tissues have been valuable for identifying biomarkers and therapeutic targets. In addition, we discuss the challenges the field of network medicine faces toward the translation of network-based findings into clinically actionable tools for personalized medicine applications.

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Daniel Ortuño-Sahagún, Centro Universitario de Ciencias de la Salud, Mexico Cheng-Chang Lien, National Yang-Ming University, Taiwan

#### \*Correspondence:

Judith A. Potashkin judy.potashkin@rosalindfranklin.edu

> Received: 18 October 2016 Accepted: 12 May 2017 Published: 29 May 2017

#### Citation:

Santiago JA, Bottero V and Potashkin JA (2017) Dissecting the Molecular Mechanisms of Neurodegenerative Diseases through Network Biology. Front. Aging Neurosci. 9:166. doi: 10.3389/fnagi.2017.00166 Keywords: Alzheimer's disease, Parkinson's disease, Huntington's disease, network biology, molecular mechanisms

#### INTRODUCTION

Neurodegenerative diseases are usually sporadic in nature and commonly influenced by a wide range of genetic, epigenetic and environmental factors. With the advent of new high-throughput technologies such as RNA sequencing, it has become essential to develop methods beyond the classical pathway analysis to systematically interpret large amounts of data in the context of health and disease. Despite the progress of high-throughput genomic studies the precise pathogenic mechanisms leading to the most prevalent neurodegenerative diseases remain elusive. To this end, the applications of network biology have been successful to provide biological insight and to decipher the molecular underpinnings of neurodegenerative diseases. Network biology is based on the premise that complex diseases, like neurodegenerative diseases, are frequently caused by alterations in many genes comprising multiple biological pathways. A network consists of nodes and edges that may represent genes, proteins, miRNAs, noncoding RNAs, drugs, or diseases connected through a wide range of interactions including, but not limited to physical, genetic, co-expression and colocalization. An example of network analysis of Alzheimer's disease (AD) that identifies central hubs is shown in **Figure 1**. Integration of multi-omic information coupled with network-based approaches is becoming

an essential step towards the advancement of personalized medicine (**Figure 2**). Some of the frequently used terms in network biology approaches are defined in **Table 1**.

Seminal work in network biology including the construction of the human disease network (Goh et al., 2007), the human functional linkage network (Linghu et al., 2009), the discovery of causal genes of obesity (Chen et al., 2008), and clinical biomarkers for cancer (Taylor et al., 2009), prompted efforts to study many different diseases using network-based approaches. In the last few years, there has been a steady growth in studies exploiting the concepts of network biology to understand neurodevelopmental and neurodegenerative diseases (Santiago and Potashkin, 2014a). For example, network approaches have successfully identified putative diagnostic biomarkers for Parkinson's disease (PD; Santiago and Potashkin, 2013a, 2015; Santiago et al., 2014, 2016), and progressive supranuclear palsy (Santiago and Potashkin, 2014b) reviewed in Santiago and Potashkin (2013b, 2014a,c). In addition, network-based approaches have provided insights into the molecular mechanisms underlying co-morbid diseases associated with PD including diabetes (Santiago and Potashkin, 2013a) and cancer (Ibáñez et al., 2014). In this review article, we highlight the most recent advances in network biology applications to understand the most common neurodegenerative diseases with an emphasis on brain specific networks.

### NETWORK-BASED APPROACHES IDENTIFIES PATHWAYS SPECIFIC TO ALZHEIMER'S DISEASE (AD)

AD is the most prevalent neurodegenerative disease, responsible for the majority of the cases of dementia, affecting more than 44 million people worldwide with an estimated global cost of more than 600 billion dollars<sup>1</sup> . Although the exact mechanism

<sup>1</sup>http://www.alzheimers.net/resources/alzheimers-statistics/

targets.

of disease remains unclear, a complex combination of genetic, epigenetic, lifestyle, environmental factors and aging are believed to be responsible for most of the cases. Pathological features of AD include the accumulation of amyloid beta (Aβ) plaques and protein tau in neurofibrillary tangles (NFT). While most of the AD cases are late onset (LOAD) and sporadic, some genetic mutations in the amyloid precursor protein (APP), presenilin 1 (PSEN1) and presenilin 2 (PSEN2) are documented to cause early onset AD, which accounts for approximately 2% of the cases with symptoms appearing before the age of 65 (Goate et al., 1991; Levy-Lahad et al., 1995; Janssen et al., 2003). The apoliporotein E-ε4 (APOEε4) is the only genetic factor identified in more than 60% of the sporadic AD cases, however, it has also been found in healthy individuals thus suggesting that other genetic factors may be responsible for the disease (Coon et al., 2007). To date, emerging high-throughput genomic technologies have reported more than 2900 genetic variations associated with AD<sup>2</sup> .

Although these studies have been valuable to understand the genetic diversity associated with AD, the multi-factorial mechanisms leading to the disease are unclear. Network-based approaches have been successful to systematically interpret these results and to gain insight into the mechanisms of disease. In particular, integrative approaches combining multi-omic data in networks have been employed to identify susceptibility genes and pathways in AD. For example, combinatorial network analysis of proteomic and transcriptomic data revealed subnetworks enriched in pathways associated with the pathogenesis of AD including the downregulation of genes associated with the MAPK/ERK pathway and the upregulation of genes associated to the clathrin-mediated receptor endocytosis pathway (Hallock and Thomas, 2012; **Table 2**). In this regard, disruption of the clathrin-mediated receptor pathway can lead to increased levels of APP thereby contributing to disease progression (Schneider et al., 2008; Hallock and Thomas, 2012). Integrative approaches have led to the identification of potential additional genetic risk factors and biomarkers for AD. For instance, integration of genome wide association studies (GWAS), linkage analysis and expression profiling in a protein-protein interaction (PPI) network yielded a 108 potential risk factors for AD including EGFR, ACTB, CDC2, IRAK1, APOE, ABCA1 and AMPH. Among these genes, EGFR, APOE and ACTB were found to overlap with proteomic data from cerebrospinal fluid of AD patients (Talwar et al., 2014) thus providing potential biomarker candidates. Collectively, these studies reinforce the power of integrative network approaches to identify pathways, genetic risk factors and biomarkers for AD.

Weighted gene coexpression networks analysis (WGCNA) are increasingly being used to find highly co-expressed gene modules associated with a particular biological pathway or a clinical trait of interest (Langfelder and Horvath, 2008).

<sup>2</sup>http://www.alsgene.org/



For example, construction of gene co-expression networks from 1647 postmortem brain tissues from LOAD patients highlighted immune and microglia enriched modules, containing a key regulator of the immune system, known as TYROBP (Zhang et al., 2013). Likewise, WGCNA analysis uncovered astrocyte-specific and microglia-enriched modules in vulnerable brain regions that associated with early tau accumulation (Miller et al., 2013). Implementation of WGCNA in RNA-sequencing data using brain samples obtained from the temporal lobe of subjects with dementia with Lewy body (DLB), LOAD and cognitively normal patients, identified network modules specific to each disease. For example, two network modules enriched in myelination and innate immune response correlated with LOAD whereas network modules associated with synaptic transmission and the generation of precursor metabolites correlated with DLB and LOAD (Humphries et al., 2015). Further, genes previously implicated in LOAD including FRMD4B and ST18 (Miller et al., 2008; Zhang et al., 2013) were prominent hubs within the myelination network (Humphries et al., 2015). Together, these findings suggested the involvement of microglia and myelination in the pathogenesis of AD and established differences in biological pathways between LOAD and DLB. Besides innate immunity pathways, network analysis of transcriptomic data from the brain hippocampus of normal aged and AD subjects identified key transcriptional regulators related to insulin (INS1, INS2) and brain derived neurotrophic factor (BDNF) interacting with the retinoic acid receptor related orphan receptor (RORA, Acquaah-Mensah et al., 2015) previously implicated in autoimmunity and diabetes (Solt and Burris, 2012).

With the growing interest in personalized medicine, it has become essential to develop tools to stratify patients according to symptoms, prognosis, and disease stage. This is highly important due to the fact that some subgroups of patients within a specific disease may experience a faster disease progression or respond to therapy differently. There are several documented examples on how networks could accelerate individualized treatment. For instance, analysis of protein interaction networks identified unstable network modules in different brain regions, in particular, in the entorhinal cortex of AD patients. Specifically, several protein interactions were present or absent at different Braak stages thus providing network modules characteristic of disease progression in AD (Kikuchi et al., 2013). Interestingly, the network modules with the largest number of disappearing protein interactions at late stage were associated with the histone acetyltransferase and the proteasome complexes. These modules were interacting via UCHL5 thereby indicating the perturbation of the ubiquitin-proteosome system in AD. Likewise, network analysis of six relevant brain regions affected in AD uncovered 136 hub genes of which 72 correlated with the Mini Mental State Examination (MMSE) and NFT scores, both widely utilized indicators of disease severity in AD (Liang et al., 2012). Among these genes, there were important transcription factors and kinases associated with AD including LEF1, SOX9, YY1, TCF3, TFDP1, CDK5, CSK and MAP3K3. Among these TABLE 2 | Brain network-based analysis of the most common neurodegenerative diseases.


genes, overactivation of CDK5 is a major trigger of tau hyperphosphorylation and NFT formation in AD suggesting it may be a target for therapeutic intervention (Wilkaniec et al., 2016).

Since some genetic risk factors have a stronger influence in the disease than others, patient categorization and stratification according to the genetic basis would be advantageous in personalized medicine. In the context of AD, APOEε4 is the strongest risk factor for LOAD accounting for more than 50% of the cases. APOEε4 carriers display different clinical and pathological features than those of non-carriers. For example, APOEε4 carriers perform worse on memory tasks (Marra et al., 2004) and have a higher amyloid beta deposition than non-carriers (Kandimalla et al., 2011; Jack et al., 2015). Moreover, APOEε4 carriers respond to treatment differently than noncarriers. For instance, a neuroprotective agent improved MMSE scores in APOEε4 carriers but not in non-carriers (Richard et al., 1997). WGCNA on a transcriptomic dataset from human cerebral cortex of LOAD identified distinct co-expression modules based on APOEε4 stratification (Jiang et al., 2016). Co-expression modules of APOEε4 carriers were enriched in hereditary disorders, neurological diseases, and nervous system development and function whereas modules of non-carriers were enriched in immunological and cardiovascular diseases thus suggesting that different biological processes could play a role in LOAD with different APOEε4 status (Jiang et al., 2016).

#### NETWORK-BASED APPROACHES IN PARKINSON'S DISEASE (PD)

PD is the second most prevalent neurodegenerative disease after AD, affecting more than 10 million people worldwide. Pathological features include the accumulation of aggregated alpha synuclein (SNCA) in intraneuronal cytoplasmic inclusion known as Lewy bodies and the progressive loss of dopaminergic neurons in the substantia nigra pars compacta. Dopamine restorative drugs and deep brain stimulation are current therapies to treat patients, however, these treatments only alleviate motor symptoms but do not impact disease progression. Mutations in the LRRK2, PARK2, PARK7, PINK1 and SNCA genes are known to cause familial PD. Most of the PD cases, however, are sporadic resulting from a complex interplay between genetics and environmental factors. In fact, some of the same genetic variants including SNCA and LRRK2 implicated in familial PD have been also associated with sporadic PD (Satake et al., 2009; Simón-Sánchez et al., 2009; Lin and Farrer, 2014). To date, advances in genomics have identified 28 genes associated with PD (Lin and Farrer, 2014).

Several pathways have been linked to the pathogenesis of PD including mitochondrial dysfunction, endoplasmic reticulum stress, autophagy, inflammation and impaired insulin signaling (Mercado et al., 2013; Nolan et al., 2013; Santiago and Potashkin, 2013b; Lin and Farrer, 2014). Despite this progress, the precise disease-causing mechanisms of PD are not fully understood. Complementation of genomic and transcriptomic studies with system biology approaches have provided insights into some novel mechanisms of disease. For instance, differential co-expression network analysis (DCA) performed on transcriptomic data from PD substantia nigra at autopsy, uncovered a transcript isoform of SNCA with an extended 3<sup>0</sup> untranslated region, termed aSynL, which influenced SNCA accumulation (Rhinn et al., 2012). Interestingly, the pattern of expression of the long aSynL isoform relative to the short isoforms was also observed in unaffected individuals harboring a PD risk variant in the SNCA locus (Rhinn et al., 2012).

Understanding the molecular events associated with the progression of PD could help delineate a timeline for effective therapeutic intervention. Gene co-expression network analysis showed differences in gene modules between PD and controls for different anatomic brain regions (Corradini et al., 2014). In PD, hub modules in the motor vagal nucleus, locus coeruleus, and substantia nigra were enriched in pathways related to stress response and neuron survival/degeneration mechanisms whereas in control samples gene modules were associated with neuroprotection and aging homeostasis. Interestingly, one of the main hubs in the substantia nigra of control samples was SIRT1, which has been widely implicated in neuroprotection in several neurodegenerative diseases (Donmez and Outeiro, 2013; Herskovits and Guarente, 2013). Analysis of PPI networks representing autophagy and mitochondrial dysfunction pathways identified key protein targets including p62, GABARAP, GBRL1 and GBRL2 that modulated 1-methyl-4-phenylpyridinium (MPP+) toxicity (Keane et al., 2015), a widely used toxin to mimic PD in animal and cellular models. Strikingly, overexpression of these proteins combined, but not each one alone, provided rescue of MPP<sup>+</sup> toxicity. This result further strengthens the notion that targeting a cluster of genes rather than a single gene may be the route to an effective treatment.

Integrative system biology approaches incorporating network analyses have been valuable in identifying potential therapeutic targets in PD. For instance, construction of networks integrating genetic information from Gene expression Omnibus (GEO), the Parkinson's disease database (ParkDB) and the Comparative Toxicogenomics Database (CTD), identified alvespimycin (17-DMAG) as a candidate neuroprotective agent for PD (Gao et al., 2014). Experimental validation showed that 17-DMAG attenuated rotenone-induced toxicity in vitro. Another approach combined human brain and blood transcriptomic data and identified RGS2 as a key regulator of LRRK2 function (Dusonchet et al., 2014), one of the most common genetic risk factor of PD. Of note, RGS2 protected against neuronal toxicity in a Caenorhabditis elegans model expressing wild type LRRK2. Combination of -omics data from different tissues, for example brain and blood, may be advantageous to understand neurodegeneration in light of the recent finding that demonstrated that cell types outside the brain contain genetic risk factors associated with PD (Coetzee et al., 2016) and thus may help uncover new putative therapeutic targets.

### NETWORK ANALYSIS IN HUNTINGTON'S DISEASE (HD)

Huntington's disease (HD) is one of the most common dominantly inherited neurodegenerative disorders. The symptoms include some motor symptoms, such as chorea and dystonia, as well as non-motor symptoms, including psychological changes, and cognitive decline leading to dementia (Ross et al., 2014). These symptoms are correlated with a selective degeneration of the striatal and cortical neurons (Ehrlich, 2012). Currently, there are no therapies to prevent the onset or slow the progression of HD.

This progressive and fatal disease is caused by abnormal extension of the CAG repeat coding for a polyglutamine (polyQ) tail in the huntingtin gene (HTT, MacDonald et al., 1993). Unaffected individuals have fewer than 36 repeats, whereas affected patients can have as many as 250 CAG repeats. It has been shown that the length of the polyQ extension is inversely proportional to the age of the disease onset (Orr and Zoghbi, 2007). Vesicle and mitochondrial transport, transcription regulation, neurogenesis and energy metabolism are among the cellular functions of the normal HTT protein (Borrell-Pagès et al., 2006). Both lost of function of the normal protein and gain of toxic properties of mutant HTT leads to HD pathology. In fact, it has been shown that whereas the normal HTT protein is neuroprotective, the mutant HTT is neurotoxic. Despite this progress, the molecular mechanisms involved in the complex phenotype of the disease are still largely unknown.

In order to understand the role of HTT in HD pathology, (Langfelder et al., 2016) expressed HTT with different CAG length in a mice model. They demonstrated that the length of the CAG repeats modified the transcriptome of the striatum, and to a lesser extent, the cortex. WGCNA allowed the identification of 13 striatal and five cortical gene coexpression modules that were strongly associated with Htt CAG length. Interestingly, cadherin and protocadherin (Pcdh) genes expression were dysregulated in four of the modules, indicating that regulatory factors of these genes, such as Rest, Ctcf and Rad21, could be involved in HTT toxicity in mice (Langfelder et al., 2016).

Similarly, WGCNA was performed on transcriptomic HD post mortem tissues including the frontal cortex, cerebellum and caudate nucleus regions. The authors found that genes involved in metalloprotein, stress response and angiogenesis were positively regulated in all the networks whereas genes involved in mitochondrion, glycolysis, intracellular protein transport, proteasome and synaptic vesicle were downregulated (Neueder and Bates, 2014). Analysis of the human transcriptome from HD patients compared to healthy samples confirmed that protein modification, vesicles transport, cell signaling and synaptic transmission are important pathways involved in HD (Mina et al., 2016). Interestingly, these modules were also found in a blood transcriptomic study (Mina et al., 2016). Despite the fact that dysregulation of similar pathways was observed in the blood and brain, there was no overlap in any of the individual genes common between the two tissues (Mina et al., 2016).

A system-based approach performed on human transcriptomic datasets from post mortem human cerebellum, frontal cortex and caudate nucleus from HD patients and controls showed that an astrocyte module is the network whose connectivity and expression is most altered in HD (Scarpa et al., 2016). This astrocyte module was located downstream of TGFβ -FOXO3 signaling. In this regard, the TGFβ pathway was upregulated in neural stem cell differentiated from HD patient induced pluripotent stem cells (iPSC; Ring et al., 2015). Analysis of corrected iPS cells expressing shorter polyQ tails showed a downregulation of TGFβ pathway target genes, including cyclindependent kinase inhibitor 2B (CDKN2B), inhibitor of DNA binding 2 (ID2), inhibitor of DNA binding 4 (ID4), paired-like homeodomain transcription factor 2 (PITX2), thrombospondin 1 (THBS1), and left-right determination factor 2 (LEFTY2, An et al., 2012). In addition, valproic acid and lithium, both affecting TGFβ signaling, have been shown to improve mood in HD patients (Grove et al., 2000; Liang et al., 2008; Watanabe et al., 2011; Scheuing et al., 2014; Raja et al., 2015).

HD non-motor symptoms such as stress-related psychiatric and sleep disturbances often precede the onset of motor symptoms (Duff et al., 2007) and system-based approaches have proposed that sleep and stress traits emerge from shared genetic and transcriptional networks (Jiang et al., 2015). Interestingly, the astrocyte network expression described by Scarpa et al also correlated with stress and sleep phenotype in a chronically stressed mouse model (Scarpa et al., 2016). Collectively, these results suggest that targeting components of the TGFβ signaling pathway may provide novel therapeutics for HD.

### NETWORK APPROACHES TO UNDERSTAND THE CONNECTION AMONG NEURODEGENERATIVE DISEASES

Widespread protein misfolding and aggregation is a hallmark of neurodegenerative diseases. Despite the fact that neurodegerative diseases are defined by a set of characteristic pathological and clinical features, there is some overlap in pathology, genetic risk factors, and mechanisms of disease. For example, accumulation of SNCA and Lewy body pathology, central in the pathogenesis of PD, are present in the brains of human AD and implicated in aberrant synapse formation (Hamilton, 2000; Kim et al., 2004). Several studies have identified Single nucleotide polymorphisms (SNPs) in the MAPT locus associated with PD and AD thus suggesting that a common genetic factor may put an individual at risk for both diseases (Desikan et al., 2015). In addition to MAPT, other genetic variants including PON1, GSTO, and NEDD9 have been associated with the risk of PD and AD thus strengthening the genetic overlap between both diseases (Xie et al., 2014). Not surprisingly, shared mechanisms related to oxidative stress, neuroinflammation, impaired insulin signaling, mitochondrial dysfunction, iron dyshomeostasis and nicotinic receptors have been implicated in the pathogenesis of AD and PD (Xie et al., 2014). Therefore, a system-level understanding of the disease-disease connections could accelerate the discovery of novel treatments for both neurodegenerative diseases.

A systems-based approach combining expression quantitative trait loci (eQTL) studies from cerebellum and frontal cortex of AD patients, GWAS from AD and PD and PPI networks indicated that some PD variants (cisSNPs, cis-acting SNPs) were associated with the expression of CRHR1, LRRC37A4 and MAPT located at 17q21 and suggestive of AD risk (Liu et al., 2015). Similarly, shortest path analysis on a network constructed from literature mining identified known genes that already have an association with AD and PD and seven previously unknown genes including ROS1, FMN1, ATP8A2, SNORD12C, ERVK10, PRS and C7ORF49 that may link both diseases (Kim et al., 2016). Besides finding shared genetic associations, network analysis employing the computation of a similarity matrix identified gene clusters related to DNA repair, RNA metabolism, and glucose metabolism shared in AD and PD (Calderone et al., 2016). Importantly, these pathways were not detected using the conventional gene ontology (GO) analysis thus highlighting the power of networks to uncover novel pathways.

In addition to the studies focused on AD and PD, recent network-based approaches have been applied to understand the molecular networks shared among other neurodegenerative diseases. One study focused on the dorsolateral prefrontal cortex (DLPFC) which is commonly affected in both AD and HD to construct coexpression networks using genome wide expression data from 600 postmorterm DLPFC tissues from AD, HD, and non-dementia controls. Differential coexpression analysis revealed a subnetwork of 242 genes enriched in pathways related to neuron differentiation, apoptosis, gap junction trafficking, and cellular metabolic processes (Narayanan et al., 2014). Interestingly, the 242 gene subnetwork overlapped with genes downregulated in postmortem brains of major depressive disorder, a condition that is associated with other neurodegenerative disorders including PD (Aarsland et al., 2012). Further inspection of this subnetwork identified a gained/lost gene coexpression patterns associated with chromatin organization and neural differentiation.

### NETWORK-BASED APPROACHES TO UNDERSTAND AGING-ASSOCIATED NEURODEGENERATION

Aging is one of the most common risk factors associated with neurodegeneration. With an average age of onset of 60 for the most common neurodegenerative diseases, the risk of developing PD or AD significantly increases with age. Dopamine synthesis, a crucial neurotransmitter that becomes depleted in the brain of PD, declines with age (Ota et al., 2006) and amyloid deposits, characteristic pathology in AD, are found in the aging brain of non-demented individuals (Pike et al., 2007). Beyond the overlap in pathological features, aging and neurodegenerative disorders share several dysregulated pathways. A system-based approach that identifies molecular networks shared between aging and neurodegeneration should reveal shared mechanisms, some of which may be targets for slowing disease progression. Discovering unique dysregulataed pathways that are not aging-associated could pinpoint potential therapeutics targets unique for a particular neurodegenerative disease.

Several studies have employed system biology tools to better understand age-related neurodegeneration. For example, a comparative pathway and network analysis of the brain transcriptome revealed shared networks and pathways between aging and PD including inflammation, mitochondrial dysfunction and metal ion homeostasis (Glaab and Schneider, 2015). Interestingly, the expression of the most significant shared gene, NR4A2, gradually declined with aging and PD. They found that this aging-associated gene expression changes in NR4A2 might increase the risk of PD by mechanisms similar to gene mutations linked to PD (Glaab and Schneider, 2015).

Proteostasis functional decline is common in aging and neurodegenerative diseases. In fact, several studies have proposed a mechanistic link between aging and loss of protein homeostasis leading to protein aggregation and toxicity. In this context, chaperones play a pivotal role in protein assembly and folding and its dysregulation may lead to protein aggregation and proteotoxicity. A recent study identified a chaperone subnetwork that exhibited concordant repression and induction expression patterns in brain tissues from human aging, AD, HD and PD patients. Subsequent investigation led to the discovery of a subnetwork comprising HSC70, HSP90, the CCT/TRiC complex and HSP40 and TPR-domain related co-chaperones with aberrant expression that were required to prevent Aβ and polyQ-associated proteotoxicity in C. elegans (Brehme et al., 2014). This shared chaperome subnetwork in aging and neurodegeneration, which is critical to maintain protein homeostasis, provides new targets for therapeutic intervention in neurodegenerative diseases. Similarly, a recent meta-analysis of about 1600 microarrays from brain tissue of AD patients revealed a set of downregulated genes corresponding to metastable proteins prone to aggregation (Ciryam et al., 2016). Thus, targeting components of the proteome homeostasis network may enable novel therapeutic opportunities for neurodegenerative diseases.

#### EPIGENETICS, AGING AND NEURODEGENERATIVE DISORDERS

Gene expression is temporally and spatially regulated by DNA methylation or histone modifications. These epigenetic changes could influence a global gene expression or target some specific genes. A role for epigenetic changes in gene expression has been proposed in aging and neurodegenerative disorders. Interestingly, many studies have reported a genome-wide tendency to DNA hypomethylation with age in different organs including the brain in aging animal models (Wilson et al., 1987; Brunet and Berger, 2014). These changes are proposed to play a role in the progression of aging (Benayoun et al., 2015; Zampieri et al., 2015). Interestingly, Humphries et al. (2015) has shown that hypomethylation was observed in a myelination network dysregulated in AD.

DNA methylation has been proposed as a biomarker for aging in cells, tissues and organs (Horvath, 2013). An acceleration of the epigenetic clock has been proposed in different neurodegenerative disorders. In this context, epigenetic age acceleration correlated with AD neuropathological markers such as neuritic plaques and amyloid load (Levine et al., 2015). In addition, an association between epigenetic age acceleration with episodic memory, working memory and cognitive decline was observed among individuals with AD (Levine et al., 2015). Histones modifications, such as acetylation and methylation, have been observed in AD models and patients (for review see Fischer, 2014). Interestingly, the epigenetic clock is also accelerated in brain regions from HD patients (Horvath et al., 2016).

Epigenetic modification is also proposed to contribute to neurodegeneration in PD. A genome wide DNA methylation and transcriptomic study in iPSC-derived dopaminergic neurons from LRRK2-associated PD patients identified common DNA methylation changes in LRRK2 and sporadic PD (Fernández-Santiago et al., 2015). DNA methylation changes in PD dopaminergic neurons correlated with the downregulation of RNA and protein expression of a network of transcription factors FOXA1, NR3C1, HNF4A and FOSL2, which have been implicated in PD. For instance, FOXA1 is a key determinant in the molecular and physiological properties of dopaminergic neurons (Pristerè et al., 2015) and HNF4A expression in blood has correlated with disease progression in PD (Santiago and Potashkin, 2015).

Several computational tools have been developed to facilitate the integration of epigenetic data in networks. For example, EpiRegNet is a publicly available web server that allows the construction of epigenetic regulatory networks from human transcriptomic data (Wang et al., 2011). Another model, the Artificial Epigenetic Regulatory Network (AERN) incorporated DNA methylation and chromatin modification in addition to genetic factors for the analysis of epigenetic networks (Turner et al., 2013). More recently, another computational model, the Biological Expression Language (BEL)<sup>3</sup> , enabled the analysis of functional consequences of epigenetic modifications in the context of disease mechanisms (Khanam Irin et al., 2015). Because BEL integrates literature-derived cause and effect relationships into networks, researchers can formulate novel hypotheses of disease mechanisms. Notably, BEL network modeling has been used to integrate epigenetic and genetic factors in a functional context in PD. Using this approach, SNCA, MAPT, DNMT1, CYP2E1, OLFR151, PRKAR2A and SEPW1, were found to be hypomethylated in PD and suggested to cause overexpression of genes that disrupt normal biological functions. Further, two SNPs, rs3756063 and rs7684318, were associated with hypomethylation of SNCA in PD patients (Khanam Irin et al., 2015). Collectively, these models demonstrate that the integration of epigenetic factors into networks can uncover novel mechanisms of disease.

<sup>3</sup>http://www.openbel.org/

### CHALLENGES AND FUTURE DIRECTIONS IN NETWORK MEDICINE APPLICATIONS TOWARDS PERSONALIZED TREATMENT

The field of network medicine has undoubtedly accelerated the understanding of the molecular mechanisms leading to neurodegeneration. The most significant brain network-based studies of the most common neurodegenerative diseases are summarized in **Table 2**. While network-based methods provide an unbiased approach to decode complex diseases and generate novel hypothesis, experimental validation is essential for network findings to be translated into useful diagnostics and therapeutic applications. In this regard, a growing number of studies have successfully identified blood-based biomarkers with potential clinical applicability. For instance, network analysis identified SOD2, APP, HNF4A, PTBP1 and NAMPT as useful to distinguish PD patients from HC in blood samples obtained from two independent cohorts (Santiago and Potashkin, 2013a, 2015; Santiago et al., 2014, 2016). Among these biomarkers, HNF4A and PTBP1, showed a dynamic expression pattern in longitudinal samples thus showing potential to track the clinical course of PD patients. Likewise, network analysis identified PTPN1 as a useful blood biomarker to distinguish PD from progressive supranuclear palsy, an atypical parkinsonian disorder commonly misdiagnosed as PD (Santiago and Potashkin, 2014b). Despite the success in PD studies, experimental validation of network-based findings in AD and HTT studies in clinically relevant studies is mostly lacking. For example, a systems medicine approach identified TYROBP as a promising target for therapeutic intervention in AD but to the best of our knowledge there are no follow-up studies (Zhang et al., 2013). Similarly, the involvement of RORA (Acquaah-Mensah et al., 2015) and other potential targets in AD are yet to be validated.

Besides experimental validation, another aspect for consideration is the cell-type and tissue specific analysis. This is important since the analysis of gene expression studies from whole brain sections might lead to misleading results that are not relevant to the specific cell type affected in the disorder. To circumvent this problem, recent studies have successfully employed high-throughput technologies that enable a single-cell resolution. A notable example studied the changes in astrocyte and microglia reactivity in AD. They observed that genes within the immune response pathway were more pronounced in astrocytes than in microglia thus demonstrating that cell-type specific characterization of the molecular changes may be more informative (Orre et al., 2014). More details about limitations in system-biology approaches in the context of neurodegenerative diseases have been well described recently (De Strooper and Karran, 2016).

FIGURE 3 | Disease-drugs networks. Interaction among different diseases, drugs and genes can be represented in a multi-level network model. For example, network-based approaches have been used to understand shared dysregulated pathways in Parkinson's disease (PD) and diabetes. For instance, some drugs to treat diabetes patients have shown neuroprotective effects in PD and the observed neuroprotection may be mediated through their interaction with the peroxisome proliferator-activated receptor gamma (PPARG). Blue and gray lines represent drug interactions and disease interactions, respectively. This network was retrieved by iCTNet application in Cytoscape v3.1.1. using genetic associations from genome wide association studies (GWAS) and drug interactions from the Comparative Toxicogenomics Database (CTD) as of September 2016.

Another emerging area of research in the network biology field is the study of disease comorbidities. Several conditions including, diabetes, cancer, major depressive disorder and cardiovascular disease, for example, have been associated with neurodegenerative diseases. For example, insulin resistance and diabetes have been linked to AD and PD and drugs to treat diabetic patients have shown promising results in both disorders (Santiago and Potashkin, 2013b). In addition, analyses of shared networks between PD and diabetes have elucidated potential blood biomarkers for PD (Santiago and Potashkin, 2013a, 2014c; Santiago et al., 2014). More recently, an integrative transcriptomic meta-analysis of PD and major depression identified NAMPT as a potential blood biomarker for de novo PD patients (Santiago et al., 2016). Furthermore, treatment with an enzymatic product of NAMPT elicited neuroprotective effects via activation of SIRT1 in an in vitro model of PD (Zou et al., 2016). Therefore, understanding the molecular networks shared between comorbid diseases could reveal novel diagnostics and therapeutic targets. Network analysis of gene-drug interactions in PD and diabetes demonstrates that some drugs may be beneficial for treating both diseases (**Figure 3**). In this context, treatment with commonly prescribed drugs to treat diabetes including rosiglitazone, metformin, pioglitazone and exenatide have shown neuroprotective effects in PD models (Santiago and Potashkin, 2013b, 2014c; Aviles-Olmos et al., 2014; Carta and Simuni, 2015). In particular, treatment with exenatide improved motor and cognitive function in PD patients (Aviles-Olmos et al., 2013, 2014). Treatment with pioglitazone, however, did not result in disease-modifying benefits in PD patients (NINDS Exploratory Trials in Parkinson Disease (NET-PD) FS-ZONE Investigators, 2015; Simon et al., 2015). Nonetheless, it has been noted that a longer exposure to pioglitazone may have been required to observe an improvement in PD patients (Brundin and Wyse, 2015). As shown in **Figure 3**, interaction of these drugs with the peroxisome proliferatoractivated receptor gamma (PPARG) may provide a mechanistic explanation to their neuroprotective effect. Some of these drugs are currently in clinical trials to determine if they are neuroprotective.

Nutrition is also recognized as an important component in the development and treatment of neurodegenerative diseases (Seidl et al., 2014). Given the promise of neuroprotective agents, the field of nutrigenomics is gaining interest among neuroscientists that are seeking to understand the complex nutrient-genetic

#### REFERENCES


interactions underlying neurodegeneration and neuroprotection. A recent example conducted a transcriptomic and epigenomic sequencing of the hypothalamus and hippocampus from a rodent model exposed to fructose consumption, which has been shown to contribute to the metabolic syndrome (Meng et al., 2016). Gene network analysis identified Bgn and Fmod as key genes involved in the observed metabolic alterations induced by fructose in mice. Strikingly, administration of docosahexaeonic acid (DHA) reversed the gene network changes elicited by fructose (Meng et al., 2016). This study provides evidence that integration of nutrigenomics coupled with network analysis can facilitate the identification of neuroprotective agents. Likewise, resveratrol, an antioxidant present in red wine, may also provide neuroprotection in PD patients and thus, could be tested in clinical trials (**Figure 3**).

In addition to a nutrient-rich diet, both physical exercise and cognitive training promote healthy aging (Kraft, 2012; Bamidis et al., 2014). It has been proposed that a combination of both together may be best to prevent cognitive decline and pathological aging (Kraft, 2012; Bamidis et al., 2014). In this regard, network analysis could be a useful tool to characterize the effects of physical exercise and cognitive training in the aging brain. Future studies directed at identifying gene expression changes associated with these lifestyle changes would be advantageous. Collectively, a multidimensional network approach that includes information about symptoms, drug treatments, comorbidities, nutrigenomics, physical exercise and cognitive training will be valuable to accelerate personalized treatment.

### AUTHOR CONTRIBUTIONS

JAS and VB wrote the first draft of the manuscript. JAS, VB and JAP edited and reviewed the final draft of the manuscript.

#### ACKNOWLEDGMENTS

This work was funded by the US Army Medical Research and Materiel Command under awards number W81XWH-09- 0708 and W81XWH13-1-0025 and the National Institute of Neurological Disorders and Stroke grant number U01NS097037 to JAP. The funding agencies had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

pluripotent stem cells. Cell Stem Cell 11, 253–263. doi: 10.1016/j.stem.2012. 04.026


aging. Neurosci. Biobehav. Rev. 44, 206–220. doi: 10.1016/j.neubiorev.2014. 03.019


genes in Alzheimer's disease. BMC Genomics 15:199. doi: 10.1186/1471-2164- 15-199


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Santiago, Bottero and Potashkin. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Racial Differences in Insular Connectivity and Thickness and Related Cognitive Impairment in Hypertension

#### Ganesh B. Chand<sup>1</sup> \*, Junjie Wu<sup>2</sup> , Deqiang Qiu2,3† and Ihab Hajjar1,4†

<sup>1</sup> Division of General Internal Medicine and Geriatrics, Department of Medicine, Emory University School of Medicine, Atlanta, GA, United States, <sup>2</sup> Department of Radiology and Imaging Sciences, Emory University School of Medicine, Atlanta, GA, United States, <sup>3</sup> Department of Biomedical Engineering, Georgia Institute of Technology, Emory University, Atlanta, GA, United States, <sup>4</sup> Department of Neurology, Emory Alzheimer's Disease Research Center, Emory University School of Medicine, Atlanta, GA, United States

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Anna Emmanouel, CITY College, International Faculty of the University of Sheffield, Greece Luis J. Fuentes, Universidad de Murcia, Spain

#### \*Correspondence:

Ganesh B. Chand ganesh.chand@emory.edu ganeshchand64@gmail.com

†These authors have contributed equally to this work.

> Received: 27 January 2017 Accepted: 18 May 2017 Published: 31 May 2017

#### Citation:

Chand GB, Wu J, Qiu D and Hajjar I (2017) Racial Differences in Insular Connectivity and Thickness and Related Cognitive Impairment in Hypertension. Front. Aging Neurosci. 9:177. doi: 10.3389/fnagi.2017.00177 Hypertensive African–Americans have a greater risk of cognitive impairment than hypertensive Caucasian–Americans. The neural basis of this increased risk is yet unknown. Neuroimaging investigations suggest that the normal neural activity comprises complex interactions between brain networks. Recent studies consistently demonstrate that the insula, part of the salience network, provides modulation effects (information flow) over the default-mode and central-executive networks in cognitively normal subjects, and argue that the modulation effect is declined in cognitive impairment. The purpose of this study is to examine the information flow at the nodes of three networks using resting state functional magnetic resonance imaging (MRI) data in cognitively impaired hypertensive individuals with the African–Americans and the Caucasian– Americans races, and to compare the thickness of impaired node between two racial groups. Granger causality methodology was used to calculate information flow between networks using resting state functional MRI data, and FreeSurfer was used to measure cortical thickness from T1-weighted structural images. We found that negative information flow of the insula in both African–Americans and Caucasian–Americans, which was in contrast with previously reported positive information flow in this region of normal individuals. Also, significantly greater negative information flow in insula was found in African–Americans than Caucasian–Americans (Wilcoxon rank sum; Z = 2.06; p < 0.05). Significantly, lower insula thickness was found in African–Americans compared with Caucasian–Americans (median = 2.797 mm vs. 2.897 mm) (Wilcoxon rank sum; Z = 2.09; p < 0.05). Finally, the insula thickness correlated with the global cognitive testing measured by Montreal cognitive assessment (Spearman's correlation; r = 0.30; p < 0.05). These findings suggest that the insula is a potential biomarker for the racial disparity in cognitive impairment of hypertensive individuals.

Keywords: cognitive impairment, cognitive racial disparity, salience network, default-mode network, centralexecutive network, Granger causality, cortical thickness, functional magnetic resonance imaging (fMRI)

## INTRODUCTION

fnagi-09-00177 May 30, 2017 Time: 16:59 # 2

It is estimated that above 30% of population and 65% of older population have hypertension worldwide (Novak and Hajjar, 2010). Previous studies suggest that hypertensive individuals have a greater chance of occurring the dementia and physical disability than normotensive individuals (Johnson et al., 2008; Elias et al., 2012; Faraco and Iadecola, 2013). Cognitive functions, especially the executive functions (Chand and Dhamala, 2016b), are widely reported of being impaired in hypertensive individuals (Vicario et al., 2005; Gorelick and Nyenhuis, 2012; Hajjar et al., 2016). Individuals with hypertension have higher chances of occurring executive dysfunction earlier than individuals with normotension, which indicates a potential vascular-cognitive association (Oveisgharna and Hachinski, 2010). However, there are very limited neuroimaging studies that investigated the neural bases of hypertension in cognitive decline (Li et al., 2015). Former non-neuroimaging case studies further report that the African–Americans bear the greater burden of hypertension in the United States and have earlier onset of hypertension and larger hypertension-associated cognitive symptomatology and mortality than the other racial groups, including the Caucasian–Americans (Redmond et al., 2011; Hajjar et al., 2016). The neural mechanisms underlying this racial disparity is largely under-investigated so far. A triple brain network model—a model consisting of the brain's default-mode, salience, and central-executive networks and their interactions—has been recently employed in the neuroimaging field to elucidate the differences in the connectivity patterns associated with the different levels of cognitive impairments such as higher versus lower impairment (Menon, 2011; Uddin, 2015). We therefore use this model to investigate whether there is a difference in the impairment at the nodes of triple network between the African–Americans and the Caucasian–Americans hypertensive cognitively impaired individuals.

Resting state functional MRI (rsfMRI) has been used to investigate the functional brain areas or neurocognitive networks (Biswal et al., 1995; Raichle, 2015). Recent neuroimaging investigations suggest that the neural basis of cognitive activity is related to a dynamically modulating interaction between multiple networks, including the salience, default-mode, and central-executive networks (Bressler and Menon, 2010; Menon, 2015; Uddin, 2015; Chand and Dhamala, 2016a). The key nodes of the default-mode network include the posterior cingulate and the ventromedial prefrontal cortices, the salience network encompasses the insula and the dorsal anterior cingulate cortices, and the central-executive network comprises the posterior parietal and the dorsolateral prefrontal cortices (Chen et al., 2013). It has been demonstrated that the insula and dorsal anterior cingulate of salience network are anatomically connected (Bonnelle et al., 2012; Jilka et al., 2014) and consist of a special type of neurons named von Economo neurons that relay information processed within these regions to other brain regions, including the nodes of default-mode and central-executive networks (Allman et al., 2005, 2010; Watson et al., 2006; Sridharan et al., 2008). This control signal by the insula and the dorsal anterior cingulate cortex has been

suggested to be crucial for cognitive maintenance, including a rest, in cognitively normal individuals (Sridharan et al., 2008; Goulden et al., 2014; Chand and Dhamala, 2016a). Alternation in insula connectivity has been consistently implicated in diseases, including autism, frontotemporal dementia, and schizophrenia (Menon, 2011; Uddin, 2015), but it has not been elucidated in cognitively impaired hypertensive patients. Literature suggests that the insula and anterior cingulate cortex—the key regions of salience network—respond as the racially biased brain regions (Cao et al., 2015), such as the greater activity of insula to faces of foreign races than faces of the same race of the subject (Lieberman et al., 2005; Liu et al., 2015). However, the difference in information flow—a measure from information theory that can be quantified using Granger Causality analysis (Dhamala et al., 2008b; Chand and Dhamala, 2017)—in these regions between the African–Americans and the Caucasian–Americans themselves has not been previously investigated. Previous investigations consistently report that the functional changes of brain regions (or networks) are associated with the underlying structural changes of those regions (or networks) with the progression of diseases (Xie et al., 2012; Menon, 2015). Specifically, recent studies suggest that the insula thickness decreases with the progression of cognitive decline in mild cognitive impairment (MCI) patients (Hartikainen et al., 2012; Moretti, 2015). However, whether there is a difference in insula thickness of the cognitively impaired hypertensive patients between the African– Americans and the Caucasian–Americans races has not been reported.

Here, we seek to examine the difference in information flow using Granger causality (Dhamala et al., 2008b; Chand and Dhamala, 2017) at the default-mode, salience and central-executive nodes between the African–Americans and the Caucasian–Americans hypertensive cognitive impaired individuals. As the African–Americans have higher hypertension-associated cognitive symptomatology and mortality than the other racial groups (Redmond et al., 2011; Hajjar et al., 2016), we hypothesized that (1) the control signal of the insula of salience network over the default-mode and central-executive nodes is more impaired (more negative value) in the African–Americans than in the Caucasian– Americans. We further seek to examine the structural difference that could substrate this racial disparity by comparing the insula thickness between the two racial groups. To test this, we further hypothesized that (2) the insula thickness is lower in the African–Americans than in the Caucasian–Americans, and finally (3) lower insula thickness is associated with poorer cognitive performance.

#### MATERIALS AND METHODS

#### Participants

This study was carried out in accordance with the recommendations of "Institutional Review Board (IRB) of Emory University" with written informed consent from all subjects. All subjects gave written informed consent in

accordance with the Declaration of Helsinki. The study and the protocol were reviewed and approved by Institutional Review Board of Emory University. The informed written consent was provided by the participants before data collection. We recorded magnetic resonance imaging (MRI) data from 78 individuals who had hypertension and MCI. The inclusion criteria were: (a) age ≥ 55 years, (b) hypertension defined by systolic blood pressure ≥ 140 mm Hg or diastolic blood pressure ≥ 90 mm Hg and (c) MCI was assessed based on previously defined Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria (Chao et al., 2009; Pa et al., 2009): Montreal cognitive assessment (MoCA) ≤ 26, clinical dementia rating score of 0.5, minimal functional limitation as reflected by the functional assessment questionnaire ≤ 7, and cognitive performance at the 10th percentile or below till the 2nd percentile on at least one of four screening tests—trail marking test B (Reitan, 1958; Tombaugh, 2004), Stroop interference (Stroop, 1935; Kimble et al., 2009), digit span forward and digit span backward (Humstone, 1919; Pa et al., 2009), verbal fluency and abstraction (Henley, 1969; Troyer et al., 1997). Trail marking test B screens the participant's executive ability to draw a line from a 'number' to a 'letter' in ascending order such as '1' to 'A', 'A' to '2', '2' to 'B', and so on. Strop interference effect measures the interference of predominant response in the reaction time of a task such as when the name of a word (say 'red') is printed in a different color (say 'blue'), it takes longer time to name the color of that word compared to when the word ('red') matches the name of color ('red'). Digit span forward and digit span backward test consists of two parts: first, the participant listens to and repeats a sequence of numbers, and second, the participant listens to a sequence of numbers and repeats those numbers in reverse order. The former part screens the short-term auditory memory while the latter part screens the participant's ability to manipulate the verbal information based on the auditory information. The verbal fluency screens the participant's fluency such as the participant is asked to tell as many words as the participant can that begin with certain letter (say 'A') in a limited time and the abstraction test screens the participant's ability to deal with ideas such as how an orange and a banana are alike. These tests have been commonly used to screen the MCI patients (Chao et al., 2009; Pa et al., 2009). The participants exclusion criteria were: (a) systolic blood pressure > 200 mm Hg or diastolic blood pressure > 110 mm Hg, (b) renal disease or hyperkalemia, (c) active medical or psychiatric problems, (d) uncontrolled congestive heart failure (shortness of breath at rest or evidence of pulmonary edema on exam), (e) history of stroke in the past 3 years, (f) ineligibility for MRI (metal implants or cardiac pacemaker), (g) inability to complete cognitive test and MRI scan, (h) women of childbearing potential and (i) diagnosis of dementia (self-reported or care-giver reported). In total sample, the mean age was 66.9 years (SD: 9.7), 55.1% were women, mean education was 14.9 years (SD: 2.6), mean systolic blood pressure 144.4 mm of Hg (SD: 22.6), mean diastolic blood pressure 86.4 mm of Hg (SD: 12.8), and MoCA ranged from 11 (minimum value) to 26 (maximum value) with mean score of 21.9 (SD: 3.1). Out of 78 participants, there were 50 African–Americans and 28 Caucasian–Americans. In the African–Americans, the mean age was 66.3 years (SD: 8.9), 60% were women, mean education was 14.9 years (SD: 2.5), mean systolic blood pressure 141.8 mm of Hg (SD: 21.3), mean diastolic blood pressure 85.4 mm of Hg (SD: 12.0), and mean MoCA score was 21.2 (SD: 3.2). In the Caucasian–Americans, the mean age was 68.5 years (SD: 10.7), 46.4% were women, mean education was 15.0 years (SD: 2.9), mean systolic blood pressure 147.6 mm of Hg (SD: 24.0), mean diastolic blood pressure 87.0 mm of Hg (SD: 12.7), and mean MoCA score was 23.4 (SD: 2.4). The age, sex, education year, systolic blood pressure, and diastolic blood pressure were not statistically significant different between the African–Americans and the Caucasian–Americans, but the MoCA score was significantly lower in the African–Americans compared to the Caucasian–Americans (p < 0.003) as shown in **Table 1**.

#### MRI Acquisition

Magnetic resonance imaging data were acquired on a SIEMENS Trio 3-Tesla scanner available at Center for Systems Imaging of Emory University, Atlanta, GA, United States. Foam padding and ear forms were used to limit head motion and reduce scanner noise to the participants. High-resolution 3D anatomical images were acquired using sagittal T1-weighted magnetization-prepared rapid gradient echo with repetition time = 2300 ms, echo time = 2.89 ms, inversion time = 800 ms, flip angle = 8 ◦ , resolution = 256 × 256 matrix, slices = 176, thickness = 1 mm. The rsfMRI were collected axially for 170 volumes during 7.14 min by using an echo-planar imaging (EPI) sequence with repetition time = 2500 ms, echo time = 27 ms, flip

TABLE 1 | Mean scores (standard deviations) and statistical comparison between the African–Americans (AA) and the Caucasian–Americans (CA) regarding their age, sex, education, blood pressure (BP), and Montreal cognitive assessment (MoCA) score.


p-Value indicates the level of statistical significance of the comparisons between the African–Americans and the Caucasian–Americans using non-parametric Wilcoxon rank sum for age, education, blood pressure and MoCA, and chi-square test for sex.

angle = 90◦ , field of view = 22 cm, resolution = 74 × 74 matrix, slices = 48, thickness = 3 mm and bandwidth = 2598 Hz/pixel. We requested the participants to hold still, keep their eyes open and think nothing during the rsfMRI scan.

### Image Preprocessing and Time Series Extraction

Images were preprocessed for slice-timing correction, motion correction, co-registration to individual anatomical image, normalization to the Montreal Neurological Institute (MNI) template, and spatial smoothing of the normalized images with a 6 mm isotropic Gaussian kernel. The SPM12 (Wellcome Trust Centre for Neuroimaging, London, United Kingdom<sup>1</sup> ) was used to perform those steps. We defined spherical regions of interest with 6 mm radius based on MNI coordinates centered at the posterior cingulate cortex (7, −43, 33) and ventromedial prefrontal cortex (2, 36, −10) of default-mode network, the insula (37, 25, −4) and dorsal anterior cingulate cortex (4, 30, 30) of salience network, and the posterior parietal cortex (54, −50, 50) and dorsolateral prefrontal cortex (45, 16, 45) of central-executive networks (see **Figure 1**) similar to the previous studies (Sridharan et al., 2008; Chand and Dhamala, 2016a). We selected the nodes only in the right hemisphere based on most prior neuroimaging studies that report the right-lateralized activations (Sridharan et al., 2008; Chen et al., 2013; Chand and Dhamala, 2016a). The MarsBaR software package<sup>2</sup> was used to extract the voxel time courses of those nodes.

#### Granger Causality Analysis

Multivariate analysis has become a commonplace to investigate the information flow between the brain areas and to study how such coordinated brain activity disrupts in diseases (Dhamala et al., 2008a,b; Chiong et al., 2013; Friston et al., 2013). Here we used Granger causality although other methods such as dynamic causal modeling, directed transfer function, and partial directed coherence provide the similar goals and results (Bajaj et al., 2016; Chand et al., 2016). The main benefits of using Granger causality are that it is a data-driven method and therefore

<sup>2</sup>http://marsbar.sourceforge.net

FIGURE 1 | Selection of the (A) default-mode (VMPFC, ventromedial prefrontal cortex; PPC, posterior cingulate cortex), (B) salience (AI, insula; DACC, dorsal anterior cingulate cortex), and (C) central-executive (DLPFC, dorsolateral prefrontal cortex; PPC, posterior parietal cortex) networks (P, posterior; A, anterior; L, left; R, right).

computes the information flow based on the data itself at the nodes and networks level, relies on fewer assumptions about the underlying interactions, and does not need computationally intensive time/efforts as opposed to other methods such as dynamical causal modeling (Stephan et al., 2010; Chand and Dhamala, 2016b, 2017). Recent studies by our group and by other groups have successfully applied Granger causality to resting state and/or task fMRI data in both health and disease and have produced meaningful results in terms of information flow at the brain nodes and networks (Sridharan et al., 2008; Chiong et al., 2013; Liang et al., 2014; Bajaj et al., 2016; Chand and Dhamala, 2016a).

Granger causality can be mathematically expressed by considering simultaneously measured time series. Suppose we have two simultaneously recorded time series represented as, (1) X1(1), X1(2),..., X1(t),... and (2) X2(1), X2(2),..., X2(t). Granger causality analysis in the frequency (f) domain examines the strengths, directions, and frequencies of interactions between dynamic processes. Granger causality from the second time series '2' to the first time series '1' (i.e., from brain region '2' to brain region '1') is computed as (Dhamala et al., 2008a,b; Chand and Dhamala, 2016b),

$$M\_{2\to 1}(f) = -\ln\left(1 - \frac{(\sum\_{22} - \sum\_{12}^2 / \sum\_{11}) |H\_{12}(f)|^2}{\mathcal{S}\_{11}(f)}\right) \tag{1}$$

where H is a transfer function, f represents a frequency-domain, S is spectral power, and P is noise covariance. The value of Granger causality (M) varies between 0 and +∞, representing zero connectivity strength and maximum connectivity strength, respectively. If there are 'N' numbers of brain areas, the information outflow (F) at a node i can be calculated as,

$$F\_i = \frac{1}{N-1} \sum\_{j}^{N} (M\_{i \to j} - M\_{j \to i}). \tag{2}$$

In our case, we have six nodes (two key nodes from each network). Therefore, an index j can be 1, 2, 3, 4, 5, and 6 nodes. The Granger causality outflow (also referred as the net information outflow) from the first node is F<sup>1</sup> = [(M1→<sup>2</sup> – M2→1) + (M1→<sup>3</sup> – M3→1) + (M1→<sup>4</sup> – M4→1) + (M→<sup>5</sup> – M5→1) + (M1→<sup>6</sup> – M6→1)]/5, where M1→<sup>2</sup> is Granger causality from the first node to the second node, and M2→<sup>1</sup> is Granger causality from the second node to the first node. Similarly, we calculate Granger causality outflows for other nodes. If the net outflow is negative at a node instead of the previously reported positive value in healthy individuals, then that node is said to have impaired directional connections.

#### Cortical Thickness Calculation

FreeSurfer version 5.3<sup>3</sup> was used to calculate the cortical thickness from T1-structural images. Briefly, this technique included spatial and intensity normalization, skull stripping, and an automated segmentation of cerebral white matter to locate the gray–white boundary (Dale et al., 1999). Cortical thickness

<sup>1</sup>www.fil.ion.ucl.ac.uk/spm/software/spm12

<sup>3</sup>https://surfer.nmr.mgh.harvard.edu/

was then computed from the distance between the gray–white boundary and the pial-surface (Fischl and Dale, 2000).

#### Cognitive Test

fnagi-09-00177 May 30, 2017 Time: 16:59 # 5

We assessed the MoCA (Nasreddine et al., 2005) of each participant. The MoCA is a 30-point scale test administered in 10 min and assesses the global cognitive abilities. It encompasses the following sub-tests. The short-term memory recall (five points) consists of two learning trials of five nouns and delayed recall after 5 min. Visuospatial test includes a clock-drawing (three points) and a three-dimensional cube copy (one point). Executive functions test comprises the trail-making time B (one point), a phonemic fluency (one point), and a two-item verbal abstraction (two points). Attention, concentration, and working memory include a sustained attention (one point), a serial subtraction (three points), and digit forward (one point) and digit backward (one point). Language test consists of naming animals (three points), repetition of two syntactically complex sentences and fluency (two points). Orientation test comprises orientation to time and place (six points). Moreover, if participant's formal education is 12 years or less, one point is added to his/her score. Previous studies (Nasreddine et al., 2005; Hachinski et al., 2006; Dong et al., 2010; Rossetti et al., 2011) suggest that the MoCA is more sensitive screening tool to define the MCI as compared with other existing screening tools such as mini-mental state examination (MMSE).

#### Statistical Analysis

We first checked whether the data are normally distributed or not using Kolmogorov–Smirnov test. The age, education, MoCA, insular thickness and net flow were negatively skewed and the systolic and diastolic blood pressures were positively skewed (asymptotic p-value in the range 6.17 × 10−<sup>71</sup> to 8.42 × 10−14) at the 0.05 significance level. As data variables did not show normal distributions before and after applying the appropriate transformations (square root, log, and reciprocal), we therefore chose the non-parametric alternative. We compared the sample characteristics between the African–Americans and the Caucasian–Americans using non-parametric Wilcoxon rank sum and/or chi-square test for discrete variables (e.g., sex). Net information outflows between the nodes of three networks were compared using non-parametric Wilcoxon rank sum test. The correlation analysis was performed using Spearman's correlation. A p-value less than 0.05 was considered a statistically significant. MATLAB (Natick, MA, United States<sup>4</sup> ) was used for analyzing the data.

### RESULTS

### Interactions among the Salience, Central-Executive, and Default-Mode Nodes

We computed the Granger causality between all possible pairs of the salience, central-executive, and default-mode

<sup>4</sup>https://www.mathworks.com

nodes and calculated the net information outflow from each node. Our net outflow calculation showed that the nodes of salience network have significantly lower outflow than that of the central-executive and default-mode nodes (**Figure 2**) (Wilcoxon rank sum; Z = 10.52; p < 0.05). This negative information outflow of salience network nodes compared to previously reported positive flow in healthy individuals implied the impaired salience network nodes in our overall cohort. We compared the net information outflow between African Americans and Caucasian–Americans and found that the insula of salience network has significantly lower (negative value) in the African–Americans than in the Caucasian–Americans (Wilcoxon rank sum; Z = 2.06; p < 0.05) as shown in **Figure 3**.

### Cortical Thickness Comparison and Correlation

To evaluate whether structural alterations explain these connectivity differences, we measured the insular cortical thickness and compared the values between African–Americans and Caucasian–Americans as displayed in **Figure 4**. Cortical thickness in the African–Americans (median = 2.797 mm) and in the Caucasian–Americans (median = 2.897 mm) was significantly different (Wilcoxon rank sum; Z = 2.09; p < 0.05). To investigate how the thickness of impaired insular cortex relates with the performance on global cognitive testing, we performed a correlation analysis between the thickness and the MoCA scores. We found that a lower thickness was associated with lower performance reflected by MoCA scores (r = 0.30; p < 0.05). These results are provided in **Figure 5**. Furthermore, we found that the lower cortical thickness was correlated with the lower net information outflow at the insular cortex (r = 0.31; p < 0.05) as shown in **Figure 6**.

## DISCUSSION

Here, we investigated the pattern of connectivity among the key brain areas of the salience, central-executive, and default-mode networks in hypertensive individuals with MCI and we compared African–Americans to Caucasian–Americans within this group. We found larger impairment in the control signal of the insula (of salience network) in the African Americans than in the Caucasian–Americans, as reflected by negative net information outflow metrics measured by Granger causality analysis. Although, there are very limited neuroimaging studies that investigate the neural bases of hypertension in cognitive decline (Li et al., 2015), our findings of greater information flow impairment in the African–Americans were in line with the previous non-neuroimaging reports that the African–Americans bear a greater risk of hypertension-associated cognitive impairment than the Caucasian–Americans (Redmond et al., 2011; Hajjar et al., 2016). We further examined the cortical thickness of impaired insula between the African– Americans and the Caucasian–Americans and found that the insula thickness of the African–Americans is significantly lower than that of the Caucasian–Americans. The insula thickness

FIGURE 2 | Net flow (information outflow minus information inflow) of the key nodes of the salience (IN, insula; DACC, dorsal anterior cingulate cortex), central-executive (DLPFC, dorsolateral prefrontal cortex; PPC, posterior parietal cortex) and default-mode (VMPFC, ventromedial prefrontal cortex; PPC, posterior cingulate cortex) networks. The AI and DACC of the salience network had a significantly lower net information flow compared with the central-executive and default-mode nodes (<sup>∗</sup> indicates statistical significance).

was found to be correlated with the behavior performance and with the net information flow at the insula cortex, respectively. Those results about insula thickness were also consistent with the existing literature that insula thickness decreases with cognitive decline (Hartikainen et al., 2012; Moretti, 2015), however the difference in insula thickness between the African–Americans and the Caucasian–Americans has not been previously explored.

Former studies consistently reported that the nodes of salience network render modulation effects over the default-mode and central-executive in healthy individuals (Sridharan et al., 2008; Goulden et al., 2014; Chand and Dhamala, 2016a). The controlling role of salience nodes over the other two networks has been argued to be structurally supported by direct white matter connections between the insula and the dorsal anterior cingulate cortex (Bonnelle et al., 2012; Jilka et al., 2014) and by their unique sharing of cytoarchitecture at neuronal level, i.e., only these regions consist of special type of neurons—von Economo neurons—that relay information processed within those nodes to other nodes, including the default-mode and executive nodes (Allman et al., 2005, 2010; Watson et al., 2006; Sridharan et al., 2008). Literature also shows that the insula of salience network is functionally connected to the central-executive network (Vincent et al., 2008), and has direct white matter connections to the other areas, including the inferior parietal lobe (Uddin et al., 2010), and temporo-parietal junction (Kucyi et al., 2012). These structural and functional settings show the great involvement of insula in many cognitive processes such as in the evaluation of task performance across varying perceptual and response demands (Uddin et al., 2010), the reorientation of attention in conscious error perception (or error awareness) (Ullsperger et al., 2010), and the switching between available cognitive resources to integrate external sensory information with internal states (Uddin and Menon, 2009). The dorsal anterior cingulate cortex of salience network is known for enhanced cognitive control

(Egner, 2009) such as in switching activity in association with the insula during behaviorally harder tasks (Chand and Dhamala, 2016a). The above mentioned neural basis of control signal of the insula and the dorsal anterior cingulate cortex network (salience nodes) has been suggested to be crucial for cognitive maintenance in both task and resting states in cognitively healthy individuals, whereas impairment to such control activity might be caused by the underlying neuroanatomical changes, including the injuries to the highly sensitive/vulnerable von Economo neurons (Allman et al., 2005, 2010; Watson et al., 2006; Sridharan et al., 2008; Bonnelle et al., 2012).

Emerging evidence suggest atypical engagement of the insula of salience network in disease, including frontotemporal dementia, autism, schizophrenia, and Alzheimer's disease (Menon, 2015; Uddin, 2015). The structural changes of the cortex, including the cortical thinning of insula with disease progression to MCI and/or Alzheimer's disease, is widely reported in elderly people (Singh et al., 2006; Hartikainen et al., 2012; Moretti, 2015). Previous studies also suggested the link between the underlying structural changes and the corresponding functional changes of brain nodes and networks (Xie et al., 2012; Menon, 2015). Our connectivity findings indicated that the control mechanism of the salience network was impaired (negative value) in both the African–Americans and the Caucasian–Americans, and the insula was more impaired in the African–Americans than in the Caucasian–Americans. Although, the cortical thinning of insula has been consistently reported in cognitive impairment (Hartikainen et al., 2012; Moretti, 2015), there are no neuroimaging studies so far to our knowledge that report the racial disparity of insula thickness. Neuroimaging literature suggests that the insula and dorsal anterior cingulate cortex—the key regions of salience network respond as the racially biased brain regions (Cao et al., 2015), especially the greater activity of insula to out-group race than ingroup race (Lieberman et al., 2005; Liu et al., 2015), but there are no previous reports about the difference in connectivity patterns of those regions between the African–Americans and the Caucasian–Americans groups with hypertension and cognitive impairment. Prior non-neuroimaging studies repeatedly report that the African–Americans bear the greater risk of hypertensionassociated cognitive impairment than the Caucasian–Americans (Redmond et al., 2011; Hajjar et al., 2016). Thus, our findings and existing neuroimaging/non-neuroimaging evidence taken together suggest that the insula is crucial in racial disparity in cognitively impaired individuals with hypertension. This study can be further extended in the future by including the groups of African–Americans and Caucasian–Americans with MCI but without a history of hypertension and by also including the cognitively healthy hypertensive and normotensive individuals to better dissociate the effect of hypertension alone in the brain nodes and networks.

In summary, we evaluated the patterns of interactions among the salience, default-mode, and central-executive nodes in the African–Americans and the Caucasian–Americans race groups, who had hypertension and cognitive impairment. We found that the insula of the salience network was functionally impaired greater, and had lower thickness in the African–Americans than in the Caucasian–Americans. Existing literature and our findings taken together thus suggest that the insula is potential biomarker in the cognitive disorders, including the racial disparity of cognitively impaired hypertensive population. It is worth nothing that the future research should direct toward dissociating the role of hypertension alone between the race groups in the insula and the other possible regions that are functionally and/or structurally connected to the insula.

#### AUTHOR CONTRIBUTIONS

Conceived and designed the experiment: IH, DQ. Performed the experiment: GC, JW, DQ, IH. Analyzed the data: GC, DQ, IH. Wrote the paper: GC, DQ, IH. Participated in the discussion and provided the comments: GC, JW, DQ, IH.

#### FUNDING

This research was conducted with NIA/NIH grants RF1AG051633 and R01AG042127 to IH. DQ is supported by NIH grants AG25688, AG42127, AG49752, AG51633, and has received research support from Medtronic and Siemens Medical Solutions.

#### ACKNOWLEDGMENT

We would like to thank the members of our team and the participants who volunteered for the study.

### REFERENCES

fnagi-09-00177 May 30, 2017 Time: 16:59 # 9



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer AM and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.

Copyright © 2017 Chand, Wu, Qiu and Hajjar. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation

Giovanni Assenza<sup>1</sup> , Fioravante Capone<sup>1</sup> , Lazzaro di Biase1,2 , Florinda Ferreri 1,3 , Lucia Florio<sup>1</sup> , Andrea Guerra1,2 , Massimo Marano<sup>1</sup> , Matteo Paolucci <sup>1</sup> , Federico Ranieri <sup>1</sup> , Gaetano Salomone<sup>1</sup> , Mario Tombini <sup>1</sup> , Gregor Thut <sup>4</sup> and Vincenzo Di Lazzaro<sup>1</sup> \*

<sup>1</sup>Clinical Neurology, Campus Biomedico University of Rome, Rome, Italy, <sup>2</sup>Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom, <sup>3</sup>Department of Clinical Neurophysiology, Kuopio University Hospital, University of Eastern Finland, Kuopio, Finland, <sup>4</sup>Centre for Cognitive Neuroimaging (CCNi), Institute of Neuroscience and Psychology, University of Glasgow, Glasgow, United Kingdom

Non-invasive brain stimulation (NIBS) has been under investigation as adjunct treatment of various neurological disorders with variable success. One challenge is the limited knowledge on what would be effective neuronal targets for an intervention, combined with limited knowledge on the neuronal mechanisms of NIBS. Motivated on the one hand by recent evidence that oscillatory activities in neural systems play a role in orchestrating brain functions and dysfunctions, in particular those of neurological disorders specific of elderly patients, and on the other hand that NIBS techniques may be used to interact with these brain oscillations in a controlled way, we here explore the potential of modulating brain oscillations as an effective strategy for clinical NIBS interventions. We first review the evidence for abnormal oscillatory profiles to be associated with a range of neurological disorders of elderly (e.g., Parkinson's disease (PD), Alzheimer's disease (AD), stroke, epilepsy), and for these signals of abnormal network activity to normalize with treatment, and/or to be predictive of disease progression or recovery. We then ask the question to what extent existing NIBS protocols have been tailored to interact with these oscillations and possibly associated dysfunctions. Our review shows that, despite evidence for both reliable neurophysiological markers of specific oscillatory dis-functionalities in neurological disorders and NIBS protocols potentially able to interact with them, there are few applications of NIBS aiming to explore clinical outcomes of this interaction. Our review article aims to point out oscillatory markers of neurological, which are also suitable targets for modification by NIBS, in order to facilitate in future studies the matching of technical application to clinical targets.

#### Keywords: neuromodulation, non-invasive brain stimulation, oscillations, TMS/tDCS, EEG

## INTRODUCTION

Oscillatory activities in neural systems may play a functional role (Gray, 1994) and abnormalities in neural synchronization mechanisms might be involved in the pathophysiology of several neuropsychiatric disorders (Uhlhaas and Singer, 2006). Thousands of neurons synchronize their activity to generate a typical oscillatory pattern that can be measured either through an electroencephalogram (EEG) from scalp electrodes or through local field potentials (LFPs) or intracranial EEG recordings from small-sized, implanted electrodes in the brain. The recording of

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Mihai Moldovan, University of Copenhagen, Denmark Giulia Cartocci, Sapienza Università di Roma, Italy

> \*Correspondence: Vincenzo Di Lazzaro v.dilazzaro@unicampus.it

Received: 14 December 2016 Accepted: 26 May 2017 Published: 13 June 2017

#### Citation:

Assenza G, Capone F, di Biase L, Ferreri F, Florio L, Guerra A, Marano M, Paolucci M, Ranieri F, Salomone G, Tombini M, Thut G and Di Lazzaro V (2017) Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation. Front. Aging Neurosci. 9:189. doi: 10.3389/fnagi.2017.00189 the magnetic field induced by the same activity is referred to as magnetoencephalography (MEG). A common obstacle in interpreting these signals arises because a given macroscopic extracellular signal can be generated by diverse cellular events. Indeed, deriving macroscopic variables from their elementary causal constituents requires to solve the inverse problem (Nunez and Srinivasan, 2006). As a consequence, explaining the physiology of neural oscillations at the macroscale is complex because each rhythm is the resultant of the physiology of specific neural assemblies, in particular a mix of their spontaneous and evoked activity (Buzsáki et al., 2012). Many non-invasive and invasive studies with a clinical focus described abnormal oscillatory activities in different neurological disorders typical of elderly patients, such as Parkinson's disease (PD), stroke, dementia and epilepsy. However, it is still unclear whether these abnormalities have a pathophysiological role, and by extension, whether these disorders can be considered ''oscillopathies'', or whether they represent merely epiphenomena of the neural changes causally underlying the symptoms.

Non-invasive brain stimulation (NIBS) techniques are able to induce functional changes in the brain, by inducing and modulating ongoing oscillatory activity (Krawinkel et al., 2015). Thus, these techniques might have a potential role in both the diagnosis and treatment of neurological disorders, revealing an abnormal oscillatory response, and/or be of use in the attempt to rebalance the activity in abnormally functioning neural circuits.

The aim of this review article is two-fold: to (1) critically analyze spontaneous and evoked neural oscillations in neurological disorders of elderly in terms of their possible pathophysiological role; and (2) evaluate the effects of NIBS on these neuronal oscillations and their possible therapeutic promise.

### SPONTANEOUS AND TASK-RELATED OSCILLATORY ACTIVITY

Brain connectivity is a dynamic process, which changes with aging, and can be modulated by cognitive and physical training (Bamidis et al., 2014). Different groups of neurons tend to synchronize their activity at specific frequencies, defined as rhythms, which have been categorized in five canonical frequency bands: delta (<4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–30 Hz) and gamma (30–90 Hz). Higher frequencies (>90 Hz) have been subsumed as high-frequency oscillatory (HFO) activity (Schomer and Lopes da Silva, 2012). EEG can easily detect oscillations from the delta to the beta band, while it is less suited for recordings of gamma and HFO activity, since the signal intensity of these activities in EEG does not emerge from the abundant artifactual activity (mainly muscular and surrounding direct current) of the scalp. The limited influence of muscular activity on MEG and intracranial recordings render these two techniques suitable to analyze also very high frequency oscillations (HFO). In the following sections, we will survey the existing literature as to evidence for a physiological and pathophysiological role of these different rhythms.

#### Delta

In physiological conditions, delta activity is the most prominent EEG feature of human non-rapid eyes movement (NREM) sleep. It originates in cortical neurons and has been proposed as possible mediator of sleep-dependent synaptic plasticity (for a review, see Tononi and Cirelli, 2012), synchronizing the excitability state of huge groups of cortical neurons to facilitate cortico-hippocampal memory processes (Abel et al., 2013; **Figure 1**). During wakefulness, delta activity is almost absent in physiological conditions, but it appears both after subcortical brain lesion sparing cerebral cortex (Gloor et al., 1977; Steriade et al., 1993, 2001) and after the induction of cortical plasticity (Assenza et al., 2013a). Clinical studies in acute stroke patients suggest that delta activity of the affected hemisphere (AH) is related to both the lesion volume and the acute neurological deficit (Assenza et al., 2009), and that its spreading from the AH to the unaffected hemisphere (UH) is associated with poor prognosis (Finnigan et al., 2004). Advanced EEG analyses demonstrate that delta activity of the UH results from an interhemispheric communication breakdown of electrical signals between the two hemispheres and that this might, in turn, interfere with the UH contribution to recovery, i.e., plasticity processes (Graziadio et al., 2012; Assenza et al., 2013b). Patients with focal epilepsy show an increase in delta activity during daytime and sleepiness, but its biological meaning is uncertain (Pellegrino et al., 2017).

### Theta

Human EEG experiments reliably demonstrated the relevance of frontal (Klimesch, 1999) and hippocampal (Colgin, 2016) theta power in memory tasks. One theory posits that items presented according to a theta rhythm can induce Hebbian plasticity and thus long-term potentiation (LTP) or depression favoring memory retention (Jensen and Lisman, 1996). More specifically, LTP is generated by the activation of slow NMDA channels, which own a 150 ms time constant. Therefore, repetitive activity of these channels can produce theta oscillations supporting episodic memory (Jensen and Lisman, 1996). Others have suggested that cortical and hippocampal theta activity guides cortico-hippocampal synchronization to realize consolidation processes in NREM and REM sleep phases (Abel et al., 2013; **Figure 1**). Clinical studies show that theta band power increases typically after acute brain lesions. Activity in this frequency band is very sensitive to acute neural damage induced by perfusion reduction (Astrup et al., 1979). However, its role in neural disorders is controversial, as in sub-acute and chronic stages after stroke, the persistence of theta oscillations can be a sign of a damaged network or alternatively of its attempt to reorganize itself, in analogy to delta activity (Assenza et al., 2013b). Accordingly, the enhanced presence of this slow oscillation may suggest a network disassembly, but also signal a role in promoting plasticity in the context of neural network reorganization.

presumably different functions. Slower rhythms (delta-theta) are associated with higher power spectra generated by the synchronization of a big number of neurons or of networks of neurons. They prevail during sleep, when memory consolidation phenomena occur. The alpha rhythm is the dominant oscillation in the awake state has been associated with inhibitory functions to gate information flow. Beta activity is a rhythm of the motor system (pyramidal and extra-pyramidal) and inhibits the changing of motor activity. Gamma and higher oscillations are resident in intracortical activity synchronizing small group of neurons. See the text of manuscript for more details.

#### Alpha

Alpha-band activity is the dominant oscillation in the awake human brain. It is prominently observed over areas of the visual and attention network, where it is negatively related to visual perception (Hanslmayr et al., 2005; Thut et al., 2006), which is in line with the hypothesis that this frequency band plays an inhibitory role (Klimesch et al., 2007). Conversely, posterior alpha-band activity over visual/attention areas is positively related to non-perceptual functions such as memory (Hanslmayr et al., 2005; see also Jokisch and Jensen, 2007). This suggests that the inhibition of external (visual) input is helping performance of internal (memory) tasks, and has led to the hypothesis that alpha activity may shape functional network architecture through inhibiting task-irrelevant areas. A further relevant feature of alpha-band activity is that it is the only rhythm (with the exception of slow beta activity) that can respond to a stimulus and/or task demand by either decreasing or increasing its amplitude/power, namely showing event-related desynchronization (ERD) and eventrelated synchronization (ERS; Klimesch, 2012). More specifically, brain regions that are activated during a task exhibit ERD, whereas regions associated with processing of irrelevant and/or potentially interfering tasks exhibit ERS (Pfurtscheller and Lopes da Silva, 2004). Such opposite changes in posterior alphapower (i.e., ERD vs. ERS) have been observed with tasks varying in stimulus modality (e.g., visual vs. sounds), stimulus processing domain (color vs. motion) or stimulus side (left vs. right), depending on which sensory feature needs to be processed or suppressed, and on the areas engaged in the processing of the relevant or irrelevant feature (Foxe and Snyder, 2011). This supports the theory that alpha ERS is the EEG correlate of cortical inhibition, while alpha ERD is a reduction of this inhibition. In brief, alpha activity may promote selection of cortical networks by inhibiting task-irrelevant areas (Pfurtscheller and Lopes da Silva, 2004; Klimesch, 2012; **Figure 1**).

In cognitive disorders, the corruption of the alpha band activity is a prominent finding. In Alzheimer's disease (AD), there is strong evidence in favor of impaired resting state cortical alpha activity. In comparison to healthy age-matched controls, AD patients show a decrease of posterior alpha power along with a significant anterior shifting of the maximum alpha peak, mainly during oscillatory activity at rest (Huang et al., 2000; Babiloni et al., 2004, 2013a; Jeong, 2004). Notably, this decrease in alpha band activity directly correlates with the cognitive deficits and the severity of the disease (Jeong, 2004). Moreover, longitudinal follow-up of those patients reveals that the aforementioned changes are short-term predictors of progression from Mild Cognitive Impairment to dementia (Jelic et al., 1996, 2000; Huang et al., 2000; Rossini et al., 2006). In contrast, the reduction of parietal-occipital alpha-beta power is a marker of dementia progression (Coben et al., 1985; Soininen et al., 1989, 1991). Furthermore, a constant and reliable hallmark of AD across EEG and MEG studies is the significant decrease of coherence at alpha frequency in temporo-parietal areas (Leuchter et al., 1992; Locatelli et al., 1998; Wada et al., 1998; Jelic et al., 2000; Adler et al., 2003; Montez et al., 2009). Similarly, in stroke patients, a prominent ipsilesional alpha-decrease is a predictor of poor outcome (de Vos et al., 2008) and conversely its preservation is a marker of good prognosis. In unilateral middle/anterior cerebral artery ischemic stroke, the alpha band decrease is more prominent in the brain regions responsible for those behavioral deficits that will still be present 3 months after stroke (van Putten and Tavy, 2004). Moreover, a high ratio of delta/alpha power during subacute stroke is associated with higher scores of NIHSS at 30-days post-stroke. The relevance of alpha rhythm in clinical symptoms has also been highlighted in extrapyramidal disorders. After levodopa administration in PD patients, alpha-activity is increased in the pedunculopontine nucleus (PPN), bidirectionally-coupled with similar changes in cortical EEG, when participants perform self-paced movements (Androulidakis et al., 2008). These findings suggest a possible physiological role of these oscillations in the PPN area, such as promoting motor related attentional processes, which can be affected in non-treated PD.

#### Beta

#### Beta Activity in the Cortico-Spinal System

Voluntary movements, or even cues that predict the need for a voluntary movement, are preceded and accompanied by suppression of the beta rhythm (Pfurtscheller and Lopes da Silva, 2004). However, the motor relevance of beta-activity is not confined to movement coding, but becomes also evident when a voluntary isometric contraction is sustained (Brown and Marsden, 1998). In this task, cortical beta activity synchronizes with electromyographic oscillations, as evidenced in the so-called cortico-muscular coherence. During a sensorimotor task, as in isometric muscle contraction, cortico-muscular coherence can also recruit primary sensory areas, for which synchronized oscillatory activity in the beta band correlates with performance (Tecchio et al., 2008; Chakarov et al., 2009). Thus, beta band is a prominent rhythm of the corticospinal system, tracking the efficient flow of motor information between the cortex and the periphery.

Clinical modulation of beta-activity occurs in acute and chronic, vascular and degenerative pyramidal system lesions. Compared to healthy controls, acute and chronic ischemic stroke patients with motor deficits have lower bi-hemispheric beta activity (EEG and MEG studies; Tecchio et al., 2005; Dubovik et al., 2012; Graziadio et al., 2012) and reduced beta ERS after a somatosensory input (Rossiter et al., 2014). While a more prominent ERS reduction in the AH is associated with bigger lesions (Laaksonen et al., 2012), the preservation of beta activity in both hemispheres in the acute phase correlates with a better motor outcome (Tecchio et al., 2005, 2007). In addition, beta ERS after a tactile stimulation is typical of patients with greater motor dexterity (Gerloff et al., 2006) and the increase of beta ERS in the AH from acute phase to chronic phase predicts hand dexterity recovery (Laaksonen et al., 2012). Patients affected by amyotrophic lateral sclerosis (ALS) show a significantly smaller beta ERD compared to controls during motor imagery as revealed by EEG studies in ALS patients with an emphasis on brain computer interface technology (Kasahara et al., 2012). Furthermore, ALS patients exhibit reduced beta ERS after movements (Riva et al., 2012) which in addition is merely unilaterally localized (beta rebound), as compared to a bilateral phenomenon in controls (Bizoviˇcar et al., 2014). These results are in line with alterations of the beta rhythm observed in stroke patients with motor deficits. Because observed in a pathophysiological setting purely affecting the pyramidal system, this confirms the relevance of oscillatory beta activity in pyramidal tract physiology and pathology.

#### Beta Activity in Cortico-Basal Ganglia Loops

There is wide agreement on the association of beta activity in cortico-basal ganglia loops (13–30 Hz) with static motor control, such as tonic or postural contraction (Jenkinson and Brown, 2011). In PD, in the absence of levodopa therapy, the cortical-subcortical motor loops tend to synchronize within the beta band. After levodopa treatment however, they tend to synchronize within higher frequencies (>70 Hz; Brown et al., 2001; Williams et al., 2002; Foffani et al., 2003). Deep brain stimulation (DBS) at 20 Hz (beta band) of the subthalamic nucleus (STN) synchronizes GP internus (GPi) at the same frequency, whereas high frequency (>70 Hz) STN DBS suppresses beta band GPi oscillations (Brown et al., 2004). In line with (in)direct modulation of these oscillations having a clinical effect, STN or Gpi high frequencies stimulation improves PD motor symptoms (Brown et al., 2004), while beta frequency stimulation of STN has an antikinetic effect in PD patients (Timmermann et al., 2004; Fogelson et al., 2005; Chen et al., 2011). Levodopa administration, which is effective in treating bradykinesia, decreases basal ganglia beta oscillation (Kühn et al., 2005, 2009; Weinberger et al., 2006; Ray et al., 2008; Zaidel et al., 2010). Conversely, anticholinergic drugs that are more effective in treating tremor do not affect basal ganglia beta band activity (Priori et al., 2004). Furthermore, cortical beta oscillations are inversely correlated with movement acceleration (Gilbertson et al., 2005). Thus, it can be argued that enhanced basal ganglia beta band activity may be an indirect marker of bradykinesia, adjustable by levodopa therapy (Brown et al., 2001; Levy et al., 2002; Priori et al., 2004; Brown and Williams, 2005). The relevance of beta band in basal ganglia motor control is corroborated by data on dystonia, where the GPi LFP presents a lower beta band power than in PD (Silberstein et al., 2003). Therefore, excessive beta synchronization in the basal ganglia circuit has an antikinetic effect, as occurs in untreated PD patients, while its reduction with treatment leads to bradykinesia improvement, but excessive reduction can lead to hyperkinetic disorders, such as dystonia.

Besides mere motor control, a modulation of beta-band (specifically, an increase in functional connectivity between bilateral occipital, parietal, temporal and prefrontal regions) was also observed after a mixed physical-cognitive training in elderly patients with mild cognitive impairment (Klados et al., 2016).

In conclusion, in the pyramidal and extra-pyramidal motor network, brain oscillations play a key role in motor control, with beta activity reflecting a residential rhythm of the motor system, necessary for its proper functioning. Beside its antikinetic role, beta activity should be considered in a wider context of information gating favoring the maintenance of the status quo of the selected neuronal system (pyramidal and extrapyramidal; Engel and Fries, 2010; Brittain and Brown, 2014; **Figure 1**).

#### Gamma and High Frequency Oscillations

Because of the broad frequency spectrum covered by Gamma activity and HFOs, there are many types of oscillations in these higher frequency bands (>30 Hz). Those in the normal brain appear to facilitate synchronization and information transfer necessary for cognitive processes, memory and sensory-motor integration (Tecchio et al., 2008; Vinck et al., 2013), while other classes of HFOs reflect fundamental mechanisms of epileptic phenomena and of basal ganglia movement disorders in patients. We will focus on the latter as the aim of the present article is to provide basic neurophysiological background in a clinical perspective.

As discussed above, in PD patients on levodopa therapy, the cortico-subcortical motor network tends to synchronize into gamma and higher frequencies (Brown et al., 2001, 2004; Williams et al., 2002; Foffani et al., 2003). Moreover, while LFPs recordings from the STN of dopamine treated PD patients show a dopamine- and movement-dependent 300-Hz rhythm (Foffani et al., 2003), no consistent rhythm is found in the absence of dopaminergic medication at rest in the 100–1000 Hz frequency band in most cases. More specifically, levodopa or apomorphine administrations elicits a 300-Hz rhythm, which is modulated by voluntary movements. The dopamine-dependent 300-Hz rhythm therefore probably reflects a bistable compound of STN activity supporting high-resolution information processing in the basal ganglia circuit. The switching from low to high frequencies oscillations (>300 Hz), has a neurophysiological prokinetic role, related to the motor improvement that follow levodopa treatment in PD. Furthermore, the absence of subthalamic 300-Hz may represent a pathophysiological clue in PD and thus also provide the rationale for an excitatory and not only inhibitory use of DBS mechanisms of action in patients (Foffani et al., 2003; Özkurt et al., 2011).

In epilepsy, pathognomonic oscillatory activity can be subsumed in two highly specific patterns: the spike-wave complex (SWC), and HFO. The SWC is the EEG marker of the paroxysmal depolarization shift occurring in neuronal cells. It is maximally expressed in absence epilepsy, where 3 Hz SWCs are reverberating in the thalamo-corticothalamic loop and cause a breakdown of all higher cognitive functions (Niedermeyer and Lopes da Silva, 2005). HFOs on the other hand are generated locally (as mechanisms must be fast enough to synchronize activity within 2 ms to 5 ms) and are frequently observed in epileptic patients. Candidate mechanisms of HFOs are ephaptic interactions, electrotonic coupling via gap junctions, or fast synaptic transmission (Zijlmans et al., 2012). HFO activity has recently been proven to be an excellent biomarker for the epileptogenic zone (Zijlmans et al., 2012), and can be sub-classified in ripples (80–250 Hz) and fast ripples (250–600 Hz; Bragin et al., 1999; Jirsch et al., 2006). Interictal HFOs mostly occur during slow wave sleep. In epilepsy surgery, removal of tissue with HFOs seems to predict good surgical outcome, even better than removal of the ictal onset zone (Jacobs et al., 2010), indicating that HFOs may mark cortex that needs to be removed to achieve seizure control. In brief, both HFO and SWC are excellent markers of epileptic disease activity.

### OSCILLATORY ACTIVITIES EVOKED BY STIMULATION OF THE PHERIPHERAL AND CENTRAL NERVOUS SYSTEM

In addition to examining intrinsic spontaneous and task-related brain oscillations across disorders, electrophysiology can be used to probe oscillatory brain responses and reverberations provoked by external stimuli such as tactile stimuli presented to the hand, the presentation of visual stimuli or transcranial magnetic stimulation (TMS) pulses applied directly to the cortex. This approach is routinely used for diagnostic purposes (see e.g., sensory or motor evoked potentials, i.e., SEPs or MEPs) in various peripheral and central nervous system diseases. More recently, the approach has been flagged to be of interest also for the characterization of network organization in the normal and dysfunctional brain, in particular when single pulse TMS is used over different cortical area in combination with multichannel EEG recordings (e.g., Rosanova et al., 2009; Bortoletto et al., 2015).

#### Sensory-Motor Cortex Stimulation

To assess the integrity of the ascending and descending tracts and cortical projection zones, the sensorimotor cortex can be stimulated by either direct electrical or magnetic stimulation or by peripheral nerve stimulation, and the evoked activity can be recorded from the cerebral cortex or higher cervical segments using spinal electrodes located close to the axons of the corticospinal tract.

Stimulation of the primary motor cortex (M1) by TMS evokes a high frequency discharge in a cluster of cortical pyramidal cells, both in animals and humans. For example, epidural recordings have revealed that a single M1-TMS pulse evokes a complex pattern of discharge in the cortico-spinal (CS) projections, as compared to the single action potential evoked by stimulation of a peripheral nerve. This evoked response has an oscillatory nature, being composed of a series of descending waves that are separated from each other by about 1.5 ms (i.e., ∼670 Hz; Di Lazzaro et al., 2012). Based on the hypothetical site of TMS activation of the CS system, the components of this CS volley can be subdivided into three components. The first component (D-wave) originates from CS axon stimulation, the second component (I1-wave) from monosynaptic activation of CS cells, and the later components (late I-waves) is believed to originate from the activation of cortical interneurons at greater anatomical or functional distance from the body of the CS cell. Also, D-, I1 and late I-waves are differentially activated depending on TMS intensity and orientation of the induced electric field (Di Lazzaro et al., 2004a). In neurological patients, an abnormality of this high frequency oscillatory activity has been reported in two rare single cases who had an electrode implanted in the high cervical epidural space for pain treatment (one patient had a stroke; Di Lazzaro et al., 2006) and the other a cerebral cortex atrophy due alcohol abuse (Di Lazzaro et al., 2004b; **Figure 2**).

In routine SEP diagnostics, the latency and amplitude of low frequency components are evaluated in a large variety of neurological disorders to assess the integrity of the somatosensory system. Peripheral nerve stimulation also elicits HFO activity (>400 Hz; SEP-HFOs) in different relay stations all along the somatosensory pathways (Klostermann et al., 1999). The frequency of these SEP-HFOs is not identical along the whole system, but varies according to the region: very high frequency

FIGURE 2 | Epidural activity recorded in a patient with cerebral cortex atrophy. Descending volleys evoked by latoro-medial magnetic stimulation and posterior-anteriormagnetic stimulation at 120% resting motor threshold (RMT) in five patients with no abnormality of the central nervous system and at the maximum stimulator output in one chronic alcoholic patient. The grand averages of epidural volleys recorded in patients with no abnormality of central nervous system are shown on the left and the averages of epidural volleys (of 10 sweeps) recorded in the patient with cerebral cortex atrophy are shown on the right. The latencies of the D and I1 waves evoked by LM and PA magnetic stimulation are indicated by vertical dotted lines. In control subjects, LM stimulation evokes a large D wave followed by 5 I waves; PA stimulation evokes only I waves. In the patient with cerebral cortex atrophy the output evoked by LM and PA magnetic stimulation is similar. Both techniques evoke a large D wave and no clear I waves but only two very small and delayed peaks (Di Lazzaro et al., 2004b).

components (>1000 Hz) are generated in subcortical structures, while slower frequencies (about 600 Hz) are recorded within the cerebral cortex (Klostermann et al., 2002; Hanajima et al., 2004; Insola et al., 2004). A great deal of work has focused on cortical components of upper limb SEPs (for a review, see Curio, 2000). These components decrease during sleep (Yamada et al., 1988; Hashimoto et al., 1996; Halboni et al., 2000) and are powerfully enhanced by arousal (Gobbelé et al., 2000; Restuccia et al., 2004), thus suggesting that they can be utilized to evaluate the activity of arousal-related structures (Restuccia et al., 2004). In idiopathic generalized epilepsy patients, SEP-HFOs are abnormally enhanced with respect to those obtained from healthy volunteers (Restuccia et al., 2007). Furthermore, the same authors observed increased SEP-HFOs in seizure-free childhood and juvenile absence epilepsy (CAE–JAE) patients, whereas they were normal in drug-resistant patients and in all patients with juvenile myoclonic epilepsy (JME), which is an idiopathic epilepsy that is usually drug-resistant. These results might reflect the hyperactivity of arousal-related brainstem structures and this hyperactivity may account for the different clinical outcome among IGE sub-syndromes (Restuccia et al., 2007), i.e., the better prognosis of CAE–JAE with respect to JME patients.

#### Visual Cortex Stimulation

Among the peripherally evoked oscillations, it is worth to also mention the steady state visually evoked potentials (SSVEPs). These consist of oscillatory brain responses generated by visual stimuli that are presented at a constant flicker rate. Clinical SSVEPs are usually (although not exclusively) recorded by EEG and their frequency equals the stimulus frequency plus its even harmonics. Typically, SSVEPs demonstrate an excellent signal-to-noise ratio and a distinct spectrum with characteristic peaks (Vialatte et al., 2010). Although they are thought to be generated in the occipital cortex, extracortical sources (e.g., subcortical structures) have been reported, in particular for the low-frequency components of SSVEPs. These oscillatory potentials have been successfully used in cognitive neuroscience as a probe for physiological brain processes (such as visual attention—(Morgan et al., 1996); working memory—(Silberstein et al., 1995); body perception—(Giabbiconi et al., 2016)), as well as in clinical neuroscience to study age-related diseases. In AD, the study of medium and high frequency components of SSVEPs allowed to demonstrate an alteration in the magnocellular pathway of the visual system (Jacob et al., 2002). In PD, which is often associated with visual abnormalities due to retinal dopaminergic deficiency (Stanzione et al., 1992), these evoked oscillations have proven useful for clinical assessment and monitoring of dopaminergic treatments (Tagliati et al., 1996). In epilepsy diagnostics, arguably the most famous application of SSVEPs, intermittent photic stimulation is commonly used for photic driving to study photosensitive epilepsy since it can induce epileptiform waves in the EEG (Fisher et al., 2005).

#### TMS-EEG

TMS-EEG allows the co-registration of EEG during TMS, thus providing the opportunity to noninvasively and directly probe the brain's cortical excitability and time-resolved connectivity, as a function of instantaneous state (Ilmoniemi and Kici´c, 2010; Ferreri and Rossini, 2013). Furthermore, analysis of the spectral components of the TMS-evoked EEG responses has revealed that TMS consistently evokes dominant alpha-band oscillations in the occipital cortex (Herring et al., 2015), beta-band oscillations in the parietal cortex, and beta/gammaband oscillations in the frontal cortex (Rosanova et al., 2009). This has been interpreted to suggest that each cortical area tends to preserve its own natural frequency, even when indirectly engaged by TMS through network effects or when stimulated at different intensities (Rosanova et al., 2009). Moreover, through this methodology, new information has emerged on how ongoing oscillatory activity maps to behavioral output measures. In the healthy motor system, for instance, TMS-EEG has revealed that MEP amplitude (an index of corticospinal excitability) is inversely correlated with spontaneous fluctuations of rolandic alpha power (Zarkowski et al., 2006; Sauseng et al., 2009) and positively correlated with ipsilateral prefrontal beta activity as well as with bilateral centro-parietal-occipital delta activity (Mäki and Ilmoniemi, 2010; Ferreri et al., 2014). Interestingly, in older adults this pattern is only partially preserved, possibly reflecting a compensatory phenomenon to physiological aging (Ferreri et al., 2017). In PD patients who underwent unilateral thalamotomy (ventrolateral nucleus), a preliminary TMS-EEG study demonstrated higher TMS-induced beta amplitudes in the intact hemisphere, relative to stimulation of the AH. This may confirm a significant contribution of the motor thalamus in the facilitation of cortically generated oscillation through cortico-subcortico-cortical feedback loops. Moreover, these patients showed abnormal TMS-induced beta oscillation in the intact hemisphere when compared to young healthy controls (Werf et al., 2006), supporting that TMS-EEG can be a useful technique to non-invasively monitor pathological oscillatory activity in patients.

In summary, several neurological conditions of elderly are known to be associated with abnormal oscillatory profiles in specific frequency bands, depending on the disorder (**Figure 3**). Many of these oscillations normalize with treatment, and/or are predictive of disease progression or recovery. Importantly also, some of these oscillations could be modulated with a functional neurosurgery approach. In the following, we ask the question to what extent existing NIBS protocols may be tailored to interact with these pathological oscillations and possibly also with associated dysfunctions.

#### NIBS MODULATION OF OSCILLATORY ACTIVITIES

#### Brief Introduction of Main NIBS Techniques

TMS is a non-invasive and painless method to stimulate excitable neuronal tissue with an electric current induced by an external time-varying magnetic field (Rossini et al., 1994). Its key feature

is the unique ability to probe changes in excitability, connectivity, as well as global plasticity of intracortical circuits in specific brain areas (particularly motor and visual cortex) or even to induce plasticity by means of non-physiologically induced neuronal activity using repetitive TMS (rTMS) paradigms (Di Lazzaro and Ziemann, 2013). The latter, rTMS protocols were introduced in 1989 using consecutive TMS pulses with a progressively shorter interpulse interval (as short as 10 ms, for a review, see Fitzgerald et al., 2006; Di Lazzaro et al., 2011). For rTMS, a specific set of stimulators is needed to overcome the recharging time and maintain constant TMS output. rTMS has a modulatory effect on cortical excitability which outlasts the stimulation period and can be used in a variety of ways both in motor and non-motor brain regions and with local and nonlocal effects on brain activity. The aftereffects of rTMS might relate to activity-dependent changes in the effectiveness of synaptic connections, reflecting plasticity mechanisms of the brain (Di Lazzaro et al., 2011). The majority of research on the effects of rTMS on cortical activity has focused on M1-TMS, as the excitability of this brain region can be easily measured via MEPs amplitude modulation. In terms of the directionality of aftereffects, there is a relative consensus that rTMS frequencies below 1 Hz are mainly inhibitory, while repetition rTMS rates of 5 Hz and beyond are mainly facilitatory at least with regard to the corticospinal motor output. In addition, other patterned protocols such as theta burst stimulation (TBS) or quadripulse stimulation (QPS) have been introduced, which opens new intriguing horizon for TMS (Di Lazzaro et al., 2011). Moreover, in the last two decades, advances in the coupling of TMS with functional imaging techniques has allowed to study the temporo-spatial patterns of local and distal plastic changes in the brain following TMS. This has provided a sensitive means for identifying brain regions where changes in regional neuronal activity correlate with TMS-outcome on motor performance and behavior in both healthy and pathologic conditions. When declined in its specific paradigms—such as rTMS or coupled with different kinds of neuroimaging techniques, TMS has the potential for sophisticated uses to study and/or modify the regional cortical involvement in processes as diverse as motor and cognitive functioning (Rossini et al., 2015; Thut et al., 2017). For example, if the modulation of a target cortex is the objective, TMS can be declined in its repetitive paradigms (see e.g., Di Lazzaro et al., 2011). If the understanding of temporal aspects of neuronal processing or interventions is the main focus of an investigational study, TMS can be coupled with neuroimaging methods with a high temporal resolution such as EEG for guiding TMS or documenting its effects (see e.g., Ferreri and Rossini, 2013). For instance, new EEG/MEG-guided approaches explore the benefits of tuning rTMS to underlying brain oscillations to enhance TMS efficacy, e.g., by triggering to high excitability states or tuning to oscillatory frequencies (Thut et al., 2017).

Along with the TMS methods, transcranial electrical stimulation methods have been developing in the last decades. The first to be introduced was transcranial direct current stimulation (tDCS) about 20 years ago. tDCS requires the application of a weak and constant direct electrical current (1–2 mA) through two or more electrodes placed on the scalp (Priori et al., 1998). The main advantage of tDCS is its ability to achieve relatively long-lasting cortical changes after the stimulation has ended, with the duration of these changes depending on the length of stimulation as well as its intensity. tDCS does not induce activity in resting neuronal networks, but modulates spontaneous neuronal activity (Fritsch et al., 2010). In addition, the amount and direction of its effects critically depend on the previous physiological state of the target neural structures (Antal et al., 2014). Its biological effects include changes in neurotransmitters, effects on glial cells and on microvessels, modulation of inflammatory processes and, most importantly, the subthreshold modulation of neuronal membrane potentials which is capable to vary cortical excitability and activity depending on the direction of current flow through the target neurons (for a review, see Woods et al., 2016). Recently, an additional type of transcranial electrical stimulation method has been introduced, namely transcranial alternating current stimulation (tACS). tACS is able to induce or entrain brain oscillations by causing coherent changes in the firing rate and timing of neuronal populations (Antal and Paulus, 2013). It is capable of modulating perception, cognitive functions, and motor performance (for a review, see Woods et al., 2016). If the stimulation is strong enough, such behavioral effects can be achieved simply through the rhythmic modulation of excitability in the form of alternating ''Up'' and ''Down'' states, being this due to alternating relative neural depolarization and hyperpolarization. However, only low-current densities are used in studies in humans and so successful stimulation is often thought to leverage the resonance characteristics of the underlying brain. For this to occur, the stimulation frequency must approximate the natural resonance frequency of local neural circuits, so that spontaneous network oscillations are preferentially entrained (Helfrich et al., 2014b; Guerra et al., 2016). Accordingly, tACS effects tend to be frequency and area selective (Feurra et al., 2011). In the case of sensorimotor cortical areas, convergent evidence suggests a tendency for resonant activity to occur at about 10 Hz and 20 Hz (Hari and Salmelin, 1997). Beta activity, centered on 20 Hz, is focused anterior to the central sulcus and stimulation at or near 20 Hz can thus synchronize the activity of populations of pyramidal neurons so that there is increased corticomuscular coherence at the stimulation frequency (Pogosyan et al., 2009). At present, this represents one of the main field of application of this technique (for a review, see Woods et al., 2016).

#### Modulation of Brain Oscillations by NIBS in Healthy Subjects

NIBS techniques reach both cortical and subcortical structures (e.g., through anatomical connections; Ruff et al., 2009) and consequently modulate neural oscillations in both the directly stimulated cortex and the associated network areas/structures (Lang et al., 2005; Fregni and Pascual-Leone, 2007; Zaghi et al., 2010; Fox et al., 2014). Therefore, therapeutic application of NIBS could conceptually rely on the interference with subcortical oscillators through cortical targets (Fregni and Pascual-Leone, 2007). However, NIBS effects on brain oscillations seem to be topographical selective. Indeed, the different frequencies at which transient oscillations can be triggered by single pulses of TMS are a function of cortical site (Rosanova et al., 2009), as is the frequency that can be manipulated by rTMS (Thut et al., 2011; Bortoletto et al., 2015; Romei et al., 2016). This suggests frequency specificity of particular functional networks (in line with some of the studies reviewed above). Similar considerations apply to rhythmic stimulation with tACS (see below). The numerous rTMS studies aimed at establishing modulatory effects on different resting EEG rhythms however obtained inhomogeneous results so far, possibly related to the heterogeneity of the stimulation protocols and the investigated cognitive or behavioral tasks (Okamura et al., 2001; Strens et al., 2002; Klimesch et al., 2003; Schutter et al., 2003; Griškova et al., 2007; Brignani et al., 2008; Fuggetta et al., 2008; Grossheinrich et al., 2009; Thut et al., 2011; Noh et al., 2012; Vernet et al., 2013). The most reliable findings are in favor of changes in delta band activity evoked by stimulating frontal areas, which are the main source of the so-called slow traveling wave of sleep (Ferri et al., 2005; Murphy et al., 2009). Indeed, when a single TMS pulse is applied over these regions, both during sleep (Massimini et al., 2004) and wakefulness (Manganotti et al., 2013), cortical activity in the delta band typically dominates the response. Furthermore, spontaneous delta band activity can be increased by intermittent TBS (iTBS), a rTMS paradigm inducing LTP-like activity (Assenza et al., 2013a; Di Pino et al., 2014) and by frontal 10-Hz rTMS (Griškova et al., 2007) over frontal areas. This suggests that frontal stimulation may trigger and or modify the activation of cortical delta oscillators, perhaps favoring cortical plasticity (Assenza and Di Lazzaro, 2015). However, it is worth noting that none of these attempts leads to a reliable improvement of cognitive performance in healthy individual.

Electric NIBS research has recently developed the use of sine-wave electric current, namely tACS (see above) with the goal of modulating endogenous brain oscillations (Antal and Paulus, 2013). Several intuitive and promising findings have been obtained (reviewed in Herrmann et al., 2013; Frohlich, 2015). However, very little is known about how such stimulation modulates neuronal activity (that in turn guides behavior) in the long run (Veniero et al., 2015). It seems that when tACS is delivered with a stimulation frequency in the alpha band, robustly enhanced spontaneous alpha oscillations are observed after stimulation (aftereffects) in comparison to a sham control (Helfrich et al., 2014b; Vossen et al., 2015). However, the mechanisms of these aftereffect, in particular relative to the correlation between stimulation frequency and endogenous frequency, and whether these effects modulate behavior remain unknown. It is unclear whether the aftereffect on brain oscillations can emerge due to an adjustment of the endogenous frequency to the stimulation frequency (entrainment), or whether it is the endogenous frequency that is modulated in amplitude, although initial results would favor the latter scenario (Vossen et al., 2015). Similarly, convincing results are available for enhancement of gamma oscillations (Helfrich et al., 2014a; Strüber et al., 2014) by gamma-tACS in humans. However, a long-lasting modulatory aftereffect has not yet been achieved in this frequency bands (Nowak et al., 2017). Overall, given the small number of tACS studies, many gaps need to be addressed, including the development of principled, computational models and parameterized experiments to transform speculations about underlying mechanisms into carefully tested hypotheses (Frohlich, 2015).

A final NIBS technique that is promising for modulating EEG activity is stimulation with low-intensity, low-frequency (0–300 Hz) magnetic fields (ELF-MFs). This can induce measurable changes in electrical brain activity and influence neuronal functions such as motor control, sensory perception, and cognitive activities. In healthy humans, Bell et al. (1994) compared the effects of 1.5 and 10 Hz MF, while Cvetkovic and Cosic (2009) analyzed the effects of several ELF-MF frequencies (4, 8.33, 10, 13, 16.6 and 50 Hz) on the power of the corresponding EEG bands. In both cases, spectral analysis demonstrated that specific EEG frequencies can be influenced by stimulation at matching MF frequencies. However, similarly to tACS, the mechanisms of the interaction between ELF-MFs and the brain are unclear (Di Lazzaro et al., 2013).

In the following paragraphs, we critically revise the experiments that have been conducted so far in different patient groups.

## NIBS and Brain Oscillations in Neurological Disorders

#### Parkinson's Disease and Tremor

The neurophysiological circuit underlying parkinsonian tremor showed to be very stable, in frequency domain, indicating a strong oscillating system (di Biase et al., 2017). However, this circuit can be perturbed by NIBS. rTMS is capable to briefly alleviate PD motor symptoms (Zanjani et al., 2015). One possibility is that rTMS could disrupt the excessive beta activity that is characteristic of cortico-subcortical motor network activity in PD patients, similarly to what is supposed to be the mechanism of action of STN DBS. Accordingly, Doyle Gaynor et al. (2008) analyzed the effects of single pulse TMS on pathological STN oscillations recorded from intracranial electrodes in DBS patients who were on medication. Prior to the stimulation, all but two patients presented excessive 13–35 Hz activity in STN, when the patient were at rest. A delayed, transient beta activity suppression was found after ipsilateral and contralateral M1 and SMA (supplementary motor area) TMS (30–50 single TMS pulses delivered every 5 s). This was observed with sub-threshold as well as supra-threshold TMS, suggesting that changes in oscillatory activity in the STN were centrally driven and not due to peripheral afferent inputs secondary to TMS-evoked muscle responses. No effect on alpha band activity was found, in line with the frequency specificity of the target area/function. Importantly, the authors have found that the duration of the beta activity suppression was of about 400 ms following each single TMS pulse. Future studies should therefore investigate whether this duration can be prolonged when rTMS is used, such as 5 Hz rTMS. Longer-term effects would be required to pave the way for NIBS secondary therapeutic implications.

Krause et al. (2014) used tACS over motor cortex in PD patients to study its effects on brain rhythms and motor performance using MEG and behavioral measures during motor tasks (i.e., static isometric contraction, fast dynamic finger tapping, diadochokinesia). The results revealed that 20 Hz tACS attenuated cortico-muscular coupling in the beta band during static isometric muscle contraction in PD patients. Moreover, the same patterned stimulation reduced amplitude variability during finger tapping. These finding were not present when the PD patients were stimulated with 10 Hz tACS, nor in a control group of healthy subjects. The authors concluded that the clinical improvement was possibly frequency specific because of the pathological motor-cortical synchronization in the beta band (i.e., 13–30 Hz).

Tremor at rest is an invalidating PD sign and typically responds weakly to pharmacological therapies. STN DBS efficacy on PD tremor is well known (Okun, 2012). In a recent tACS study, Brittain et al. (2013) investigated the potential of this technique to reduce tremor in a group of tremordominant PD patients by stimulating the motor cortex at tremor frequency. In their experiment, for each patient, the authors identified first the most effective phase lag of motor cortex tACS relative to the peripherally recorded rest tremor, showing that, among patients, the mean phase of peak tremor suppression is −139◦ . Following this first finding, the patients' motor cortex was then stimulated by closed-loop phase-locked tACS leading to a 50% tremor intensity reduction, likely through a mechanism of phase cancellation. This supports the notion that central oscillators play a key role in tremor pathophysiology, and represents promising targets for NIBS interventions. The same group investigated also the effectivity of tACS in modulating ''physiological tremor'' showing that while M1 tACS was able to alleviate postural tremor, it did not affect kinetic tremor in healthy subjects (Mehta et al., 2014).

#### Alzheimer's Disease

Several groups investigated the effects of NIBS on cognitive deficits, but none of them evaluated the NIBS effects on electrophysiological markers of the disease. In mild to severe AD patients, high-frequency (HF) rTMS (10 Hz or higher) over the prefrontal cortex induces a transient improvement in associative memory, and in naming of actions and objects, possibly due to the compensatory recruitment of contralateral prefrontal and bilateral posterior cortical regions (Cotelli et al., 2006, 2008; Solé-Padullés et al., 2006). In eight AD patients undergoing HF-rTMS (five sessions/week for six consecutive weeks followed by maintenance sessions) applied over Broca and Wernicke areas, right and left DLPFC, or right and left parietal somatosensory association cortex, a significant improvement in the ADAS-cog score was found both after 6 weeks and 4.5 months (Bentwich et al., 2011; Rabey et al., 2013). However, despite these initial promising rTMS results on cognitive AD deficits, and the emerging knowledge regarding the alteration of neuronal oscillatory activity in AD (see ''Alpha'' Section), no study has yet attempted to examine a possible link of cognitive improvements with restoration of oscillatory activity. A challenging open question is whether NIBS could be rendered more effective by optimizing rTMS parameters in order to restore the affected oscillatory properties at the stimulation site and functionally connected brain networks, which would require simultaneous TMS-EEG/MEG recordings.

#### Stroke

Oscillatory MEPs are a reliable marker of NIBS-induced neurophysiological cortical changes. For instance, iTBS over the damaged cortex in a monohemispheric stroke patient (same patient as reported in ''Sensory-Motor Cortex Stimulation'' Section) resulted in an increase in corticospinal activity evoked from the damaged cortex, as demonstrated by an amplitude increase of evoked I-waves and by the appearance of a further descending wave (Di Lazzaro et al., 2006). In contrast, the total corticospinal volley evoked from the unaffected motor cortex decreased in this patient by about 40%, demonstrating a decrease in the excitability of the corticospinal output of the hemisphere opposite to iTBS. Other studies have considered changes in spontaneous electrical brain activity that occur after a stroke in relation to influences of rehabilitation and NIBS protocols. A positive correlation has been found between better rehabilitation outcome and greater ERD after different rehabilitation programs (Altenmüller et al., 2009; Fujioka et al., 2012). A MEG study found a correlation between motor recovery and reduction in bilateral post-movement beta-ERS as well as in gamma-ERS of the AH during movement (Wilson et al., 2011). Besides the effects on power, reductions in interhemispheric coherence in the high beta and gamma bands have also been reported after a 12-week rehabilitation training (Pellegrino et al., 2012). It's worth noting that, so far, only two studies have carried out analysis on the modulation of oscillatory activity by rTMS/tDCS enriched rehabilitation in stroke (Krawinkel et al., 2015). In a single post-stroke aphasic patient, 3 weeks of high frequency rTMS over the AH improved naming and comprehension on a repetition tasks, while increasing functional theta and high beta connectivity between the damaged left inferior frontal gyrus and its controlateral homologous area (Dammekens et al., 2014). Furthermore, anodal tDCS over M1 of the AH, has been shown to enhance alpha-band ERD in chronic stroke patients (Kasashima et al., 2012). A recent meta-analysis suggested a beneficial tDCS effect in stroke rehabilitation (Kang et al., 2016), but the mechanisms behind this effect are not understood. In particular, to the best of our knowledge, a possible link between modulated brain rhythms and clinical improvements has not been explored yet.

#### Epilepsy

The incidence of epilepsy peaks in patients younger than 15 years and in those aged 65 or older (Forsgren et al., 2005). Given this epidemiological data, we included the epilepsy in the present review article, although the neuromodulation experiments reported mostly young adult epileptic patients.

Studies reporting rTMS application in epilepsy are encouraging but need to be interpreted cautiously given their often uncontrolled design (Lefaucheur et al., 2014). In the context of randomized, controlled trials, the antiepileptic effects of active rTMS varied widely from no beneficial effects (Theodore and Fisher, 2007) to significant clinical and electroencephalographic improvement (Fregni et al., 2006; Cantello et al., 2007; Sun et al., 2012). Therefore, availably data still do not provide conclusive evidence in favor or against the efficacy of this emerging treatment modality. Limitation in sample size and lack of a control condition could account for the heterogeneity of published results and for the difficulty of drawing definitive conclusions. Consequently, recommendations for the use of rTMS in epilepsy do not exceed Level C recommendation (''possible efficacy''), at least for focal low frequency rTMS of the epileptic focus.

tDCS has been applied in a few dozens of epileptic patients with the main objective of diminishing seizures and/or electroencephalographic epileptiform activity and evaluating the safety of the procedure (reviewed by San-juan et al., 2015). Fregni et al. (2006) conducted the first exploratory randomized sham controlled study of the effects of tDCS in refractory epileptic patients. Cathodal tDCS over the epileptogenic zone reduced the number of EEG epileptiform discharges and number of seizures immediately after and 15 and 30 days relative to the baseline. In a distinct group of pediatric patients, cathodal tDCS suppressed epileptiform discharges in most of the patients for 48 h, but the effect of a single session on EEG abnormalities was not sustained for 4 weeks. In both these studies, the seizure reduction rate was clinically negligible. Faria et al. (2012) reported that cathodal tDCS produces a reduction in epileptiform activity in pediatric patients with continuous spikes and wave syndrome during slow sleep (CSWS). However, this effect was not observed in an experiment conducted on the same syndrome by Varga et al. (2011). We very recently reported a group of ten adult patients with drug-resistant temporal lobe epilepsy showing a clinically significant amelioration of seizure frequency after ctDCS relative to sham tDCS (Assenza et al., 2017a).

Even if the safety of NIBS has been clearly documented in elderly people (Bikson et al., 2016), the vast majority of the above studies are on young and adult patients. Older patients with epilepsy (>65 years) are very rarely included in the trials protocols of rTMS and tDCS. Moreover, even when they were considered (Rotenberg et al., 2009; Morrell, 2011; Stefan et al., 2012; San-juan et al., 2015), these trials or case reviews did not separately analyze or extrapolate data regarding only elderly patients. Thus, conclusive data about specific or different NIBS effects on young vs. elderly groups of patients is missing, although to the best of our knowledge, results in elderly patients did not differ from those in young adults.

Although there is evidence that tDCS and rTMS reduces epileptiform activity, i.e., impacting on a pathophysiological marker of the disease, current protocols do not evoke suitably strong effects to reduce clinical seizures for application in the clinical context. However, cortical excitability is not only affected by direct stimulation of the cortex, but also by chronic peripheral stimulation of cranial nerves, in particular those with conspicuous and widely diffusing vegetative afferences. Invasive vagal nerve stimulation (iVNS) is an extra-cranial neurostimulation protocol that is FDA-approved as an add-on therapy in patients with drug-resistant partial-onset epilepsy. The stimulator is placed along the cervical branch of the left vagal nerve in the neck. Recently, transcutaneous vagal nerve stimulation (tVNS) has been proposed as a non-invasive alternative to iVNS. tVNS consists of stimulating the auricolar conchae in order to activate the auricular branch of the vagus nerve to then reach the brainstem nuclei, which are responsible for the antiepileptic effect of iVNS. A recent neurophysiology study from our group (Capone et al., 2015) demonstrated that, similarly to what has already been demonstrated for the iVNS (Di Lazzaro et al., 2004c), tVNS produces an increase of intracortical inhibition, namely short-interval-intracortical-inhibition (SICI) as evidenced by the paired-pulse TMS paradigm. Animal studies showed equivalent anti-seizure effects of iVNS and tVNS (He et al., 2013b). A recent human study reported an in vivo human evidence that tVNS and iVNS engage the same neural pathways (Assenza et al., 2017b). These results corroborate the hypothesis of a biological similarity between the effects of iVNS and tVNS and warrants further studied aimed at evaluating the clinical efficacy of tVNS in epilepsy. Pilot trials of tVNS in epileptic patients (Stefan et al., 2012; He et al., 2013a) reported an improvement of the number of seizure episodes in about 50% of the patients, but they do not allow conclusive results on efficacy so far because of the small number of recruited patients. Stefan et al. (2012) observed a reduction of seizure frequency and of epileptiform EEG discharges in five of the seven patients who underwent tVNS for 9 months. He et al. (2013a) administered bilateral tVNS for 24 weeks to 15 pediatricepileptic patients, and at the end of the observational period, observed a 50% or more reduction of seizures' frequency in six patient, with the greatest response registered in the first 2 months. In conclusion, tVNS is a promising technique to replicate iVNS results, but more studies are needed to assess clinical efficacy.

Overall, these studies indicate that NIBS in epilepsy is promising, given the solid evidence base for a potential clinical effectivity, but that it is not yet sufficiently refined to achieve clinically perceptible results. On the other hand, invasive direct cortical stimulation is undoubtedly effective in controlling seizures and epileptiform activity. Indeed, Jacobs et al. (2014) found a clear decrease of HFOs after electrical stimulation (with a diagnostic purpose) during surgical exploration in drug-resistant epileptic patients, and this reduction was not limited to the seizure onset zone. Furthermore, stimulation of the epileptic focus with varying pulse frequencies also led to a significant reduction of seizure activity (Morrell, 2011). This further adds to the arguments that there is a promise also in using transcranial electric or magnetic stimulation techniques to interact with epileptiform activity for clinical purposes. Further studies led by neurologist expert in both epilepsy and NIBS will be necessary to improve tDCS and rTMS efficacy in epilepsy.

#### Safety of NIBS Application

To date NIBS techniques are overall considered safe, as they have been used since their introduction without serious collateral effects. However, there are some general and specific points that need to be considered depending on NIBS protocol and the study population (e.g., healthy volunteers or patients). Safety issues on the use of rTMS and tDCS in adults and adolescents are reviewed extensively elsewhere (Rossi et al., 2009; Krishnan et al., 2015; Bikson et al., 2016; Palm et al., 2016) and are summarized below, as a full description of them is beyond the scope of this review article.

A major reason of concern that arises from interfering with brain oscillatory activity and that applies to all NIBS techniques is the possibility of increasing excitability and synchronizing neuronal discharge leading to epileptic activity. Although the risk of seizures is reported to be very low and applies to any protocol increasing cortical excitability (beyond transcranial brain stimulation), it must be taken into consideration especially for protocols delivering high intensity pulses at high frequency capable of inducing neuronal spikes, such as high frequency rTMS (Bae et al., 2007; Pereira et al., 2016). Other NIBS protocol inducing subthreshold depolarization such as tDCS or tACS are less of a concern.

Further frequent side effects common to all NIBS techniques are headache and syncope, usually due to psycho-physical discomfort. High intensity TMS has a risk for impairing hearing that can be mitigated by using earplugs. Heating/ burns must also be considered, especially when TMS is applied over EEG electrodes or with deep brain electrodes, which needs to be mitigated by assessing the compatibility of biomaterials with magnetic and electric fields and reducing the duration of stimulation. tDCS and tACS can provoke skin redness or tingling under the electrodes. Moreover, tACS is known to evoke phosphenes during stimulation that are believed to be of retinal origin (Antal and Paulus, 2013). Finally, contraindications to TMS include cardiac pacemakers or other objects or internal devices that are not compatible with magnetic fields. TMS is also best avoided during pregnancy as insufficient data are available on effects on the fetus. However, a recent review on mothers treated with rTMS for major depression during pregnancy reports no complications to the unborn children (Felipe and Ferrão, 2016).

#### CONCLUSIONS AND FUTURE DIRECTIONS

Over the last two decades, NIBS techniques have been under intense investigation as to their potential for clinical applications, led by the discovery of the influence of electromagnetic fields on human brain functions and their potential to induce cerebral plasticity. Unfortunately, progress has been limited despite the huge experimental efforts spent to identify NIBS applications that can be used during rehabilitation for additive treatment of motor and linguistic deficits in stroke, Parkinson disease, or for drug-resistant epilepsy, Alzheimer disease and psychiatric disorders. So far, the only FDA-approved clinical use of NIBS is for drug-resistant depression (by rTMS).

Taking into account the numerous clinical studies that have been conducted, the issues of low sample sizes and heterogeneity of protocols might account for the failure of some trials, but cannot explain the overall picture of a lack of proven evidence of NIBS effect in neurological disorders. One clue may reside in the fact that, in most of the clinical trials, the only biological neurophysiological variable targeted by neuromodulation is cortical excitability and its marker, i.e., the amplitude of MEP. This trend is likely explained by cortical excitability being an easy-measurable parameter that is susceptible to modification by NIBS. More recently, the question has been raised whether NIBS, which often uses oscillatory pulses or waveforms, can interact with a specific aspect of human brain activity, namely brain oscillations. Initial results are promising. They show that NIBS in its rhythmic form can entrain EEG activity, and in particular induce specific patterns of oscillatory activity by modulating the power of the rhythms residing in specific areas (e.g., alpha for posterior cortex, beta for motor cortex). Other ideas are based on optimizing the timing of TMS application by aligning it to periods of high excitable states as inferred from EEG measures (e.g., specific EEG phase or power values). Although the induction of a consistent and long-lasting clinical effect, even with multiple stimulating sessions, is far from been proven, the reviewed data illustrates that brain oscillations represent informative and traceable biological markers for pathophysiological conditions in different neurological disorders, and potentially effective targets for NIBS intervention. We find solid evidence for very specific alteration of cortical oscillatory activities in the aforementioned diseases, as well as promise from the few therapeutic trials that started to build on the modulation of these pathological oscillatory activities by NIBS. Research on extrapyramidal movement disorders and epilepsy in particular gathered exemplar evidence on how research on a biological marker of a disease can lead to therapeutic research about the disease itself. In epilepsy, it is foremost the fact that its main EEG biomarker (i.e., epileptiform activity) is also pathognomonic of the disease, which facilitates the design and documentation of interventions in NIBS trials. In addition, even if past clinical results of rTMS and tDCS experiments in epilepsy were overall of variable success, some studies reliably demonstrated the reduction of EEG epileptiform activity. In our opinion, these two factors make epilepsy research one of the most promising fields for working towards clinical NIBS applications using more tailored designs, e.g., considering the patients clinical (type, duration and frequency of seizures) and neurophysiological heterogeneity (position, orientation, frequency and strength of the dipole of the epileptic source). Moreover, in extrapyramidal movement disorders, the modulation of beta activity is a consolidated marker for the success of invasive brain stimulation in PD, and preliminary results of NIBS studies seem promising as to the ability of NIBS to modulate basal ganglia activity.

In conclusion, a large number of research groups are working intensively towards tailoring NIBS to induce long-lasting clinical effect. In parallel, new views on brain oscillations as pathophysiological biomarkers are emerging. These biomarkers vary between disorders, suggesting that pre-set NIBS intervention strategies are unlikely to be the panacea for each patient and disease. We conclude that research should focus on the core of the cerebral disease, i.e., the distortion of cortical activity and its pathophysiological meaning, and tailor NIBS to modulate this activity taking into account the functional neural reserve of the elderly patient and the severity of the disease.

### AUTHOR CONTRIBUTIONS

GA coordinated the manuscript writing among authors, reassembled the single parts and designed the structure of the review article. FC wrote the part on evoked MEP. LB wrote the part on Parkinson. FF wrote the part of EEG/TMS. GA wrote the part on the Alzheimer's disease. LF wrote the part on SEP. MM wrote the part of tremor. MP wrote the part on stroke. FR wrote the part on clinical cases. GS wrote the part on dystonia. MT wrote the part on epilepsy. GT and VDL scientifically supervised the work of the group and revised the english language. VDL designed this review article.

### REFERENCES


somatosensory cortex. Electroencephalogr. Clin. Neurophysiol. 100, 189–203. doi: 10.1016/0168-5597(95)00244-8


meta-analysis. J. Neurol. Neurosurg. Psychiatry 87, 345–355. doi: 10.1136/jnnp-2015-311242


connectivity in Alzheimer-type and multi-infarct dementia. Brain 115, 1543–1561. doi: 10.1093/brain/115.5.1543


improvements after robotic stroke rehabilitation. Restor. Neurol. Neurosci. 30, 497–510. doi: 10.3233/RNN-2012-120227


I.F.C.N. Committee. Clin. Neurophysiol. 126, 1071–1107. doi: 10.1016/j.clinph. 2015.02.001


perspectives. Prog. Neurobiol. 90, 418–438. doi: 10.1016/j.pneurobio. 2009.11.005


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Assenza, Capone, di Biase, Ferreri, Florio, Guerra, Marano, Paolucci, Ranieri, Salomone, Tombini, Thut and Di Lazzaro. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Corrigendum: Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation

Giovanni Assenza<sup>1</sup> , Fioravante Capone<sup>1</sup> , Lazzaro di Biase1, 2, Florinda Ferreri 1, 3 , Lucia Florio<sup>1</sup> , Andrea Guerra1, 2, Massimo Marano<sup>1</sup> , Matteo Paolucci <sup>1</sup> , Federico Ranieri <sup>1</sup> , Gaetano Salomone<sup>1</sup> , Mario Tombini <sup>1</sup> , Gregor Thut <sup>4</sup> and Vincenzo Di Lazzaro<sup>1</sup> \*

*<sup>1</sup> Clinical Neurology, Campus Bio-medico University of Rome, Rome, Italy, <sup>2</sup> Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom, <sup>3</sup> Department of Clinical Neurophysiology, Kuopio University Hospital, University of Eastern Finland, Kuopio, Finland, <sup>4</sup> Centre for Cognitive Neuroimaging, Institute of Neuroscience and Psychology, University of Glasgow, Glasgow, United Kingdom*

Keywords: neuromodulation, non-invasive brain stimulation, oscillations, TMS/tDCS, EEG

#### Edited and reviewed by:

*Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece*

> \*Correspondence: *Vincenzo Di Lazzaro v.dilazzaro@unicampus.it*

Received: *06 July 2017* Accepted: *14 July 2017* Published: *02 August 2017*

#### Citation:

*Assenza G, Capone F, di Biase L, Ferreri F, Florio L, Guerra A, Marano M, Paolucci M, Ranieri F, Salomone G, Tombini M, Thut G and Di Lazzaro V (2017) Corrigendum: Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation. Front. Aging Neurosci. 9:252. doi: 10.3389/fnagi.2017.00252* **Oscillatory Activities in Neurological Disorders of Elderly: Biomarkers to Target for Neuromodulation**

by Giovanni, A., Capone, F., di Biase, L., Ferreri, F., Florio, L., Guerra, A., et al. (2017). Front. Aging Neurosci. 9:189. doi: 10.3389/fnagi.2017.00189

An author name was incorrectly spelled as Giovanni A. The correct spelling is Giovanni Assenza; the correct surname (family name) is Assenza; the correct name is Giovanni; thus when abbreviated it should be cited as Assenza G.

The authors apologize for this error and state that this does not change the scientific conclusions of the article in any way.

The original article has been updated.

**A corrigendum on**

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Assenza, Capone, di Biase, Ferreri, Florio, Guerra, Marano, Paolucci, Ranieri, Salomone, Tombini, Thut and Di Lazzaro. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Inhibition of PirB Activity by TAT-PEP Improves Mouse Motor Ability and Cognitive Behavior

Ya-Jing Mi 1† , Hai Chen1,2† , Na Guo1† , Meng-Yi Sun<sup>3</sup> , Zhao-Hua Zhao<sup>1</sup> , Xing-Chun Gao<sup>1</sup> , Xiao-Long Wang<sup>1</sup> , Rui-San Zhang<sup>1</sup> , Jiang-Bing Zhou<sup>3</sup> \* and Xing-Chun Gou1,3 \*

1 Institute of Basic and Translational Medicine, and School of Basic Medical Sciences, and Shaanxi Key Laboratory of Brain Disorders, Xi'an Medical University, Xi'an, China, <sup>2</sup>Department of Anesthesiology, Tangdu Hospital, Fourth Military Medical University, Xi'an, China, <sup>3</sup>Department of Neurosurgery, School of Medicine, Yale University, New Haven, CT, United States

Paired immunoglobulin-like receptor B (PirB), a functional receptor for myelin-associated inhibitory proteins, plays an important role in axon regeneration in injured brains. However, its role in normal brain function with age has not been previously investigated. Therefore in this study, we examined the expression level of PirB in the cerebral cortex, hippocampus and cerebellum of mice at 1 month, 3 months and 18 months of age. The results showed that the expression of PirB increased with age. We further demonstrated that overexpression of PirB inhibited neurite outgrowth in PC12 cells, and this inhibitory activity of PirB could be reversed by TAT-PEP, which is a recombinant soluble PirB ectodomain fused with TAT domain for blood-brain barrier penetration. In vivo study, intraperitoneal administration of TAT-PEP was capable of enhancing motor capacity and spatial learning and memory in mice, which appeared to be mediated through regulation of brain-derived neurotrophic factor (BDNF) secretion. Our study suggests that PirB is associated with aging and TAT-PEP may be a promising therapeutic agent for modulation of age-related motor and cognitive dysfunctions.

Keywords: PirB, motor capacity, cognitive behavior, TAT-PEP, BDNF

### INTRODUCTION

Paired immunoglobulin-like receptor B (PirB) is a functional receptor for myelin-associated inhibitory proteins, including Nogo, myelin-associated glycoprotein (MAG) and oligodendrocytemyelin glycoprotein (OMgp), which inhibit axonal regeneration and functional recovery after brain injury (Gou et al., 2014). The expression level of PirB in neurons has been shown to increase after neurological injuries, including spinal cord injuries (Zhou et al., 2010), hypoxic-ischemic brain damage (Adelson et al., 2012; Wang et al., 2012; Guo et al., 2013), encephalitis (Deng et al., 2012), hippocampal aging (VanGuilder Starkey et al., 2012) and retinopathy (Cai et al., 2012). Blocking PirB activity through antibody antagonism or genetic approaches allowed the promotion of axon regeneration and synapse plasticity (Adelson et al., 2012; Wang et al., 2012; Kim et al., 2013; Bochner et al., 2014).

Recent emerging evidence further suggests that PirB may regulate cognitive functions in addition to neurite regeneration. It was found that the expression of PirB in the cortex and hippocampus was significantly upregulated in rats with memory deficits induced by lipopolysaccharide (Deng et al., 2012). The expression of PirB was also found to gradually increase in the CA1 and DG subregions of the hippocampus in aging mice with cognitive behavior

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Daniel Ortuño-Sahagún, Universidad de Guadalajara, Mexico Linda Ann Bean, Rush University Medical Center, United States

#### \*Correspondence:

Jiang-Bing Zhou jiangbing.zhou@yale.edu Xing-Chun Gou gouxingchun@189.cn

†These authors have contributed equally to this work.

> Received: 09 March 2017 Accepted: 02 June 2017 Published: 20 June 2017

#### Citation:

Mi Y-J, Chen H, Guo N, Sun M-Y, Zhao Z-H, Gao X-C, Wang X-L, Zhang R-S, Zhou J-B and Gou X-C (2017) Inhibition of PirB Activity by TAT-PEP Improves Mouse Motor Ability and Cognitive Behavior. Front. Aging Neurosci. 9:199. doi: 10.3389/fnagi.2017.00199 deficits (VanGuilder Starkey et al., 2012). Blocking PirB receptor could enhance both the short-term and long-term cognitive functions after bilateral common carotid artery occlusion in mice (Deng et al., 2016; Li et al., 2017). However, neither the association between PirB and cognition in normal brain development nor whether PirB can be used as a pharmacological target for modulating cognitive functions has been previously studied.

In this study, we examined the pattern of PirB expression in the brain of mice at different developmental stages. We found that the expression of PirB in the cerebral cortex, cerebellum and hippocampus increased with age. Then, we explored whether PirB could be targeted to improve age-related cognitive dysfunctions by using TAT-PEP, a recombinant protein consisting of soluble PirB ectodomain fused with TAT. TAT is a cell penetration peptide derived from the TAT-protein in the human immunodeficiency virus and was previously successfully used for facilitating drug delivery to the brain by us and others (Aarts et al., 2002; Wang et al., 2008). We demonstrated that treatment with TAT-PEP significantly increased the length of axons in neurons in vitro and improved exhaustive swimming capacity, spatial learning and memory in mice. Our study suggests that PirB plays an important role in aging and is a promising target for pharmacological modulation of cognitive function.

### MATERIALS AND METHODS

#### Animals

All animal procedures were approved by the Animal Care and Ethical Committee at Xi'an Medical University (Permit Number: 2012-8, 7 March 2012). Male C57BL/6 mice of 1 month, 3 months and 18 months were supplied by the Experimental Animal Center of Xi'an Jiao Tong University. Efforts were made to reduce the number of animals used in the study by following the 3Rs (reduction, refinement and replacement). The total number of mice used in our experiments was 180.

### Construction, Expression and Purification of TAT-PEP

TAT-PEP (PirB extracellular peptide) construction, expression and purification were carried out as previously reported (Deng et al., 2016). Briefly, cDNA of extracellular domain of PirB was synthesized and cloned into expression vector pTAT-HA-6xHis. For protein expression, pTAT-PEP was transformed into BL21 (DE3). Protein production was induced with 100 mM isopropyl β-D-1-thiogalactopyranoside (TaKaRa, Tokyo, Japan), and purified by Ni-NTA-agarose chromatography (Merck, Darmstadt, Germany). The size and purity were confirmed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). As a control peptide, the scrambled PEP fusion protein named as TAT-mPEP was expressed and purified according to the same procedure used for TAT-PEP (Deng et al., 2016).

### Overexpression PirB in PC12 Cells

PC12 cell lines derived from rat adrenal gland pheochromocytoma were cultured in Dulbecco's Modified Eagle's Medium (DMEM) with 10% horse serum, 5% FBS and 1% penicyline/streptomycine at 37◦C and 5% CO2. Cells with overexpressed PirB were obtained through lentiviral transduction of mouse PirB, which was constructed by HANBIO company (Shanghai, China), and then selected with puromycin (8 µg/ml) according to our recently published procedures (Chen et al., 2015). These cells were named PC12PirB cells.

### Determination of the Length of Axons

PC12 cells and PC12PirB cells were plated on glass coverslips, and one half of PC12PirB cells were treated with 150 µg/L TAT-PEP. After 24 h, cells were stained with the anti-β-tubulin antibody (Catalog#: MA5-16308, 1:200, Thermo Scientific) and then the Alexa-594-labeled donkey anti-mouse IgG secondary antibody (1:800, Thermo Scientific). Cells were imaged using a fluorescence microscope (OLYMPUS IX73). The length of an axon was determined as the linear distance from the point of exit to the end of the longest branch of the neurite according to a previous report (Richardson et al., 2007). In every experimental group, 150–200 cells were analyzed.

### Design of Animal Studies

A scheme for animal studies was shown in **Figure 4A**. Mice at the selected ages were randomly divided into two groups,

with 12 mice in each group. One group received treatment of TAT-PEP in saline, another group received TAT-mPEP in saline as control. Treatments were carried out through intraperitoneal administration of TAT-PEP or TAT-mPEP at 8 mg/kg/injection, twice a day, for 30 or 60 days, as indicated in **Figure 4A**. Behavior training and evaluation were performed according to the time points indicated in **Figure 4A**.

### Morris Water Maze

Water maze tests were performed as previously described (Torres et al., 2015). The diameter and the depth of the circular pool were 90 cm and 50 cm, respectively. The inner wall was carefully cleaned to eliminate any local cues. The temperature of the room and water was kept at 23 ± 1 ◦C, and the pool was filled with water to a depth of 40 cm and rendered opaque by the addition of milk to hide the escape platform. The plexiglas platform of 9 cm in diameter was placed in the pool 1 cm below the water surface. The recording camera was connected to a digital tracking device, and the water-maze software was used to process the tracking information. The platform position remained constant throughout the training period.

In the training period, each animal performed four trials per session, one session per day, for 4 days. The intertrial interval (ITI) was 1 min. During each session, the mouse started from a random position in the pool, and the maximum time allowed for swimming and searching was 60 s. Mice were allowed to remain on the platform for 10 s after they had found it. The subjects that failed to find the platform were gently guided to it and placed on the platform for 10 s. The average distance traveled to the platform per day was recorded. On the fifth day, probe test was performed with the platform

removed from the pool. Mice were allowed to swim for 60 s. The distance traveled and the time spent in target quadrant were recorded, and the percentages were calculated from the recorded data.

TAT-PEP protein provided as the positive control.

P121-P124 (3 months), P571-P574 (the first 18 months group) and P621-P624 (the second 18 months group). Probe tests (<sup>∗</sup> ) were performed on the next day after the training. Swiming-to-exhaustion test (\$) was performed at P66 for 1 month mice, P126 for 3 months mice, P576 for the first 18 months group mice and P626 for the second 18 months group mice. Mice in the second 18 months group were sacrificed at P627 for ELISA analysis of brain-derived neurotrophic factor (BDNF). (B) Evaluation of exhaustive swimming capacity was performed in the next day of Probe test. Mice were forced to swim until they were exhausted. The length of swimming time was recorded. (C–E) The distance traveled to the platform of different age mouse was recorded in the training of the spatial reference memory task in the Morris water maze. (F,G) The spatial learning and memory were measured by Probe test using water maze. Percentage of distance in the target quadrant is relative to the total distance in the water maze (F) in indicated groups, and percentage of time in the target quadrant is relative to the total time in the water maze (G). During the above experiments (B–G), there was no significant difference between the 18 months group received TAT-PEP and TAT-mPEP for 30 days. Data was not shown. <sup>∗</sup>p < 0.05, ∗∗p < 0.01 for mice with TAT-PEP treatment compared with TAT-mPEP treatment.

### Swimming-to-Exhaustion Test

On the next day after the probe test, a weight (5% body weight) was attached to the tail of each mouse for the swim-to-exhaustion test (You et al., 2015). The swimming exercise was carried out in a tank filled with water in 38 cm depth and at 34 ± 1 ◦C. Mice were defined to be exhausted when they failed to rise to the surface of the water to breathe within 5–7 s period. Swimming time was recorded for each mouse.

### Quantitative Real-Time PCR (qPCR)

Total RNA was isolated using the Trizol reagent (Invitrogen), and treated with RNase-free DNase I (Roche) to remove residual DNA. cDNA was obtained using ReverTra Ace quantitative real-time PCR (qPCR) RT Kit (TOYOBO). qPCR was carried out using ABI Stepone plus and Realtime PCR Master Mix (SYBR Green; TOYOBO). Primers for mouse PirB were: 5<sup>0</sup> -TACAAGGAAGTACCACGCCC-3<sup>0</sup> (forward) and 5<sup>0</sup> -GGTTCAGCCTTGATGGTTGG-3<sup>0</sup> (reverse). GAPDH was used as the endogenous control.

#### Western Blot

Proteins were prepared using RIPA lysis buffer containing protease and phosphatase inhibitors (Beyotime Biotechnology). Protein concentrations were determined using BCA protein assay (Thermo Scientific). Proteins were separated using 8% SDS-PAGE and transferred to the polyvinylidene fluoride membranes (Millipore). The membranes were blocked for 1 h, followed by incubation with the primary antibodies overnight, and then incubated with the HRP-conjugated anti-mouse antibody afterwards (1:10,000, Roche). The primary antibodies were anti-6xHis antibody (1:1000, Abcam) and anti-PirB (1:500, Thermo Scientific). GAPDH was used as loading control and determined using the anti-GAPDH antibody (1:2000, Santa Cruz). Blots were detected using BM Chemiluminescence Western Blotting kit (Roche). Densitometry quantification was analyzed using IPP6.0 software.

#### Flow Cytometry

The cell surface expression level of PirB in transduced cells was determined using flow cytometry. Briefly, the cells were centrifuged and washed with phosphate buffered saline (PBS), followed by incubation with the primary antibody against PirB (1:200, Thermo Scientific) for 1 h on ice. After washing 3 times with PBS, the cells were treated with a 1:50 dilution of FITC-conjugated secondary antibody (1:800, Thermo Scientific). The cells were then washed 3 times with PBS and analyzed using a Accuri C6 flow cytometer (BD Biosciences), and data analysis was performed with Flowjo software.

#### ELISA Assay

BDNF concentration was determined using the BDNF Emax ImmunoAssay (Promega, Madison, WI, USA) according to the manufacturer's instructions. First, a 96-well microplate was sealed, incubated with anti-BDNF antibody (1:1000) overnight at 4◦C, and washed with Tris-buffered saline (TBS). On the next day, the plate was blocked at RT for 1 h and washed. Samples and BDNF standards were added into the plate and incubated at RT for 2 h. After an extensive wash, the anti-BDNF antibody (1:500) was added to each well and the plate was incubated at RT for 2 h. After an additional wash, the HRP-conjugated anti-IgY (1:200) was added. One hour later, the plate was washed and incubated with TMB One solution for 10 min at RT. The reaction was ended by adding 1 M HCl. The absorbance was measured at 450 nm. The same procedures were used to determine the level of nerve growth factor (NGF).

#### Data Analysis

The data is expressed as the mean ± SD and analyzed with SPSS 13.0 statistical software. The data in **Figures 4C–E** was analyzed using two-way ANOVA with Bonferroni post-test analysis, with the treatments as the between-subject factor and the trial days as the within-subject factors. The data in **Figures 1B**, **2C** were analyzed using One-way ANOVA with Dunnett's test. The data in **Figures 2B**, **4B,F,G**, **5** were analyzed using Student's t-test.

### RESULTS

### Expression of Endogenous PirB

To determine the temporal and spatial expression pattern of endogenous PirB in the brain, we harvested the cerebral cortex, cerebellum and hippocampus from mice at the age of 1 month, 3 months and 18 months. The expression of PirB in different regions was determined by qPCR and western blot. Results in **Figure 1** showed that the expression of PirB increased gradually with age in all compartments at both the mRNA and protein levels, suggesting that PirB is associated with aging.

### Over-Expression of PirB Inhibits Axon Outgrowth, Which Can be Reversed by TAT-PEP Treatment

We studied the impact of PirB on axon outgrowth using PirB-overexpressed PC12 cells, PC12PirB, which were generated through lentiviral transduction. The results from the flow cytometry showed that the expression level of PirB in PC12PirB cells was 2.4-fold of that in PC12 cells (**Figures 2A,B**). As shown in **Figure 2C**, overexpression of PirB significantly inhibits axon outgrowth. The average length of an axon in PC12PirB cells is 7 µm, compared to 53 µm in control PC12 cells. This result is consistent with previous reports that PirB inhibits axon outgrowth (Gou et al., 2014; Liu et al., 2015).

Next, we investigated whether blocking PirB activity could reverse the axon outgrowth inhibition caused by PirB overexpression. TAT-PEP, a recombinant protein consisting of PirB extracellular motif fused with TAT, which demonstrated the ability to rescue neurite outgrowth inhibition induced by Nogo, MAG and OMgp in stroke (Deng et al., 2016), was used to treat PC12PirB. Results in **Figure 2C** showed that treatment with TAT-PEP effectively prolonged the growth of axon.

### TAT-PEP Penetrates the Brain

In our previous work, we demonstrated that the TAT peptide mediated efficient delivery of a 40 aa peptide, NEP1-40, to the brain (Han et al., 2016). To evaluate whether the conjugation of TAT enhances brain penetration of PEP, we injected TAT-PEP intraperitoneally into mice. At 0.5 h, 2 h, 12 h, 24 h, 48 h and 72 h after injection, the hippocampus was collected and processed. The presence of TAT-PEP was detected by western blot. Results in **Figure 3** showed that TAT-PEP was detectable at 0.5 h after administration. The concentration of TAT-PEP in the hippocampus peaked at 12 h. The protein sample of the first lane was TAT-PEP synthesized in vitro, and here it was used as a positive control.

### Intraperitoneal Administration of TAT-PEP Enhances Mouse Motor Capacity, Spatial Learning and Memory

Motor deficit is one of the major results of aging. As the PirB expression is associated with aging, we set to study whether suppression of PirB activity by TAT-PEP could enhance mouse motor capacity. Based on the pharmacokinetic result described in **Figure 3**, TAT-PEP or TAT-mPEP was administered in mice every 12 h. Treatment lasted for 30 days for 1 month, 3 months and 18 months groups. Exhaustive swimming capacities in the three groups of mice were examined (**Figure 4A**). Compared to TAT-mPEP treatment, the time of swimming significantly increased in both 1 month and 3 months groups mice that received the 30-day TAT-PEP treatment (**Figure 4B**). With the same treatment regimen, the difference in the 18 months group between mice with TAT-PEP and TAT-mPEP treatments was insignificant (data not shown). However, the difference reached significance when the 60-day treatment was carried out (**Figure 4B**).

The spatial learning ability and memory in mice was examined using the Morris water maze evaluation method. **Figures 4C–E** showed the performance of both TAT-PEP and TAT-mPEP treated animals on the spatial reference memory tests. The group of 1 month mice injected with TAT-PEP exhibited no significant difference in the acquisition of the spatial reference memory task, as revealed by distance traveled to the platform in **Figure 4C**. For 3 months and 18 months groups, the TAT-PEP treatment notably increased the acquisition of the spatial reference memory task (**Figures 4D,E**). **Figures 4F,G** showed that treatment with TAT-PEP significantly improved the percentage of time and distance traveled in the target quadrant for 3 months and 18 months groups. Similarly, the data for 18 months groups in the Morris water maze evaluation was presented as a 60-day treatment of TAT-PEP, because no significant difference was detected for the 30-day treatment (data not shown).

### BDNF Expression but Not NGF Is Upregulated in TAT-PEP Treated Mice

We explored the potential mechanism for TAT-PEP-mediated motor and cognitive behavior enhancement by analyzing the expression of brain-derived neurotrophic factor (BDNF) and NGF in the brain of 18 months mice. Both BDNF and NGF are known to play major roles in promoting neuronal survival, neurite growth and cognitive ability (Lu et al., 2014; Lin et al., 2015). At the end of the behavioral test, the cerebral cortex, cerebellum and hippocampus were harvested and subjected to an ELISA assay. Results in **Figure 5** showed that the 60-day TAT-PEP treatment significantly enhanced the level of BDNF in all three sub-regions. By contrast, no significant difference in NGF expression was seen between mice with TAT-PEP and TAT-mPEP treatment. These results suggest that the motor and cognitive behavior enhancement effect of TAT-PEP is likely mediated by BDNF.

### DISCUSSION

As a functional receptor for the myelin inhibitors of axonal regeneration, PirB has been previously characterized in the neurological system with injuries, including spinal cord injuries (Zhou et al., 2010), optic nerve injuries (Cai et al., 2012), hypoxic-ischemic damage (Wang et al., 2012), stroke and lipopolysaccharide-induced chronic neuroinflammation (Deng et al., 2012). Across all these disease conditions, the expression level of PirB was found to be significantly elevated. In this study, we examined both the mRNA and protein levels of PirB in young (1 month), adult (3 months) and aged (18 months) mice, and found that the expression of PirB in the cerebral cortex, cerebellum and hippocampus increased with brain development (**Figure 1**). Our results are consistent with a previous report by VanGuilder Starkey et al. (2012) in which they found that both the mRNA and protein levels of PirB were upregulated in subregions of hippocampus in Fischer 344–Brown Norway rats with advanced aging.

The association of PirB with motor and cognitive functions has been previously reported. In a mouse model of stroke, it was found that knockout of the PirB gene enhanced corticospinal projections and motor recovery (Adelson et al., 2012). However, similar findings were not seen in mouse models of cortical injury (Omoto et al., 2010) nor spinal cord injury (Nakamura et al., 2011). Genetic deletion of PirB leads to activation of alternative

The extracellular segment of PirB consists of six extracellular Ig-like domains, D1 to D6 from the N- to the C-terminus (Takai, 2005). Among them, D1-D2 and D3-D6 have high affinities with MHCI and Nogo-66, respectively (Matsushita et al., 2011). It was previously reported that, during the process of age-related hippocampal changes, which were accompanied with cognitive decline, the expression levels of MHCI and PirB were elevated, suggesting that D1-D2 of PirB might promote cognition deficits through interaction with MHCI (VanGuilder Starkey et al., 2012). By contrast, Nogo-66 is correlated with motor behavior, thus we speculate that D3-D6 domain mainly mediates PirB-related motor deficits (Fouad et al., 2001). In this study, TAT-PEP was designed to include all of the six domains. Therefore, the effects of TAT-PEP treatment observed in this study is likely due to the interaction of TAT-PEP with both MHCI and Nogo-66.

We found that the therapeutic benefit of TAT-PEP treatment might be mediated by BDNF. BDNF is known to play an important role in the regulation of learning and memory (Komulainen et al., 2008; Cowansage et al., 2010; Lu et al., 2014; Tong et al., 2015), and can be used as a therapeutic agent to alleviate cognitive impairment (Wu et al., 2015). Recently, Raiker et al. (2010) reported that myelin inhibitors antagonize BDNF-induced signaling cascades through attenuation of Erk1/2 activation. This result suggests that myelin inhibitors

#### REFERENCES


and their receptors, such as PirB, may coordinate structural and functional neuronal plasticity in CNS health and disease through regulation of BDNF signaling, and thus, could be targeted for improvement of neurological functions. This hypothesis is well supported by the findings described in this study.

In summary, we found that PirB is correlated with age and might be a promising molecular target for modulation of motor and cognitive dysfunctions. Our results suggest TAT-PEP as a promising therapeutic agent in the future. Of course, TAT-PEP only antagonizes the extracellular segments of PirB and partly inhibits the signaling pathway of axon regeneration. Besides PirB, there are still other receptors such as NgR, integrin and Ephrin 4A that exert similar functions. Therefore, combined antagonists targeting all the receptors will be a more expected treatment strategy.

### AUTHOR CONTRIBUTIONS

X-CGou and J-BZ designed the experiments. Y-JM, NG, Z-HZ, X-CGao, R-SZ and M-YS performed the experiments. HC and X-LW did the data analyses. Y-JM, X-CGou and J-BZ wrote the article, with the help of the co-authors.

### FUNDING

This work was supported by NSFC grant, 81471415, 31400913, 81402063 and 81271290. Shaanxi Scientific Research Program fund 16JK1655 and 16JK1645, Shaanxi Natural Science Basic Research Plan fund 2016JQ8022, NIH Grants NS095817 and NS095147, and the Leading Disciplines Development Government Foundation of Shaanxi. University's Key Disciplines of Molecular Immunology.

deficit in rats. Neuroscience 209, 161–170. doi: 10.1016/j.neuroscience.2012. 02.022


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Mi, Chen, Guo, Sun, Zhao, Gao, Wang, Zhang, Zhou and Gou. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Age-Related Decline in the Variation of Dynamic Functional Connectivity: A Resting State Analysis

Yuanyuan Chen1,2 , Weiwei Wang2,3 , Xin Zhao2,3\*, Miao Sha2,3 , Ya'nan Liu2,3 , Xiong Zhang2,3 , Jianguo Ma<sup>1</sup> , Hongyan Ni <sup>4</sup> and Dong Ming2,3 \*

<sup>1</sup>College of Microelectronics, Tianjin University, Tianjin, China, <sup>2</sup>Tianjin International Joint Research Center for Neural Engineering, Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin, China, <sup>3</sup>Department of Biomedical Engineering, College of Precision Instruments and Optoelectronics Engineering, Tianjin University, Tianjin, China, <sup>4</sup>Department of Radiology, Tianjin First Center Hospital, Tianjin, China

Normal aging is typically characterized by abnormal resting-state functional connectivity (FC), including decreasing connectivity within networks and increasing connectivity between networks, under the assumption that the FC over the scan time was stationary. In fact, the resting-state FC has been shown in recent years to vary over time even within minutes, thus showing the great potential of intrinsic interactions and organization of the brain. In this article, we assumed that the dynamic FC consisted of an intrinsic dynamic balance in the resting brain and was altered with increasing age. Two groups of individuals (N = 36, ages 20–25 for the young group; N = 32, ages 60–85 for the senior group) were recruited from the public data of the Nathan Kline Institute. Phase randomization was first used to examine the reliability of the dynamic FC. Next, the variation in the dynamic FC and the energy ratio of the dynamic FC fluctuations within a higher frequency band were calculated and further checked for differences between groups by non-parametric permutation tests. The results robustly showed modularization of the dynamic FC variation, which declined with aging; moreover, the FC variation of the inter-network connections, which mainly consisted of the frontal-parietal network-associated and occipital-associated connections, decreased. In addition, a higher energy ratio in the higher FC fluctuation frequency band was observed in the senior group, which indicated the frequency interactions in the FC fluctuations. These results highly supported the basis of abnormality and compensation in the aging brain and might provide new insights into both aging and relevant compensatory mechanisms.

Keywords: aging, dynamic functional connectivity, functional connectivity variation, functional connectivity fluctuation frequency, resting-stated fMRI

#### INTRODUCTION

Normal aging in the human brain refers to degradation phenomena that occur in brain structures, brain function and brain morphology with increasing age, indicating that a certain degree of senior brain dysfunction will occur (Hedden and Gabrieli, 2004; Fjell et al., 2014). Considering the increasing size of the aging population, the incidences of diseases that are highly associated with age, such as Alzheimer's (Kern and Behl, 2009; Mosher and Wyss-Coray, 2014) and Parkinson's (Xu et al., 2012; Reeve et al., 2014), are also increasing. Until now, the mechanism of aging has remained unclear, and further investigation of brain aging could greatly help in managing problems associated with both aging and disease.

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Ana Maria Buga, University of Medicine and Pharmacy of Craiova, Romania Charis Styliadis, Aristotle University of Thessaloniki, Greece

#### \*Correspondence:

Xin Zhao zhaoxin@tju.edu.cn Dong Ming richardming@tju.edu.cn

Received: 01 December 2016 Accepted: 06 June 2017 Published: 30 June 2017

#### Citation:

Chen Y, Wang W, Zhao X, Sha M, Liu Y, Zhang X, Ma J, Ni H and Ming D (2017) Age-Related Decline in the Variation of Dynamic Functional Connectivity: A Resting State Analysis. Front. Aging Neurosci. 9:203. doi: 10.3389/fnagi.2017.00203

Functional connectivity (FC) based measures of the resting state functional magnetic resonance imaging (rs-fMRI), which reflect the coherence between temporal fluctuations across brain regions, are organized into distinct systems or networks (Damoiseaux et al., 2008; Zuo et al., 2010). Most studies of FC have focused on the decline of specific functional systems, such as the default mode network (DMN; Raichle et al., 2001; Damoiseaux et al., 2008; Wang et al., 2010; Ferreira and Busatto, 2013), or have focussed on other specific brain networks or regions, such as the language system (Zou et al., 2012), subcortical regions (Yi et al., 2015) or the motor system (Coynel et al., 2010; De Vico Fallani et al., 2013). Increasing evidence has shown that the decline in cognitive function associated with aging is related to changes in communication between different brain regions and subsystems (Andrews-Hanna et al., 2007; Sambataro et al., 2010), even in the resting state (Shehzad et al., 2009; Meindl et al., 2010; Guo et al., 2012; Zuo et al., 2013). Despite this progress, how brain systems cooperate to handle aging-associated declines remains unclear, especially considering the averaging of complex spatiotemporal phenomena during a period of time (Hutchison et al., 2013a).

Traditionally, functional connectivities derived from fMRI data are computed using signals across the entire scan time; it is assumed that the functional connectivities among the brain regions are static during the duration of the resting time (Handwerker et al., 2012; Zuo et al., 2013). However, recent work has shown that FC is temporally dynamic (Chang and Glover, 2010; Calhoun et al., 2014) even at rest. This dynamic FC, which varies over a timeframe of seconds, could be highly related to unconstrained mental activity during the resting state (Hutchison et al., 2013a; Allen et al., 2014; Zalesky and Breakspear, 2015) and even under anesthesia (Hutchison et al., 2013b). A widely applied method for analyzing temporal dynamics is the sliding window correlation method (Sakolu et al., 2010; Hutchison et al., 2013a; Di and Biswal, 2015). A series of FC matrices was obtained using this method, which showed the time-varying connectivity network. Thus, researchers have started to perform dynamic FC investigations of mild cognition impairment (Wee et al., 2016), epilepsy (Liu et al., 2016), schizophrenia (Rashid et al., 2014; Du et al., 2016), major depressive disorder (Demirta et al., 2016) and normal development (Sakolu et al., 2010; Rashid et al., 2014; Qin et al., 2015).

The convergent results of previous studies have suggested that the dynamic resting state FC is highly intrinsic and physiologically relevant. Several studies have reported that the FC states revealed by changes in connectivity over the course of the scan can be sensitive to changes related to neurological disorders (Sakolu et al., 2010; Li et al., 2014; Leonardi and Van De Ville, 2015; Ou et al., 2015; Shakil et al., 2016). Increasing efforts have been directed toward using functional microstates and their transmissions to depict the working mechanisms of the brain (Allen et al., 2014; Shakil et al., 2016). Microstate transmissions could be the bases of integration and segregation between different brain networks or cognitive resources (Hansen et al., 2015; Yu et al., 2015; Shakil et al., 2016). Previous work has shown that aging impacts not only within-network connectivity but also the integration and segregation of different brain networks (Ferreira and Busatto, 2013). Advancing age induces increased reorganization to establish compensatory mechanisms or plasticity that counteract the aging process (Meunier et al., 2014; Sala-Llonch et al., 2015; Sugiura, 2016). Segregation and integration are the bases of reorganization of the brain connectivity network. The investigation of dynamic FC in the resting state may provide new insights into communication between various cognitive resource pools in the aging brain. The derived patterns of temporal variation in FC thus reflect the interactions of the brain functional networks and are therefore expected to facilitate our understanding of the mechanisms that underlie mental diseases.

We expected that the fluctuations of resting FC comprise a dynamic balance that maintains the intrinsic connectivity patterns in the brain. The dynamic balance of FC allows us to capture the interactions between all of the subsystems and the basic states of brain connectivity. Since the FC levels and patterns are age-related, this dynamic balance and the serial connectivity networks must also change with increasing age. A previous study (Leonardi and Van De Ville, 2015) suggested that the spontaneous fluctuations in the FC have frequency dependence and result from the interactions of various frequency components associated with neural activities. Limited resources and decreased processing speed can indicate performance during aging; thus, we hypothesized that the dynamic FC can provide clues for the capacity and efficiency of the connectivity states that transfer and present aging features. This article focuses on revealing the effects of aging on the time-varying FC of the brain in the resting state. Using the sliding window correlation method, the resting state fMRI data from two groups of young and senior healthy individuals were processed to construct dynamic FC matrices. We expected that the variation and the frequency spectrum of the FC fluctuations were the important bases of the dynamic balance and were highly related to aging.

## MATERIALS AND METHODS

### Participants and fMRI Data Acquisition

All resting-state fMRI data used in this study were obtained from the NKI-Rockland Sample (NKI-RS<sup>1</sup> ), which is provided by the Nathan Kline Institute (NKI, Orangeburg, NY, USA) and is available online in a public database. To study the changes in the dynamic characteristics of FC that resulted from normal brain aging, we collected fMRI data from 68 healthy subjects who were organized into two groups: 36 young subjects were assigned to one group (mean age, 28.1 years; range, 20–35 years; 24 male), and 32 senior subjects were assigned to the other group (mean age, 70.6 years; range, 60–85 years; 15 male). According to the demographic information provided by the NKI-RS data set, there was a remarkable difference in the age of the participants, but no significant differences in either gender or hand dominance between the two groups.

<sup>1</sup>http://fcon\_1000.projects.nitrc.org/indi/pro/nki.html

Resting-state fMRI data were collected using an echo-planar imaging (EPI) sequence on a 3.0 T SIMENS Trio scanner. The scanning parameter settings were as follows: TR/TE = 2500/30 ms, flip angle (FA) = 80◦ , field of view (FOV) = 216 × 216 mm<sup>2</sup> , voxel size = 3 × 3 × 3 mm<sup>3</sup> , number of slices = 38, scan time = 650 s time points = 260. During the data acquisition, the subjects were instructed to keep their eyes closed and to stay awake. High-resolution T1-weighted images were also acquired using the magnetization-prepared rapid gradient echo (MPRAGE) sequence. The acquisition parameter settings were as follows: TR/TE = 2500/3.5 ms, FA = 8◦ , FOV = 256 × 256 mm<sup>2</sup> , voxel size = 1 × 1 × 1 mm<sup>3</sup> , slice = 192.

#### Data Preprocessing

Functional images were preprocessed using the Connectome Computation System (CCS<sup>2</sup> ). The CCS designed by Zuo et al. (2013) provides a computational platform for multimodal neuroimaging brain connectomics computations by integrating the functionalities of AFNI, FSL and FreeSurfer (Zuo et al., 2013; Betzel et al., 2014; Cao et al., 2014). The functional preprocessing included the following: the first ten functional volumes were discarded to allow for signal equilibration; slice timing was corrected using the middle slice as the reference frame; 3D geometrical displacement was used to correct for head motion; and 4D global mean-based intensity correction was performed. In addition, the Friston-24 model was used to remove micro-level motion artifacts (Friston et al., 1996) and nuisance regressors; for instance, the individual white matter and the cerebrospinal fluid (CSF) mean signals were regressed out. The functional data were also temporal band-pass filtered (0.01–0.1 Hz) and detrended (both linear and quadratic trends). Finally, spatial smoothing was performed with a Gaussian filter kernel (FWHM = 6 mm). The structural processing steps were as follows: the image noise was removed using a spatially adaptive non-local means filter and brain surface reconstruction; the individual functional space was spatially normalized to the MNI152 standard brain space; a customized group T1 template in the standard space was generated to reduce the error term that resulted from the image registration and bias in the template selection; and the fMRI images in the native space of each subject was registered to the standard space with a final resolution of 3 mm.

### Dynamic Functional Connectivity Network Construction

Because time-varying FC is complicated and differs from static FC, the recruitment of more regions in the associated networks could help to provide more precise information. A total of 142 regions that covered the cingulo-opercular network (CON), DMN, fronto-parietal network (FPN), occipital network (OCC) and sensorimotor network (SMN) as defined by Dosenbach et al. (2010) were selected. The cerebellum network was neglected because we sought to examine only the effects of aging on the higher-order brain network interactions and the dynamics of brain cognition and perception. Among these networks, OCC and SMN are involved in perception and primary visual and motion processing, respectively; the other three networks are important to higher-order cognitive functions. In each of these brain regions, time courses were extracted and averaged over a spherical region of interest (ROI) with a diameter of 6 mm. Then, a dynamic FC network was estimated using the sliding window Pearson correlation method, which yielded a series of 142 × 142 correlation matrices. We used a fixed-length rectangle window (width = 24 × TRs = 60 s), and the window was shifted by 1 TR. The obtained correlation series were then Fisher-Z transformed and low-pass filtered with a cut-off frequency of 1/w Hz. All of these network matrices were vectorized to simplify the analysis.

### Phase Randomization

As suggested previously (Hutchison et al., 2013a; Hindriks et al., 2016), phase randomization analysis was used to explore the dependability of the dynamic FC fluctuation. The processed rs-fMRI time courses from the senior and young groups were phase randomized into new time courses in which the frequency spectra of the bold signals were invariable. We called the phase processed data the null group, which was then compared with the senior and young groups. Phase randomization was conducted for all parameters except amplitude (Friston et al., 1994; Handwerker et al., 2012), which could preserve the temporal correlation properties. These steps were taken to allow for assessing the dependability of the dynamic fluctuations and to verify whether the FC fluctuations over the rs-fMRI involved specific neural activities.

### Functional Connectivity Variation (FCV)

The dynamic functional connectivity variation (FCV) was calculated as the standard variation of the dynamic FC series. In this approach, the stability of the FC fluctuation over time is quantitatively measured and compared between brain region pairs. Previous studies of resting-state fMRI have demonstrated that some intrinsic neural activities are related to the variations in FC. These findings likely suggest the internal mechanism of the resting-state fMRI; thus, the FCV matrix was calculated for each subject. The original FCV matrices and phase randomized FCV matrices of both groups were statistically analyzed and compared with the averages of each group using a one-sample t-test. Thus, we could easily examine network modularization and the FCV of each network or connection.

#### Frequency Spectrum Analysis

A sliding rectangle window was used with a low-pass filtering effect on the functional fluctuations. The cut-off frequency was 1/w (1/60 Hz = 0.018 Hz). We assumed that the frequency spectrum of the dynamic FC fluctuation would change with aging; thus, we sought to specify the age-related changes in frequency or energy as age increased. The frequency band of the FC fluctuation was divided equally into two frequency bands, 0–1/2w Hz and 1/2w–1/w Hz, within which the fluctuation energies were calculated with a Fourier transform. Then, the energy ratio of the two frequency bands was calculated as the

<sup>2</sup>http://lfcd.psych.ac.cn/ccs.html

energy of the lower frequency band dividing the energy of the higher frequency band.

### Permutation Tests

To obtain robust results on aging-related variations within and between groups, a non-parametric permutation test with 5000 randomizations of the group labels was utilized for all measures described above. We defined the t-statistics between the two groups as the difference measurement, yielding a distribution of t-statistics after 5000 randomizations. Then, p = 0.001 was set as the threshold of significance. To better understand the results, both connectivity-based and network-averaged indices were examined. With the non-parametric approach, the size of the Type I error was guaranteed to be set at the prescribed significance level. The permutation test demonstrated an excellent ability to differentiate between different profiles, even when those profiles appeared to be highly similar.

### Sliding Window Length Analysis

Previous works (Hutchison et al., 2013a; Hindriks et al., 2016) have much discussed the influence of sliding window parameter settings. However, there have been no definite conclusions about the optimal window length. In this article, we also carefully assessed the influence of the window length on the variation and frequency spectrum of the dynamic FC. A sequence of window lengths from 2 to 256 time points was selected to examine the FC variation between the mean dynamic FC and the static FC. Another sequence of window lengths from 10 to 70 time points was selected to obtain the frequency characteristics of the dynamic function connectivity time series.

### RESULTS

### Phase Randomized and Within-Group Analyses

The within group and between group comparisons were conducted after passing the normality tests. Both groups lost the network organization pattern after phase randomization, and all of the within- and between-network connections showed similar FC variations (**Figures 1A,D**). From the original data obtained from both groups, clear modularization could be identified from the lower within-network variation and the higher betweennetwork variation (**Figures 1B,E**). In the young group, higher variation of the dynamic FC was found in the DMN-related inter-network connections and the connections between the FPN and OCC. By contrast, in the senior group, only the DMN-related inter-network connections showed high variation. In both groups, however, the inter-network variations of the connections between the CON and SMN were clearly lower than the corresponding group averaged values; a similar result was obtained for the CON and OCC. Both the within- and betweennetwork variations based on the original data were significantly higher than those of the phase randomized data for almost all of the connections (**Figures 1C,F**).

### Age-Related Changes in Static FC

Compared with the young group, all connections indicated decreased within-network FC in the senior group, and fewer within-network functional connections showed increases within the CON and OCC (**Figure 2A**). Most of the between-network connections, especially between the CON, DMN, FPN and OCC, showed increased FC in the senior group compared with those of the young group. From the 3D view in **Figure 2B**, more of the connections crossing the cerebral hemispheres were changed compared with the connections within one side. Most of the connections that both crossed hemispheres and occurred within a hemisphere were located between the posterior-anterior brain; in these connections, several distinct, intensively connected nodes were found. In the networks that were averaged and assessed, all or most of the between-network connectivities increased and the inner network functional connectivities decreased in the senior group compared with those of the young group (**Figure 2C**). However, only the DMN and SMN showed significantly decreased within-network FC.

### Age-Related Changes in FC Variation

Most of the significantly changed connections shown in **Figure 3** were between-network connections. The few connections with increased connectivity were located between the hemispheres, and the changed connections shared a similar location with the FC connection, which crossed both hemispheres, showed a posterior-anterior distribution and was also changed (**Figure 3B**). Several intensively connected nodes were also obvious in the prefrontal and occipitotemporal regions. The averaging analysis of the networks showed that these decreases occurred only in the inter-networks, including all of the FPN-associated inter-networks, the OCC-DMN inter-network and the OCC-CON inter-network. The solid black lines in **Figures 2**–**4** mark the subcortical regions.

### Age-Related Changes in Frequency Spectra

The connections that covered all networks showed increased energy ratios between the energies of the higher and lower frequency bands (**Figure 4A**). On average, the within-network connections, which included the DMN, OCC and SMN, and the inter-network connections CON-OCC and OCC-SMN indicated significantly increased ratios in the senior group. The 3D view revealed that these changed connections had both crosshemisphere and anterior-posterior distributions (**Figure 4B**).

## Influence of the Sliding Window Length

The absolute difference between the mean dynamic FC and the static FC over the scan time was calculated and illustrated (**Figure 5**). Five network differences between the dynamic FC and the static FC are shown in the chart in **Figure 5**, in which the red central lines are the mean values and the gray lines indicate the values for different subjects. All five networks followed similar trends of variance, and when the window length was increased, the dynamic deviation of the static FC also varied. At approximately 50–60 s this difference reached a

FIGURE 1 | Illustration of the dynamic functional connectivity (FC) variation patterns within groups. Pictures (A,D) show the one-sample t-test results of senior and young null groups with phase randomization processed data (FDR p-value < 0.05; actual p-value < 0.0010); Pictures (B,E) show the one-sample t-test results of senior and young groups with the original data (FDR p-value < 0.05; actual p-value < 0.0011); Pictures (C,F) show two-sample t-test results between the original and phase randomized data within the senior and young groups (FDR p-value < 0.05; actual p-value < 0.0008). Red indicates a higher-than-average level in the one-sample t-test and a higher-than-null group in the two-sample t-test.

minimum, and at approximately 300 s the difference was at its maximum.

Accordingly, the frequency spectrums of the five networks from one typical subject also varied when the sliding window length increased from 10 to 70 time points (25–175 s **Figure 6**). The white line in **Figure 6** is at 1/60 Hz and indicates the cut-off frequency of the low-pass filtering on the dynamic FC time series. With the increasing length of the sliding window, the energy of the higher frequency attenuated faster than the energy of the lower frequency. At frequencies above 1/60 Hz,

the energy was very small, and at window lengths below approximately 60 s the energy difference was sufficiently stable for examination.

### DISCUSSION

In this study, we showed the pattern of age-related changes in the dynamic connectivity profile between and within the whole-brain resting state networks. The originality of this study consisted of characterizing the age-related variation and frequency transition of whole-brain dynamic FC with the sliding window correlation approach. Several interesting findings were as follows: (1) in both groups, the FC variation showed distinct organization and age-related modularization, which were missing after phase randomization; (2) the FC variation indicated a significant decrease between networks, which increased with age and was dramatic in the FPN-associated and OCC-associated inter-networks; and (3) at a higher frequency, the dynamic FC showed an increased energy ratio in the senior group. These results not only shed light on the mechanisms of dynamic FC but also add to our understanding of normal brain aging. Here, we carefully discuss the results described above as well as the methodology of the sliding window correlation.

Similar to the functional specialization in FC, temporal variation of the connectivity also revealed a similar pattern or modularization, as illustrated in **Figure 1**. The whole-brain variations in the dynamic FC were not uniform across all systems and fell into two sides of the mean level within each group. The within-network connectivity showed a lower variation, and most of the between-network connectivity showed higher variation compared with the mean except for the CON-SMN and

FIGURE 6 | Illustration of the frequency spectrum, which varied with the sliding window length, for all five networks of one typical subject. The results of five networks

CON-OCC connections. Many previous studies have reported that the states of the dynamic FC matrices showed high variation in the between-network connectivity (Allen et al., 2014; Hansen et al., 2015). Allen et al. (2014) reported that seven reproducible states could be differentiated by connectivity between the DMN regions, indicating great variation. Those findings are consistent with the current results (**Figures 1B,E**), in which the inter-networks of the DMN showed high variation in both groups. No type of pattern among the networks was indicated in the phase randomized analysis (**Figures 1A,D**). This modularization of the FC variation was most likely associated with the intrinsic neural activities and interactions between the systems.

were calculated from the individually averaged FC sequence of the within-network connections.

The pattern and modularization of the dynamic FC variation showed a clear decline in the senior group. The connections with higher and lower levels of variation both tended toward the mean, which indicated that less diversity in the variation between the subsystems occurred with increased age. A previous study (Du et al., 2016) reported that the FC of some connections showed less variation in schizophrenic patients. The changed modularizations of the FC variation in the two groups provided clear evidence that the modularization of the connectivity variation also declined with age, which also reflected functional integration and segmentation in aging brains (Hagmann et al., 2008; La Corte et al., 2016). Dedifferentiation of cognitive functions occurs in the aging brain, and many regions are reconfigurable to compensate for declines in other regions that occur with increased age (Sleimen-Malkoun et al., 2014). The dynamic variation in connectivity reflected that reconfiguration occurs all the time and follows some patterns, and when age increases, this pattern slowly changes. Previous studies have also reported that the brain regions dynamically participate or reconfigure into different modules during the scan time at rest (Bassett et al., 2011; Schaefer et al., 2014). The declining modularization of the FC variation in the senior group was likely related to aging and compensatory mechanisms.

The presence of FC variation, even within a single brain state (including the resting state), has been increasingly recognized (Chang and Glover, 2010; Hutchison et al., 2013a; Allen et al., 2014) and has been established as clinically relevant (Damaraju et al., 2014; Kucyi and Davis, 2014; Elton and Gao, 2015). A recent direct comparison of the awake, resting state with the anesthetized state has revealed a dramatic reduction in the connectivity variation during unconsciousness, which suggests that the connectivity variation is at least partly related to conscious operations (Barttfeld et al., 2015). One potential source of connectivity variation during consciousness is ''mind wandering'', in which the brain consciously engages in different mental operations that produce fluctuations in the FC. However, we expected that the FC variation is a capacity of elasticity or operations to maintain states or connectivity transitions. This concept is similar to the cognition capacity resource, in which cognitive resources are limited and reduced in the aging brain. High variation of FC leads to the increased possibility of reconfiguration to address the loss of some resources. Qin et al. (2015) used the amplitude of the low frequency fluctuations (ALFF) of the dynamic FC to predict brain maturation between 7 years and 30 years of age. The findings of that study suggested that the increased variation was highly related to maturation. The ALLF was essentially the same with respect to the FC variation, both of which could be associated with the plasticity of the brain.

All decreased dynamic FC variations in the senior group were located in the between-network connections, especially in the FPN-linked and OCC-linked inter-networks. The internetwork between FPN and OCC was also indicated to be missing in the senior group (**Figures 1B,E**). These findings might indicate a posterior-anterior shift in aging (Davis et al., 2008; Vinette and Bray, 2015), which has been interpreted as compensatory in that higher-order cognitive processes are recruited to offset deficits in sensory processing. In all of the 3D views of changed connections, many posterior-anterior connections were also obvious. Cognitive and perceptual changes could be linked because they are susceptible to the same age-related factors, and a perceptual system decline could have an impact on the cognition outcome (Allen and Roberts, 2016). In young adults, visual learning engages an extended network of occipito-temporal, parietal and frontal regions, which is known to be involved in perceptual decisions (Kim and Shadlen, 1999; Shadlen and Newsome, 2001; Heekeren et al., 2004; Mayhew and Kourtzi, 2013). The subcortical regions, including the basal ganglia and thalamus, have been shown to be associated in cognition processing (Koziol and Budding, 2009). The results for the subcortical regions showed mainly increased connectivity in the DMN and decreased variation in the FPN and OCC. Considering the central or hub roles of both functional and structural networks, the subcortical regions might regulate passing signals or communication between the perception and cognitive systems, such as the OCC and FPN, the OCN and OCC, or the CON and SMN (Marchand et al., 2011). The increased fluctuation energy ratio of the high-frequency connected networks of the subcortical region might further reflect an age-related dynamic regulation, which requires further research in terms of behavioral and task fMRI experiments. With limited cognitive resources of the cognitive and perceptual systems in the aging brain, the interaction between these two systems declines as brain processing capacity is reduced.

Except for the declining capacity of cognitive resources, the changed FC fluctuation energy ratio of the higher frequencies could allow for a deeper inspection of the dynamic communication between regions. Shakil et al. (2015) found that the dynamic FC has a frequency dependance. Fluctuations within the frequency of 0–1/w were also suggested, which implied a real, physiologically dynamic connectivity (Leonardi and Van De Ville, 2015). The connectivity fluctuation is a type of intrinsic interaction of different frequency components involved in different neural activities. The speed of processing in the aging brain (Park and McDonough, 2013) is also an important property that reflects cognitive resources. Aging-related decreases in the amplitudes of low-frequency BOLD signal fluctuations have been observed, suggesting that the low-frequency fluctuations of neuroactivities are more vulnerable to aging-associated declines (Hu et al., 2014). We thought that the increased energy ratio of the fluctuating high frequencies was most likely caused by increased damage to the fluctuations of the low frequencies. The decreased variation and increased speed of fluctuations in the dynamic FC would result in disorder of the dynamic communications between different brain regions in senior individuals.

Logically, in the sliding window correlation method, the window length should have a substantial effect on the captured connectivity fluctuation. This factor was the most important consideration in terms of the overall accuracy of the technique (Leonardi and Van De Ville, 2015; Hindriks et al., 2016; Shakil et al., 2016). However, there is still no clearly determined standard for window length selection. Convergent results suggested that a window length between 50 s and 60 s is optimal. As we expected, the dynamic FC during rest was found to be a dynamic balance that maintained intrinsic connectivity patterns and even vigilance for cognitive tasks. Moreover, the variation in the FC was predicted to be centered around the static connectivity level. This expectation was supported by the appearance of a minimum on the difference curve between the mean dynamic and the static FC, as shown in **Figure 5**. The maximum was located at approximately 300 s (5 min), which was consistent with the suggestion of a longer scan time for reliably detecting resting state connectivity (Heekeren et al., 2004; Zuo et al., 2013). The individual frequency spectrum changed with the window length, as shown in **Figure 6**; this result indicated that a window shorter than 60 s retained potentially important energy for all of the networks, and the lower frequency energy contributed most of the connectivity fluctuation. A previous study (Leonardi and Van De Ville, 2015) suggested a meaningful frequency of under 1/w, which was even lower than the commonly considered frequencies for BOLD fluctuations. A fluctuation frequency of 0–1/w was suggested previously (Leonardi and Van De Ville, 2015) and employed in a later study (Qin et al., 2015; Liu et al., 2016). In the present article, the frequency was even lower than the commonly considered 0.01–0.10 Hz of the BOLD signal fluctuation, which was in accordance with the neural activity. This type of mismatch could be understood given that unlike BOLD signal fluctuations, which indicate direct neural metabolic activity, the fluctuation in the functional connectivities reflected interactions between the different regions. The interactions were also revealed in the current results, which showed that more of the affected connections were located in inter-network and inter-hemispheric connections. Furthermore, we believe that the FC variation must depended more on the fluctuation frequency, which is influenced by various factors, including anatomy, cognition, physiology and disease. The more that we understand the physiology of dynamic FC and the methodologies used to study these phenomena, the more insights we will have into the mechanisms of aging and disease.

#### CONCLUSION

In conclusion, this study presented a resting-stated dynamic FC analysis of normal brain aging. All of the results converged to expound compensation and reorganization of the networks during aging. We examined the modularization of the FC variations in the brain and decreased modularization in the aging brain. Additionally, decreased variation and increased damage to the low frequency fluctuations of the dynamic FC with aging were detected; these changes were interpreted to be associated with declining cognitive resources and limited processing speeds in the senior brain. In this article, we provided and applied new insights into FC analysis for use in aging research. The results indicated that the dynamic features of the resting-state FC were

#### REFERENCES


actually the intrinsic interactions between regions and cognitive resources. When some cognitive resources were reduced in aging, this type of dynamic mechanism acts to reconfigure or even train a new cognitive resource. Our conclusions in this article were fully supported by dependable results; we believe that the dynamic FC can potentially capture the intrinsic rules of compensatory processes in the aging brain, and that the present results will promote insightful understanding of spontaneous fluctuations in FC as well as aging mechanisms.

#### AUTHOR CONTRIBUTIONS

MS, YL and XiongZ processed all the image data and conducted some analysis work; YC and WW were in charge of the analysis work and wrote the manuscript; JM and HN provide some useful guidance and ideas; YC, XinZ and DM designed and provided the original idea; XinZ and DM sponsored the whole research.

#### FUNDING

This research was supported by National Natural Science Foundation of China (No. 81571762, 81630051, 91520205), National Key Technology R and D Program of the Ministry of Science and Technology of China (No. 2012BAI34B02) and Tianjin Key Technology R and D Program (No. 15ZCZDSY00930).

during motor sequence learning. Neuroimage 49, 759–766. doi: 10.1016/j. neuroimage.2009.08.048


and brain states. Neuroimage 133, 111–128. doi: 10.1016/j.neuroimage.2016. 02.074


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Chen, Wang, Zhao, Sha, Liu, Zhang, Ma, Ni and Ming. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Altered Neuronal Activity Topography Markers in the Elderly with Increased Atherosclerosis

Takashi Shibata<sup>1</sup> \*, Toshimitu Musha<sup>2</sup> , Yukio Kosugi <sup>2</sup> , Michiya Kubo<sup>1</sup> , Yukio Horie<sup>1</sup> , Naoya Kuwayama<sup>3</sup> , Satoshi Kuroda<sup>3</sup> , Karin Hayashi <sup>4</sup> , Yohei Kobayashi <sup>2</sup> , Mieko Tanaka<sup>2</sup> , Haruyasu Matsuzaki 2, 5, Kiyotaka Nemoto<sup>6</sup> and Takashi Asada<sup>7</sup>

<sup>1</sup> Department of Neurosurgery, Stroke Center, Saiseikai Toyama Hospital, Toyama, Japan, <sup>2</sup> Brain Functions Laboratory Inc., Yokohama, Japan, <sup>3</sup> Department of Neurosurgery, Graduate School of Medicine and Pharmacological Science, University of Toyama, Toyama, Japan, <sup>4</sup> Department of Neuropsychiatry, Toho University Medical Center Sakura Hospital, Chiba, Japan, <sup>5</sup> Department of Medical Course, Teikyo Heisei University, Tokyo, Japan, <sup>6</sup> Department of Neuropsychiatry, Institute of Clinical Medicine, University of Tsukuba, Tsukuba, Japan, <sup>7</sup> Department of Neuropsychiatry, University of Tokyo Medical and Dental University, Tokyo, Japan

Background: Previously, we reported on vascular cognitive impairment (VCI) templates, consisting of patients with VCI associated with carotid stenosis (>60%) using a quantitative electroencephalographic (EEG) technique called neuronal activity topography (NAT). Here using the VCI templates, we investigated the hypothesis that internal carotid artery–intima-media thickness (ICA–IMT) is associated with EEG spectrum intensity (sNAT) and spectrum steepness (vNAT).

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

Reviewed by:

Manousos A. Klados, Technische Universität Dresden, Germany Fu-Jung Hsiao, National Yang-Ming University, Taiwan

> \*Correspondence: Takashi Shibata sibata@dj8.so-net.ne.jp

Received: 06 September 2016 Accepted: 20 June 2017 Published: 06 July 2017

#### Citation:

Shibata T, Musha T, Kosugi Y, Kubo M, Horie Y, Kuwayama N, Kuroda S, Hayashi K, Kobayashi Y, Tanaka M, Matsuzaki H, Nemoto K and Asada T (2017) Altered Neuronal Activity Topography Markers in the Elderly with Increased Atherosclerosis. Front. Aging Neurosci. 9:216. doi: 10.3389/fnagi.2017.00216 Methods: A total of 221 community-dwelling elderly subjects were recruited. Four groups were classified according to quartiles of ICA–IMT as assessed by ultrasonography: control group A, normal (≤0.9 mm); group B, mild atherosclerosis (1−1.1 mm); group C, moderate atherosclerosis (1.2−1.8 mm); and group D, severe atherosclerosis (≥1.9 mm). EEG markers of power ratio index (PRI), and the binary likelihood of being in the VCI group vs. the that of being in control group A (sLx:VCI−A, vLx:VCI−A) were assessed, respectively. Differences in mean total scores for PRI, sLx:VCI−A, vLx:VCI−A, between control group A and the other groups were compared using Dunnett's test, respectively.

Results: The mean total scores of the PRI were 3.25, 3.00, 2.77, and 2.26 for groups A, B, C, and D, respectively. There was a significant decrease in the PRI in group D compared with group A (P = 0.0066). The mean total scores of the sLx:VCI−<sup>A</sup> were −0.14, −0.11, −0.1, and −0.03 for groups A, B, C, and D, respectively. The sLx:VCI−<sup>A</sup> in group D was significantly higher compared to that in group A (P < 0.0001). The mean total scores of the vLx:VCI−<sup>A</sup> were −0.04,−0.01, 0.01, and 0.06 for group A, B, C, and D, respectively. The vLx:VCI−<sup>A</sup> in group D and group C was significantly higher compared to that in group A, respectively (P < 0.0001, P = 0.02).

Conclusion: Community-dwelling elderly subjects in the increased carotid atherosclerosis of ICA–IMT (≥1.9 mm) were at greatest risk of an EEG change as assessed by NAT.

Keywords: EEG, vascular cognitive impairment, neuronal activity topography, atherosclerosis, elderly

## INTRODUCTION

The carotid bifurcation and the proximal part of the internal carotid artery (ICA) are predilection sites for atherosclerotic plaques. Changes in the ICA–intima-media thickness (IMT) in this area, as assessed by ultrasonography, are a first sign of subclinical atherosclerosis (Polak et al., 2010; Bauer et al., 2012). Ultrasonography is an easily accessible and noninvasive method to measure different stages of the carotid artery atherosclerotic process, and it is widely used in clinical assessments and for epidemiological and clinical research (Arntzen and Mathiesen, 2011).

Carotid disease is a known risk factor for stroke and vascular cognitive impairment (VCI), but the relationship between carotid artery stenosis and cognitive function in asymptomatic people is unclear. The role of subclinical atherosclerosis in cognitive function can be studied by ultrasound measurement of the carotid arteries and neuropsychological tests. Previous studies indicate that patients with carotid stenosis have markedly poorer scores on cognitive tests compared with control subjects (Rao, 2001; Johnston et al., 2004; Mathiesen et al., 2004; Arntzen and Mathiesen, 2011). One cross-sectional study found that subjects with carotid stenosis (>35%) had lower levels of performance on cognitive function tests than subjects without stenosis (Mathiesen et al., 2004). In a cohort from the Cardiovascular Health Study (Haan et al., 1999), cognitive decline was markedly increased in subjects with an ICA–IMT > 2.01 mm. Most patients with subclinical carotid atherosclerosis have only minor impairments of cognitive function, and standard tests (e.g., the Mini-Mental State Examination: MMSE) are not sufficiently sensitive to detect such impairments. However, early detection of VCI is of particular importance because pharmacological intervention to prevent or delay dementia will prove effective for most patients with subclinical carotid atherosclerosis.

Electroencephalographic (EEG) signals are generated by electrical activity in the brain and are rich in information regarding cerebral function. In 2013, Musha et al developed a neuroimaging tool called Neuronal Activity Topography (NAT), which gives direct information on neuronal activity using quantitative EEG analysis (Musha et al., 2013). This tool categorizes cerebral neuronal activity by EEG spectrum intensity (sNAT) and spectrum steepness (vNAT). NAT consists of 210 submarkers referring to 10 frequency components ranging from 4 to 20 Hz (more precisely, from 4.7 to 18.8 Hz). Each submarker has its own role in the characterization of cerebral neuronal activity, which is represented by the 210-dimensional NAT spaces. The NAT system has been used to detect Alzheimer's disease (AD) and to discriminate AD from other forms of dementia, VCI, dementia with lewy bodies, and is currently undergoing testing for its practical use in the clinical setting (Musha et al., 2002, 2004, 2013). Recently, we reported on VCI templates, consisting of patients with VCI associated with moderate carotid stenosis (>60%) and normal controls (NLc) (Shibata et al., 2014). In brief, the binary likelihood of being in the VCI group vs. that of being in NLc group (sLx:VCI−NLc, vLx:VCI−NLc) was assessed in each of the sNAT and vNAT spaces. Separation of the VCI group and NLc group was made with a sensitivity of 92 and 88%, as well as a false-positive rate of 8 and 12% forsLx:VCIc−NLc and vLx:VCI−NLc, respectively. Therefore, the VCI templates based on NAT might be applied to communitydwelling elderly people to detect an EEG change reflecting VCI.

If EEG markers of NAT combined with ICA–IMT measurement could detect cognitive decline in subclinical atherosclerosis, then a combination of EEG and ultrasonography could be used to detect cognitive decline during routine medical checkups, because both EEG and ultrasonography are inexpensive, reliable, and noninvasive. To the best of our knowledge, the relationship between EEG finding and ICA–IMT has not yet been explored in elderly subjects. In the present study, using the VCI templates previously obtained from Toyama city, we investigated the hypothesis that increased ICA–IMT are associated with altered EEG markers of NAT for community-dwelling elderly people in Tsukuba city.

### SUBJECTS AND METHODS

### Characteristics of Survey Subjects

The present investigation is part of the Tsukuba epidemiological investigation project for the prevalence of dementia among inhabitants older than 65 years of age in Tsukuba, Ibaraki, Japan. This study was approved by the ethical committees from the University of Tsukuba (Tsukuba, Japan). All subjects gave written informed consent in accordance with the Declaration of Helsinki. As part of this project, a three-phase survey was carried out in the Tsukuba area. The survey protocol has been described in detail in a previous report (Ikejima et al., 2012). Between February 2012 and October 2012, 221 community-dwelling elderly subjects were enrolled. The 221 subjects underwent ultrasonography, EEG recording, and magnetic resonance imaging (MRI; 1.5T ECHELON RX; Hitachi Medical Corporation, Tokyo, Japan). Deep and periventricular white matter hyperintensities (WMH) were coded from 0 to 3 according to the Fazekas scale (Fazekas et al., 1987). Each EEG finding was independently interpreted by two EEG specialists who were blind to other data about the patients except their age and sex. All subjects underwent the MMSE to screen for cognitive function (cut-off score for cognitive impairment = 26/27).

### Characteristics of VCI Group (Shibata et al., 2014)

The selected 55 VCI inpatients previously admitted at our hospital in Toyama city included 47 men and 8 women aged 58–87 years [mean ± standard deviation (SD), 72.6 ± 7.1 years]. Patients were selected based on the following criteria: (i) evidence of unilateral carotid stenosis of >60% (symptomatic or asymptomatic) confirmed with conventional angiography or computed tomography (CT) angiography and a degree of carotid artery stenosis determined using the criteria of the North American Symptomatic Carotid Endarterectomy Trial and (ii) mild cognitive impairment, considered as a score of <90 (low average) on the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) (Takaiwa et al., 2006, 2009). Of the 55 VCI patients with proximal carotid

stenosis, 23 had right-sided lesions (stenosis or occlusion), 19 had left-sided lesions, and 13 had bilateral lesions. The mean score (± SD) on the MMSE and RBANS for the 55 VCI patients was 27.1 ± 1.5 and 74.5 ± 13.6, respectively. Exclusion criteria included (i) evidence of a previous major stroke and brain damage revealed through MRI; (ii) rapidly evolving symptoms with any hemiparesis, aphasia, and apraxia; and (iii) evidence of dementia, considered as MMSE score of <24, (iv) history of cerebral surgery, obvious psychiatric or neurological disorders, or (v) uncontrolled or malignant general complications. All VCI patients had good levels of daily living activities.

We showed a flow chart for understanding two groups: (1) elderly group: community-dwelling elderly subjects from Tsukuba city, n = 221, (2) VCI group: VCI patients from Toyama city, n = 55 (**Figure 1**).

#### EEG Recording and Preprocessing

Electroencephalographic (EEG) recordings (EEG-1200/9100; Nihon Kohden Corporation, Tokyo, Japan) were performed in an awake, resting state with eyes closed for 10 min. Scalp potentials were recorded with 21 electrodes arranged according to the international 10–20 System (Fp1, Fp2, F3, F4, F7, F8, Fpz, Fz, T3, T4, T5, T6, C3, C4, Cz, P3, P4, Pz, O1, O2, and Oz). The contact impedance between the electrode and the scalp was < 50 k. The reference electrode was on the right earlobe, and then the average reference for NAT analysis is computed as a mean of all electrodes. The raw EEG data were recorded at 0.08∼300 Hz with 1,000 Hz sampling rate, and then the converted EEG data for NAT analysis were sampled at 200 Hz per channel and bandpass filtered to pick up signals in a frequency range of 4– 20 Hz, which minimizes the effect of artifacts contaminating the recorded signals. The recorded signal sequence was divided into 0.64-s segments. To make the EEG data as high quality as possible and exclude the artifacts, each EEG finding was independently interpreted by EEG specialist who was blind to other data about the subjects except their age and sex. We carefully avoided particular epochs containing ocular movements, baseline shifts, drowsiness signs, and muscle or cardiac contamination.

#### EEG Data Analysis

The average power ratio index (PRI) was calculated for the 221 subjects by dividing the total sum of low frequency (4–8 Hz) power spectrum by the total sum of high frequency (8–20 Hz) power spectrum at each electrode (Cz, C3, C4, Pz, P3, P4, Oz, O1, and O2), respectively, because of an efficient calculation of dominant parieto-occipital rhythm prior to Normalized Power Spectrum (NPS) analyses using the following formula:

$$PRI \equiv \Sigma PS\_h / \Sigma PS\_l \tag{1}$$

The PS<sup>l</sup> and PS<sup>h</sup> represent power spectrum of low frequency (4–8 Hz) and that of high frequency (8–20 Hz), respectively.

The mathematical background of NAT (Brain Functions Laboratory, Inc., Yokohama, Japan) was previously described in detail (Musha et al., 2013).

#### 1) Normalized Power Spectrum

We are going to derive dimensionless markers from the EEG signals which characterize stochastic nature of the cerebral neuronal activities generating these signals. The discrete power spectrum PS consists of the ten frequency components D <sup>X</sup>j,<sup>m</sup> 2 E seg for m = 3–12 on signal channel j. The subscript attached to the averaging symbol denotes, hereafter, that the averaging is performed on this variable. In the present case, the averaging is performed across all of the segments. Dependence of the EEG signal level on individual subject is eliminated by normalizing it to its mean level. The NPS consists of 10 (m = 3–12) such components NPSj,<sup>m</sup> defined as,

$$\text{NPS}\_{j,m} = \left\langle \left| X\_{j,m} \right|^2 \right\rangle\_{\text{seg}} \left\langle \left| X\_{j,m'} \right|^2 \right\rangle\_{\text{seg},m'} \tag{2}$$

They make a set of 10 submarkers for each signal channel. This marker characterizes the fractional partition of the EEG power over the 10 frequency components.

Another role of the collective neuronal activities is the signal transmission through the neuronal networks. The biological signal is encoded as the modulation of the occurrence rate of neuronal activities. The organized modulation mode introduces coherence in the collective neuronal activities, and the coherence causes spiky variations of the power spectrum. The modulation is characterized by a ratio pj,<sup>m</sup> between the adjacent power components given as pj,<sup>m</sup> = D <sup>x</sup>j,m+<sup>1</sup> 2 E seg / D <sup>x</sup>j,<sup>m</sup> 2 E seg . Such ratios derive the 10 (m = 3–13) dimensionless submarkers NPVj,<sup>m</sup> on the signal channel j,

$$NPV\_{j,m} = \frac{4p\_{j,m}}{\left(1 + p\_{j,m}\right)^2} \,\text{}.\tag{3}$$

#### 2) Zero-Level Resetting

Averages (NPSj l,m)seg,jl and (NPVj l,m)seg,jl across the signal channels are left with offset values. The offset values lower the quality of discrimination between different brain diseases, and new markers sNATj,<sup>m</sup> and vNATj,<sup>m</sup> are introduced after removing the offset as,

$$\text{sNAT}\_{j,m} \equiv \left< \text{NPS}\_{j,m} \right>\_{\text{seg}} - \left< \text{NPS}\_{j',m} \right>\_{\text{seg},j'} \tag{4}$$

$$\nu \text{NAT}\_{j,m} = \left< \text{NPV}\_{j,m} \right>\_{\text{seg}} - \left< \text{NPV}\_{j',m} \right>\_{\text{seg},j'}.\tag{5}$$

#### 3) Likelihood

Briefly, a template marker sNAT map is prepared from data for 55 previously diagnosed VCI patients using the mean and SD of a number of sNAT states. The sNAT state of a test subject was assigned to a point in the 210-dimensional sNAT space, whereas the template state of the VCI patients was assigned to another

FIGURE 1 | A flow chart for understanding two groups: (1) elderly group: community-dwelling elderly subjects from Tsukuba city, n = 221, (2) VCI group: VCI patients from Toyama city (Shibata et al., 2014), n = 55. MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rating; WMS-R, "logical memory A" from the Wechsler Memory Scale-Revised; PAS, Psychogeriatric Assessment Scales; RBANS, Repeatable Battery for the Assessment of Neuropsychological Status; GDS, Geriatric Depression Scale-Short Form; DSM-IV, Diagnostic and Statistical Manual of Mental Disorders, 4th edition; AD, Alzheimer disease; VaD, Vascular dementia; DLB, Dementia with Lewy Bodies; FTLD, FrontoTemporal Lober Degeneration; MCI, Mild Cognitive Impairment.

point in this space. The states of patients with VCI are distributed within these spaces making clusters around their mean states, which are regarded as the template states sNATVCI j,m together with the standard deviations sσ VCI j,m around them. sNATVCI j,m of VCI is defined as sZx:VCI <sup>j</sup>,<sup>m</sup> which is

The likelihood of the test subject x being in the VCI group, sLx:VCI, was given as a function of the effective distance between the two points properly normalized in terms of the mean and SD related to the VCI template. The likelihood sLx:VCI of a test subject x to be in VCI is defined as,

$$\mathrm{sZ}\_{j,m}^{\mathrm{x:VCI}} \equiv \left( \mathrm{sNAT}\_{j,m}^{\mathrm{x}} - \mathrm{sNAT}\_{j,m}^{\mathrm{VCI}} \right) / \mathrm{s\sigma}\_{j,m}^{\mathrm{VCI}} \tag{6}$$

$$sL\_{\mathbf{x}:VCI} \equiv \exp\left\langle -\left(sZ\_{j,m}^{\mathbf{x}:VCI}\right)^2 \right\rangle\_{j,m} \tag{7}$$

The differential (binary) likelihood of being in the VCI group vs. the control group A, sLx:VCI−<sup>A</sup> was defined as,

$$sL\_{\mathfrak{x} : V \!CI - A} \equiv sL\_{\mathfrak{x} : V \!CI} - sL\_{\mathfrak{x} : A} \tag{8}$$

Similarly, another differential likelihood in reference to vNAT was introduced as vLx:VCI−A. The subject was more likely to be in the VCI group than in the control group A when this value was more positive and vice versa.

#### Ultrasonography Protocol

The same sonographer, who was blinded to the subjects' case status and risk factor levels, carried out all ultrasound examinations. High-resolution B-mode ultrasonography of the ICA was performed with an EUB-5500 ultrasound machine (Hitachi Medical Corporation, Tokyo, Japan). Subjects were examined in the supine position for about 15–20 min. Longitudinal images of the ICA were obtained by combined B-mode and color Doppler ultrasound examinations. With this technique, two parallel echogenic lines separated by an anechoic space can be visualized at the level of the artery wall. These lines are generated by the blood-intima and mediaadventitia interfaces. The distance between the two lines gives a reliable index of the thickness of the intima-media complex. The posterior (far) wall IMT was measured with the electronic calipers of the ultrasound machines, as described by Pignoli et al. (1986). On a longitudinal, 2-dimensional ultrasound image of the ICA, images of the posterior wall are displayed as two bright white lines separated by a hypoechoic space. The distance between the leading edge of the first bright line and the leading edge of the second bright line indicates the IMT. We assessed the maximum IMT, which was defined as the single thickest part of the wall among the near and far right and left walls of the ICA.

#### Statistical Analysis

All data were analyzed using JMP <sup>R</sup> 12 (SAS Institute Inc., Japan). Mean ± SD values of age, total MMSE score, and years of education were used as descriptive measures of normally distributed variables. The results were compared among the 4 groups included VCI group by analysis of Dunnett's test as control group A. Tukey's HSD (honest significant difference) was used for quantitative variables (age, total MMSE score and years of education), and the χ 2 test was used for categorical variables (Fazekas score, EEG finding). Linear regression analysis among PRI, the binary likelihoods sLx:VCI−<sup>A</sup> and vLx:VCI−A, and the ICA−IMT was performed. Differences with a P-value of <0.05 were considered statistically significant.

#### RESULTS

The clinical characteristics of the 221 study subjects grouped according to quartiles of ICA−IMT are shown in **Table 1**. The 221 subjects included 110 men and 111 women with a mean (± SD) age of 74.9 ± 6.5 years. The mean MMSE score (± SD) was 28.8 ± 1.2, and the mean education period (± SD) was 12.7 ± 3.1 years. The quartiles of ICA−IMT were defined based on the maximum ICA−IMT and were 1, 1.2, and 1.9 TABLE 1 | Clinical and radiological characteristics of the study population.


Mean values ± standard deviation of demographic characteristics (age, MMSE score, years of education, and ICA–IMT, EEG finding, Fazekas score on MRI). Age and education are expressed in years. Group A, normal (ICA–IMT ≤ 0.9 mm); Group B, mild atherosclerosis (1–1.1 mm); Group C, moderate atherosclerosis (1.2–1.8 mm); Group D, severe atherosclerosis (≥1.9 mm). MMSE, Mini-Mental State Examination; IMT, intima-media thickness; and SD, standard deviation.

mm. The groups were as follows: Group A, normal (ICA−IMT ≤0.9 mm); group B, mild atherosclerosis (1−1.1 mm); Group C, moderate atherosclerosis (1.2−1.8 mm); and Group D, severe atherosclerosis (≥1.9 mm). There were no significant differences in age, total MMSE score, years of education among the four groups. Although the cerebral white matter lesions on Fazekas class four and borderline on EEG findings were associated with an increased IMT, there were no significant differences in Fazekas score and EEG finding among the four groups (χ 2 test).

The characteristics distributions of sNAT in the elderly with increased IMT have appeared at a specific frequency range of 7.8 Hz, 10.9 Hz as follows. (1) At theta frequency of 7.8 Hz, there were statistically significant increased activities over occipitaltemporal areas (O1, O2, T5, T6, F7, and F8) and decreased activities over bilateral frontal areas (F7, F8) in comparison with control group A (**Figure 2A**). Mean z-score sNAT map at 7.8 Hz frequency range was shown for group B, group C, group D, VCI group in comparison with control group A, respectively (**Figure 3**). (2) At alpha frequency range of 10.9 Hz, there were statistically significant decreased activities over occipital areas (O1, O2, and Oz) and increased activities over bilateral frontotemporal areas (F7, F8, T3, and T4) in comparison with control group A (**Figure 2B**).

The characteristics distributions of vNAT in the elderly with increased IMT have appeared at a specific frequency range of 7.8 Hz, 9.4 Hz, respectively as follows. (1) At theta frequency range of 7.8 Hz, there were statistically significant increased activities over occipital areas and decreased activities over bilateral fronto-temporal areas in comparison with control group A (**Figure 4A**). (2) At alpha frequency range of 9.4 Hz, there were statistically significant decreased activities over occipital areas

and increased activities over bilateral parieto-temporal areas in comparison with control group A (**Figure 4B**). Mean z-score vNAT map at 9.4 Hz frequency range was shown for group B, group C, group D, VCI group in comparison with control group A, respectively (**Figure 5**).

A comparison of the PRI among the normal group A and the subclinical atherosclerosis groups (B, C, D), VCI group is shown in **Figure 6A**. The mean total scores (± SD) for the PRI were 3.25 ± 1.64, 3.00 ± 1.87, 2.77 ± 1.8, 2.26 ± 1.17, and 1.41 ± 0.69 for group A, B, C, D, and VCI, respectively. In results of Dunnett's test as control group A, there was a significant decrease of the PRI in the group D and VCI compared with the normal group A, respectively (P = 0.0018, P < 0.0001).

Comparison of the binary likelihood sLx:VCI−<sup>A</sup> among the control group A and the subclinical atherosclerosis groups (B, C, D), VCI group is shown in **Figure 6B**. The mean total scores (± SD) for the binary likelihood sLx:VCI−<sup>A</sup> were −0.14 ± 0.1, −0.11 ± 0.13, −0.1 ± 0.12, −0.03 ± 0.13, and 0.09 ± 0.08 for group A, B, C, D, and VCI, respectively. In results of Dunnett's test as control group A, there was a significant increase in the binary likelihood sLx:VCI−<sup>A</sup> in the group D and VCI in comparison with control groups A, respectively (P < 0.0001, P < 0.0001).

FIGURE 3 | Mean sNAT map at 7.8 Hz frequency range for (A) group B, (B) group C, (C) group D, (D) VCI group, respectively. The colorbar indicates the range of z-scores in comparison with control group A: Green indicates normal spectrum intensity. Red and blue indicate hyperspectrum intensity and hypospectrum intensity, respectively. L, left; R, right.

Comparison of the binary likelihood vLx:VCI−<sup>A</sup> among the control group A and the subclinical atherosclerosis groups (B, C, D), VCI group is shown in **Figure 6C**. The mean total scores (± SD) for the binary likelihood vLx:VCI−<sup>A</sup> were −0.04 ± 0.09, −0.01 ± 0.09, 0.01 ± 0.1, 0.06 ± 0.1, and 0.1 ± 0.07 for groups A, B, C, D, and VCI, respectively. In results of Dunnett's test as control group A, there was a significant increase in the binary likelihood vLx:VCI−<sup>A</sup> in group C, D, and VCI in comparison with control group A, respectively (P = 0.021, P < 0.0001, P < 0.0001).

The correlation of the ICA–IMT and the PRI, the binary likelihood sLx:VCI−<sup>A</sup> and the binary likelihood vLx:VCI−<sup>A</sup> was examined, respectively (**Figure 7**). Linear regression analysis showed a weak negative correlation between the ICA–IMT and the PRI (r = −0.25, P = 0.0002), and a weak positive correlation between ICA–IMT and the binary likelihoods sLx:VCI−NLc (r = 0.31, P < 0.0001) and vLx:VCI−NLc (r = 0.3, P < 0.0001).

#### DISCUSSION

In the present study, a relationship between subclinical atherosclerosis and EEG markers of PRI, sNAT, vNAT in an elderly population was investigated, respectively. In results, we identified 2 important findings. First, it was confirmed that EEG

spectrum intensity (sNAT) and spectrum steepness (vNAT) are related to pathological changes indicative of subclinical carotid atherosclerosis. In particular, the binary likelihoods sLx:VCI−<sup>A</sup> and vLx:VCI−<sup>A</sup> in group D (ICA–IMT, ≥1.9 mm) suggests a risk of an EEG pathological change. Second, NAT might visually show topographic information about EEG alterations in the elderly with increased ICA–IMT. Although there was difficult to interpret an association between all the 210-dimensional NAT spaces and EEG alterations regarding subclinical atherosclerosis, the characteristics distributions of sNAT and vNAT in the elderly with increased ICA–IMT might have appeared at a specific frequency range (e.g., 7.8 Hz). As far as EEG alterations of a dominant parieto-occipital alpha rhythm goes, in sNAT, there were decreased activities over occipital areas at alpha frequency range of 10.9 Hz, and increased activities over occipital areas at theta frequency range of 7.8 Hz, suggesting the alterations of EEG spectrum intensity from alpha to upper theta ranges. Similarly, in vNAT, there were decreased activities over occipital areas at alpha frequency range of 9.4 Hz and increased activities over occipital areas at theta frequency range of 7.8 Hz, suggesting the alterations of EEG spectrum steepness, that is, more spectrum blur (i.e.,

FIGURE 5 | Mean vNAT map at 9.3 Hz frequency range for (A) group B, (B) group C, (C) group D, (D) VCI group, respectively. The colorbar indicates the range of z-scores in comparison with control group A: Green indicates normal spectrum steepness. Red and blue indicate spectrum blur and spectrum sharpness, respectively. L, left; R, right.

undersynchrony) at upper theta frequency range and more spectrum sharpness (i.e., oversynchrony) at alpha frequency range.

There were many quantitative EEG markers of PRI, EEG global power, an asymmetry-based EEG marker, and NAT for detecting mild cognitive impairment (MCI). Previous findings in patients with MCI secondary to ischemic vascular damage, have demonstrated an increase in low frequency power and a decrease in high frequency power (Nagata, 1988; Nagata et al., 1989). Other studies of subjects with cognitive decline have identified an increase in theta relative power and a decrease in gamma relative power (Moretti et al., 2007a,b, 2009). Recently, Sheorajpanday et al reported about EEG global power in subcortical VCI, no dementia independently predicts vascular impairment, and brain symmetry index reflects severity of cognitive decline (Sheorajpanday et al., 2014). Although in this study, NAT was not compared with other markers of EEG global power or an asymmetry-based EEG marker, we have previously demonstrated a decrease of the PRI in VCI patients compared with normal controls, namely an increase in low frequency power and a decrease in high frequency power (Shibata et al., 2014). In the present study, we confirmed a decrease of the PRI in severe atherosclerosis group D compared

with normal control group A. Further investigations are needed in order to widen the clinical applicability of EEG markers, that is, what is a best practical and cost-effective analysis for detecting MCI among many EEG markers of PRI, EEG global power, asymmetry-based EEG marker, NAT (sNAT, vNAT), and so on.

Carotid atherosclerosis might act as a marker of intracerebral and generalized atherosclerosis and small vessel disease, and has been associated with increased cognitive decline. Some studies have suggested that stenosis of the ICA may be an independent risk factor for cognitive impairment (Rao, 2001; Johnston et al., 2004; Mathiesen et al., 2004). High-grade stenosis of the ICA may be associated with MCI, even without evidence of infarction on MRI (Sztriha et al., 2009). In a large cohort study, high-grade stenosis was seen as an important predictor of cognitive decline (Johnston et al., 2004). Several

vertical line indicates 95% confidence intervals. \*P < 0.05, \*\*P < 0.01.

population-based studies of elderly subjects (aged > 65 years) have found associations between carotid IMT and subsequent cognitive decline (Haan et al., 1999; Sander et al., 2010). However, the pathophysiology of VCI in carotid atherosclerosis without evidence of infarction on MRI is unclear (Mathiesen et al., 2004). In a previous study using magnetoencephalography (MEG), a theta rhythm (6–8 Hz) over parieto-temporal areas, which was separated from a occipital alpha rhythm, appeared in patients with internal carotid artery occlusive disease (Seki et al., 2005). Although conventional EEG in general may not be suitable to separate the upper theta rhythm of 6–8 Hz from the occipital alpha rhythm of 8–12 Hz, NAT might detect the increase of the upper theta frequency bands (more precisely, from 6.3 to 7.8 Hz) and the decrease of the alpha frequency bands over temporo-occipital areas in the elderly with increased atherosclerosis, which of regions might be in part corresponding to the characteristic findings of parietotemporal upper theta activity (6–8 Hz) measured by MEG (Seki et al., 2005). Although another EEG study (Hsiao et al., 2016) pointed out that a carotid stenosis <50% did not alter theta (4–8 Hz) oscillations, the present study suggests that, at a specific frequency band from alpha to upper theta range, a subtle EEG alternation might appear in the elderly at an early stage of atherosclerosis (ICA–IMT ≥ 1.9 mm) before misery perfusion. Therefore, EEG markers included MEG might be more useful for detecting subtle cognitive decline, rather than MRI and conventional visual EEG analysis. To detect subtle cognitive decline with alternation of EEG, prospective studies are needed to investigate the precise association between EEG markers included MEG and several neuropsychological tests (e.g., Wechsler Adult Intelligence Scale -III, Raven's progressive matrices, Rey–Osterrieth complex figure test, and Montreal Cognitive Assessment) (Larner, 2012; Kirkpatrick et al., 2014; Sheorajpanday et al., 2014).

In general, EEG oscillations in the alpha and theta band reflect cognitive and memory performance (Klimesch, 1999). A recent study confirms the major role of the interplay of theta (5–7 Hz) and alpha (8–12 Hz) frequency in the cognitive impairment, that is, the local compensation in the baseline activity at a theta and alpha frequency range (Abuhassan et al., 2014). The structure of the model suggests that cortical oscillations respond differently to compensation mechanisms in the cognitive impairment. In the present study, changes of characteristics distributions regarding sNAT in the elderly with increased IMT have appeared at specific frequency ranges (anchor frequencies) of theta (7.8 Hz) and alpha (10.9 Hz), suggesting an insufficient interplay between a theta and alpha frequency band in the Default Mode Network (DMN). Similarly, changes of characteristics distributions regarding vNAT in the elderly with increased IMT have appeared at specific frequency ranges of theta (7.8 Hz) and alpha (9.3 Hz) (anchor frequencies), suggesting that disruptions of a balance in the DMN might be projected on the mapping of vNAT through a spectrum steepness on EEG.

Although NAT might detect abnormal cortical neuronal activity in VCI patients, our results should be interpreted with caution based on the following limitations. One limitation is that it seems to be very far away from understanding a concept and getting a clinical merit of sNAT and vNAT markers. It is easier to understand a concept of sNAT than that of vNAT, because sNAT abnormality in the power partition over the spectrum of EEG signals (i.e., hyperspectrum intensity or hypospectrum intensity) is partially similar to PRI (see Supplementary Material). On the other hand, it is likely to confuse a concept of vNAT characterized by a ratio of power spectrum between the adjacent power component, with a well-known EEG coherence which indicating the spectral correlation between electrodes. Although, at this time, we have no obvious findings about the relation between vNAT and the well-known EEG coherence, Musha thought that the coherence causes spiky variations of the PS, suggesting a relationship between EEG coherence and spectrum steepness in vNAT. When vNAT is larger or smaller than that of the NLc, the collective neuronal activities are in the undersynchrony or oversynchrony, in other words, power spectrum blur or power spectrum sharpness, respectively. In case of flat variations or gentle gradient of PS (i.e., power spectrum blur), the collective randomly activated neurons have no modulation and no meaningful biological signals are transmitted, because the neuronal activities are generated at random. In case of spiky variations or steep gradient of PS at a specific frequency band (i.e., power spectrum sharpness), the collective neuronal activities are partly coherent and partly random, and some signal contents are transmitted through the collective neuronal activities. However, not to confuse vNAT with the well-known EEG coherence, we would like to change a mode of expression from the EEG coherence to EEG spectrum steepness regarding vNAT in this paper. Therefore, a precise relationship between vNAT marker and EEG coherence needs to be investigated. Second, EEG analysis is characterized by low spatial resolution (several centimeters) when compared to structural MRI. However, NAT includes 10 frequency bands ranging from 4.7 to 18.8 Hz, which might convey peculiar physiological information on cortical activity beyond MRI. Therefore, we would like to investigate the integration of brain structure based on MRI with brain function based on EEG at a specific frequency range (e.g., 7.8 Hz). Anatomical MRI and functional EEG might provide complementary information into the process of MCI in future.

The prevention of carotid atherosclerosis could protect against vascular cognitive decline, but properly designed intervention studies are needed to demonstrate whether treatment of carotid atherosclerosis could lower the risk of cognitive decline in people without prior cerebrovascular disease. In previous studies, the highest quintile of carotid IMT was associated with dementia risk (Van et al., 2007; Wendell et al., 2012). Although ICA–IMT must reach a criterion thickness to become predictive of dementia from large-scale epidemiologic investigations, the present study did not prove that the abnormality of NAT is an independent risk factor for cognitive decline in future. Therefore, longitudinal studies are needed to examine these associations between EEG markers and ICA–IMT, and neuropsychological tests in the elderly with subclinical carotid atherosclerosis. In a first checkup for cognitive decline in community-dwelling elderly people, a combination with neuropsychological test, ultrasonography, and EEG might be useful for screening subclinical cognitive impairment.

Electroencephalographic (EEG) analysis has practical advantages for neuronal activity assessment because it is inexpensive, portable, noninvasive, sensitive to MCI, and it provides direct, real-time physiological information regarding 4–8 Hz (theta), 8–13 Hz (alpha), 13–30 Hz (beta) frequency range. To widen the applicability of EEG for detecting MCI, reliable, standardized, and user-friendly methods should be needed and developed. NAT could provide information on pathophysiology at a specific frequency band, and asymmetrical visual information of cortical neuronal activity to assess cognitive decline. The present study suggests that physiological aging accompanied by EEG alpha (8–13 Hz) decreasing (Rossini et al., 2007) and MEG theta (6–8 Hz) increasing over the posterior parietal and occipital areas (Puligheddu et al., 2005) might be affected by a increased atherosclerosis to some extent. Bamidis have proposed the neuroscience of physical and cognitive interventions in aging (Bamidis et al., 2014). Further improvements in the NAT system are needed to detect prior cognitive decline in aging and support a monitor of physical and cognitive interventions to maintain a healthy brain.

#### CONCLUSION

Community-dwelling elderly subjects in the upper quintile for ICA–IMT (≥1.9 mm) were at greater risk of an EEG change reflecting VCI as assessed by NAT. Our study highlights the importance of early intervention for carotid atherosclerosis to minimize the risk of an EEG change that might be related to subsequent VCI.

#### ETHICS STATEMENT

The ethical committees in the University of Tsukuba. All participants signed a consent form approved

#### REFERENCES


by the ethical committees in the University of Tsukuba.

## AUTHOR CONTRIBUTIONS

The authors contributed to this manuscript in the following manner: study design (TS, TM, YK, HM, and TS), data acquisition and analysis (YK, MT, HM, and KN), interpretation of results (MK, YH, NK, and SK). All authors contributed to revise and approve the final version of the manuscript and agree to be accountable for this work.

### FUNDING

Part of the present study was financially supported by SENTAN, JST (Japan Science and Technology Agency).

#### ACKNOWLEDGMENTS

We acknowledge collaboration of the members involved in the TONE epidemiological project and the Tsukuba project, in particular, Masanori Ishikawa. We thank Kazuo Ogino, chairman and CEO of Nihon Koden Corporation, and Kaoru Imajo, its deputy senior manager, for technical support.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2017.00216/full#supplementary-material

Supplementary Figure1 | Two normalizations for understanding sNAT figuratively. Normalization of sNAT is often likened to a photograph processing. (a) PS, (b) NPS, (c) sNAT corresponds to (d–f), respectively. (d) means a original subject (e.g., ship) in a frame, (e) means the adjustment of the size to frame the subject, and (f) means that the centroid of the subject is moved to the center of the frame.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Shibata, Musha, Kosugi, Kubo, Horie, Kuwayama, Kuroda, Hayashi, Kobayashi, Tanaka, Matsuzaki, Nemoto and Asada. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Investigating Focal Connectivity Deficits in Alzheimer's Disease Using Directional Brain Networks Derived from Resting-State fMRI

Sinan Zhao<sup>1</sup> , D Rangaprakash1, 2, Archana Venkataraman<sup>3</sup> , Peipeng Liang4, 5, 6 \* and Gopikrishna Deshpande1, 7, 8 \*

*<sup>1</sup> AU MRI Research Center, Department of Electrical and Computer Engineering, Auburn University, Auburn, AL, United States, <sup>2</sup> Department of Psychiatry and Biobehavioral Sciences, University of California, Los Angeles, Los Angeles, CA, United States, <sup>3</sup> Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, United States, <sup>4</sup> Department of Radiology, Xuanwu Hospital, Capital Medical University, Beijing, China, <sup>5</sup> Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing, China, <sup>6</sup> Key Laboratory for Neurodegenerative Diseases, Ministry of Education, Beijing, China, <sup>7</sup> Department of Psychology, Auburn University, Auburn, AL, United States, <sup>8</sup> Alabama Advanced Imaging Consortium, Auburn University and University of Alabama Birmingham, Auburn, AL, United States*

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Karl Friston, University College London, United Kingdom Luca Cecchetti, IMT School for Advanced Studies Lucca, Italy*

\*Correspondence:

*Peipeng Liang p.p.liang@163.com Gopikrishna Deshpande gopi@auburn.edu*

Received: *17 February 2017* Accepted: *15 June 2017* Published: *06 July 2017*

#### Citation:

*Zhao S, Rangaprakash D, Venkataraman A, Liang P and Deshpande G (2017) Investigating Focal Connectivity Deficits in Alzheimer's Disease Using Directional Brain Networks Derived from Resting-State fMRI. Front. Aging Neurosci. 9:211. doi: 10.3389/fnagi.2017.00211* Connectivity analysis of resting-state fMRI has been widely used to identify biomarkers of Alzheimer's disease (AD) based on brain network aberrations. However, it is not straightforward to interpret such connectivity results since our understanding of brain functioning relies on regional properties (activations and morphometric changes) more than connections. Further, from an interventional standpoint, it is easier to modulate the activity of regions (using brain stimulation, neurofeedback, etc.) rather than connections. Therefore, we employed a novel approach for identifying focal directed connectivity deficits in AD compared to healthy controls. In brief, we present a model of directed connectivity (using Granger causality) that characterizes the coupling among different regions in healthy controls and Alzheimer's disease. We then characterized group differences using a (between-subject) generative model of pathology, which generates latent connectivity variables that best explain the (within-subject) directed connectivity. Crucially, our generative model at the second (between-subject) level explains connectivity in terms of local or regionally specific abnormalities. This allows one to explain disconnections among multiple regions in terms of regionally specific pathology; thereby offering a target for therapeutic intervention. Two foci were identified, locus coeruleus in the brain stem and right orbitofrontal cortex. Corresponding disrupted connectivity network associated with the foci showed that the brainstem is the critical focus of disruption in AD. We further partitioned the aberrant connectomic network into four unique sub-networks, which likely leads to symptoms commonly observed in AD. Our findings suggest that fMRI studies of AD, which have been largely cortico-centric, could in future investigate the role of brain stem in AD.

Keywords: Alzheimer's disease, functional MRI, effective connectivity, disease foci, brain stem, orbitofrontal cortex

## INTRODUCTION

Alzheimer's disease (AD) is a progressive neurodegenerative disorder with a long pre-morbid asymptomatic period (Caselli et al., 2004) which affects millions of elderly individuals worldwide (Blennow et al., 2006). The disease is initially characterized by the presence of neuronal and synaptic loss, β-amyloid (Aβ) production which results in the formation of intracellular neurofibrillary tangles and senile plaques (Buerger et al., 2006), thereby resulting in memory loss, cognitive decline, etc. Structural and functional decline are inevitable with age and the existing treatment options for AD are highly limited. Therefore, determining neural aberrations underlying AD are an important step in addressing this challenge.

Resting-state functional magnetic resonance imaging (RSfMRI) is a promising neuroimaging technique that can non-invasively characterize underlying brain networks. This technology has been widely used to identify biomarkers of AD based on brain network alterations (Wang et al., 2007; Agosta et al., 2012; Sui et al., 2015). Seed-based approaches (Fox et al., 2009), independent components analysis (ICA) based approaches (Lee et al., 2015) and graph theory (Zhang et al., 2011) have been the three primary methods used in the study of resting-state functional connectivity (FC) in the brain. The seed-based approach involves predefining a region of interest (ROI) and extracting the BOLD signal from it; then a map of FC is obtained by calculating the cross-correlation between the time series extracted from the seed ROI and all other voxels in the brain. Previous studies in AD employing seedbased FC revealed decreased connectivity between the posterior cingulate cortex seed and regions spread across the whole brain in subjects with AD compared to healthy aging, with the Default Mode Network (DMN) being the most affected system (Zhang et al., 2009; Dennis and Thompson, 2014). Rather than define prior seeds, the ICA approach is model-free, which identifies independent components or co-activation networks throughout the brain. Damoiseaux et al. (2012) examined the components corresponding to the DMN for AD patients, and found significantly decreased FC in the posterior DMN and increased connectivity in ventral and anterior DMN in the AD group. Graph theoretic analysis is typically performed using FC matrices, revealing the topological properties and organization of the underlying brain network. For example, Brier et al. (2014) found that AD impacted the clustering coefficient and modularity in resting-state networks before the onset of the symptoms, suggesting that there might be a network-level pathology even in the preclinical stage. In summary, a profile of decreased connectivity has been consistently observed in AD.

However, most of the existing works on connectivity analyses have relied on FC or co-activation patterns, the literature on directed or effective connectivity (EC) patterns in AD is comparatively limited (more on this in the next paragraph). It is noteworthy that synchronization and causality in fMRI time series both represent distinct mechanisms in the brain (Friston, 2011), hence investigating EC aberrations in AD deserves attention. Motivated by this, we employed EC modeling to investigate aberrations in causal relationships between brain regions in AD. EC is often obtained using either of the two popular approaches, Granger causality (GC) (Granger, 1969; Deshpande et al., 2008, 2010a) and dynamic causal modeling (DCM) (Friston et al., 2003). DCM is highly dependent on prior assumptions concerning the underlying connectomic architecture and is therefore not generally considered suitable for analyses of large graphs. On the other hand, GC is a data-driven approach that does not need a predefined model (Deshpande et al., 2012; Sathian et al., 2013; Grant et al., 2014; Kapogiannis et al., 2014; Lacey et al., 2014; Wheelock et al., 2014; Chattaraman et al., 2016). Recent developments have demonstrated that GC is a viable technique for obtaining EC networks from fMRI data (Katwal et al., 2013; Wen et al., 2013). Therefore, in this study, we used a GC-based analysis framework. Strictly speaking, GC measures directed functional connectivity because it does not appeal to an underlying model of causal influences. In other words, GC tests for temporal precedence, thereby endowing functional connectivity with a direction. However, to emphasize the distinction between directed and non-directed connectivity, we will refer to our GC measures as effective connectivity (see Friston et al., 2013) for further discussion on this issue).

There have been several studies investigating EC-related aberrations in AD (Liu et al., 2012; Li et al., 2013; Chen et al., 2014; Zhong et al., 2014). These studies have reported distributed increases as well as decreases in directed relationships among brain regions in AD compared to healthy controls. However, these studies performing conventional GC analysis assume connectivity to be stationary over time, wherein only one connectivity value is obtained from the whole scan (Hampstead et al., 2011; Krueger et al., 2011; Lacey et al., 2011; Preusse et al., 2011; Sathian et al., 2011; Strenziok et al., 2011). However, connectivity, specifically the non-directed FC, has been shown to be non-stationary across time (Chang and Glover, 2010; Hutchison et al., 2013). Recent works suggests that connectivity varies over time, and that the temporal variability of connectivity is sensitive to human behavior in health and disease (Garrett et al., 2013; Jia et al., 2014; Rashid et al., 2016; Rangaprakash et al., 2017). Therefore, in addition to studying the conventional static effective connectivity (SEC), we also estimated dynamic effective connectivity (DEC; Grant et al., 2015; Hutcheson et al., 2015; Bellucci et al., 2016; Feng et al., 2016; Hampstead et al., 2016) from the resting-state fMRI data acquired from participants with AD as well as healthy controls (HC).

Traditionally, univariate statistical tests are performed for analyzing connectivity differences in population studies. Based on the statistical score, connectivity paths that differ from HC are ascertained. However, it is not straightforward to interpret such connectivity results, because traditionally our knowledge of brain functioning relies more on region-based properties (activations and morphometric changes) than connectivities. Further, from an interventional standpoint, it is easier to modulate the activity of brain regions (using brain stimulation, neurofeedback, etc.) rather than connections. With these viewpoints, Venkataraman et al. (2013) recently introduced a technique for identification of focal regions of functional disruption based on non-directed FC differences between populations. In this work, we extend this technique for identifying focal regions of disruption based on static as well as dynamic directed/effective connectivity aberrations in AD compared to HC.

We constructed brain networks using strength (SEC) and temporal variability (variance of DEC [vDEC]). After certain modifications to the connectivity measures, we fed them into the foci-identification model to obtain disrupted foci. The foci obtained independently from SEC and vDEC networks were then overlapped (intersection) to identify the common foci which exhibited impairments in both static and time-varying EC. Reduced temporal variance in dynamic connectivity is often associated with psychiatric disorders (Miller et al., 2016; Rangaprakash et al., 2017), and a relatively low variability of connectivity has been associated with poor behavioral performance in healthy individuals (Jia et al., 2014). Recall that a profile of decreased static connectivity has been consistently found in AD as discussed above. Taken together, we hypothesized that AD is characterized by dysfunctional disease foci, and that these foci are associated with connectivity paths that exhibit lower strength (SEC) as well as lower variability (vDEC) of effective connectivity.

#### MATERIALS AND METHODS

#### Participants

Data used in this study were obtained from the ADNI database (http://www.loni.ucla.edu/ADNI). Resting state fMRI data of 30 participants diagnosed with Alzheimer's disease (AD), along with 39 matched healthy controls (HC) were obtained through ADNI-2 cohort. Participants in this study were recruited between 2011 and 2013 through the ADNI-2 protocol, and we selected participants who had completed both 3D MPRAGE and restingstate fMRI data. Functional MRI data were obtained from a 3.0 Tesla Philips MR scanner with repetition time (TR) = 3,000 ms, echo time (TE) = 30 ms, flip angle (FA) = 80 degrees, field of view (FOV): RL (right-left) = 212, AP (anterior-posterior) = 198.75 mm, FH (foot-head) = 159 mm, voxel size: RL = 3.3125 mm, AP = 3.3125 mm, slices = 48, thickness = 3.3125 mm. 140 temporal volumes were acquired for each participant in a single scanning session. All data available from the ADNI database was acquired in accordance with the recommendations of local IRBs with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by local IRBs. More specific information can be obtained from the ADNI website (http://www.loni.ucla.edu/ADNI). The data was subjected to a standard resting-state preprocessing pipeline using SPM12 (Friston et al., 1995) and DPARSF toolboxes (Chao-Gan and Yu-Feng, 2010), including slice timing correction, realignment and motion correction, normalization to MNI space, and spatial smoothing with a Gaussian kernel of 4 × 4 × 4 mm<sup>3</sup> full width at half maximum (FWHM). Six rotation and translation parameters were first tested individually. Except rotation in Y axis (P < 0.05), there were no significant differences between the groups (P > 0.05). Then, all the six head motion parameters were aggregated into a single metric (i.e., framewise displacement), and no significant differences in framewise were found between the groups (P > 0.05). Nuisance variables such as the mean white matter signal, mean cerebrospinal fluid signal, and six head motion parameters were regressed out of the BOLD time series. It should be noted that band-pass filtering was not performed during pre-processing since it will likely impact deconvolution. Mean time series were extracted from 200 functionally homogeneous ROIs identified via spectral clustering (Craddock et al., 2012).

#### Connectivity Analysis

SEC was obtained using Granger causality (GC) analysis. However, before GC analysis is performed, it is necessary to acknowledge the impact of hemodynamic response function (HRF) on connectivity modeling, which is known to vary across different regions within a participant, as well as vary across participants (Handwerker et al., 2004). Previous studies have shown that results obtained by using GC analysis on HRFcorrupted fMRI data can be confounded by the variability of the HRF (David et al., 2008; Deshpande et al., 2010b). Hence, a blind deconvolution technique, proposed by Wu et al. (2013), was employed to minimize the non-neural variability of the HRF and estimate the latent neuronal time series from the observed fMRI data. In brief, the resting-state data was modeled as spontaneous event-related data (Tagliazucchi et al., 2012), and the HRF of each voxel was estimated by Wiener deconvolution (Glover, 1999). The estimated neural time series were then used in further GC analysis.

The underlying concept of GC is that a directed causal influence from time series X to time series Y can be inferred if the past values of time series X improves the prediction of the present and future values of time series Y (Granger, 1969). Let q time series X(t) = [x1(t), x2(t),...,xq(t)] be the latent neural time series obtained after HRF deconvolution of selected ROI fMRI time series, with q being 200 ROIs in this study. Then the multivariate autoregressive (MVAR) model with order p is given by

$$X(t) = A(1)X(t-1) + A(2)X(t-2) + \dots + A(p)X(t-p) + E(t) \tag{1}$$

Where A(1)...A(p) are the model parameters, and E(t) is the vector of the residual error.

To remove the zero-lag correlation effect (i.e., ignore coactivations), the time series were input into a modified multivariate autoregressive model which included the zero-lag term used by Deshpande et al. (2009) shown as follows:

$$X(t) = A'(0)X(t) + A'(1)X(t-1) + \dots + A'(p)X(t-p) + E(t) \tag{2}$$

The diagonal elements of A ′ (0) were set to zero, to model only the instantaneous cross-correlation rather than zero-lag autocorrelation. The off-diagonal elements of A ′ (0) corresponded to the zero-lag cross-correlation (Deshpande et al., 2009). It is to be noted that the coefficients in Equation (1) A(1),...A(p) would not be the same as A ′ (1)...A ′ (p) as in Equation (2), because the modified zero-lag term affects other coefficients since it removes the zero-lag cross correlation effects from them. Accordingly, the correlation-purged granger causality (CPGC) from time series i to time series j was obtained using the following equation

$$\text{CPGC}\_{\vec{\eta}} = \sum\_{n=1}^{p} \left( a'\_{\vec{\eta}} \right)^2 \langle n \rangle \tag{3}$$

Where a ′ ij are the elements of A'. It is well-known that the coupling among brain areas is time-varying and contextsensitive. Indeed, the most interesting parameters of dynamic causal models are the fluctuations in effective connectivity (induced by experimental manipulations or time). In recent years, the functional connectivity (resting state) community has dubbed these fluctuations in coupling as "dynamic functional connectivity." In our work, we characterized DEC using a temporally adaptive modified MVAR model:

$$X(t) = A'(0, t)X(t) + A'(1, t)X(t - 1) + \dots + A'(p, t)X(t - p)$$
 
$$+ E(t) \tag{4}$$

In this model, the coefficients A ′ (p) were allowed to vary over time, thus "dynamically" estimating EC.

The parameters A ′ (n,t), n = 0,...,p were estimated in a Kalman filter framework using variable parameter regression (Arnold et al., 1998; Büchel and Friston, 1998). The Kalman filtering is a recursive process, where new information is added when it arrives. Thus, estimates taken from early steps are less reliable compared to later ones. A forgetting factor (FF) is introduced to circumvent this problem by taking recent past Kalman filter estimates into account during current estimation in order to control smoothness and enhance stability. The forgetting factor was determined by minimizing the variance of estimated error energy (Havlicek et al., 2010) and was found to be equal to one in our study. In brief, Kalman filtering treats the underlying MVAR coefficients as slowly fluctuating states. This enables the estimation of time varying directed connectivity that was used for subsequent modeling at the between-subject level. The DGC is estimated as:

$$DGC\_{i\hat{\jmath}}(t) = \sum\_{n=1}^{p} \left( a'\_{i\hat{\jmath}}(n, t) \right)^2 \tag{5}$$

Where DGCij (t) is the dynamic Granger causality value from time series i to time series j at time point t. Given that the neural delays of interest are of the order of a TR or less (Deshpande et al., 2013), and that previous literature supports using a first order model to capture most relevant causal information (Deshpande and Hu, 2012), we employed a first order model for estimating both SEC and DEC in this work.

#### Identification of Disease Foci

Connectivity studies often report aberrations in functional connections between brain regions. While this is useful, it does not provide a comprehensive characterization of the underlying connectomics. First, it is likely that several aberrations in connectivity are the after-effects arising from disruptions in certain focal brain regions. Second, our knowledge about brain functioning is centered on functions of regions rather than connections. Therefore, it is advantageous to identify certain focal regions of disruption using connectivity data. Thus in this study, we sought to identify diseased foci in AD. A recent study introduced a novel technique for the identification of disease foci (Venkataraman et al., 2013) based on non-directed FC differences between populations. Here we generalize this technique to the identification of diseased foci from effective connectivity as well as dynamic connectivity data.

The model proposed by Venkataraman et al. (2012) considers the connectivity measure (C M ij for HC group and P M ij for the AD group) as a noisy observation of the latent connectivity (C L ij for HC group and P L ij for the AD group). The model is illustrated in **Figure 1** and consists of several parts.

The first part defines a binary indicator vector that selects disrupted regions, and a binary graph characterizes corresponding abnormal connectivity. Let N be the total number of regions in the brain being considered. The model assumes a the random variable R = [R1,...,RN] is a binary vector (i.e., brain regions are either healthy with R<sup>i</sup> = 0 or disrupted with R<sup>i</sup> = 1, where i = 1 .. N) indicates the state of each region in the brain. Elements of R follow an independent, identically distributed (i.i.d.) Bernoulli distribution model Q b (R) where Q( . ) denotes the posterior distribution and superscript b indicates a Bernoulli distribution. Then, an underlying binary graph G which characterizes the network of abnormal connectivity can be defined as follows: a connection between two healthy regions is always healthy with probability equal to 1, a connection between two disrupted regions is always abnormal with probability equal to 1, and a connection between a healthy region and a disrupted region is abnormal with probability η. The second part specifies the latent connectivity for controls (C L ) as a tri-state variable from a multinomial distribution with parameter π<sup>k</sup> (k denotes three different states), positive connectivity with probability

the variables.

π1, little or no functional connection (0) with probability π0, and negative connectivity with probability π−1. Given the binary graph G and latent connectivity for controls C L , the tristate latent connectivity of the AD population can be defined. Specifically, the latent connectivity from the control group C L ij equals to P L ij with probably ǫ if the binary graph connection between regions i and j is abnormal, C L ij equals to P L ij with probably 1 − ǫ if the connection between regions i and j is healthy. The third part characterizes the observed connectivity measures C M ij and P M ij as Gaussian random variables whose mean and variance (µ and σ) depend on the value of C L ij and P L ij. Then, the joint likelihood of all configurations of latent connections between regions can be modeled as an 9-state multinomial distribution model Q <sup>m</sup>(C, P) (superscript m denotes that Q( . ) is a multinomial distribution).

The model in Venkataraman et al. (2013) was applied in the case of functional connectivity, i.e., the Pearson's correlation coefficient between regions. However, EC is not a bounded measure, a small number of outliers is to be expected. In our EC data, we found a small portion of connectivity values which were >1 or < −1 (0.3%), wherein these outliers indicate stronger causal information flow between regions. To maintain the importance of those stronger effective connections and minimize its negative impact on model evaluation, inverse Fisher transformation was used to render the EC values as a bounded measure within [−1 1]. For the variance of dynamic EC, the latent tri-states of variance of connectivity vFij can be considered as follows: little variability or stationary connection, modest variability and strong variability. It is to be noted that static FC is direction-less, hence only the upper or lower triangle of the symmetric connectivity matrices were needed to fit the model in Venkataraman et al. However, in our case, both SEC and vDEC are directed with asymmetric connectivity matrices, and hence the whole matrices were used in the model. Taken together, these modifications permitted the model to be applied to both static and dynamic EC.

After initiating the prior parameters (such as the Bernoulli prior for binary state vector R, prior for latent connectivity for controls π<sup>k</sup> , etc.) for the model, a variational expectation maximization (EM) algorithm (Dempster et al., 1977) was adopted for estimating the latent connectivity and model parameters from the observed connectivity measures (C <sup>M</sup> and P <sup>M</sup>). Technically, we inverted the (between subject) model of disconnection using variational Bayes. This scheme is formally similar to an EM algorithm that uses a variational update for all the factors of an approximate posterior. These included an approximate posterior distribution over model parameters (π<sup>k</sup> , η, ǫ, µ, and σ), latent connectivity for both groups of subjects [Q <sup>m</sup>(C, P)] and regional pathology [Q b (R)]. In brief, this variational scheme optimizes the sufficient statistics of each marginal distribution or density with respect to variational free energy (FE), under the expected values of the remaining factors. The variational EM alternates between updating the latent posterior distribution and estimating the nonrandom model parameters. Convergence was based on the relative change in free energy of the model of <10−<sup>4</sup> between consecutive iterations. Disrupted focal regions and latent abnormal connectivity would then be identified from the posterior probabilities for each region and each connection. **Figure 2** illustrates the flow chart of the algorithm.

The significance of the resulting foci was estimated using nonparametric permutation tests. Specifically, the group label of each participant was randomly permuted for 1,000 times. For each permutation, we fit the data to the model and obtained the posterior probability of disrupted foci for each region. This provided an empirical null distribution from which the p-value of the significance was obtained. The method also identified the affected connections associated with the disrupted foci. Among such connections, we retained those that were also in accordance with our hypothesis (paths exhibit lower SEC, as well as lower vDEC of effective connectivity in AD compared to healthy controls with a threshold of p < 0.05).

### RESULTS

We identified two disrupted foci which were common to both SEC and vDEC networks: (1) Locus Coeruleus (LC) in the Brainstem (p = 0.003 for SEC and 0.006 for vDEC), (2) Right orbitofrontal cortex or R OFC (p = 0.007 for SEC and 0.002 for vDEC). Disrupted connectivity paths associated with these foci exhibited higher strength and larger temporal variability in HC as compared to AD (in accordance with our hypothesis). Furthermore, they exhibited a unique pattern of disrupted connectivity—those associated with the LC in the brain stem emanated from it, while connectivity paths associated with R OFC converged onto it (**Figure 3**).

Five of the ten connectivity paths emanating from the LC resulted in connectivity paths terminating in the R OFC, with four of these five paths being indirect pathways via the L MFG, L MTG, R MOG, and L Calcarine, and one path being a direct connection from LC to R OFC. All connectivity paths exhibited lower SEC and lower vDEC in AD compared to HC.

Further clarity on the corresponding aberrant connectomic network was obtained by partitioning the network into four unique subnetworks: (**Figure 4A**) LC-PFC working memory system, (**Figure 4B**) LC-PHG emotional memory system, (**Figure 4C**) LC-visual cortex sensory system, and (**Figure 4D**) LC-MTG language system. Note that this partitioning is based on different functions performed by the locus coeruleus norepinephrine system and is not based on any analytical strategy. Taken together, the disruption of these networks likely leads to working memory deficits, difficulties in processing emotional memories, and several other symptoms commonly observed in those with AD. The relevance of these subnetworks to AD pathology are discussed in detail in the next section.

### DISCUSSION

In this study, we estimated static and dynamic measures of directed influences between 200 ROIs covering the entire brain in both AD and HC participants taken from the ADNI database. SEC and vDEC connectivity data were fed into a

probabilistic model to identify regions with focal connectivity deficits in AD, with the hypothesis that connections associated with those regions would be weaker in strength and lower in temporal variability (i.e., rigid) in AD. We identified two such foci, brain stem and orbitofrontal cortex, which were affected significantly by the disease. The aberrant connections emanating from LC suggested a widespread dysregulation originating from the brainstem, part of which terminated into the other focus (orbitofrontal cortex).

Interestingly, all connectivity paths corresponded with the directed influence of the LC (in the brain stem) on mostly cortical (and few sub-cortical) regions. This corroborates with previous studies that have shown progressive damage (Kienzl et al., 1999) in the brain stem during early periods of AD. Further, LC in the brain stem is the largest repository of Norepinephrine (NE) in the human brain (Herregodts et al., 1991). Noradrenergic neurons in LC have projections to several parts of the brain including olfactory, limbic, prefrontal, and other cortical regions (Sara, 2009; Sara and Bouret, 2012). NE is known to suppress neuroinflammation (Weinshenker, 2008). This purported role has been hypothesized to be a protective factor against AD. In fact, Heneka et al. (2010) showed that NE stimulation of mouse microglia suppressed Aβ-induced cytokine and chemokine production and increased microglial migration and phagocytosis of Aβ. Induced degeneration of the brain stem increased the expression of inflammatory mediators in amyloid precursor protein (APP)-transgenic mice and resulted in elevated Aβ deposition. Kelly et al. (2017) suggesting that the decrease of NE in the brainstem facilitates the inflammatory reaction of microglial cells in AD and impairs microglial migration and

phagocytosis, thereby contributing to reduced Aβ clearance. The Aβ is the critical initiating event in AD, starting with the aberrant clearance of Aβ-peptides followed by consecutive peptide aggregation and disruption of neural activity (Selkoe, 2002). Moreover, a post-mortem study has found significant volume decreases in the LC during AD progression, highlighting the importance of this region in AD (Theofilas et al., 2016). These findings indicate that the depletion of NE in LC is an etiological factor in the development of MCI and progression to AD. The studies discussed above provide some basis for the important

role of brainstem in AD. Further, an animal study has found that boosting NE transmission can lead to increased functional connectivity (Guedj et al., 2016), suggesting that the reduction of NE could potentially result in lower connectivity between LC and cortical regions.

Several previous studies have suggested that OFC may be important for understanding the mechanisms for putative spreading of AD pathology in the brain (Van Hoesen et al., 2000; Sepulcre et al., 2013). Robust correlation has been found between Aβ deposition levels and volume in the orbitofrontal area (Ishibashi et al., 2014). In fact, the amyloid precursor protein (APP) gene contains the sequence for the Aβ peptide, which is concentrated in the senile plaques (SPs) (Cras et al., 1991). During AD progression, the SPs appear first in the orbitofrontal and temporal cortices and later extend to the whole cortex (Braak and Braak, 1999). Further, SPs and Aβ deposition has been associated with reduced connectivity at the synaptic level (Yeh et al., 2011), suggesting a potential mechanism that might link SPs and Aβ deposition with directed connectivity estimated from fMRI. While we discuss the role of temporal regions later in this section, the findings presented above highlight the importance of the role of OFC in AD.

Connectivity paths from LC to the prefrontal cortex (PFC) in general, and OFC in specific (note that OFC is a region in the PFC), can be considered as an aberrant LC-PFC working memory system (**Figure 4A**). Given that many studies have referred to the PFC in general without specifying sub-regions, and hence we are going to use the same nomenclature in the ensuing discussion. Previous studies have indicated that NE is instrumental in enhancing working memory through actions within the prefrontal cortex (PFC). PFC underlies the encoding of task-relevant information in working memory (Baddeley, 2003), and it has been shown that damage to the noradrenergic innervation of the PFC impairs performance in working memory (Brozoski et al., 1979). The stimulation of α2-adrenergic receptors in the PFC of nonhuman primates has been shown to improve performance in working memory tasks (Li et al., 1999) while α1-adrenergic receptors impaired the working memory (Arnsten and Jentsch, 1997). α2-adrenergic receptors have a higher affinity for NE compared to α1-adrenergic receptors, thus under normal conditions, NE facilitates working memory performance via actions at α2-adrenergic receptors in general and also in the PFC. However, dysfunction in noradrenergic pathways emanating from LC may result in low PFC NE levels, affecting working memory (O'Rourke et al., 1994).

The connectivity from LC to PHG can be considered as a LC-PHG emotional and spatial memory system (**Figure 4B**). The LC-NE system modulates emotional memories, and studies have suggested that emotional memories induce the activation of LC and subsequent NE release (Weiss et al., 1980). Corticotropinreleasing hormone (CRH) receptors are known play an important role in the coordination of autonomic and electrophysiological responses associated with emotional memories (Koob and Bloom, 1985; Dunn and Berridge, 1990). CRH-immunoreactive fibers were observed in the LC, suggesting that CRH may modulate LC neuronal activity (Merchenthaler et al., 1982; Cummings et al., 1983). In fact, many studies (Valentino et al., 1983; Finlay et al., 1997; Jedema et al., 2001) have shown that CRH administered locally into the LC increases LC discharge activity and NE release in its terminal fields. Moreover, an abundant expression of CRH was found in PHG (Wong et al., 1994). The first sign of emotional memories was also observed in PHG, and was found to then gradually spread to PFC and other cortical regions (Sotiropoulos et al., 2011). On the other hand, PHG is known to be involved in spatial memory (Bohbot et al., 1998). Noradrenergic neurons within LC have widely distributed, ascending projections to the limbic system including PHG (Szabadi, 2013). Thus, the LC-NE system may help trigger the involvement of the PHG in spatial memory. An animal study has indicated that the LC-NE system is necessary for the acquisition of spatial memories (Gertner and Thomas, 2006). These evidence suggest that the decrease of NE in LC could likely cause dysregulation of the emotional and spatial memory system in the LC-PHG network.

Connectivity paths from LC to the frontal cortex, mediated by sensory visual regions, can be considered as a LC-visual sensory system (**Figure 4C**). Previous works in animal models have shown that the LC-NE system can alter receptive field properties such as velocity tuning, direction selectivity, etc. (Waterhouse et al., 1990; McLean and Waterhouse, 1994). Malfunction of the LC-visual sensory network may contribute to deficits in visual assessment (Johnson et al., 2012).

Connectivity paths from LC to the OFC mediated by MTG can be considered as a LC-MTG language system (**Figure 4D**). A previous study has shown decreased regional cerebral blood flow (rCBF) after ingestion of an α2-adrenergic agonist drug in the MTG (Swartz et al., 2000). Given that the noradrenergic system in the brain originates from LC, this suggests that there might exist a noradrenergic pathway between LC and MTG which is impaired in AD. The malfunction of the LC-MTG language system may cause language impairments often observed in AD (Ferris and Farlow, 2013; Szatloczki et al., 2015).

It is evident that most of the disrupted connectivity paths emanating from the LC in the brain stem drive OFC either directly or via other systems. OFC is known to play a critical role in memory, emotions, reward, as well as decision-making mechanisms (Rolls, 2004; Rempel-Clower, 2007). Disrupted connectivity paths that converge into the OFC were observed in three of the subnetworks, and could potentially underlie behavioral deficits in these domains.

Taken together, we identified LC in the brainstem and OFC as the foci of network disruption in AD. The dysregulation of LC-NE neurotransmission likely contributes to behavioral deficits observed in AD. In corroboration, previous literature has pinpointed the same regions (Heneka et al., 2010; Ishibashi et al., 2014) to be affected in AD. Our identification of the LC in the brain stem as the disease focus in AD supports these previous observations and suggests that functional MRI studies of AD, which have been largely cortico-centric (Dennis and Thompson, 2014; Li et al., 2014), must in future investigate the role of this structure in AD.

Previous studies have also identified some other regions to be crucial to AD pathology (Brier et al., 2014; Dai et al., 2015; Mutlu et al., 2016). In fact, our foci-identification technique did identify some of the regions reported in these papers. Specifically, we also identified parahippocampal gyrus, middle frontal gyrus, and precuneus as foci only considering DEC networks. Further, middle temporal gyrus, lateral occipital cortex and cerebellum posterior lobe were identified as foci in SEC networks. However, these regions were not identified as foci in both DEC and SEC networks. Acknowledging that previous studies reported regions as having significantly different static connectivity between the groups, in this study we only reported the foci and the associated connectomic network that were found as having impairments in both static and dynamic EC.

Next, we report a few noteworthy limitations of this work. We have based our interpretation on the efferent projections of neurotransmitters arising out of LC. We employed this logic since functional imaging studies of the brain stem (and LC) in AD are limited, with the existing literature employing functional imaging in AD being cortico-centric. However, we have not directly measured norepinephrine in the brain, as it is difficult to do so using MRI. Therefore, our results form the basis for a hypothesis regarding dysfunction in the noradrenergic pathways in AD. Future studies must employ other modalities such as positron emission tomography for in vivo imaging of noradrenergic pathways (not just NE deficits) in AD. This could potentially open up possibilities for therapeutic interventions in AD. Further, the proposed methodology of combining static as well as DEC analysis with probabilistic modeling for identifying dysfunctional foci and associated dysfunctional networks could provide novel insights into the pathophysiology of other brainbased disorders.

## AUTHOR CONTRIBUTIONS

GD and PL designed the study; AV and DR contributed analysis tools; SZ performed data analysis; All authors interpreted the results and wrote the paper.

### FUNDING

The work described in this paper was supported by a grant from the National Natural Science Foundation of China (61473196). The authors also acknowledge support from the Auburn University MRI Research Center. The funders had no role in study design, data collection, and analysis, decision to publish, or preparation of the manuscript.

## ACKNOWLEDGMENTS

Data used in this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). Investigators within ADNI contributed to design and implementation of ADNI and provided data but did not participate in analysis or writing of this report. Complete listing of ADNI investigators: http://adni.loni.usc.edu/ wp-content/uploads/how\_to\_apply/ADNI\_Acknowledgement\_ List.pdf. Data collection and sharing for this work was funded by ADNI (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for NIH (www.fnih.org). The grantee is the Northern California Institute for Research and Education, and the study is coordinated by

#### REFERENCES


Alzheimer's Disease Cooperative Study at the University of California, San Diego. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California, Los Angeles, USA.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Zhao, Rangaprakash, Venkataraman, Liang and Deshpande. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Attentional Orienting and Dorsal Visual Stream Decline: Review of Behavioral and EEG Studies

#### Evatte T. Sciberras-Lim\* and Anthony J. Lambert

Department of Psychology, University of Auckland, Auckland, New Zealand

Every day we are faced with an overwhelming influx of visual information. Visual attention acts as the filtering mechanism that enables us to focus our limited neural resources, by selectively processing only the most relevant and/or salient aspects of our visual environment. The ability to shift attention to the most behaviorally relevant items enables us to successfully navigate and interact with our surroundings. The dorsal visual stream is important for the rapid and efficient visuospatial orienting of attention. Unfortunately, recent evidence suggests that the dorsal visual stream may be especially vulnerable to age-related decline, with significant deterioration becoming evident quite early in the aging process. Yet, despite the significant age-related declines to the dorsal visual stream, the visuospatial orienting of attention appears relatively well preserved in older adults, at least in the early stages of aging. The maintenance of visuospatial orienting of attention in older adults appears to be facilitated by the engagement of compensatory neural mechanisms. In particular, older adults demonstrate heightened activity in the frontal regions to compensate for the reduced activity in the posterior sensory regions. These findings suggest that older adults are more reliant on control processes mediated by the anterior regions of the frontoparietal attention network to compensate for less efficient sensory processing within the posterior sensory cortices.

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

James Danckert, University of Waterloo, Canada Christopher S. Y. Benwell, University of Glasgow, United Kingdom

#### \*Correspondence:

Evatte T. Sciberras-Lim esci911@aucklanduni.ac.nz

Received: 16 May 2017 Accepted: 14 July 2017 Published: 27 July 2017

#### Citation:

Sciberras-Lim ET and Lambert AJ (2017) Attentional Orienting and Dorsal Visual Stream Decline: Review of Behavioral and EEG Studies. Front. Aging Neurosci. 9:246. doi: 10.3389/fnagi.2017.00246 Keywords: aging, visuospatial orienting of attention, compensation, dorsal visual stream, magnocellular pathway

### INTRODUCTION

Everything from playing tennis to simply walking down the street require the successful orienting of attention from one location to the next. This process of shifting attention is referred to as the visuospatial orienting of attention. A process which relies on an interacting network of structures, which include regions of the lateral prefrontal cortex (Corbetta and Shulman, 2002, 2011; Corbetta et al., 2008; Schall, 2009; Vandenberghe and Gillebert, 2009; Asplund et al., 2010), working in conjunction with the dorsal visual stream (Siegel et al., 2008; Lambert and Shin, 2010; Marrett et al., 2011; Capilla et al., 2014) to guide the deployment of attentional resources. The involvement of dorsal visual stream processing in visuospatial orienting is consistent with both Goodale and Milner's (1992) duplex model of vision which suggests that the dorsal visual stream is responsible for visually guided actions (including eye movements; Milner and Goodale, 2008), and with the premotor theory of attention which suggests that there is considerable overlap in the neural mechanisms responsible for the overt and covert orienting of attention (Rizzolatti et al., 1987). Unfortunately, recent evidence suggests that the dorsal visual stream may be vulnerable to age-related declines from relatively early on in the aging process (Langrová et al., 2006). This mini-review article examines the impact of these age-related declines on the visuospatial orienting of attention.

#### AGE-RELATED DECLINE OF THE DORSAL VISUAL STREAM

Recent studies suggest that the dorsal visual stream may be vulnerable to age-related atrophy and decline due to its cellular composition. The dorsal visual stream which extends from area V1 to the dorsolateral occipital cortex (Maunsell, 1987), and regions of the posterior parietal cortex (Kravitz et al., 2011), receives input primarily from the magnocellular layers of the lateral geniculate nucleus (LGN; Celesia and DeMarco, 1994; Grill-Spector and Malach, 2004). In contrast, the ventral visual stream which extends from area V1 to the inferior temporal cortex (Kravitz et al., 2011), receives input from both the magnocelluar and parvocellular layers (Celesia and DeMarco, 1994; Grill-Spector and Malach, 2004) of the LGN. Braddick et al.'s (2003) dorsal stream vulnerability hypothesis argues that, since magno cells are larger than parvo cells, and neurons with larger cell bodies and axon diameters are more susceptible to damage, magno cells are more susceptible to degeneration. In addition, losses in the magnocellular pathway may be more readily apparent as there are far fewer magno cells than parvo cells (Skottun, 2000). Thus, even when similar numbers of neurons are lost in both the magnocellular and parvocellular pathways, functional declines may be more apparent in the magnocellular pathway. Since the dorsal visual stream's primary input comes from magno cells (Celesia and DeMarco, 1994), the dorsal visual stream is in turn, more susceptible to degeneration (Braddick and Atkinson, 2011). Although Braddick's proposal of dorsal visual stream vulnerability was based on an examination of dorsal visual stream deficits in early childhood development; it does highlight the possibility that the dorsal visual stream may be more susceptible to decline in general.

Evidence of age-related dorsal visual stream declines has come from studies examining the age-related declines to neural structures. In a study by Ziegler et al. (2012), 547 participants between the ages of 19–86 years old were recruited, to perform a large-sample cross-sectional voxel-based morphometry (VBM) study to examine reductions in gray matter volume with advancing age. The results revealed that the dorsal visual stream exhibited substantially larger reductions in gray matter volume, compared to the ventral visual stream and the primary visual areas. These declines have been localized to the superior parietal cortex (Driscoll et al., 2009), as well as, the superior and inferior parietal gyri (Crivello et al., 2014; Fjell et al., 2014), all of which are crucial for the visuospatial orienting of attention (Corbetta and Shulman, 2002). Moreover, in elderly subjects, neural declines in the parietal regions (Resnick et al., 2003; Driscoll et al., 2009; Thambisetty et al., 2010; Crivello et al., 2014; Fjell et al., 2014) are compounded by a reduction in cerebral blood flow to the region (Martin et al., 1991), resulting in less efficient processing within the dorsal visual stream.

#### IMPACT OF DORSAL VISUAL STREAM DECLINES (BEHAVIORAL)

This mini-review article will focus primarily on the impact of these declines on the age-related changes to performance within the spatial cueing paradigm. In these studies, a cue elicits a shift of attention to its location (in the case of a peripheral cue; Eriksen and Hoffman, 1974; Theeuwes, 1991; Rafal and Henik, 1994; Yantis and Hillstrom, 1994; Oonk and Abrams, 1998), or to the location signaled by the cue stimulus (in the case of a central cue; Jonides, 1981; Shepherd and Müller, 1989; Cheal and Lyon, 1991; Theeuwes, 1991; Friesen et al., 2004). Subsequently, a target appears in either the location signaled by the cue stimulus (valid trial), or at an un-cued location (invalid trial). Participants are tasked with making a response to the onset of the target stimulus. The cueing benefit is indexed by a decrease in reaction times when targets appear in the location signaled by the cue stimulus and is believed to reflect the engagement of attentional resources at the target location (e.g., Jonides, 1981; Shepherd and Müller, 1989; Theeuwes, 1991; Yantis and Hillstrom, 1994; Friesen et al., 2004). Conversely, the cueing cost is indexed by an increase in reaction times when the target appears at an un-cued location, and is believed to reflect the disengagement and shifting of attentional resources from the invalidly cued location to the target location (Posner, 1980).

It appears that despite the declines to the dorsal visual stream, the visuospatial orienting of attention appears relatively well-preserved in older adults (Nissen and Corkin, 1985; Hartley et al., 1990). In spite of the slower reaction times seen in older adults, the magnitude of cueing benefit for both peripherally (Folk and Hoyer, 1992, Experiment 1; Greenwood et al., 1993; Olk and Kingstone, 2009) and centrally (Greenwood et al., 1993; Lincourt et al., 1997; Curran et al., 2001; Lorenzo-López et al., 2002; Olk and Kingstone, 2009) presented cues is similar for both younger and older adults. However, some studies have noted an increase in cueing costs following invalidly cued targets (Hartley et al., 1990; Greenwood and Parasuraman, 1994). Taken together, these results suggest that aging may be selectively associated with reduced efficiency in disengaging and shifting attentional resources from one location to the next, while the ability to effectively utilize the cues to guide the deployment of attentional resources remains intact.

One potential shortcoming of the aforementioned studies is that in these studies participants are required to make a detection or discrimination response to a target in an otherwise empty screen. In contrast, visual scenes are typically cluttered and we need to be able to rapidly identify a relevant item (target) from surrounding items (distractors). A process which has typically been studied using the visual search paradigm (Treisman and Gelade, 1980). In these studies, participants are required to identify a target that differs from surrounding items by a single feature (feature search) or by a conjunction of features (conjunction search; Treisman and Gelade, 1980). Results indicate that aging appears to selectively impair conjunction search while feature search remains intact (Plude and Doussard-Roosevelt, 1989). An elegant study by Greenwood and Parasuraman (1999) examined if the selective age-related impairment in conjunction search could in part be accounted for by an age-related reduction in the ability to flexibly expand and contract the focus of attention. To do so, they employed the use of precues that indicated the location of an upcoming target with a varying degree of precision (element-size precue: highlights a single possible location; column-size precue: highlights a column of possible locations; array-size precue: highlights a whole array of potential locations). The results indicated that age-related impairments in conjunction search can be alleviated by the use of precise and valid precues. Although this benefit was somewhat smaller in the oldest group of participants (above 76 years old). They suggest that aging impairs the ability to flexibly expand and contract the focus of attention and that elderly participants are more reliant on the precues to adjust the scale of their attention. But in the oldest participants the ability to utilize the precues to adjust the scale of attention is reduced. These findings suggest that although the visuospatial orienting of attention remains relatively well preserved in older adults, its flexibility is somewhat reduced.

#### IMPACT OF DORSAL VISUAL STREAM DECLINES (EEG)

Additionally, some studies have also employed the use of electroencephalographic (EEG) recordings in order to examine the electrophysiological correlates of attention shifts. These event-related potential (ERP) studies have most commonly focused on the P1 (80–130 ms) and the N1 (140–200 ms) components in relation to processing of target stimuli (Mangun and Hillyard, 1991; Wright et al., 1995). These studies demonstrate that shifting attention to a particular location of the visual field, increases the amplitude of P1 and N1 components (Mangun and Hillyard, 1991; Eimer, 1994; Luck et al., 1994; Mangun, 1995; Anllo-Vento et al., 1998) evoked by stimuli within the attended location. The amplification of these components appears to involve a selective enhancement of the signal to noise ratio of stimuli within the attended area, thereby strengthening the perceptual representation of stimuli located within that region (see Carrasco, 2011, for review; Heinze et al., 1990, 1994). Consequently, stimuli within the attended area are more rapidly detected; giving rise to more rapid response times when targets appear at the cued location (valid trials) compared to when targets appear at un-cued locations (invalid trials). Results from ERP studies examining the age-related changes to the underlying neural substrates of visuospatial orienting closely parallel the results of behavioral studies. These studies demonstrate that despite delayed latency of the P1 (80–130 ms) and N1 (140–200 ms) components, the augmentation of these early components elicited by attended relative to unattended targets is similar across both younger and older adults (Yamaguchi et al., 1995; Curran et al., 2001; Nagamatsu et al., 2011). With some studies showing linear correlations between the latency of ERP components and the mean reaction times of elderly participants (Li et al., 2013), which further bolsters the proposal that reductions in transmission efficiency of neural signals may account for slowed visuospatial orienting in elderly subjects (Hong and Rebec, 2012).

More recently, researchers have begun to focus on the neural activity elicited during the cue-target interval; much of this work has focused upon examining the modulation of prestimulus alpha activity along the fronto-occipital axis during the cue-target interval (Foxe et al., 1998; Worden et al., 2000; Babiloni et al., 2006; Rihs et al., 2009; Foxe and Snyder, 2011). Alpha-band desynchronization of the contralateral frontooccipital axis appears to increase perceptual sensitivity by causing a baseline shift in the sensitivity of the neurons representing the to-be-attended location (Sauseng et al., 2005; Rihs et al., 2007; Siegel et al., 2008; Capotosto et al., 2009; Kelly et al., 2009; Capilla et al., 2014), while a concurrent alpha-band synchronization of the ipsilateral fronto-occipital axis inhibits processing of unattended regions (Kelly et al., 2006; Klimesch et al., 2007; Rihs et al., 2007, 2009; Capotosto et al., 2009; Foxe and Snyder, 2011; Bengson et al., 2012). The lateralization of prestimulus alpha reflects the top-down attentional modulation of neural processing in the visual cortices, and is referred to as proactive attentional control (Braver, 2012). It is termed proactive as it reflects the ability of the attentional system to bias perceptual processing in favor of an upcoming target before it is presented. In contrast, reactive attentional control refers to resolving interference between target and potentially distracting information at later stages of the processing hierarchy (Geerligs et al., 2014).

Recent studies indicate that the level of alpha power lateralization during the cue-target interval is significantly reduced in older adults (Vaden et al., 2012; Hong et al., 2015; Li and Zhao, 2015), and this reduction is most prominent along parietal-occipital regions (Zanto et al., 2011; Deiber et al., 2013). Specifically, older adults showed significant reductions in the level of event-related synchronization of prestimulus alpha (ipsilaterally) along these sites (Karrasch et al., 2004; Deiber et al., 2010; Vaden et al., 2012), which is consistent with earlier reports suggesting that older adults face significant difficulty with distractor suppression (Gazzaley et al., 2008; Schmitz et al., 2010; Haring et al., 2013). The decreased modulation of prestimulus alpha is proposed to be the result of the overall decline in alpha power in older adults, which renders the modulation of prestimulus alpha a less efficient means of attentional control (Vaden et al., 2012; Deiber et al., 2013; Hong et al., 2015; Li and Zhao, 2015). Older adults may compensate for this deficit with stronger early engagement of motor areas (Deiber et al., 2013), and an increase in reactive control (Paxton et al., 2008; Geerligs et al., 2014) mediated by the anterior nodes of the frontoparietal attention network (De Fockert et al., 2009; Schmitz et al., 2010; Haring et al., 2013; Li et al., 2013; Geerligs et al., 2014). These findings suggest that the age-related deterioration of parietal-occipital regions impairs elderly participants' ability to engage in the proactive attentional biasing of early sensory regions and leads to an increased reliance on more reactive control strategies.

Additionally, while the majority of research into the patterns of age-related cerebral reorganization have centered on intrahemispheric patterns of reorganization, there is also extensive evidence to support inter-hemispheric patterns of reorganization (Cabeza et al., 1997; Madden et al., 1999; Tulving et al., 1994; Reuter-Lorenz et al., 2000; Cappell et al., 2010). It has been proposed that with the progressive age-related deterioration of specialized neural networks, the high metabolic costs of inter-hemispheric communication (Bullmore and Sporns, 2012; Liang et al., 2013) are outweighed by the benefits to behavioral performance (Banich, 1998; Cabeza, 2002). Although most studies of age-related asymmetry reduction have focused predominantly on the bilateral recruitment of the prefrontal cortices (Madden et al., 1999; Reuter-Lorenz et al., 2000), there is also evidence for age-related asymmetry reduction within the parietal cortices (Garavan et al., 1999). This implies that while the control of spatial attention may be strongly right lateralized in young adults (Corbetta et al., 1993; Foxe et al., 2003; Thiebaut de Schotten et al., 2011), older adults may maintain their performance on visuospatial orienting tasks by the bilateral recruitment of the posterior parietal cortex. Evidence for the hemispheric asymmetry reduction in visuospatial attention comes from studies demonstrating age-related attenuation of pseudoneglect in healthy older adults (Schmitz and Peigneux, 2011; Benwell et al., 2014; Learmonth et al., 2017). Pseudoneglect refers to the consistent attentional bias to the left visual field that is typically observed in healthy young adults (Bowers and Heilman, 1980; Voyer et al., 2012), and is believed to be due to the right hemisphere dominance for visuospatial processing (Waberski et al., 2008; Cavézian et al., 2012). Learmonth et al. (2017) demonstrated the typical leftward attentional bias in young adults was coupled with greater activity over the right parieto-occipital regions, and this lateralization was absent in older adults (whom also failed to show a leftward attentional bias). These results suggest that in older adults' deterioration of parieto-occipital regions may lead to compensatory increases in activity from homologous regions within the opposite hemisphere.

#### CONCLUSION

In summary, these studies suggest that in spite of the early age-related declines of the dorsal visual stream the visuospatial orienting of attention remains relatively well preserved, at

#### REFERENCES


least in the earlier stages of aging. This maintenance appears to be facilitated by the engagement of compensatory neural mechanisms; which is consistent with both the compensationrelated utilization of neural circuits hypothesis (CRUNCH; Reuter-Lorenz and Cappell, 2008), and the scaffolding theory of aging and cognition (STAC; Park and Reuter-Lorenz, 2009), which propose that the age-related deterioration of the frontoparietal attention network will result in the engagement of compensatory neural mechanisms to support the performance of visual attention tasks. Although the engagement of compensatory neural mechanisms enables older adults to shift attention in the face of widespread structural deterioration (Schneider-Garces et al., 2010; Vallesi et al., 2011; Shafto et al., 2012; Meunier et al., 2014), it results in slower and less efficient task performance (Park and Reuter-Lorenz, 2009; Meunier et al., 2014; Reuter-Lorenz and Park, 2014). Furthermore, progressive deterioration of the frontoparietal network will result in an increased reliance on compensatory neural mechanisms, but at the same time increasing levels of atrophy and structural deterioration also limits the brain's capacity for reorganization (Burke and Barnes, 2006). As age-related atrophy proceeds, the brain will eventually reach its limits for functional reorganization and result in the more apparent attentional declines seen in the later stages of old age (80 years and above; Daffner et al., 2011).

#### AUTHOR CONTRIBUTIONS

The mini review was written by ETS-L with assistance and feedback from AJL.

#### FUNDING

Funding Body: Marsden Fund of New Zealand. Project Number: 3711736. Project Title: Sight unseen: penetrating the enigma of unconscious vision.


high-density electrical mapping and source analysis. Neuroimage 19, 710–726. doi: 10.1016/s1053-8119(03)00057-0


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Sciberras-Lim and Lambert. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Disconnectivity between Dorsal Raphe Nucleus and Posterior Cingulate Cortex in Later Life Depression

Toshikazu Ikuta<sup>1</sup> \*, Koji Matsuo<sup>2</sup> , Kenichiro Harada<sup>2</sup> , Mami Nakashima2,3 , Teruyuki Hobara2,4 , Naoko Higuchi <sup>2</sup> , Fumihiro Higuchi <sup>2</sup> , Koji Otsuki <sup>2</sup> , Tomohiko Shibata2,5 , Toshio Watanuki <sup>2</sup> , Toshio Matsubara2,6 , Hirotaka Yamagata<sup>2</sup> and Yoshifumi Watanabe<sup>2</sup>

<sup>1</sup>Department of Communication Sciences and Disorders, School of Applied Sciences, University of Mississippi, University, MS, United States, <sup>2</sup>Division of Neuropsychiatry, Department of Neuroscience, Graduate School of Medicine, Yamaguchi University, Ube, Japan, <sup>3</sup>Nagato-Ichinomiya Hospital, Shimonoseki, Japan, <sup>4</sup>Department of Psychiatry, Yamaguchi Grand Medical Center, Hofu, Japan, <sup>5</sup>Shinwaen Hospital, Onoda, Japan, <sup>6</sup>Health Administration Center, Yamaguchi University Organization for University Education, Yamaguchi City, Japan

The dorsal raphe nucleus (DRN) has been repeatedly implicated as having a significant relationship with depression, along with its serotoninergic innervation. However, functional connectivity of the DRN in depression is not well understood. The current study aimed to isolate functional connectivity of the DRN distinct in later life depression (LLD) compared to a healthy age-matched population. Resting state functional magnetic resonance imaging (rsfMRI) data from 95 participants (33 LLD and 62 healthy) were collected to examine functional connectivity from the DRN to the whole brain in voxel-wise fashion. The posterior cingulate cortex (PCC) bilaterally showed significantly smaller connectivity in the LLD group than the control group. The DRN to PCC connectivity did not show any association with the depressive status. The findings implicate that the LLD involves disruption of serotoninergic input to the PCC, which has been suggested to be a part of the reduced default mode network in depression.

#### Edited by: Christos Frantzidis,

Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Bruno Pierre Guiard, University of Toulouse, France Andrew James Greenshaw, University of Alberta, Canada

#### \*Correspondence:

Toshikazu Ikuta tikuta@olemiss.edu

Received: 20 April 2017 Accepted: 06 July 2017 Published: 02 August 2017

#### Citation:

Ikuta T, Matsuo K, Harada K, Nakashima M, Hobara T, Higuchi N, Higuchi F, Otsuki K, Shibata T, Watanuki T, Matsubara T, Yamagata H and Watanabe Y (2017) Disconnectivity between Dorsal Raphe Nucleus and Posterior Cingulate Cortex in Later Life Depression. Front. Aging Neurosci. 9:236. doi: 10.3389/fnagi.2017.00236 Keywords: depression, neuroimaging, geriatric psychiatry, magnetic resonance imaging, dorsal raphe nucleus

## INTRODUCTION

Disruption of the dorsal raphe nucleus (DRN) is implicated in depression, given that the DRN serves as the major source of 5-HT release in the brain. In neuroanatomical studies, the area of the DRN is reduced in major depression (Matthews and Harrison, 2012). The number and density of neurons in this region is higher in suicide victims (Underwood et al., 1999). Physiologically, serotonin transporter dysfunction has been found in the DRN of patients with major depression (Hahn et al., 2014), and binding potentials for the 5-HT1A receptor have been found to be reduced in depression (Drevets et al., 1999). Polymorphisms in 5-HT1A gene have been found to be associated with antidepressants (Kato et al., 2009). Genetic variations of tryptophan hydroxylase 2 (TPH2) are also associated with major depression (Zill et al., 2004), and are further found to be associated with TPH2 mRNA expression in the pons including the DRN (Lim et al., 2006). Single nucleotide polymorphisms of the 5-HT1A receptor and TPH2 have been suggested to interact with the severity of depression and respond to SSRIs (Serretti et al., 2011; Jacobsen et al., 2012). Meanwhile, the DRN is understood to be critical for the actions of antidepressants including SSRIs and tricyclics (Briley and Moret, 1993), as well as for the pathophysiology of depression.

Forebrain 5-HT projections originate in the DRN (Törk, 1990; Baker et al., 1991). Rodent models further support disruption of the DRN in depression. Genetically introducing serotonin transporter deficiency in the DRN induces depression-like behavior in mice (Lira et al., 2003). Social stress has also been shown to influence serotoninergic neurons in the DRN of rats that adopted a proactive coping strategy against stress (Wood et al., 2013). More recently, the volume of the raphe nuclei, as well as other monoaminergic regions, have been shown to be negatively correlated with social avoidance, which indexes stress susceptibility (Anacker et al., 2016) The DRN cells have inhibitory projections to the cortex, including the cingulate gyrus (Olpe, 1981), which is also implicated in depression (Bae et al., 2006; Wise et al., 2016).

The pathophysiology of Later-Life Depression (LLD) is considered to be in parallel to the pathology of depression in younger ages, though LLD has certain defining characteristics, such as less genetic influence (Brodaty et al., 2001; Alexopoulos, 2005). Antidepressants have been shown to be especially effective with LLD (Salzman et al., 2002; Schneider et al., 2003). Therefore, it can be hypothesized that the role of the DRN is substantial in the pathology of LDD. The role of the DRN in LLD, however, is not clearly understood.

Neuropathological change of the raphe nuclei has been implicated in depression. Neuronal loss in the raphe nucleus of older patients with depression has been found in a postmortem study (Tsopelas et al., 2011). Lesser gray matter concentration of the DRN was found in patients with major depression compared to the healthy control group (Lee et al., 2011). Patients who had transient relapse of depression induced by tryptophan depletion showed greater correlation between habenula-DRN co-activity compared to those patients who did not show relapse (Morris et al., 1999). Recently, studies in the functional connectivity of the DRN in humans have shown positive connectivity to cortical regions, illuminating 5-HT associated regions (Beliveau et al., 2015). Yet, functional connectivity of the DRN in depression or LLD has not been extensively studied. Here, in order to elucidate the functional connectivity of the DRN, we examined the resting state of functional connectivity in LLD in comparison to an age-matched control cohort. We hypothesized that the disrupted DRN connectivity would be found in regions whose connectivity manifests a signature role of the DRN in neuropathophysiology of LLD.

### MATERIALS AND METHODS

#### Participants

We examined a total of 95 individuals, including 33 LLD patients and 62 healthy age-matched participants. All patients met the DSM-IV criteria for major depressive disorder. Mean age of the LLD group was 60.40 (SD = 7.82) years, and healthy group 62.73 (7.43) years. Patients were recruited from Yamaguchi University Hospital; referred by area clinics and hospitals; and diagnosed by clinical interviews performed by senior psychiatrists, case conferences with psychiatrists, and structured interviews using the Mini International Neuropsychiatric Interview (MINI, Japanese version 5.0.0; Otsubo et al., 2005). Patients' clinical demographics were obtained through a clinical interview. Patients with current or a history of substance abuse/dependence or other psychotic illnesses were excluded from the study. Healthy control participants were recruited from the community through local advertisements and word of mouth. Any healthy subject who had psychiatric illness was excluded through the MINI and clinical interviews. Participants with an immediate family member with a psychiatric disorder were also excluded. Exclusion criteria included ambidextrous or left handedness (Oldfield, 1971), MR imaging contraindications, presence of serious medical conditions, and hospitalization in the prior 6 months. Current mood states were assessed using the Structured Interview Guide for the Hamilton Depression Rating Scale (SIGH-D; Williams, 1988). Social functioning was assessed by the Global Assessment of Functioning scale (GAF; American Psychiatric Association, 2000). Participants with scores of 23 or below on the Mini-Mental State Examination were considered demented and excluded (Folstein et al., 1975). Through interviews, blood tests, and physical examinations, subjects with endocrinological disease, head trauma, neurological disease, a family history of hereditary neurological disorders, or other medical conditions (i.e., hypertension, diabetes, active liver disease, kidney problems, or respiratory problems) were also excluded from the study. LLD was defined as depression in patients over 50 years old at the time of study participation. The mean number of depressive episodes was 2.3 (1.8) in patients with LLD. Thirty patients were under medication at the time of study participation. Five patients were treated only with antidepressants; eight with antidepressants and firstgeneration antipsychotics; 12 with antidepressants and secondgeneration antipsychotics; two with an antidepressant and a mood stabilizer; with antidepressant, a second-generation antipsychotic, and a mood stabilizer; one with a secondgeneration antipsychotic; and one was with a mood stabilizer. Among the remaining three patients, one was drug-naïve, and the other two had been medication-free at least for the previous 4 years. The mean imipramine-equivalent dose among all the patients was 174.5 (151.3) mg. Most patients were under one or more antidepressant medications: SSRI-16; SNRI-14; Tricyclics-8; Others (mirtazapine, trazodone and mianserin)-13. This study was conducted in accordance with the latest version of the Declaration of Helsinki. This study was approved by the Yamaguchi University Institutional Review Board and written informed consent was obtained from all study participants.

#### Clinical Assessments

Current mood states were examined using the SIGH-D (Williams, 1988). To assess social functioning, the GAF (American Psychiatric Association, 2000) was used. Self-reported depression status was measured by the Beck Depression Inventory (BDI; Beck et al., 1961). To uniformly assess various antidepressant dosage Imipramine Equivalent Potency was calculated.

#### MRI Data Acquisition

A total of 204 echo-planner imaging (EPI) volumes were acquired on a Siemens Magnetom Skyra system (TR = 2500 ms, TE = 30 ms, matrix = 64 ∗ 64, FOV = 220 mm, slice thickness = 4 mm, 34 continuous axial slices) as well as the MPRAGE structural volume (TR = 2300 ms, TE = 2.95 ms, flip angle = 9◦ , matrix = 256 × 256, FOV = 270 mm, 1.2 mm thick sagittal images).

#### MRI Data Processing

Script package from the 1000 Functional Connectomes Project<sup>1</sup> (Biswal et al., 2010) was used for preprocessing and seed-voxel correlation matrix production whereby FSL<sup>2</sup> and AFNI<sup>3</sup> are called. Motion correction and spatial smoothing (6-mm FWHM Gaussian kernel) were carried out through preprocessing. Each individual rs-fMRI series was registered and normalized to MNI152 2 mm space via T1 volume. Through this registration, 12 affine parameters were created between rs-fMRI volume and MNI152 2 mm space, so that a seed ROI can be later registered to each individual rs-fMRI space. The rs-fMRI time series were band-pass filtered (between 0.005 Hz and 0.1 Hz). Each resting state volume was regressed by white matter and cerebrospinal fluid signal fluctuations as well as the six motion parameters.

The DRN seed was defined as a 32 mm<sup>3</sup> region centered at [MNI: 0, −27, −9] following the literature (Kranz et al., 2012; Beliveau et al., 2015). From the DRN seed, voxel-wise connectivity analyses were conducted by the ''singlesubjectRSFC fcon'' script, whereby the time course is spatially averaged within the ROI so that correlations from the ROI to each individual voxel across the brain.

To compare the two groups, a Z statistic image was estimated where clusters were determined by a family-wise, error-corrected cluster significance threshold of p < 0.01 assuming a Gaussian random field for the Z-statistics.

In order to examine the association between the functional connectivity differences and status of depression, clusters that showed significant effect of group were tested for further association with depression scales within the LLD group. The mean Z-scores of a cluster were calculated and correlations between the mean score and each of the depression scales was tested.

### RESULTS

The LLD group showed significant indications of depression; BDI (control: 7.15 ± 4.71; LDD: 29.53 ± 11.45), GAF (control: 93.02 ± 4.72; LDD: 52.79 ± 6.19), and SIGH-D 17 (control: 0.81 ± 1.07; LDD: 21.87 ± 3.49), as well as their antidepressant medications assessed as imipramine equivalent potency (control: 0 ± 0; LDD: 193.64 ± 145.75).

The LLD group showed significantly lower functional connectivity between the DRN and bilateral posterior cingulate cortex (PCC). There was no other cluster that showed

lower functional connectivity in the LLD group. The cluster had 301 voxels and the peak voxel (corrected p < 0.001) was at [MNI: −4 −30 26] (**Figure 1**). No region showed greater connectivity in the LLD group than the control group.

In the LLD group, none of these depression scales (SIGH, BDI and GAF) or imipramine equivalent dose showed significant correlation to the functional connectivity between the DRN and cluster in the posterior cingulate cortices.

### DISCUSSION

control group.

This study aimed to isolate the functional connectivity patterns of the DRN distinct in LLD. Here, the DRN showed deficient functional connectivity to the PCC in LLD. The results illuminate the serotoninergic dysfunction that may influence LLD pathology, which is independent from the depressive status.

The current results are solely based on the functional connectivity between two distant regions, and therefore does not grant anatomical connections between two regions. However, serotoninergic projections from the DRN to PCC has been well-documented in rodents (Olpe, 1981; Finch et al., 1984; Kosofsky and Molliver, 1987). In humans, depressive symptoms in Parkinson's disease have been found to be associated with 5-HT transporter binding in the raphe nuclei and PCC, as well as other limbic structures (Politis et al., 2010), suggesting the association between depression and raphe-cingulate 5-HT projections. The current findings are consistent with raphecingulate serotoninergic projections that have been shown to be anatomically present.

<sup>1</sup>http://www.nitrc.org/projects/fcon\_1000

<sup>2</sup>http://www.fmrib.ox.ac.uk

<sup>3</sup>http://afni.nimh.nih.gov/afni

Dysfunctions of the PCC have repeatedly been found in depression (Zhou et al., 2010; Berman et al., 2011; Leech and Sharp, 2014). At the same time, SSRIs have been shown to ameliorate PCC deficits. Sertraline and fluoxetine increase glucose metabolism in the PCC, correlating with clinical improvement (Buchsbaum and Hazlett, 1997; Mayberg et al., 2000). In addition, administration of SSRIs has been found to increase the PCC volume in the non-depressed, healthy population (Kraus et al., 2014). The current results are consistent with the previous PCC findings in depression. In the nondepressed, human population, the PCC has been shown to mediate emotion and memory related processes (Maddock et al., 2003), implicating its importance to mood disorders whereby impacts of emotional input are amplified. The PCC has been suggested to be associated with social/contextual selfreflection, duties or obligations, and autobiographical memory, suggesting PCC's role in self rumination (Johnson et al., 2006, 2009; Svoboda et al., 2006; Herwig et al., 2012). The current results support the understanding that serotoninergic dysfunction at the DRN underlies LLD by showing the disruption of the connectivity between the DRN and PCC in LLD.

The PCC in non-depressed population has also been shown to play critical roles in the default mode network (Fox et al., 2005; Buckner et al., 2008; Fransson and Marrelec, 2008; Uddin et al., 2009). The PCC is one of the regions that is more active at rest than during a task (Buckner et al., 2008). Furthermore, the PCC activity negatively predicts the motorcontrol network (Uddin et al., 2009), implicating the importance of the PCC during rest. It may imply that the PCC may mediate cognitive process of rest through the default mode network. Remarkably, the default mode network has been found to be suppressed in major depressive disorder (Sheline et al., 2009).

The functional connectivity between the DRN and PCC was not significantly associated with any of depression scales or antidepressant dose, while LLD group showed significantly lesser connectivity compared to the control group. It may suggest that the reduced DRN-PCC functional connectivity is a trait characteristic of depression independent from the depressive state. The association between depressive traits and the PCC has previously been implicated. In a task fMRI study, greater activation of the PCC in depression was found during negative rumination tasks, compared to the healthy control group (Cooney et al., 2010). Increased dominance of the default mode network in depression, compared to the task positive network, was associated with higher levels of depressive rumination (Hamilton et al., 2011). Decreased connectivity of the PCC to the default mode network in depression has also been found in a resting state fMRI study and was associated with over general autobiographical memory (Zhu et al., 2012). Taken together, the PCC in depression is associated with negative rumination. Recall that the raphe serotoninergic projection to the PCC is inhibitory (Olpe, 1981). Our results can imply that decreased DRN-PCC connectivity reflect lesser inhibitory inputs to the PCC, which would downregulate negative rumination.

The imprecision of the DRN seed needs to be addressed as the limitation of this study. The DRN seed was defined in 2 mm<sup>3</sup> space, which would not have sufficient resolutions to exclusively isolate the DRN. It should also be recalled that 6 mm smoothing was applied. Therefore, the seed may contain other structures around the DRN. Specifically, the median raphe nucleus (MRN) may not be excluded. In the previous study that differentiated resting state connectivity of DRN and MRN (Beliveau et al., 2015), [11C]DASB PET data were acquired in each subject and the DRN and MRN were localized in each individual brain. As we did not obtain PET data, we have far coarse localization of the DRN. The current finding might not be exclusively attributed to the DRN but it may also include the greater raphe nuclei including MRN. It would be ideal to conduct a similar study with PET data for more precise localization of the DRN.

The MRN also possesses serotoninergic characteristics which implicates an association with depression (Bach-Mizrachi et al., 2008). MRN has also been shown to modulate emotional behaviors (López Hill et al., 2013). The association with the hippocampus has been found more in the MRN than DRN (Jacobs et al., 1974). Also, the MRN has known efferent to the cingulate cortex (Azmitia and Segal, 1978). It might be more appropriate to apprehend the finding as Raphe—PCC disconnectivity in LDD.

There are other limitations of this study that need to be addressed. The relationship between medication and functional connectivity is unclear. The findings may be the effect of medication as opposed to the effect of LLD. Studies examining a non-medicated population could better discriminate these possibilities. Though no association was found between the DRN-PCC connectivity and imipramine equivalent dose, it does not exclude the possibility for the influence of medication. In addition, the experience of being in an MRI scanner may have influenced the groups differently. LLD may trigger differential neuronal activation as a function of the experience.

In conclusion, this study found disconnectivity between the DRN and PCC in the LLD group that is consistent with serotoninergic dysfunction of the DRN in depression.

### AUTHOR CONTRIBUTIONS

KM and YW conceived and designed the experiments. KH, MN, TH, NH, FH, KO, TS, TW, TM and HY performed the experiments. TI, KM and KH analyzed the data and wrote the article.

### FUNDING

This study was supported, in part, by the ''Integrated Research on Neuropsychiatric Disorders'' conducted under the Strategic Research Program for Brain Sciences from the MEXT and AMED (HY), Japan Society for the Promotion of Science (JSPS) KAKENHI (16K10189 to HY, 15K09832 to KM), and the GlaxoSmithKline (GSK) Japan Research Grant (to HY).

#### REFERENCES


**Conflict of Interest Statement**: TI has received speaker's honoraria from Eli Lilly, Daiichi Sankyo, and Dainippon Sumitomo. KM has received research donations from GlaxoSmithKline and Ohtsuka Pharmaceutical. YW has received research donations from MSD, GlaxoSmithKline, Eli Lilly and Company, Yoshitomiyakuhin, Shinogi, Pfizer, Janssen Pharma, Meiji Seika Pharma, FujiFilm RI Pharma, Takeda Pharmaceutical, Astellas, Dainippon Sumitomo Pharma, and Ohtsuka Pharmaceutical.

The other authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Ikuta, Matsuo, Harada, Nakashima, Hobara, Higuchi, Higuchi, Otsuki, Shibata, Watanuki, Matsubara, Yamagata and Watanabe. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Beta-Band Functional Connectivity Influences Audiovisual Integration in Older Age: An EEG Study

Luyao Wang<sup>1</sup>† , Wenhui Wang<sup>2</sup>† , Tianyi Yan<sup>2</sup> \*, Jiayong Song<sup>3</sup> , Weiping Yang<sup>4</sup> , Bin Wang<sup>5</sup> , Ritsu Go1,6, Qiang Huang1,7 and Jinglong Wu1,6 \*

1 Intelligent Robotics Institute, School of Mechatronical Engineering, Beijing Institute of Technology, Beijing, China, <sup>2</sup> School of Life Science, Beijing Institute of Technology, Beijing, China, <sup>3</sup> The Affiliated High School of Peking University, Beijing, China, <sup>4</sup> Department of Psychology, Hubei University, Wuhan, China, <sup>5</sup> College of Computer Science and Technology, Taiyuan University of Technology, Shanxi, China, <sup>6</sup> International Joint Research Laboratory of Biomimetic Robots and Systems, Ministry of Education, Beijing, China, <sup>7</sup> Key Laboratory of Biomimetic Robots and Systems, Ministry of Education, Beijing, China

Audiovisual integration occurs frequently and has been shown to exhibit age-related differences via behavior experiments or time-frequency analyses. In the present study, we examined whether functional connectivity influences audiovisual integration during normal aging. Visual, auditory, and audiovisual stimuli were randomly presented peripherally; during this time, participants were asked to respond immediately to the target stimulus. Electroencephalography recordings captured visual, auditory, and audiovisual processing in 12 old (60–78 years) and 12 young (22–28 years) male adults. For non-target stimuli, we focused on alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz) bands. We applied the Phase Lag Index to study the dynamics of functional connectivity. Then, the network topology parameters, which included the clustering coefficient, path length, small-worldness global efficiency, local efficiency and degree, were calculated for each condition. For the target stimulus, a race model was used to analyze the response time. Then, a Pearson correlation was used to test the relationship between each network topology parameters and response time. The results showed that old adults activated stronger connections during audiovisual processing in the beta band. The relationship between network topology parameters and the performance of audiovisual integration was detected only in old adults. Thus, we concluded that old adults who have a higher load during audiovisual integration need more cognitive resources. Furthermore, increased beta band functional connectivity influences the performance of audiovisual integration during normal aging.

Keywords: functional connectivity, EEG, audiovisual integration, aging, beta band

### INTRODUCTION

In daily life, our brain must constantly combine all kinds of information in one or more cues from different sensory modalities. A large body of evidence from daily life has suggested that cognitive functions decline during normal aging. This decline brings trouble during elderly life. As auditory and visual information become more important, the study of age-related differences in audiovisual integration helps us understand the aging process.

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Zhao Wang, Tsinghua University, China Winfried Schlee, University of Regensburg, Germany Evangelos Paraskevopoulos, Aristotle University of Thessaloniki, Greece

#### \*Correspondence:

Tianyi Yan yantianyi@bit.edu.cn Jinglong Wu wujl@bit.edu.cn

†These authors have contributed equally to this work.

Received: 15 February 2017 Accepted: 07 July 2017 Published: 07 August 2017

#### Citation:

Wang L, Wang W, Yan T, Song J, Yang W, Wang B, Go R, Huang Q and Wu J (2017) Beta-Band Functional Connectivity Influences Audiovisual Integration in Older Age: An EEG Study. Front. Aging Neurosci. 9:239. doi: 10.3389/fnagi.2017.00239

**266**

**Abbreviations:** A, unimodal auditory stimulus; AUC, area under the curve of parameters; AV, bimodal audiovisual stimulus; CDFs, cumulative distribution functions; EEG, electroencephalography; PLI, phase lag index; V, unimodal visual stimulus.

To understand the processing of audiovisual stimuli, recent Event-related potentials (ERPs) studies analyzed the time course of visual, auditory and audiovisual stimuli (Fort et al., 2002; Stekelenburg et al., 2004). Furthermore, oscillatory responses in the alpha, beta, and gamma bands have been related to sensory processing. It may be related to harmonize activation of cell assemblies. Fu et al. (2001) showed that the alpha-suppression mechanism occurs during audiovisual stimulus with the use of auditory cues in an attention experiment. Studies with EEG and magnetoencephalogram (MEG) have suggested that the gamma band is also related to the integration of information (Basar, 2013). For the beta band, several groups discussed it during different cognitive processes. The results indicated that oscillatory beta forms an important substrate of human cognition processes, such as attention, working memory and audiovisual integration. Senkowski et al. (2006) investigated beta oscillatory facilitation behavior in an ERP study during visual, auditory and audiovisual stimuli. Sakowitz et al. (2005) found increased beta responses during audiovisual stimuli in comparison to unisensory stimuli on the basis of the intersensory component. Senkowski et al. (2008) used a sensory gating paradigm, which is an integration of meaningful semantic inputs, and they reported that crossmodal effects were related to evoked beta responses.

Recent studies have described age-related audiovisual integration. Some behavioral researches have reported enhanced audiovisual integration in older adults (Laurienti et al., 2006; Peiffer et al., 2007; Mahoney et al., 2011). However, most of the studies were behavioral studies and did not focus on different oscillatory frequency bands. There are many factors that influence the audiovisual integration, such as the location of the presented experimental (Molholm et al., 2004). When the audiovisual stimuli were presented peripherally, the integration could also be elicited. In addition, age-related differences were significant. The maximal behavioral enhancement in older adults occurred more delayer and the time window was longer than in younger adults (Wu et al., 2012). ERP and EEG studies have shown the deficits in attentional control affected the audiovisual integration (Mozolic et al., 2012). Some study using arrow as cue to investigate the age-related visuospatial attention. The results shown the performance for old and young adults is similar. In addition, old adults had slower ERP components and similarly amplitude compared to young adults (Curran et al., 2001). These findings indicate that there are some changes to audiovisual integration with aging, but the underlying neuronal mechanisms are still not fully understood.

Recent research has shown that functional interactions between brain areas are crucial for effective cognitive functioning (Wen et al., 2012), a concept referred to as "functional connectivity." Functional connections play an important role in multisensory processing. Connections not only between sensory related subcortical structures but also between cortical areas can mediate multisensory integration (Beer et al., 2011; Bishop et al., 2012; van den Brink et al., 2014). Some studies investigated the functional network affects audiovisual integration in different ways. The network could be reorganized due to long-term training (Paraskevopoulos et al., 2015). In addition, functional connectivity could be reorganized by cognitive training (Bamidis et al., 2014; Klados et al., 2016). However, the agerelated differences of functional connectivity during audiovisual integration is still unknown.

We used EEG to investigate age-related audiovisual integration; its high temporal resolution makes it rather suitable for the identification of synchronization across frequency bands. The EEG signals were recorded over brain to study the functional connectivity. The PLI, a synchronization measure, reflects the extent of inter-trial phase variability for a given frequency across time. PLI is defined as a period of phase locking between two events, and it can only be estimated in a statistical sense. It removes and attenuates the synchronization that occurs at or near the zero phase difference. From this way, we could reduce the interference of signals from common sources or volume conduction, which were regarded as spurious synchronization (Stam et al., 2007; Doesburg et al., 2013).

In this study, we sought to investigate the functional connectivity in different oscillatory frequency bands during audiovisual integration. We hypothesized that functional connection could influence the audiovisual integration and there are differences between old and young adults. To address this issue, we designed three stimuli: V, A, and bimodal audiovisual (AV) stimuli, which are presented peripherally. We combined the phase synchrony of electrode interactions and graph-theoretical metrics of network topography to investigate task-dependent functional connectivity derived from EEG data. The PLI computed for each pair of sensors was used to construct graphs in various frequency bands independently.

### MATERIALS AND METHODS

### Participants

Twelve old male adults (60–78 years, mean age ± SD, 68.6 ± 4.74) and 12 young male adults (22–28 years, mean age ± SD, 23.9 ± 1.73) participated in this study. To confirm their cognitive function, all of the participants did the mini-mental state exam to identify cognitive function. Participant who had a score more than 2.5 SD from the mean score that matched his age and level of education were excluded (Bravo and Hebert, 1997). In addition, participants were excluded if they self-reported any disease. Due to the experiment requirement, all of the participants had normal or corrected-to-normal vision (none of the participants were color blind) and normal hearing capabilities. The individuals provided written informed consent, which was previously approved by the ethics committee of Okayama University.

#### Stimuli and Task

The experiment was performed in a dimly lit, soundattenuated, electrically shielded room (laboratory room; Okayama University, Japan). Stimuli presentation and response collection were determined using the Presentation software (Neurobehavioral Systems Inc., Albany, CA, United States). A 21-inch computer monitor with a black background was

presentation of both visual and auditory target or non-target stimuli. In addition, non-target stimuli were presented at a frequency of 80% of the total stimuli.

positioned 60 cm in the front of the participant's eyes and was used to present visual stimuli. The auditory stimuli were presented through an earphone. Each block consisted of 300 visual stimuli, 300 auditory stimuli and 300 audiovisual stimuli. All of the stimuli were randomly presented and had an equal probability of appearing to the left or right of the central fixation point.

The visual target stimulus was a red and white block, and the non-target visual stimulus was a black and white block (5.2 cm × 5.2 cm with a subtending visual angle of ∼5 ◦ ). These visual stimuli were peripherally presented at an angle of ∼12◦ from a centrally presented fixation point in the lower visual fields (∼5 ◦ below the horizontal meridian) (He et al., 1996; Talsma and Woldorff, 2005). The auditory target stimulus was white noise, and the non-target auditory stimulus was a 1000 Hz sinusoidal tone (60 dB sound pressure level, 5 ms rise or fall time). The audiovisual stimulus consisted of the simultaneous presentation of both the visual and auditory target or non-target stimuli. In addition, non-target stimuli were presented at a frequency of 80% of the total stimuli (**Figure 1**).

At the beginning, each participant was required to complete five experimental blocks, and each block lasted ∼5 min. In formal experiment, each block had a 3000 ms fixation period, followed by the test stimulus. Each type of stimulus displayed 150 ms and continued with 1300 – 1800 ms interstimulus interval time. Within the interval time, participant responded to the target stimulus and the screen was cleared. The experiment continued regardless of whether the participant responded. Participants were instructed to click the left or right button with the forefinger or middle finger quickly and accurately when the target stimuli occurred. They were also instructed to stare at the central white cross during the whole experiment.

### EEG Data Collection

The EEG signals were recorded from 30 scalp electrodes mounted on an electrode cap (Easy cap, Germany), as specified by the International 10–20 System, and 2 electrooculogram electrodes that were referenced to the earlobes. Data were bandpass filtered from 0.05 to 100 Hz during the recordings and were digitized at a sampling rate of 500 Hz by BrainAmp amplifiers (BrainProducts, Munich, Germany).

## DATA ANALYSIS

#### Interregional Phase Synchronization

Our data were pre-processed with Matlab R2013a (Mathworks Inc., Natick, MA, United States) with the following open source toolboxes: EEGLAB<sup>1</sup> (Swartz Center for Computational Neuroscience, La Jolla, CA, United States). An Independent Component Analysis was used to remove artifacts (e.g., eye artifacts, muscle artifacts and electrocardiographic activity) from the data within all channels. We also corrected the baseline for each epoch.

The non-target stimuli were filtered into alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz) frequency ranges. The network synchronization of all three bands was investigated. For each subject, the Hilbert transform was employed to obtain the time series of instantaneous phase measures for each trial, source

<sup>1</sup>http://sccn.ucsd.edu/eeglab/

and frequency band. Phase locking was calculated for each EEG sensor pair and frequency with the PLI (Stam et al., 2007).

$$\text{PLI} = \left| \left\langle \text{sign}(\Delta \phi(\mathbf{t\_n})) \right\rangle \right| = \left| \frac{1}{M} \sum\_{\mathbf{k}=1}^{M} \text{sigh}(\Delta \phi(\mathbf{t\_n})) \right| \tag{1}$$

At a given time point, it measures the reliability of phase relations between two EEG sensors, which produces a sensorby-sensor adjacency matrix. Epochs were extracted from 300 ms before stimulus onset until 500 ms after stimulus onset. We removed the first and last 150 ms (75 sample points) due to the distortions caused by Hilbert transform at the edges of the epochs (Doesburg et al., 2008). These were then averaged within each group (old adults and young adults) for each trial condition (auditory, visual and audiovisual, each condition includes two orientations). The average PLI values across EEG sensors for each time point reflect task-dependent dynamic network connectivity. For each participants, we calculated it at each frequency, respectively. An two-sample t-test (sample size of bootstrap is 1000) was performed at each time point to compare the differential PLI value of the old and young adults. Time points with significant age differences were used to identify windows for further analyses.

#### Statistical Analysis of Network Dynamics

According to the results of the two-sample t-test above and previous studies, adjacency matrices for non-overlapping 150 ms time-windows after stimulus onset were extracted for each frequency: 0–150 ms for alpha and gamma bands and 50–200 ms for the beta band. To characterize task-dependent network connectivity dynamics, these adjacency matrices were averaged and represent the mean connectivity within this active window for each subject.

Task-dependent network synchronization was analyzed by Network Based Statistic (NBS), which is a data-driven approach. Statistical significance of group differences could be displayed, which was corrected for multiple comparisons. (Zalesky et al., 2010, 2012). In this study, the purpose of the NBS is to identify any connected structures that are significantly different between old adults and young adults. At first, we applied a univariate statistical threshold to each

element in the compared adjacency matrix, and multiple comparisons was achieved regardless of this threshold. In this case, a t-test was performed in the 30 × 30 adjacency matrix (p < 0.05) (Zalesky et al., 2010, 2012). Then, data surrogation was repeated 5,000 times to establish statistical confidence.

### Graph of Theoretical Analysis of Dynamic Network Topologies

Functional connectivity among sensors was measured by computing the PLI for every possible pair during the time window. The resulting non-linear correlation matrices were converted to weighted graphs. To characterize the task-dependent weighted network dynamics, we constructed network G (30 × 30) for each trial, frequency and subject by GRETNA (Wang et al., 2015) by using the time window identified in the above analysis.

For the constructed brain networks, we calculated brain network parameters (including the clustering coefficient, path length, small-worldness, global efficiency, local efficiency and degree) to examine both the global and regional topological characteristic variations. Each attribute was compared with those of 100 random networks. We applied a sparsity threshold (0 < S < 1), which normalized the networks, to examine the relative network organization. In addition, task-dependent areas under the curve (AUCs) of parameters were calculated for each network measure to provide a scalar that did not depend on the specific threshold selection.

#### Statistical Analysis

SPSS version 20.0 (SPSS, Inc., Chicago, IL, United States) was used for statistical analyses. For each frequency, repeatedmeasures ANOVAs were carried out separately for the averaged adjacency matrices of PLI and task-dependent AUC of eight network parameters. A 2 (age group: old, young) × 3 (sensory modality: A, V, or AV) × 2 (stimuli direction: left, right) repeated-measures ANOVA analysis was performed separately to examine the effects of audiovisual integration and age as well as their interaction. The Greenhouse-Geisser epsilon value was obtained in all cases in which the repeatedmeasures data failed the sphericity test (Greenhouse and Geisser, 1982). All statistical comparisons were two-tailed with α = 0.05. We used the Bonferroni correction to correct for the effect of multiple comparisons in neural oscillations.

#### Relationship between Network Topology Parameters and Behavior Data

Trials with target stimuli were extracted for behavior analysis. A race model was used to identify whether audiovisual integration occurred (Miller, 1986). For each participant, the target stimuli were analyzed with CDFs for the V, A and AV stimuli. In addition, the CDFs for the V, A and AVstimuli were generated using 10 ms time bins. At each time bin, the distribution of race model was calculated by the following formula: [P (V) + P (A)] – [P (V) × P (A)]. Each participant's race model curve was then subtracted from their AV CDF. The peak time point of each probability difference curve was recorded, which represented the response time at that the audiovisual integration most likely occurred. A one-way ANOVA was performed to compare age differences (two-tailed with α = 0.05, Bonferroni correction).

Furthermore, at each oscillatory frequency, a Pearson correlation was conducted to test the relationship between each network parameter and behavior peak time point.

### RESULTS

### Time Courses of Average PLI

We filtered the EEG data into alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz) frequency ranges. The phase synchronization of these frequencies was calculated. For each time point, we averaged the PLI of each subject within groups (**Figure 2**). For subsequent analyses, no differences were found between the left and right hemi-spaces (see Supplementary Figure 1) in the strength of PLI and network topology parameters, and we averaged the results of the two orientations.

The results show that in both the alpha and beta band, there were clear dynamic changes after stimulus onset for all condition, especially for the audiovisual stimulus. There was a significant (p < 0.05) difference within 50–200 ms for the beta band. This finding indicates that the strength of functional connectivity in old adults is higher than in young adults. In alpha and gamma band, only a small part of the time point within 0– 150 ms was different between groups. To compare the differences between groups, we performed a subtraction for each condition. The main results obtained from our studies are summarized in **Figure 3**.

### Topographical Analysis between Groups

According to time courses of average PLI and previous ERP studies (Fort et al., 2002; Stekelenburg et al., 2004), we choose the 150 ms time window after stimulus onset (0–150 ms for alpha and gamma, 50–200 ms for beta) and averaged them to characterize the task-dependent weighted network connectivity. As presented in **Figure 2**, we averaged the results of the two orientations (**Figures 4A–C**). In beta band, results of repeated-measures ANOVA shown that there is significant interaction between group and stimuli type [p = 0.02, F(2,44) = 4.800]. The simple effect results showed that there are significant differences between old and young participants during the AV target stimulus task [p = 0.03, F(1,22) = 5.39] (**Figure 4B**). The results that contains two orientations are in Supplementary Figure 2.

A strict statistical analysis was performed with Network Based Statistics to investigate group differences in phase synchrony

race model curve (race model). (B) CDFs in old adults. (C) The cumulative probability difference curve for old adults (solid line) and the young adults (gray line). The peak time point is significant difference between two groups. <sup>∗</sup> p < 0.05.

during each condition for each frequency range. In the beta band, NBS revealed induced connectivity in the old group in a distributed network of EEG sensors only in the AV stimulus condition (p < 0.05, corrected, two-tailed) in **Figure 4B**. However, a left or right stimulus activated contralateral brain results in different pairwise connections of two orientations (**Figure 4D**). No significant group differences were observed for other stimulus trials.

However, NBS only reveals connected structures that are significantly different between groups. It does not reveal any statistically significant differences in topological properties.

#### Statistical Results in Task-Dependent Network Topology Parameters

To assess age differences in brain network connectivity during the task, the AUC of the clustering coefficient (Cp), path length (Lp), small-worldness (Gamma, Lambda, Sigma), global efficiency (Eg), local efficiency (Eloc), and degree was analyzed. For each frequency, a 2 (age group: old, young) × 3 (sensory modality: A, V, or AV) × 2 (stimuli direction: left, right) repeated-measures ANOVA was performed.

The detailed results are shown in **Table 1**. Differential parameters between groups are presented in larger fonts. In the alpha band, there are significant interactions for the Cp value [p = 0.04, F(2,44) = 6.367]. In the beta band, there are significant interactions for Cp [p = 0.049, F(2,44) = 3.255], Gamma [p = 0.04, F(2,44) = 6.642], Sigma [p = 0.02, F(2,44) = 7.412], Lp [p = 0.013, F(2,44) = 4.820], Eg [p = 0.021, F(2,44) = 4.374], Eloc [p = 0035, F(2,44) = 3.739], and Degree [p = 0.014, F(2,44) = 4.721] values. In the gamma band, there are significant interactions for Gamma [p = 0.30, F(2,44) = 4.054], and Sigma [p = 0.036, F(2,44) = 3.781] values.

The simple effect results showed that there are significant trends for old individuals to have lower Cp or Gamma (normalized value of Cp) values during the visual task than young individuals for the alpha band [p = 0.045, F(1,22) = 4.53], beta band [p = 0.036, F(1,22) = 5.01], and gamma band [p = 0.02, F(1,22) = 12.60]. In addition, for the beta band, all of the parameters showed significant differences between old and young participants during the AV target stimulus task for Cp [p = 0.049, F(1,22) = 4.35], Gamma [p = 0.003, F(1,22) = 11.57], Sigma [p = 0.002, F(1,22) = 12.49], Lp [p = 0.008, F(1,22) = 8.42], Eg [p = 0.018, F(1,22) = 6.58], Eloc [p = 0.024, F(1,22) = 5.90], and Degree [p = 0.024, F(1,22) = 5.86].

#### Relation to Behavior

The results of CDFs of V, A, AV and race model was shown in **Figure 5A** (old adults) and **Figure 5B** (young adults). The distribution of CDFs revealed that the responses to the AV stimuli were faster than the response to V or A stimuli in both age groups. Furthermore, to identify whether audiovisual integration occurred, we measured the response time to the AV stimuli by subtracting the race model for each age group independently (**Figure 5C**). The time window of behavioral

facilitation in older adults was longer and more delayed than that in the younger adults. The peak time point of each probability difference curve was recorded for each participant in each age group. There was a significant difference between groups in terms of peak time point (p = 0.02). For the young adults, the average peak time point was 330 ms (SD: ± 56 ms), whereas the old adults had delayed response times at 420 ms (SD: ± 55 ms).

At each oscillatory frequency, a Pearson correlation was conducted to test the relationship between network parameters and behavior response. No significant relationships were observed for the alpha and gamma bands. For the beta band, only for the AV stimulus, the network topology parameters showed a significant relationship with the peak time point (**Figure 6**). It is interesting that only old adults showed a strong relationship.

#### DISCUSSION

#### Summary

Neuroanatomical changes have been recognized and are thought to account for the cognitive declines during aging. However, the underlying neuronal mechanisms in age-related audiovisual integration are still unclear. This study explored this question by analyzing phase synchronization and the graph-theoretical network of EEG data. Studies have shown that EEG is rather suitable for the identification of synchronization across frequency bands in functional networks (Stam et al., 2007). Synchronous oscillatory neural activity is a possible candidate mechanism for the coordination of neural activity between functionally specialized brain regions.

The goal of the present study was to clarify the age-related functional connectivity in alpha, beta and gamma bands during visual, auditory and audiovisual stimuli. We found age differences within 200 ms after stimulus onset, which is consistent with previous studies (Fort et al., 2002; Stekelenburg et al., 2004). The results show that old adults have stronger functional connectivity while performing the same tasks, especially for audiovisual stimuli. Furthermore, beta oscillatory network connectivity influences the performance of audiovisual integration during normal aging.

### Age-Related Beta Band Functional Connectivity in Audiovisual Integration

Oscillatory phenomena corresponding to the EEG frequency bands play a major role in functional communication in the brain during cognitive process. The present study is the first to analyze age-related oscillatory functional connectivity during audiovisual integration. We focused on the alpha, beta and gamma bands, which have been related to sensory processing (Fu et al., 2001;

TABLE 1 | Statistical results of network topology parameters for each condition.


The size of indexes that significantly different between groups was not larger.

Senkowski et al., 2006; Basar, 2013). The results indicated that age-related differences occurred during audiovisual stimuli only in the beta band. von Stein and Sarnthein (2000) showed that the beta band serves as a communication mechanism between distant cortical areas. These findings confirmed that the beta band connection plays an important role in visual, auditory and audiovisual processes during aging (Sakowitz et al., 2005; Senkowski et al., 2008).

Both old and young adults showed increased PLI in beta bands after stimulus onset. In addition, only in the beta band, old adults had a significantly higher PLI during audiovisual processing (**Figure 4B**), which indicates the presence of stronger phase locking while performing tasks. As presented in this study, we avoid effects of motor responses and analyzed only non-target stimuli. Some studies reported the same results in responses to both target and non-target stimulation (Missonnier et al., 2007; Kukleta et al., 2009). One explanation of our results is that the beta band connection increases with higher load during normal aging, which suggests that the amount of processing resources allocated to audiovisual tasks is larger for old than young adults. To achieve audiovisual tasks, old adults need to activate more beta band connection than young adults. Our result is in line with previous studies that old adults exhibit larger responses in the beta frequency range during cognitive processing (Sebastian et al., 2011; Sallard et al., 2016). Hong et al. (2016) used a Go/NoGo task to examine the effects of aging on brain networks, and showed increased phase synchrony in the beta band that was more robust in old adults. Zarahn et al. (2007) showed an increased beta response during the memory load task. In addition, researches have reported that lower cognitive reserve was related to higher functional connectivity (Lopez et al., 2014).

The NBS results shown the connected structures are significantly different between groups. Statistical results in task-dependent network topology parameters confirmed this difference. Furthermore, the peak time point of each probability difference curve was different between old and young adults (**Figure 5C**). The peak time point represents the likely occurrence of audiovisual integration. In the beta band, network topology parameters of audiovisual processing showed strong correlations with peak time points in old adults but not in young adults (**Figure 6**). However, no age differences were detected for unimodal stimuli. This finding indicates that beta band functional connectivity influences the performance of audiovisual integration during normal aging. Old adults need more cognitive resources to perform highly demanding tasks (Sakowitz et al., 2005), which leads to changes in communication within the cortical system. Diaconescu et al. (2013) revealed that the engaged additional regions during audiovisual stimuli compared to younger adults. Audiovisual integration requires a higher level of cognition than visual or auditory processing and requires an old adult to think. However, young adults do not need to try to achieve tasks that lead to a low relationship with behavior results. Our results are supported by the study by Steffener et al. (2014), who revealed the relationship between cognitive performance and functional brain activity. These previous findings suggested that increased functional brain activity relates to worse (slower) task performance in old adults but not in young adults.

Therefore, our study is in good accordance with previous studies, which showed that audiovisual integration is different between old adults and young adults. Furthermore, the oscillatory beta network functional connectivity increased and graph characteristics changed during normal aging, which influence the reaction to audiovisual stimuli.

In the future, we will determine how to adjust beta band functional connectivity to benefit audiovisual integration during normal aging. Because our participants included old adults who are unable to adapt the long-term experiment, we chose 30 scalp electrode channels to construct the brain network. One main limitation of this study may be that the node of the network is relatively small.

#### ETHICS STATEMENT

fnagi-09-00239 August 3, 2017 Time: 16:19 # 10

This study was carried out in accordance with the recommendations of ethics committee of Okayama University with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the ethics committee of Okayama University.

#### AUTHOR CONTRIBUTIONS

LW analyzed and interpreted the data, wrote the paper. WW analyzed and interpreted the data. JS, WY, and QH performed the experiments. BW and RG conceived and designed the

### REFERENCES


experiments. TY and JW revised the paper, approved the final version.

#### ACKNOWLEDGMENTS

We acknowledge and thank the subjects involved in the study. This study was financially supported by the National Natural Science Foundation of China (grant numbers 61473043 and 81671776), the Beijing Municipal Science & Technology Commission (grant number Z161100002616020), Beijing Nova Program (grant number Z171100001117057), and in part by the "111" Project under Grant B08043.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2017.00239/full#supplementary-material


evoked potentials. Cogn. Brain Res. 23, 316–326. doi: 10.1016/j.cogbrainres. 2004.10.012


synchronization. Int. J. Psychophysiol. 38, 301–313. doi: 10.1016/S0167- 8760(00)00172-0


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer EP and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.

Copyright © 2017 Wang, Wang, Yan, Song, Yang, Wang, Go, Huang and Wu. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnagi-09-00266 August 8, 2017 Time: 15:42 # 1

# Altered Functional and Causal Connectivity of Cerebello-Cortical Circuits between Multiple System Atrophy (Parkinsonian Type) and Parkinson's Disease

Qun Yao<sup>1</sup> , Donglin Zhu<sup>1</sup> , Feng Li<sup>1</sup> , Chaoyong Xiao<sup>2</sup> , Xingjian Lin<sup>1</sup> , Qingling Huang<sup>2</sup> and Jingping Shi<sup>1</sup> \*

<sup>1</sup> Department of Neurology, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China, <sup>2</sup> Department of Radiology, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Jesus M. Cortes, BioCruces Health Research Institute, Spain Foteini Protopapa, Scuola Internazionale di Studi Superiori Avanzati (SISSA), Italy Constantinos Siettos, National Technical University of Athens, Greece Roma Siugzdaite, Ghent University, Belgium

> \*Correspondence: Jingping Shi profshijp@163.com

Received: 28 April 2017 Accepted: 26 July 2017 Published: 10 August 2017

#### Citation:

Yao Q, Zhu D, Li F, Xiao C, Lin X, Huang Q and Shi J (2017) Altered Functional and Causal Connectivity of Cerebello-Cortical Circuits between Multiple System Atrophy (Parkinsonian Type) and Parkinson's Disease. Front. Aging Neurosci. 9:266. doi: 10.3389/fnagi.2017.00266 Lesions of the cerebellum lead to motor and non-motor deficits by influencing cerebral cortex activity via cerebello-cortical circuits. It remains unknown whether the cerebello-cortical "disconnection" underlies motor and non-motor impairments both in the parkinsonian variant of multiple system atrophy (MSA-P) and Parkinson's disease (PD). In this study, we investigated both the functional and effective connectivity of the cerebello-cortical circuits from resting-state functional magnetic resonance imaging (rs-fMRI) data of three groups (26 MSA-P patients, 31 PD patients, and 30 controls). Correlation analysis was performed between the causal connectivity and clinical scores. PD patients showed a weakened cerebellar dentate nucleus (DN) functional coupling in the posterior cingulate cortex (PCC) and inferior parietal lobe compared with MSA-P or controls. MSA-P patients exhibited significantly enhanced effective connectivity from the DN to PCC compared with PD patients or controls, as well as declined causal connectivity from the left precentral gyrus to right DN compared with the controls, and this value is significantly correlated with the motor symptom scores. Our findings demonstrated a crucial role for the cerebello-cortical networks in both MSA-P and PD patients in addition to striatal-thalamo-cortical (STC) networks and indicated that different patterns of cerebello-cortical loop degeneration are involved in the development of the diseases.

Keywords: functional connectivity, granger causality analysis, multiple system atrophy, Parkinson's disease, resting-state fMRI

### INTRODUCTION

The parkinsonian variant of multiple system atrophy (MSA-P) is a neurodegenerative disorder that is clinically difficult to differentiate from idiopathic Parkinson's disease (PD), especially in the early stages of the diseases (Wenning et al., 2000; Kim et al., 2016). Inchoate differentiation between MSA-P and PD has significant therapeutic and rehabilitative implications. At the neuronal level, these diseases are all characterized by extensive cell loss in the substantia nigra pars compacta fnagi-09-00266 August 8, 2017 Time: 15:42 # 2

(Dickson, 2012). In the past, functional brain imaging had been proved to be of some value for the differential diagnosis of parkinsonism. Positron emission tomography, for instance, disclosed decreased striatal presynaptic uptake, binding, glucose metabolism, and post-synaptic binding in both MSA-P and PD (Ghaemi et al., 2002; Bohnen et al., 2006), especially the reduced post-synaptic binding in MSA-P. Liu et al. (2013) described that the dopamine deficits in striatal subregions impair the function of striatal-thalamo-cortical (STC) networks involved in motor, cognitive and emotional processing (Braak and Braak, 2000).

Considering the demonstrated alteration of pathology was found in the basal ganglia of PD and MSA-P (Galvan et al., 2015). The cerebellum is also an important component in motor control, higher cognitive, and emotional processing (Allen et al., 2005; Schutter and van Honk, 2005; Habas, 2010). It is known to affect cerebral cortical activity by cerebello-thalamocortical (CTC) circuits (Middleton and Strick, 2000). Previous studies have demonstrated the cerebellum to be involved in these diseases. For example, in MSA-P patients, morphological and microstructural alterations of cerebellum have been reported (Nicoletti et al., 2006a,b). The cerebellar functional activation was increased in MSA-P after repetitive transcranial magnetic stimulation (rTMS) treatment (Wang H. et al., 2016). Moreover, the cerebellum is anatomically and functionally connected with the basal ganglia and its connectivity changes in PD have been discovered (Wu and Hallett, 2013). Wu et al. (2009) found increased cerebellar activity in PD by a regional homogeneity method. Another causal connectivity study has shown that the connectivity of cortico-cerebellar motor regions is strengthened in PD during the performance of self-initiated movement (Wu et al., 2011). A PET research has also revealed that the cerebellum is a crucial node in the abnormal metabolic patterns of both MSA-P and PD (Poston et al., 2012), with decreased cerebellar 18F-fluorodeoxyglucose metabolism in MSA-P and increased in PD. It has been presumed that compensatory activity in CTC circuits in PD patients may act as a compensatory mechanism to overcome the deficits in the STC circuits (Cerasa et al., 2006; Palmer et al., 2010). However, the exact role of the cerebellum in Parkinsonism, especially the MSA-P, remains unclear.

The cerebellar outputs polymerize to the dentate nucleus (DN), which successively sends neural fibers to the thalamus and cerebral cortex via the superior cerebellar peduncles, thus completing the cerebello-cortical circuits (Middleton and Strick, 2000). Histological studies have demonstrated anatomically segregated cerebello-cortical circuits including motor and non-motor loops (Clower et al., 2001; Middleton and Strick, 2001). Available evidence regarding the cerebellocortical circuits connecting the lateral cerebellum to motor and non-motor cortical areas is limited in humans, due to technical challenges in assessing the long polysynaptic connections between the cerebellum and the cerebral cortex in vivo. Although the established cerebellar involvement in Parkinsonism, subtile studies on cerebellum for the differential diagnosis of parkinsonian syndromes are still penurious to date. Recently, resting-state fMRI (rs-fMRI) has been widely used to discover abnormalities in spontaneous neuronal activity by measuring the functional connectivity between spatially distinct

brain regions (Biswal et al., 1995). Functional connectivity (FC) is defined as statistical dependencies among remote neurophysiological events. However, granger causality analysis (GCA) is another widely used method for identifying directed functional ('causal') connectivity in neural time series data (Roebroeck et al., 2005). GCA has been applied to human EEG data (Hesse et al., 2003). Moreover, GCA has recently also been applied to human fMRI data based on temporal order (Friston, 2009; Seth et al., 2013). It has been widely used in exploring cognitive functions such as working memory (Protopapa et al., 2014, 2016), as well as other neurological disorders (Brovelli et al., 2004; Jiao et al., 2011). The DN is the largest single structure linking the cerebellum to the rest of the brain. Accordingly, we selected the bilateral DN as regions of interest (ROIs) to explore the different roles of the cerebellum in PD and MSA-P.

Only a few studies have investigated DN connectivity changes in the resting state in PD (Liu et al., 2013; Ma et al., 2015). However, there has been no reported data about MSA-P. In this study, we focused on potential connectivity changes between the DN and cerebral cortices. We hypothesized that the connectivity between the DN and cortical or subcortical regions may be altered and implicate motor symptoms difference between PD and MSA-P. To test this hypothesis, the functional connectivity (FC) and multivariate granger causality analysis (mGCA) methods were combined to explore the connectivity differences within the cerebello-cortical circuits during resting state.

### MATERIALS AND METHODS

### Participants

All subjects were recruited from Nanjing Brain Hospital from June 2013 to December 2015. Twenty-six MSA-P patients, 31 PD patients and 30 normal subjects were recruited into this study. The patients were diagnosed by a movement disorders specialist using established criteria: PD based on United Kingdom PD Society Brain Bank criteria (Hughes et al., 1992) and probable MSA-P based on the American Academy of Neurology and American Autonomic Society criteria (Gilman et al., 2008). Five subjects (2 MSA-P, 2 PD, and 1 control) were excluded due to excessive head motion during the fMRI procedure or incomplete scanning data, yielding a total of 24 MSA-P patients, 29 PD patients and 29 controls for the final analysis. All subjects underwent comprehensive neuropsychological assessments. Overall cognitive condition was assessed by the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA) and Frontal Assessment Battery (FAB). The severity of motor symptoms for all patients was assessed using the motor part of Unified Parkinson's Disease Rating Scale (UPDRS-III) and the Hoehn-Yahr (H-Y) scale. In this study, we used the sum of all left hemibody Part III items 20–26 (UPDRS-III L), the sum of all right hemibody Part III items 20–26 (UPDRS-III R) and the sum of all Part III items (UPDRS-III total). In addition, MSA-P patients were evaluated by the Unified Multiple System Atrophy Rating Scale (UMSARS), which was conducive to classification. Assessments were performed on the day before fMRI scanning in all subjects. Patients who had hemorrhage, infarction, tumors, trauma, or severe white matter hyperintensity were excluded from the study. All participants had written informed consent and the study was approved by the Medical Research Ethical Committee of Nanjing Brain Hospital, Nanjing, China.

#### Image Acquisition

fnagi-09-00266 August 8, 2017 Time: 15:42 # 3

All scans were acquired using a Siemens 3.0 T singer scanner (Siemens, Verio, Germany) with an 8-channel radio frequency coil. All subjects lay supine with their head cozily fixed by sponge earplugs to minimize head movement. Participants were instructed to remain as still as possible, close their eyes, remain awake and not think of anything. Three-dimensional T1 weighted images were acquired in a sagittal orientation employing a 3D-SPGR sequence with the following parameters: TE = 3.34 ms; TR = 2530 ms; flip angle = 7 ◦ ; 128 sagittal slices; 1.33 mm slice thickness; matrix = 256 × 256. Functional images were collected using a gradient-recalled echo-planar imaging pulse sequence: 140 time points (that were sufficient to assess resting state connectivity); TE = 30 ms; TR = 2000 ms; FOV = 240 mm × 240 mm; matrix = 64 × 64; flip angle = 90◦ ; 30 axial slices; 3.0 mm thickness; section gap = 0 mm.

### Definition of Regions of Interest (ROIs)

Regions of interest of the left and right DN were defined by WFU PickAtlas<sup>1</sup> and were resliced into Montreal Neurological Institute (MNI) space. The blood oxygen level dependent (BOLD) time series of the voxels within the ROI were extracted to generate the reference time series for each ROI.

### Data Preprocessing

The fMRI data were preprocessed using Data Processing Assistant for Resting-State fMRI toolkit (DPARSF<sup>2</sup> ), which is based on the Statistical Parametric Mapping software SPM8<sup>3</sup> . The first 10 volumes of the rest session were discarded for each subject. The remaining images were corrected for slice timing and motion correction. According to the record of head motion, all participants had less than 2.0 mm maximum displacement in the x, y, or z plane and less than 2◦ of angular rotation about each axis. After spatial normalization to T1 space, all images were resampled into 3 mm × 3 mm × 3 mm voxels and spatially smoothed with a Gaussian filter of 4 mm full-width at half-maximum (FWHM). The fMRI data were then temporally band-pass-filtered (0.01–0.08 Hz) to remove low-frequency drifts and physiological high-frequency noise. To further reduce the effects of confounding factors, linear drift, six motion parameters and the mean time series of all voxels within the entire brain, white matter and cerebrospinal fluid signals were removed from the data by linear regression.

In the fMRI data, global signal can be defined as the time series of signal intensity averaged across all brain voxels, which includes both the signal of the neural activity and the noise of the non-neural activity. Global signal regression (GSR) uses linear regression to remove variance between the global signal and the time course of each individual voxel. It can improve the specificity of positive correlations and improve the correspondence to anatomical connectivity. It helps remove non-neuronal sources of global variance such as respiration and movement. With the growing use of GSR, some problems it brings have led to some controversy. Different processing techniques likely produce different complementary insights into the brain's functional organization. Whether GSR should be useful or not depend on the scientific question and how we use it. If applied and interpreted correctly, they provide complementary information (Murphy and Fox, 2016). In this study, we are concerned with the neural activity of a particular brain area, and are not interested in the noise of those non-neural activities, so we applied GSR to remove global signal variance from all voxel time series.

The voxel based morphometry (VBM) was processed and examined using SPM8 software. The 3D-T1 weighted images were segmented into gray matter, white matter and cerebrospinal fluid. The gray matter and white matter images were then normalized and resampled to MNI space in 1.5 mm cubic resolution with modulation to preserve the local tissue volumes. The resulting images were smoothed using an 8 mm FWHM Gaussian kernel.

#### Statistical Analysis

To compare certain demographic information and clinical characteristics (age, education, MMSE, MoCA, FAB), one-way analysis of variance (ANOVA) was used. Disease duration, UPDRS-III, H-Y, and levodopa equivalent daily dose (LEDD) were compared using the two-sample t-test. A chi-squared test was used to compare sex distribution among the groups. VBM analysis among MSA-P, PD, and control groups was carried out with the ANOVA, followed by Bonferroni test for post hoc comparisons. All data were statistically analyzed using SPSS19.0 (SPSS, Inc., Chicago, IL, United States). Two-sided values of p < 0.05 were considered statistically significant.

### Functional Connectivity Analysis

Functional connectivity analysis was performed between each seed reference and the entire brain in a voxel-wise manner using the REST Toolkit<sup>2</sup> . Correlation coefficients were transformed to z-values using the Fisher r-to-z transformation to enhance normality. Statistical analysis across the three groups was conducted using one-way analysis of covariance (ANCOVA), with age, gender, disease duration, and gray matter volume as covariates. Then, post hoc two-sample t-tests were followed. The multiple comparisons of ANCOVA result was AlphaSim corrected with a cluster-level significance threshold of p < 0.01 (cluster size > 32 voxels and voxel-level p < 0.01; determined by a Monte-Carlo simulation). The post hoc two-sample t-tests were performed within a mask showing conspicuous differences acquired from the ANCOVA results, with a significance threshold of p < 0.01 with AlphaSim correction (cluster size > 14 voxels; voxel-level p < 0.01; determined by a Monte-Carlo simulation). The ANOVA was also performed with the similar results of ANCOVA after correction.

<sup>1</sup>http://www.ansir.wfubmc.edu

<sup>2</sup>http://www.rest.restfmri.net

<sup>3</sup>http://www.fil.ion.ucl.ac.uk/spm/software/spm8

#### Granger Causality Analysis

fnagi-09-00266 August 8, 2017 Time: 15:42 # 4

Causal connectivity refers to the influence that one neural system exerts over another. The distinction between functional and causal connectivity is that functional connectivity is ambiguous with respect to underlying directed interactions that generated the observed correlations. While causal connectivity corresponds to the intuitive notion of coupling or directed causal influence (Friston, 2011). GCA is a feasible technique for analyzing fMRI data (Wen et al., 2013). It is based on the idea that, given two time series x and y, if knowing the past of y is useful for predicting the future of x, then y must have a causal influence on x. GC method shows distinct and complementary functions in relation to the detection of causality and does not need predefined model. To identify the informational influence on the functional interactions and infer their causal relationship within the cerebello-cortical circuits, the mGCA method was applied (Blinowska et al., 2004; Chen et al., 2014). This approach is based on the MATLAB toolbox (The MathWorks, Inc., Natick, MA, United States). Briefly, to determine the structure of the cerebello-cortical circuits, nine ROIs were selected according to the abnormal functional connectivity patterns identified in the group comparisons in this study (for details refer to **Table 1**). Each seed region was represented by a radius of 6 mm around the central coordinates. The average time series for each ROI was extracted and may be expressed in Eq. 1.

$$X(t) = (\boldsymbol{\kappa}\_1(t), \boldsymbol{\kappa}\_2(t), \dots, \boldsymbol{\kappa}\_m(t))\tag{1}$$

where m represented the number of ROIs. The values of causal connectivity strengths from all other regions to region j were measured by multivariate auto regression (MVAR) model (Eq. 2).

$$x\_{\dot{l}}(t) = \sum\_{i=1}^{p} A\_{\dot{l}}(i)X(t-i) + E\_{\dot{l}}(t) \tag{2}$$

The parameter p is the model order or the lag parameter. Aj(i) is the regression coefficient matrix, X is the time series matrix of different regions, E is the residual error



Brain regions obtained from differential functional connectivity of cerebello-cortical network among MSA-P, PD, and controls; MNI, Montreal Neurological Institute; L, left; R, right.

matrix. The optimal lag parameter p is usually determined by minimizing Akaike Information Criterion (AIC) (Akaike, 1998). For each subject, random-effect Granger causality maps were calculated. Statistical thresholds for these maps was performed in the context of the bootstrap technique with corrections for multiple comparisons based on a permutation test (p < 0.05) (Chen et al., 2014). For each group, an average Granger causality map was created to illustrate the effective connectivity influence on the paired regions. Then, group comparisons were also conducted to identify the significantly altered causal influence between paired brain areas using the Kruskal–Wallis, followed by the Dunn–Bonferroni test for post hoc comparisons.

### Connectivity-Behavior Correlation

To explore whether the alterations of causal connectivity are covariant with disease progression, a correlation analysis between altered causal connectivity and neuropsychological performance metrics, disease duration and LEDD was performed separately for MSA-P and PD patients. First, causal connectivity values for the significant group differences were extracted. Then, Pearson's correlative analysis was conducted to examine the relationships between causal connectivity and neuropsychological scores [including MMSE, MoCA, FAB, UPDRS-III (total, L, R) and H-Y staging scale] and LEDD.

#### RESULTS

#### Demographic and Clinical Characteristics

The demographic and clinical characteristics of the participants are presented in **Table 2**. There were no significant differences in age, gender, education level, MMSE, and MoCA scores among the three groups. No significant differences were found in UPDRS-III (total, L, R) and H-Y staging scores between MSA-P and PD groups. Difference in disease duration was significant between MSA-P and PD (p < 0.001). The FAB scores were significantly different between MSA-P versus control (p = 0.001) and PD versus control (p = 0.037).

#### Voxel Based Morphometry

Voxel based morphometry did not reveal significant differences between patients and controls for gray matter volume and white matter volume (for details refer to **Table 2**).

#### Functional Connectivity

The ANCOVA results revealed that the left DN had significantly different FC values in the frontal, parietal or cingulate cortices among MSA-P, PD, and controls (**Table 2** and **Figure 1**). A further detailed investigation of these alterations in the three groups showed that the FC values of both MSA-P and PD patients were significantly decreased in the left dorsolateral prefrontal cortex (DLPFC) compared with the controls. PD patients exhibited lower FC values in the bilateral inferior parietal lobe compared to MSA-P or the controls. Reduced FC values in the posterior fnagi-09-00266 August 8, 2017 Time: 15:42 # 5


GMV, gray matter volume; WMV, white matter volume; MMSE, Mini-Mental State Examination; MoCA, Montreal Cognitive Assessment; FAB, Frontal Assessment Battery; UPDRS-III, Unified Parkinson's Disease Rating Scale-motor part III; H-Y, Hoehn-Yahr staging scale; LEDD, levodopa equivalent daily dose; NA, not applicable; p < 0.05 was considered significant. <sup>a</sup>Means significant difference between the two groups (MSA-P versus PD). bcMeans significant group differences between (MSA-P versus control), (PD versus control) by post hoc comparisons.

cingulate cortex (PCC) were also observed in the PD patients relative to the controls. Moreover, abnormal FC values of the right DN were observed throughout the frontal, parietal, and cingulate cortices by ANCOVA analysis (**Table 3** and **Figure 2**). Decreased FC values in the left precentral gyrus (M1) and right DLPFC were shown in MSA-P patients compared with the controls. PD patients produced attenuated connectivity in the PCC, medial PFC and bilateral DLPFC compared to the controls, as well as decreased FC values of the PCC compared to MSA-P.

#### Causal Connectivity

fnagi-09-00266 August 8, 2017 Time: 15:42 # 6

We further attested that the DN, as core regions of the neural circuitry, is causally influenced in the cerebello-cortical circuits. As shown in **Figure 3**, MSA-P and PD patients presented with different patterns of the connectivity strength of causal flow between the previously defined paired regions compared to controls. MSA-P patients showed significantly enhanced causal information from the left DN to the PCC compared to PD or controls, as well as lower information flow from the left M1 to the right DN relative to the controls. In the PD group, the right DLPFC exhibited obvious causal interaction disruption with the left DN compared to the controls, as shown in **Figure 3B**.

### Correlation Analysis with Clinical Behavior Scores

As shown in **Figure 3D**, the connectivity of the causal path from the left M1 to the right DN was significantly decreased in the MSA-P group, which was negatively correlated with UPDRS-III R scores.

TABLE 3 | Regions showing significant differences in DN FC among MSA-P, PD, and control group.


MNI, Montreal Neurological Institute; Significance set at p < 0.01 corrected by AlphaSim; Cluster size is in mm<sup>3</sup> ; Results are given in MNI space.

### DISCUSSION

Recently, many efforts have been made to detect early differences between MSA-P and PD patients (Wang et al., 2012; Ji et al., 2015). The novelty of this study lies in the fact that we have used unidirectional and directional connectivity methods to explore the unique association between altered cerebellocortical connectivity and motor and non-motor defects in both MSA-P and PD patients. Our research yielded several major findings. First, MSA-P an PD patients exhibited both functional and causal connectivity differences within cerebello-cortical networks compared with controls. Second, the comparison of MSA-P patients with PD patients by BOLD signal mainly revealed differences in the DN-PCC connectivity. Third, in the MSA-P group, a significant correlation was found between the UPDRS-III R scores and the causal connectivity from the left M1 to right DN.

Relative to the controls, MSA-P patients displayed disrupted functional and causal coupling between the left M1 and right DN. The classical view of M1, which was based principally on direct cortical stimulation and attributed to select the muscles and force for executing an intended movement. A recent diffusion tensor imaging study has reported stronger connections between the cerebellum and the precentral gyrus as well as the superior frontal gyrus (Doron et al., 2010), which indicated that the cerebellum involved in the processing of motor, oculomotor, and spatial working memory by cerebello-cortical circuits (du Boisgueheneuc et al., 2006). Rs-fMRI in MSA has shown reduced regional homogeneity of the spontaneous BOLD fluctuations in left M1 compared with controls (You et al., 2011). Wang H. et al. (2016) also found that the induced motor improvement in MSA-P patients by rTMS over left M1 may be associated with increased activation in the cerebellum. The cerebellum receives input from the M1 and then projects to thalamus, M1 (Kelly and Strick, 2003). In this study, the causal influence from M1 to the cerebellum in MSA-P patients was attenuated. This result could imply that the cerebellum exerts its influence on cortical inhibitory activity (Picazio and Koch, 2015). It may be one explanation for the underlying pathophysiology of motor deficits. Noticeably, the decreased causal connectivity from left M1 to right DN is significantly correlated with the contralateral motor symptom scores for the MSA-P group, which emphasized that the motor impairments in MSA-P patients could be influenced by cerebello-cortical loop degeneration in addition to striatal pathology.

In this study, the cerebellar functional connectivity was disrupted in several default mode network (DMN) regions (the PCC, medial PFC, and inferior parietal lobe) in PD patients compared with MSA-P patients or controls. In the baseline state, the PCC appears a high metabolic rate (Raichle et al., 2001) and plays a crucial role in modulating the balance between internal and external information for maintaining normal brain functions (Leech and Sharp, 2014). Lin et al. (2016) found that the PCC is a core node of the DMN and it is closely related with cognitive task performance, by aggregating information to allow functional cooperation within the DMN. A previous study reported that cognitively unimpaired patients with PD fnagi-09-00266 August 8, 2017 Time: 15:42 # 7

tend to show attenuated DMN activity compared to controls (Tessitore et al., 2012). The study associated DMN deficits with cognitive decline, because cognitive function was correlated with DMN connectivity. It suggested that there is an early functional disruption of the DMN in PD prior to clinical evidence of cognitive impairment. The DMN sub-regions play important roles in remembering, self-reflection, mental imagery, and stream-of-consciousness processing (Greicius et al., 2004; Buckner and Carroll, 2007). The consolidation and maintenance of brain function might be facilitated through the DMN plasticity. However, the mechanisms underlying modulations of the DMN are not yet clear. We hypothesized that the DN dysfunction contributed to abnormal recruitment of the PCC node of the DMN. The DN, which represents the capital cerebellar output channel, is considered to be implicated with salience and sensorimotor networks (Habas et al., 2009). The DMN deactivates during externally goal-directed activity (Gusnard et al., 2001) and is causally influenced by the salience network (Chiong et al., 2013). In a recent report, a whole-brain functional connectivity analysis method was used to uncover connectivity changes following rTMS intervention over M1 for MSA-P patients. This result indicated that the rTMS-related functional links were mainly connected to the cerebellar and limbic networks from the DMN (Chou et al., 2015). It seems plausible that DMN plasticity might be sensitive to the cerebellar network effects. Only one study has described reduced DMN activity in MSA-P (with a average disease duration of 3.9 years) (Franciotti et al., 2015). In this study, MSA-P patients showed significant strengthened causal connectivity from the left DN to the PCC compared with both PD patients and controls. The DN exerted a causal influence on the PCC, perhaps as a compensatory mechanism for DMN functions in the early stages. In previous study, the DMN activity was enhanced in dementia with lewy bodies (DLB), the explanatory hypothesis was that the preserved DMN activity could depend on compensatory mechanisms attempting to maintain DMN functions, in the face of developing pathology (Kenny et al., 2012). The consolidated DMN connectivity was similar to the enhancement observed in DLB, as we found no cellular loss, and this finding supports that functional modulatory mechanisms are relevant rather than structural differences.

The inferior parietal lobe within the executive network is involved with sustained attention and working memory information for action preparation (Seeley et al., 2007). In patients with PD, the functional parieto-motor impairment could be related to bradykinesia (Palomar et al., 2013). The cerebellar damage may impair the competence to convert a programmed motion sequence into action before the launch of movement (Bhanpuri et al., 2014; Brunamonti et al., 2014). The interruption of the dynamic equilibrium between the cerebellar and executive network may weaken the ability of the motor system to prepare for future task execution. Evidence from previous studies has demonstrated the existed functional connectivity between the cerebellum and parietal cortex (Macher et al., 2014). The decline in functional connectivity between the DN and the inferior parietal lobe may contribute to the impaired function of PD patients in linking simple movements together into complicated and sequential movements (Liu et al., 2013). fnagi-09-00266 August 8, 2017 Time: 15:42 # 8

right inferior parietal lobe; LDN, left dentate nucleus; RDN, right dentate nucleus.

However, such correlations between clinical observations and neural connectivity patterns need further explorations.

Both MSA-P and PD patients showed decreased connectivity between the DN and the DLPFC compared with the controls. Reduced causal information flows from the DLPFC to the DN were also found in the PD group compared to the controls. The DLPFC, a crucial node in the cognition control network (Wang Y.L. et al., 2016), is involved in attention, working memory and executive control (Pope et al., 2015). MSA patients presented with a distinctive pattern of cognitive deficits in frontal executive dysfunction (Siri et al., 2013). A PET study suggested that MSA patients with memory and frontal executive dysfunction tended to show hypometabolism in the anterior cerebellum and frontal cortex in the early stage of the disease (Lyoo et al., 2008). In addition, the hypoactivity of the DLPFC in PD patients with depression has also been identified in previous studies (Zhu et al., 2016). The impaired striatal cells in parkinsonism could lead to secondary frontal lobe dysfunction, including disruption of the cognitive loop linking the striatum with the DLPFC (Jokinen et al., 2013). Recently, the possibility of cerebellar influence on cognitive function in parkinsonism has also been raised (Hirata et al., 2006; Wu and Hallett, 2013; Kim et al., 2015). Abundant structural and functional investigations in both human and non-human primates revealed that the cerebellum is involved in higher-order cognitive and emotional processes by sending fibers from the DN to PFC via the thalamus (Allen et al., 2005; Ramnani, 2006). The decreased DLPFC connectivity may be significantly associated with executive control and emotional processes in both MSA-P and PD patients, although no significant correlations were obtained in this study. These results illustrated that the connectivity of the cerebellum with the motor and non-motor cortical domains is significantly involved in the PD and MSA-P disease process.

This analysis demonstrated the cerebellum to be a causal flow hub of the cerebello-cortical network, with the high number of causal flow connections. A possible correspondence of tractography between cortices with similar functional roles, as reported here, suggests that the cerebellum contributes to parallel associative cerebello-cortical networks involved in various aspects of motor and cognition. The converging results strongly indicated that the causal topology of the cerebellocortical circuits may be disrupted in both MSA-P and PD patients, adding an additional hint for comprehending the neurobiology underlying patients with MSA-P and PD.

#### CONCLUSION

fnagi-09-00266 August 8, 2017 Time: 15:42 # 9

This rs-fMRI study provides evidence that the dysfunction reported within the cerebello-cortical networks, typically related to motor and cognitive defects in MSA-P and PD. It may be associated with impaired interactions between the cerebellum and key cerebral cortical regions. In conclusion, our findings indicate a crucial role for the cerebello-cortical network in both MSA-P and PD patients in addition to STC network and revealed that different patterns of cerebello-cortical loop degeneration are involved in the development of the diseases. Furthermore, the alterations of the functional link within the

#### REFERENCES


cerebello-cortical circuits, especially the DN-PCC connectivity, may facilitate early differential diagnosis and the monitoring of disease progression.

#### AUTHOR CONTRIBUTIONS

JS designed the study and revised it critically for important intellectual content. QY performed the research and drafted the manuscript, DZ and FL helped in data analyses, XL, CX, and QH help in clinical data collection, analyses and made patients follow-ups.

### FUNDING

This work was supported by the Nanjing Bureau of Science and Technology (No. 201605040).


and progressive supranuclear palsy. Brain 129, 2679–2687. doi: 10.1093/brain/ awl166


fnagi-09-00266 August 8, 2017 Time: 15:42 # 10

fnagi-09-00266 August 8, 2017 Time: 15:42 # 11

definite multiple system atrophy from Parkinson's disease?J. Neurol. Neurosurg. psychiatry 68, 434–440.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Yao, Zhu, Li, Xiao, Lin, Huang and Shi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnagi-09-00267 August 10, 2017 Time: 16:26 # 1

# Does Aerobic Exercise Influence Intrinsic Brain Activity? An Aerobic Exercise Intervention among Healthy Old Adults

Pär Flodin1,2 \*, Lars S. Jonasson1,2,3, Katrin Riklund2,3, Lars Nyberg2,3,4 and C. J. Boraxbekk1,2,5

<sup>1</sup> Center for Demographic and Aging Research, Umeå University, Umeå, Sweden, <sup>2</sup> Umeå Center for Functional Brain Imaging, Umeå University, Umeå, Sweden, <sup>3</sup> Diagnostic Radiology, Department of Radiation Sciences, Umeå University, Umeå, Sweden, <sup>4</sup> Physiology, Department of Integrative Medical Biology, Umeå University, Umeå, Sweden, <sup>5</sup> Danish Research Centre for Magnetic Resonance, Centre for Functional and Diagnostic Imaging and Research, Copenhagen University Hospital Hvidovre, Copenhagen, Denmark

Previous studies have indicated that aerobic exercise could reduce age related decline in cognition and brain functioning. Here we investigated the effects of aerobic exercise on intrinsic brain activity. Sixty sedentary healthy males and females (64–78 years) were randomized into either an aerobic exercise group or an active control group. Both groups recieved supervised training, 3 days a week for 6 months. Multimodal brain imaging data was acquired before and after the intervention, including 10 min of resting state brain functional magnetic resonance imaging (rs-fMRI) and arterial spin labeling (ASL). Additionally, a comprehensive battery of cognitive tasks assessing, e.g., executive function and episodic memory was administered. Both the aerobic and the control group improved in aerobic capacity (VO2-peak) over 6 months, but a significant group by time interaction confirmed that the aerobic group improved more. Contrary to our hypothesis, we did not observe any significant group by time interactions with regard to any measure of intrinsic activity. To further probe putative relationships between fitness and brain activity, we performed post hoc analyses disregarding group belongings. At baseline, VO2-peak was negativly related to BOLD-signal fluctuations (BOLDSTD) in mid temporal areas. Over 6 months, improvements in aerobic capacity were associated with decreased connectivity between left hippocampus and contralateral precentral gyrus, and positively to connectivity between right mid-temporal areas and frontal and parietal regions. Independent component analysis identified a VO2-related increase in coupling between the default mode network and left orbitofrontal cortex, as well as a decreased connectivity between the sensorimotor network and thalamus. Extensive exploratory data analyses of global efficiency, connectome wide multivariate pattern analysis (connectome-MVPA), as well as ASL, did not reveal any relationships between aerobic fitness and intrinsic brain activity. Moreover, fitness-predicted changes in functional connectivity did not relate to changes in cognition, which is likely due to absent crosssectional or longitudinal relationships between VO2-peak and cognition. We conclude that the aerobic exercise intervention had limited influence on patterns of intrinsic brain activity, although post hoc analyses indicated that individual changes in aerobic capacity preferentially influenced mid-temporal brain areas.

Keywords: aerobic exercise, brain plasticity, aging, fMRI, resting-state, ASL

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Hans-Peter Müller, University of Ulm, Germany Berend Malchow, Ludwig-Maximilians-Universität München, Germany

> \*Correspondence: Pär Flodin parflodin@gmail.com

Received: 09 April 2017 Accepted: 26 July 2017 Published: 11 August 2017

#### Citation:

Flodin P, Jonasson LS, Riklund K, Nyberg L and Boraxbekk CJ (2017) Does Aerobic Exercise Influence Intrinsic Brain Activity? An Aerobic Exercise Intervention among Healthy Old Adults. Front. Aging Neurosci. 9:267. doi: 10.3389/fnagi.2017.00267

### INTRODUCTION

fnagi-09-00267 August 10, 2017 Time: 16:26 # 2

Given the increasing disease burden of age related cognitive problems that comes with an aging world population (Christensen et al., 2009), physical exercise could provide a widely available and cost-effective approach to reduce age related cognitive decline at a large scale (Baker et al., 2010). Prospective studies show that low fitness in early adulthood is associated with increased risk for early onset dementia later in life (Nyberg et al., 2014; Wendell et al., 2014). Moreover, cross-sectional population based studies have confirmed that individuals that stay physically active have an improved brain-behavior relationship (Boraxbekk et al., 2016).

Human intervention studies probing the cognitive effects of exercise have, however, showed mixed findings. Meta-analyses have reported a range of effect sizes on exercise induced improvements of cognition among elderly. Smith et al. (2010) reported modest improvements in attention, processing speed (PS), executive function (EF), and memory. Colcombe and Kramer (2003) report medium improvements (of particularly) EFs, whereas a recent systematic meta-analysis concludes that there is no evidence that aerobic exercise benefit cognition among healthy older adults (Young et al., 2015). Thus, there is a need for further investigations of the extent to which physical exercise interventions among elderly could maintain, or even restore, cognitive function and brain health. For investigations of the neurophysiological mechanisms subserving the neuroprotective effects of aerobic exercise, human brain imaging plays a central role (for a review see, e.g., Stillman et al., 2016).

Evidence suggests that resting state brain activity could be sensitive to also early-stage neuroplastic brain changes (Kelly and Castellanos, 2014). To date, very few studies have investigated changes in intrinsic brain activity following a structured physical exercise intervention among healthy older adults. Voss et al. (2010) investigated longitudinal changes in three age sensitive brain networks, using a region of interested (ROI) based seed correlation analysis (SCA), following 6 and 12 months supervised cardiovascular training. Compared to an active control group, the aerobic (walking) group increased connectivity in parts of the default and frontal executive network after 12 months, although no significant group differences were observed at 6 months.

Among the cross-sectional studies linking fitness to intrinsic brain activity among elderly, a commonly reported finding is reversal of age-related changes. Voss et al. (2016) used network based statistics (NBS) on graphs based on hubs defined in age-sensitive networks. These networks were identified as the default mode- (DMN), dorsal attention-, and executive control-, salience- and sensory related networks. Whereas older subjects displayed enhanced between-network connectivity, younger subjects displayed larger within-network connectivity. They concluded that cardiorespiratory fitness among older was positively associated with connectivity within age-sensitive networks, primarily the DMN, whereas no associations were observed for self-reported physical activity. However, another large scale cross-sectional study, Boraxbekk et al. (2016) found that current and accumulated physical activity was associated with stronger integrity of the DMN in the anterior parts of posterior cingulate cortex (PCC).

In the current study, we wanted to expand on the findings relating aerobic fitness to intrinsic brain activity. In addition to investigate resting-state functional connectivity (i.e., correlations of BOLD-signal time series of distributed brain regions), we also examined fluctuations of BOLD-signal time series (BOLDSTD). Resting state BOLDSTD has previously been used as a proxy measure for vascular flexibility (Burzynska et al., 2015). Moreover, vascular stiffness has in previous studies been linked to aging and cognitive decline (Mitchell et al., 2011; Gauthier et al., 2015), and there is evidence that physical activity counteracts age related vascular stiffness. Burzynska et al. (2015) detected a positive relationship between variability in intrinsic brain activity and physical activity (measured with actigraphs), but not for cardiovascular fitness (VO2-max). The authors concluded that BOLD-signal fluctuations could provide a putative neural correlate of brain health among elderly, and that "longitudinal and intervention studies will shed more light on the potential of BOLD in detecting changes in brain function as a result of increased physical activity" (Burzynska et al., 2015). Using a more direct MRI based measure of vascular rigidity, aortic pulse wave velocity, Gauthier et al. (2015) reported an increase of vascular rigidity with age, and a negative association to VO2-max (i.e., maximum rate of oxygen consumption).

Another way to characterize vascular function (which is a likely target for aerobic exercise) is quantification of cerebral blood flow (CBF). Maass et al. (2015) used gadolinium-based perfusion imaging MRI to investigate longitudinal changes among healthy older adults that partook in a 3-months exercise intervention, and found that improvements in aerobic capacity correlated with increases in hippocampal blood flow. In cross-sectional data, Boraxbekk et al. (2016) found a positive relationship between physical activity and CBF (ASL) in PCC.

To our knowledge, the current study is the first to employ pure resting-state functional magnetic resonance imaging (rsfMRI) scans before and after an actively controlled, long-term physical exercise intervention among healthy elderly. In the current article, we present the resting-state results of the research project "Physical Influences on Brain in Aging" (PHIBRA), for which the design, cognitive performance, and structural MRIdata recently were reported (Jonasson et al., 2017). By using a comprehensive set of analytical approaches of rs-fMRI data, and additional measures of ASL-CBF, we attempted to replicate and extend on previous research that evaluates the impact of aerobic fitness on intrinsic brain activity. Based on previous studies we predicted that analyses of group by time interactions would reveal differential longitudinal changes in (1) hippocampal resting state functional connectivity and betweenness centrality, (2) DMN integrity (3) voxel wise BOLDSTD and CBF. Additionally, we aimed to characterize relationships between rs-fMRI and fitness using more exploratory approaches. These included multivariate representations of whole brain connectivity (in the following referred to as multivariate pattern analysis, connectome-MVPA), NBS and graph theoretical measures (global efficiency). In order to investigate the functional significance of any observed fitness-brain relationship stipulated above, we performed post hoc analyses relating neurophysiological changes to changes in cognition.

### MATERIALS AND METHODS

fnagi-09-00267 August 10, 2017 Time: 16:26 # 3

#### Subjects

Sixty healthy but sedentary older adults (age 64–78 years) were recruited and randomized into performing either supervised aerobic exercise training, or stretching and toning control training, for 6 months, three times a week (30–60 min per session). In the present analysis, 13 subjects were excluded due to the following reasons: one subject was excluded due to brain abnormalities; two subjects were unable to complete the intervention; one person was excluded since baseline VO2-peak exceeded the group average by three standard deviations. Another nine persons were excluded due to suprathreshold movement during resting state scans at either pre- or post intervention sessions, see criteria below. Thus, complete data were obtained from 47 subjects, 22 in the aerobic intervention arm and 25 in the active control group. The groups did not significantly differ with respect to age [t(45) = 0.92, p = 0.36], sex [χ2(1) = 0.045 p = 0.83], education [t(45) = 0.1, p = 0.36], or BMI [t(45) = 1.01, p = 0.32], see **Table 1**.

For the ASL analyses, pre- and post intervention data was available from 55 subjects (aerobic group: 30 subjects, 16 females; control group: 25 subjects, 14 females), age 68.7 ± 2.79 years. There were no significant group differences in age, [t(53) = −1.27, p = 0.21), sex χ2(1) = 0.040 p = 0.843], education [t(53) = −0.087, p = 0.93], or BMI [t(53) = −0.98, p = 0.33]. For complete sample characteristics, see Jonasson et al. (2017).

### Intervention Procedures and Behavioral Outcome Measures

For detailed descriptions of the intervention, the VO2-peak assessment, and the cognitive test battery, see Jonasson et al. (2017). In short, the intervention procedure was as follows: Subjects were recruited from local newspaper advertisement. Prior to randomization into either of the two intervention arms, all subjects underwent pre-intervention testing. The testing consisted of visits to the lab at six different days. On the first testing day participants visited the Sports Science Lab, Umeå University, where fitness measures, including VO2-peak, were assessed using a standardized graded cycle ergometer test. Neuropsychological testing was performed on three separate days. Assessments of cognitive ability included three tests designed to tax (verbal) episodic memory (EM), three test for EF, as well as other tests including assessments of working memory updating (UPD) and PS. A composite score, cognitive score (CS), was computed as a unit-weighted average of EM, EF, UPD, and PS, as a single metric of overall cognition. Finally, MRI-data of different modalities was collected, of which the rs-fMRI and the ASL-CBF data is reported here. Additional questionnaires and dopamine positron emission tomography (PET) imaging were also collected, but will be reported elsewhere. The testing was conducted both before and after the interventions. The cognitive 


testing, including the tests assessing EM, EF, and CS that are used here, is described in detail in Jonasson et al. (2017).

### Acquisition

fnagi-09-00267 August 10, 2017 Time: 16:26 # 4

Brain imaging data was acquired on a 3 T GE scanner equipped with a 32-channel head coil. High-resolution T1-weighted structural images were acquired using the following parameters: 180 slices; 1 mm thickness; repletion time (TR) 8.2 ms; time to echo (TE) 3.2 ms; flip angle 12◦ ; FOV 25 cm × 25 cm. Functional data were acquired with a gradient EPI sequence with the following parameters: 37 transaxial slices, 3.4 mm thickness, 0.5 mm gap, TR 2000 ms, TE 30 ms, flip angle 80◦ , FOV 25 cm × 25 cm, inplane resolution of 2.6 mm × 2.6 mm. During resting state scan acquisition, subjects were instructed to lie still and keep their eyes on a white fixation cross for 9 min and 40 s, resulting in 290 volumes per subject and session.

Whole brain perfusion was measured with a 3D pseudocontinuous ASL sequence. Labeling time was 1.5 s, post-labeling delay time was 2 s, field of view was 24 cm, slice thickness was 4 mm, number of averages was 3, number of control label pairs was 30, and acquisition resolution was 8 × 512 (arms × data points in spiral scheme). Forty slices covered the whole brain and the reconstructed voxel size was 1.88 mm × 1.88 mm × 4 mm. CBF maps were computed using the standard GE reconstruction, showing tissue CBF in ml/min/100 g.

#### MRI Data Preprocessing and Denoising

Rs-fMRI data preprocessing was performed in SPM12 using standard preprocessing steps. These included slice time correction, realignment and unwarping using 6th degree B-spline interpolation, and functional to structural coregistration. Structural images were segmented into gray matter, white matter and cerebral spinal fluid images. Coregistered functional and structural images were normalized to standard space using Geodesic Shooting as implemented in the Shoot toolbox in SPM12 (Ashburner and Friston, 2011). In short, a common anatomical group-average 1 mm isotropic template was generated, and subject specific deformation fields were created and used for warping T1- and T2<sup>∗</sup> -weighted images from subject space to group template space. Secondly, non-linear deformation fields pushing images from group space to MNI space were applied. Functional data was then resampled into 2 mm × 2 × 2 mm and finally smoothed using a Gaussian kernel of 8 mm.

Cerebral blood flow data was coregistered to structural data, and normalized using the same normalization procedures as applied to the structural and functional data.

At the subject level, rs-fMRI time series were denoised by controlling for: (1) suprathreshold movement (frame wise displacement >0.25 mm or >3 std change in global signal intensity, similar to, e.g., (Satterthwaite et al., 2013), (2) signals from white matter and cerebrospinal fluid (five most explanatory principal components from each tissue mask), and (3) six movement parameters obtained from spatial realignment plus their time derivatives. Subsequent to nuisance regression, time signals were filtered (band passing 0.008–0.09 Hz). Nine subjects lost >50% of their volumes either at pre- or post intervention sessions (of originally 290 volumes) due to movement censoring, and were therefore excluded from further analysis. Denoising of data was accomplished with the Conn toolbox (version 15.h).

### fMRI Analysis

Longitudinal effects of the intervention were evaluated as group by time interactions. Voxel-wise group analyses of functional connectivity were performed using the Conn toolbox, where the two experimental groups were compared in terms of longitudinal changes with regard to the resting state metrics. Likewise, 2-sample t-tests of the post minus pre intervention BOLDSTD and CBF volumes were analyzed using the 2-sample t-test as implemented in SPM12. Group level statistics of graph-theoretical indices (of betweenness centrality and global efficiency) were calculated using repeated measure ANOVA (using MatLab) with group and time as factors.

Post hoc analyses aimed to link individual longitudinal changes in VO2-peak to changes in brain activity were also conducted, ignoring experimental group-belonging. For these we performed linear regressions, where changes in brain activity were regressed on changes in VO2-peak. To guide longitudinal analyses, rsfMRI data were first related to aerobic capacity at baseline. Any obtained significant brain-fitness association at baseline was used to inform subsequent longitudinal analysis (see below).

In all group analyses, we controlled for age and sex (as e.g., done in Maass et al., 2015). For the longitudinal analyses, we also controlled for baseline ratings of VO2-capacity. Additionally, since the two experiential groups were matched with regard to sex and age, the analyses of group by time interactions were also performed without control for sex, age and VO<sup>2</sup> – peak ratings. In all second-level fMRI analyses, we controlled for mean frame wise displacement (FD).

#### Seed Based Correlation Analysis

Voxel-wise seed based correlation analyses (SCA) were performed using the Conn toolbox (15.h). For each subject and each seed region, z-transformed Pearson correlation maps were brought to second level group analysis. We defined a priori seeds in brain regions that in previous literature have been influenced by aerobic exercise. These included (a) left and right hippocampus obtained from the SPM Wake Forest University (WFU) Pickatlas toolbox (Maldjian et al., 2003), (b) left and right parahippocampus (WFU pickatlas), and (c) PCC, based on a cluster reported by Boraxbekk et al. (2016) (center of gravity MNI coordinates −6, −32, 27; 67 voxels).

#### BOLD-Signal Fluctuations

Variations in the preprocessed and cleaned resting state BOLD time series were quantified using FSL maths<sup>1</sup> , rendering one voxel-wise whole brain map of the standard deviations (BOLDSTD) for each subject and each session. As a corroborating measure, we also calculated Amplitude of Low Frequency Fluctuations (ALFF) which quantifies the total power within the lower frequency band. ALFF was calculated for the same preprocessed and cleaned volumes as for the BOLDSTD

<sup>1</sup>http://www.fmrib.ox.ac.uk/fsl

calculations, using the MatLab toolbox DPARSFA<sup>2</sup> (Chao-Gan and Yu-Feng, 2010). Thus, for the examined ALFF calculations, the examined frequency window was essentially the same as the bandpass filter used in the resting state data, i.e., 0.008–0.09 Hz.

#### Graph Analyses

fnagi-09-00267 August 10, 2017 Time: 16:26 # 5

For the network analyses, we created subject specific weighted graphs based on 264 nodes defined as spherical (4 mm radius) ROIs centered around functionally relevant coordinates (Power et al., 2011; Cole et al., 2013) (see Supplementary Table S1). Edges were defined as Pearson correlation values between each time series of each ROI–ROI pair, where negative correlations were nulled. For measures of hippocampal betweenness centrality, we added two nodes corresponding to the left and right hippocampus (see Seed Based Correlation Analysis) to the original (264 nodes) graph. A high degree of betweenness centrality imply that nodes have central positions, i.e., act as hubs in the given network. Global efficiency is a measure of how efficiently information is exchanged across a network. All graph theoretical measures (i.e., Betweenness Centrality for left and right hippocampus, and global efficiency) were calculated using the brain connectivity toolbox (Rubinov and Sporns, 2010) as implemented in GraphVar1.0 (Kruschwitz et al., 2015).

#### Independent Component Analyses

Independent component analysis (ICA) was performed using the conn toolbox (v16, Whitfield-Gabrieli and Nieto-Castanon, 2012). Comparable ICA's across subjects were obtained through a dual regression procedure, where we first calculated 20 group level spatial independent components (default value), and subsequently used the time series pertaining to each component to obtain subject-specific spatial components. Longitudinal change in five different networks were analyzed with regard to baseline and change in aerobic capacity, respectively. The ICA components of interest were chosen in an attempt to conceptually replicate the analysis of resting state networks investigated in Voss et al. (2016). Specifically, these included the default mode network (DMN), right and left frontoparietal network (FPN), dorsal attention network (DAN), and Sensorimotor Network (S1M1) (see Supplementary Figure S1).

#### Network Based Statistics

We also conducted NBS analysis (Zalesky et al., 2010), on the graph consisting of the same 264 nodes as used for graph analysis (see Graph Analyses). NBS allows for identification of groupeffects among ROI–ROI connections or subnetworks, while controlling for multiple comparisons. NBS is the conceptual equivalent to cluster statistics for voxel wise whole brain analyses. Analogous to cluster statistics, NBS yields stronger statistical sensitivity for distributed effects, which in turn is traded for spatial resolution which is limited to the size of the identified network (c.f. cluster), rather than individual ROI–ROI connections (c.f. voxels). NBS was used to identify ROI–ROI pairs or subnetworks that revealed group by time interactions, or where connectivity related to aerobic capacity (either at baseline, or in change-change regressions). For this, we employed a typically used network defining ROI– ROI ("primary") connectivity threshold (Zalesky et al., 2010); p = 0.001 uncorrected. In order to detect weaker but spatially more distributed sub-graphs for 2nd level contrasts where the original ROI–ROI connectivity threshold did not show any effect, we also investigated fitness-network associations using ROI–ROI connectivity thresholds of p = 0.01 and p = 0.05 uncorrected (two sided). For NBS (i.e., "network component intensity") threshold, we used the significance level of family-wise error correction (FWE) p = 0.05 (1000 permutations).

#### Whole Brain Pattern of Functional Connectivity

Additionally, we performed a principal component MVPA, or so called "connectome-MVPA," to detect patterns of whole brain connectivity that correlated with aerobic capacity. The MVPA analysis complements the SCA and ICA since the investigated connectivity is not restricted to pre-selected seed regions or independent components, respectively. Instead, it provides a regionally unbiased mapping of brain areas with whole brain connectivity patterns that are predicted by aerobic capacity or longitudinal change thereof. In detail, the MVPA measure was obtained by dimension reduction of the whole brain connectivity matrix of each voxel. The connectivity matrix of each subject was reshaped into a row vector and subsequently concatenated over all participants into a matrix N × V, where N was the number of subjects and V is the number of voxels within the brain mask. The dimensionality of the N × V group correlation matrix was reduced by principal component analysis (PCA). This yielded an N × C matrix, where C is the number of maintained principal components. We maintained the first seven principal components that explained the most of the variance of the connectivity matrix (C = 7) [according to the rule of thumb to maintain an approximate 1:7 ratio between number of components and subjects, as proposed by the metric implementer Nieto-Castenon (2015)<sup>3</sup> ]. In other words, the resulting seven component score volumes best represented the whole brain connectivity pattern for each subject. These volumes were included in an F-test at the 2nd level analysis. Thus, we tested for clusters that were predicted by aerobic capacity with regard to whole brain connectivity.

The reported activation maps (of contrasted regression parameter estimates) in group analyses were considered significant at a cluster-level significance of p < 0.05, falsediscovery rate (FDR) corrected, and cluster defining voxel threshold of p < 0.001, two-sided. Due to unreliable estimations of smoothness of the MVPA rendered PCA maps, we used non-parametric permutation statistics (1000 permutations) as advocated by the implementer of the metric (Alfonso Nieto-Castanon, personal communication, September, 23, 2016). Connectome-MVPA analyses were thresholded using a cluster based significance threshold of p = 0.05 (FDR). Un-thresholded statistical maps for all reported results (**Figures 1–3**) are available online at http://neurovault.org/collections/2415/.

<sup>2</sup>http://restfmri.net

<sup>3</sup>http://www.nitrc.org/forum/message.php?msg\_id=14332

fnagi-09-00267 August 10, 2017 Time: 16:26 # 6

lobe and frontal and parietal regions. (C) Change in aerobic capacity is negatively associated with connectivity between left hippocampus and contralateral precentral gyrus. Green areas indicate seed regions; blue indicate positive associations to (change in) aerobic capacity; red indicates negative associations.

For all significant findings of group by time interactions, and associations between intrinsic brain activity either with regard to baseline aerobic capacity or change in aerobic capacity, we further examined the relationships to cognition. More

specifically, we defined volumes of interest based on the brain regions that were significantly predicted by aerobic capacity. These volumes were subsequently used as inclusive masks (i.e., small volume correction) within which we investigated effects of either baseline score or longitudinal change in cognition, respectively.

experiencing decreased connectivity to the IC with higher (or gain in) VO2-peak; blue indicates areas with increased fitness predicted connectivity.

## RESULTS

## Intervention Induced Gains in Aerobic Capacity

Within the sub-sample used here we observed a significant group by time interaction, with the aerobic group displaying a significantly larger improvement in aerobic capacity compared to fnagi-09-00267 August 10, 2017 Time: 16:26 # 7

the active control group F(1,43) = 6.20, p = 0.02 (see **Table 1**). This was similar to what Jonasson et al. (2017) reported using the full sample.

### Intervention Induced Changes in Intrinsic Brain Activity

None of the analyses (described in the method fMRI Analysis) of intrinsic brain activity reveled significant group by time interactions, neither with nor without control for sex, age and baseline VO2-peak estimates.

Likewise, the NBS-analysis was non-significant using a range of network defining ROI–ROI thresholds (see Network Based Statistics). Finally, none of the graph theoretical measures (betweenness centrality of left and right hippocampus, or global efficiency) rendered significant group by time interactions, see **Table 1**.

Given non-significant significant group by time interactions with regard to intrinsic brain activity, we conducted additional exploratory analyses that in future may guide confirmatory studies. In these, we tested for relationships between longitudinal change in VO2-peak and intrinsic brain activity across all participants disregarding group belonging, similar to what has been done previously (Maass et al., 2015; Jonasson et al., 2017). Therefore, we first explored cross-sectional relationships between VO2-peak and brain activity at baseline. Any relationships observed at baseline data were subsequently used to inform regressions of longitudinal change in VO2-peak to changes in brain activity.

#### Relationship between Aerobic Capacity with Resting State Measures at Baseline BOLD-Signal Fluctuations

Variation of the BOLD signal time series (BOLDSTD) was negatively correlated with aerobic capacity in three clusters located in midbrain regions at baseline (**Figure 1** and **Table 2**). The largest cluster (973 voxels) extended over right posterior hippocampus and thalamus. A second cluster (298 voxels) covered parts of left pallidum, and to a smaller extent also left thalamus and amygdala. The third cluster (232 voxels) contained right anterior hippocampus and right amygdala. Clusters were anatomically labeled using the WFU pick atlas. BOLDSTD were not positively correlated with aerobic capacity in any brain region.

Amplitude of Low Frequency Fluctuations (ALFF), resulted in virtually identical results as BOLDSTD (see Supplementary Figure S2), and thus ALFF results are not presented further.

Follow-up analyses that tested for the association between cognitive performance and BOLDSTD were restricted to the search volume of regions where aerobic capacity predicted BOLDSTD (displayed in **Figure 1**). Given that particularly hippocampal and mid-temporal brain regions displayed fitness related BOLDSTD, we tested for EM. In addition, a meta-analysis (Colcombe and Kramer, 2003) found the largest association between improved fitness and cognitive performance to be related to EFs. Furthermore, in our previous study, improved fitness was primarily related to improved general cognitive performance (i.e., CS) (Jonasson et al., 2017). Therefore, we also tested the association between EF, CS and BOLDSTD. Neither EM, EF nor CS were significantly related to BOLDSTD. (All cognitive outcome measures are presented in detail in Jonasson et al., 2017).

#### Seed Correlation Analysis (SCA)

None of the a priori defined seeds displayed significant VO2 peak predicted connectivity at baseline. However, connectivity of the post hoc seed in the right medial temporal cortex (see BOLD-Signal Fluctuations), to frontal-, parietal-, and occipital regions displayed a positive correlation to aerobic capacity. We also detected a negative association between aerobic capacity and the connectivity between the same seed and right thalamus and right occipital cortex (**Figure 2A** and **Table 3**).

#### Independent Component Analysis (ICA)

At baseline, among the five investigated ICA-derived networks we observed a negative relationship between VO2-peak and the connectivity of a sensorimotor network and a brain region extending over calcarine sulcus and precuneus (see **Figure 3A**). No associations between VO2-peak and connectivity of the other four networks were seen.

#### Graph Measures, NBS, Connectome-MVPA and ASL

Several measures of intrinsic brain activity turned out not to be significantly related to aerobic capacity. For graph theoretical measures, neither global efficiency [t(42) = −0.06, p = 0.95] nor left or right hippocampal centrality [L: t(42) = −0.94, p = 0.34, R: t(42) = −1.23, p = 0.22] reached significance.

In the NBS analysis, no ROI–ROI connectivities or subnetworks were correlated with aerobic capacity, neither corrected nor uncorrected. Furthermore, the connectome-MVPA did not reveal any significant clusters predicted by baseline VO2-peak.

Finally, analysis of ASL perfusion data did not reveal any associations between aerobic capacity and CBF from neither whole brain nor small volume voxel based analysis.


## Relationship between Longitudinal Change in Aerobic Capacity and Change in Intrinsic Brain Activity

#### Seed Correlation Analysis

fnagi-09-00267 August 10, 2017 Time: 16:26 # 8

Longitudinal gain in aerobic capacity was positively associated with changes in functional connectivity between the BOLDSTDderived seed region in right medial temporal lobe, and right angular gyrus and left inferior frontal gyrus (**Figure 2B** and **Table 3**). However, change in aerobic capacity was negatively associated with change in connectivity between a priori defined seed regions in the left hippocampus and the right precentral gyrus (**Figure 2C** and **Table 3**). None of the other a priori defined seeds related to change in VO2 peak.

Behavioral follow-up analyses did not reveal any significant relationships between changes in the fitness related SCA results and EM, EF, or CS.

#### Independent Component Analysis

Independent component analysis identified two significant associations between changes in VO2-peak and connectivity. Firstly, there was a positive relationship between gains in VO2 peak and connectivity between DMN and left frontal pole (see **Figure 3B** and **Table 3**). Secondly, a negative relationship between change in VO2-peak and the connectivity between the S1M1-network and right hypothalamus was observed (**Figure 3C**). There was no significant relationship to changes in cognition (EM, EF, and CS).

#### Network Based Statistics

The NBS revealed a subnetwork constituted by connected nodes in the DMN, the frontal parietal-, visual- and ventral attention network (**Figure 4**) in which gains in fitness was related to changes in intra-network connectivity. Since the subgraph detected by NBS does not allow for inferences on individual edges (i.e., ROI–ROI connections), we also correlated change in VO2 peak to connectivity between all individual ROI–ROI pairs (i.e., 34584 individual connections). In this analysis, no connections were significant (Bonferroni correcting for the edges in the full adjacency matrix).

#### Graph Measures, BOLDSTD, Connectome-MVPA and ASL

For the graph analysis, neither global efficiency [F(1,42) = 0.70, p = 0.39], nor betweenness centrality of hippocampus [left: F(1,43) = 1.35, p = 0.25; right: F(1,43) = 0.45, p = 0.50] was significantly related to change in fitness.

None of the other measures of intrinsic brain activity (MVPA, BOLDSTD,) or ASL yielded significant change-change relationships to fitness.

### DISCUSSION

In the current study, we aimed to characterize the effects of exercise induced improvements in aerobic capacity on intrinsic brain activity. To accomplish this, we have employed a wide range of resting-state fMRI metrics as well as measures of CBF using ASL.

### No Longitudinal Group Differences in Brain Activity

Based on previous literature, we predicted significant group by time interactions with regard to hippocampal connectivity, DMN integrity and vascular response as revealed by BOLDSTD and CBF. In the present study, none of these measures displayed group differences in longitudinal changes.

A possible explanation may be that both groups improved in aerobic capacity. Since we stipulate that gain in VO2-peak would be the active constituent mediating improvements in cognition and modulations of intrinsic brain activity (in line with the fitness hypothesis of cognition, (North et al., 1990; Kramer et al., 1999)], we reasoned that interindividual change in aerobic capacity across both groups could be a sensitive predictor of brain changes (similar to approaches used in previous studies, e.g., Maass et al., 2015). Thus, we performed post hoc analyses

TABLE 3 | Relationships between functional connectivity and aerobic capacity at baseline and change over time.


Target regions are labeled based on the locations of the largest number of voxels within significant cluster, as identified and labeled within the Conn-toolbox. <sup>∗</sup>Seed region defined post hoc from VO2-peak predicted BOLDSTD effects at baseline.

fnagi-09-00267 August 10, 2017 Time: 16:26 # 9

investigating inter individual change–change relations. The lack of significant effects is discussed in further detail below.

### Post hoc Analyses of Fitness-Brain Relationships

#### Aerobic Capacity and the Medial Temporal Lobe

In the post hoc analyses we observed several associations between aerobic fitness and intrinsic brain activity, both at baseline and over time. Longitudinal change in aerobic capacity predicted change in hippocampal connectivity. Whereas left hippocampus displayed a decreased connectivity to precentral gyrus with increasing fitness, the right hippocampus was associated with increased connectivity to frontal and parietal areas. The right hippocampus, similar to baseline, was associated with increased frontal and parietal connectivity. Previously, Voss et al. (2013) reported enhanced connectivity between parahippocampus and bilateral temporal cortex, as well as to occipital and parietal areas following a 12 months aerobic (walking) intervention. Similar to the current findings, they observed reduced connectivity of left hippocampus, although to the right prefrontal cortex rather than to the right precentral gyrus as seen in here. Interestingly, the authors showed that increased connectivity between bilateral parahippocampal gyrus and middle temporal gyri were positively predicted by increases in neurotrophic growth factors, although no direct test of the behavioral significance of any of the connectivity changes was presented.

We detected relationships between BOLDSTD and aerobic capacity at baseline, whereas gains in aerobic capacity did not relate to changes in BOLDSTD over time. Contrary to what previously has been reported (Burzynska et al., 2015; Gauthier et al., 2015), the relationships between aerobic fitness and BOLDSTD were negative and primarily observed in mid-temporal rather than cortical areas. Burzynska et al. (2015) reported cross-sectional associations between BOLDSTD and physical activity, suggesting that BOLDSTD could reflect a long-term cardiovascular trait, which coheres with the absence of any intervention related changes in current study. Notably, they did not observe relationship between aerobic capacity and BOLDSTD, but only a positive relationship between BOLDSTD and physical activity. Our findings suggest that high aerobic capacity is associated with a more stable BOLD-signal within the temporal medial lobe. Speculative, the high fit subjects would experience relatively lower levels of physical demand, which could reconcile current findings with those reported by Burzynska et al. (2015). However, no on-line monitoring of degree of physical work during scanning was performed. Thus, further studies are required to establish the relationship between gains in aerobic capacity and BOLDSTD.

Longitudinal changes in intrinsic brain activity that were predicted by change in aerobic fitness did not relate to changes in any of the investigated cognitive domains (EM, EF, and CS). A plausible reason for this is the absent relationships between fnagi-09-00267 August 10, 2017 Time: 16:26 # 10

the cognitive variables and VO2-peak, both at baseline and in change-change scores (Supplementary Table S2). Unfortunately, this limits the conclusions of the functional meaning of the observed fitness related brain changes. Thus, any functional interpretations of these brain changes are unknown and would have to rely on reverse inference, for which we lack unique cognition-connectivity associations.

Taken together, the findings above support the notion that aerobic fitness primarily affects mid temporal brain regions. In addition to the exercise intervention studies investigating changes in intrinsic brain activity (Voss et al., 2013; Burzynska et al., 2015; Maass et al., 2015; Bär et al., 2016; Tozzi et al., 2016), corroborating evidence of hippocampal involvement is also provided by studies investigating structural changes (e.g., Erickson et al., 2011, for a review, see Erickson et al., 2014; Jonasson et al., 2017). The fact that hippocampus display a high capacity for plastic change in relation to environmental factors (for a review, see Leuner and Gould, 2010), and also undergoes age related resting state connectivity changes (Salami et al., 2016), provides further motivation to specifically probe the intrinsic brain activity in the mid-temporal lobe. By standardizing research protocols and outcome measures, the mechanisms linking hippocampal brain activity, brain structure, vascularity, molecular growth factors and behavior will hopefully be revealed (Duzel et al., 2016; Stillman et al., 2016).

#### Network Changes

Network based statistics analysis revealed a graph subcomponent that was predicted by the longitudinal change in VO2-peak. Whereas previous cross-sectional reports (Voss et al., 2016) showed a higher within network integrity (particularly in DMN) with higher fitness, our longitudinal findings suggest that increased fitness also modulate between network connectivity. Whether the explanations for the discrepant finding are found in differences in methodological approaches (e.g., graph defining set of nodes) or lack of statistical power remains to be determined.

Resting state networks derived by ICA revealed relationships between fitness and connectivity of both the S1M1 and the DMN. Decreased connectivity between M1S1 and right thalamus was associated with increased aerobic capacity. M1S1 receives input from parts of thalamus, and the decreased connectivity with gains in fitness seems counterintuitive at first. Speculatively, this could reflect enhanced neuronal efficiency, although such a hypothesis should be investigated using methods that complement the correlational fMRI approaches used here. The connectivity between the same S1M1 network and occipital cortex were negatively related to fitness at baseline. In a small exercise intervention among obese children, Krafft et al. (2014) observed that the ICA derived motor network underwent decreased connectivity to visual areas in precuneus in the exercise group, resembling current baseline finding. Although our fitness predicted S1M1 connectivity targeted different regions at baseline compared to over time, it seems that functional segregation of S1M1 reflects better fitness. However, in an aged cohort more similar to ours, Voss et al. (2010) did not see group by time changes in motor connectivity, why this finding would need to be investigated further.

Interestingly, we observed enhanced connectivity between the DMN component and the left orbitofrontal cortex that correlated with gains in fitness. Even though the orbital cluster is localized more laterally than the typical ventromedial frontal DMN hub, the observation bears similarities to previous literature of fitness related connectivity of the posterior anterior midline core of the DMN (Voss et al., 2010, 2016). Both these studies interpret the fitness-related DMN integrity as a rejuvenated connectivity pattern, typical for younger subjects. A similar interpretation could be given here, but the lack of associations between cognition and these fitness-related connectivity patters prevent any firm conclusion of their behavioral significance.

#### Limited Effects

The a priori hypothesized group by time interactions, as well as the majority of the exploratory tests that aimed to relate longitudinal change in VO2-peak to change in intrinsic brain activity, turned out to be non-significant. Despite drawing on previous literature that have investigated intrinsic capacity in relation to aerobic fitness among elderly, we were largely unable to conceptually replicate these. The replication failures could likely be attributed both to factors pertaining to cognitive brain imaging in general (see below), as well as to factors specific to this study context (e.g., differences in study design, interventions, cohorts and analyses strategies). Another potentially important issue is the fact that the active control group may have exercised more vigorously than in many previous interventions, considering that aerobic capacity increased substantially also for this group. Recent research have shown that resistance and coordination training, part of our active control training regimen, may influence the BOLD-signal during task performance (Voelcker-Rehage et al., 2011; see Voelcker-Rehage and Niemann, 2013, for a review), as well as show similar effects on hippocampus volume as aerobic exercise (Niemann et al., 2014). By comparing two training regimens which both have positive effects on brain and behavior, actual intervention effects may thus have been masked.

The effects of gains in fitness on the exploratory resting-state measures (e.g., global efficiency, connectome-MVPA, BOLDSTD) of the resting-state brain activity were likely too subtle to be detected in current study. The BOLD signal acquired during resting state fMRI is inherently noisy, and estimations based on high quality fMRI data showed that the variance associated with neuronal activity only constituted around 4% of the total variance of the BOLD-signal time series (Marcus et al., 2013). However, despite the large contribution of nonneuronal noise to the resting state BOLD signal, test–retest reliability of group average of cardinal resting state networks has been reported to be fairly reliable over time (Zuo et al., 2010; Wisner et al., 2013), although the choice of both resting state measures and preprocessing pipeline influence the degree of test–retest reliability (Yan et al., 2013). Not surprisingly, between-group comparisons of brain activity are also typically much smaller than single group averages (for an informative discussion on effect sizes in fMRI, see Poldrack et al., 2017).

### Limitations and Future Directions

fnagi-09-00267 August 10, 2017 Time: 16:26 # 11

Current study has several limitations. First, we were unable to selectively induce improvements of VO2-peak in the aerobic group, although the aerobic group improved more than the control group. This likely decreased any exercise induced group differences of the outcome measures. Although challenging, experimental designs that more efficiently prevent cardiovascular training in the active control group could potentially reveal stronger experimentally induced brain changes.

Furthermore, current study only measured two time points, i.e., before and after the 6-months intervention. Future studies would benefit from sampling data at multiple (>2) time points, for higher temporal resolution of the developmental trajectories. One motive for this is the expansion-partial normalization hypothesis of neuroplasticity (Brehmer et al., 2014), which proposes that longitudinal brain changes commonly follow inverted u-shaped temporal profiles, which could only be detected if data were acquired at multiple time points.

Resting state studies of physical interventions are still sparse. A greater understanding of the mechanisms mediating the neurocognitive effects of aerobic exercise are valuable for optimizing intervention programs to target the relevant neurophysiological processes and cognitive domains with higher precision. To enhance the reliability and replicability of imaging findings, several factors have recently been proposed (Ioannidis et al., 2014; Poldrack et al., 2017). Complete reporting of results (i.e., including null findings), as well as clear declarations of which analyses are ad hoc or exploratory and which are hypothesis driven, are both critical when aggregating research findings in meta-analyses, or for informing the design of future studies. Above all, the authors emphasized the importance of attaining proper statistical power, e.g., by enhancing cohort sizes for instance through laudable data sharing initiatives. Moreover, increased statistical power is also obtained by larger effect sizes. Stronger experimental effects could likely be achieved by longer and more intensive interventions. Potentially, longer intervention studies would also enable the observed brain changes to translate into improvements in behavior.

### CONCLUSION

We have characterized changes in intrinsic brain activity following a 6-months physical exercise intervention. None of our a priori hypothesis of group by time interactions regarding intrinsic brain activity were confirmed. However, when investigating the linear relationships between longitudinal gain in aerobic capacity and changes in functional connectivity we observed exercise modulated hippocampal connectivity. Likewise, gain in aerobic capacity was associated with increased connectivity between the DMN and prefrontal cortex, but negatively related to sensorimotor-thalamic connectivity. However, these changes did not relate to changes in cognition, possibly due to the length of the intervention, or insensitive behavioral measures. The majority of the exploratory analyses did not reveal significant associations between fitness and intrinsic brain activity, although we observed that mid-temporal BOLDSTD was negatively associated with fitness at baseline. The current study provides resting-state fMRI evidence that exercise preferentially modulate mid-temporal brain regions and hippocampus. The functional significance of these brain changes should be investigated further, preferably in high powered studies using data acquired at multiple time points.

### ETHICS STATEMENT

This study was carried out in accordance with the recommendations of the regional ethical committee in Umeå, Sweden with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the regional ethical committee in Umeå, Sweden.

### AUTHOR CONTRIBUTIONS

LJ, LN, KR, and CB designed the study. PF performed the statistical analyses. All authors contributed in revising the work and approved the final version of the manuscript.

### FUNDING

The research is part of the programme Paths to Healthy and Active Ageing, funded by the Swedish Research Council for Health, Working Life and Welfare, (Dnr 2013 – 2056). Support was also obtained from the Swedish Research Council (2012- 00530), Västerbotten County Council, the Swedish Research Council for Sport Science and Umeå School of Sport Sciences to CB, from the Kamprad Family Foundation to LN and Umeå University.

### ACKNOWLEDGMENTS

We thank all participants taking part in this study, the training supervisor Peter Lundström, the staff at Umeå Center for Functional Brain Imaging, University Hospital of Northern Sweden, Umeå, and Sport Science Lab at Umeå School of Sport Sciences Umeå University.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2017.00267/full#supplementary-material

### REFERENCES

fnagi-09-00267 August 10, 2017 Time: 16:26 # 12


fnagi-09-00267 August 10, 2017 Time: 16:26 # 13


head micromovements on functional connectomics. Neuroimage 76, 183–201. doi: 10.1016/j.neuroimage.2013.03.004


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Flodin, Jonasson, Riklund, Nyberg and Boraxbekk. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Balance Training Enhances Vestibular Function and Reduces Overactive Proprioceptive Feedback in Elderly

Isabella K. Wiesmeier <sup>1</sup> , Daniela Dalin<sup>1</sup> , Anja Wehrle2, 3, Urs Granacher <sup>4</sup> , Thomas Muehlbauer <sup>5</sup> , Joerg Dietterle<sup>1</sup> , Cornelius Weiller <sup>1</sup> , Albert Gollhofer <sup>2</sup> and Christoph Maurer <sup>1</sup> \*

*<sup>1</sup> Department of Neurology and Neurophysiology, University Hospital Freiburg, Freiburg, Germany, <sup>2</sup> Institute for Sports and Sport Science, University of Freiburg, Freiburg, Germany, <sup>3</sup> Department of Internal Medicine, Institute for Exercise and Occupational Medicine, University Hospital Freiburg, Freiburg, Germany, <sup>4</sup> Division of Training and Movement Science, University of Potsdam, Potsdam, Germany, <sup>5</sup> Division of Movement and Training Sciences, Biomechanics of Sport, Institute of Sport and Movement Sciences, University Duisburg-Essen, Essen, Germany*

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Frederick Robert Carrick, Bedfordshire Centre for Mental Health Research in Association with University of Cambridge, United Kingdom Rahul Goel, University of Houston, United States Susan Elizabeth Esposito, Life University, United States*

> \*Correspondence: *Christoph Maurer*

*christoph.maurer@uniklinik-freiburg.de*

Received: *23 May 2017* Accepted: *28 July 2017* Published: *11 August 2017*

#### Citation:

*Wiesmeier IK, Dalin D, Wehrle A, Granacher U, Muehlbauer T, Dietterle J, Weiller C, Gollhofer A and Maurer C (2017) Balance Training Enhances Vestibular Function and Reduces Overactive Proprioceptive Feedback in Elderly. Front. Aging Neurosci. 9:273. doi: 10.3389/fnagi.2017.00273* Objectives: Postural control in elderly people is impaired by degradations of sensory, motor, and higher-level adaptive mechanisms. Here, we characterize the effects of a progressive balance training program on these postural control impairments using a brain network model based on system identification techniques.

Methods and Material: We analyzed postural control of 35 healthy elderly subjects and compared findings to data from 35 healthy young volunteers. Eighteen elderly subjects performed a 10 week balance training conducted twice per week. Balance training was carried out in static and dynamic movement states, on support surfaces with different elastic compliances, under different visual conditions and motor tasks. Postural control was characterized by spontaneous sway and postural reactions to pseudorandom anterior-posterior tilts of the support surface. Data were interpreted using a parameter identification procedure based on a brain network model.

Results: With balance training, the elderly subjects significantly reduced their overly large postural reactions and approximated those of younger subjects. Less significant differences between elderly and young subjects' postural control, namely larger spontaneous sway amplitudes, velocities, and frequencies, larger overall time delays and a weaker motor feedback compared to young subjects were not significantly affected by the balance training.

Conclusion: Balance training reduced overactive proprioceptive feedback and restored vestibular orientation in elderly. Based on the assumption of a linear deterioration of postural control across the life span, the training effect can be extrapolated as a juvenescence of 10 years. This study points to a considerable benefit of a continuous balance training in elderly, even without any sensorimotor deficits.

Keywords: age, balance, vestibular, proprioception, training

## INTRODUCTION

Impairments of postural control result in increased rates of unintentional falls. In fact, falls are the leading cause of injuries and subsequent deaths among people 65 years and older, and generate a fundamental financial burden to the healthcare system (Burns et al., 2016). There is general consensus that altered postural control in elderly people is determined by degradations of the sensory channels, i.e., vestibular, visual, and proprioceptive cues (Rauch et al., 2001; Goble et al., 2009; Grossniklaus et al., 2013), of the motor system (Macaluso and De Vito, 2004), and by deficits in higher-level adaptive systems (Shumway-Cook and Woollacott, 2001). It is still under debate whether, in addition, elderly's central weighting of sensory signals is affected. While some authors reported an impaired sensory weighting (Teasdale and Simoneau, 2001; Eikema et al., 2014), others found it to be unimpaired (e.g., Allison et al., 2006; Jeka et al., 2006). This controversy is possibly caused by different experimental strategies to assess sensory weighting. For example, it is well known that sensory weighting is modified by e.g., type and size of external disturbances, available sensory information, training status etc. (Oie et al., 2002; Peterka, 2002; Maurer et al., 2006). In general, it is unclear which subsystem mainly determines the degradation of postural control, given the fact that many subsystems are altered during aging.

From a diagnostic side, postural control is often monitored via spontaneous sway measures and, more rarely, challenged by external perturbations leading to motor reactions. Some authors reported age-related changes in spontaneous sway in terms of increased mean velocity, or increased sway frequencies (Prieto et al., 1996; Qu et al., 2009). However, the diagnostic value of spontaneous sway measures has been questioned (Maurer and Peterka, 2005; Pasma et al., 2014). For a more detailed analysis of postural control, the application of external perturbations has frequently been suggested. Interestingly, postural reactions to external perturbations (proprioceptive, vestibular, or visual) have been reported to be altered in the elderly (e.g., Ghulyan et al., 2005; Maitre et al., 2013; Eikema et al., 2014). More recently, the relationship between stimulus and subsequent body motion was systematically evaluated using model simulations (e.g., Peterka, 2002; Davidson et al., 2011; Nishihori et al., 2011; van der Kooij and Peterka, 2011). Models are usually based on simple feedback mechanisms, involving inverted pendulum bodies, stiffness, damping, feedback time delay, and sensory weighting (Maurer et al., 2006; van der Kooij and Peterka, 2011; Engelhart et al., 2014; Wiesmeier et al., 2015). They have already been applied to elderly people's postural control (Maurer and Peterka, 2005; Cenciarini et al., 2010; Davidson et al., 2011; Nishihori et al., 2011). Some authors reported increased damping of the system in the elderly (Cenciarini et al., 2010; Davidson et al., 2011). Stiffness findings are inconsistent (Maurer and Peterka, 2005; Cenciarini et al., 2010; Davidson et al., 2011; Nishihori et al., 2011). Surprisingly, systematic evaluations of intervention programs like balance training are completely lacking.

The improvement of elderly people's postural control via balance training is well documented (see e.g., Nagy et al., 2007; Gillespie et al., 2012). However, evidence for an optimal training program of healthy elderly people is scarce (Lesinski et al., 2015). Recently, Lesinski et al. (2015) concluded from a systematic review and metaanalysis of numerous training studies that an optimal training should last 11–12 weeks with a training frequency of three sessions per week resulting in a total number of 36–40 training sessions. A single training session should take 31–45 min. Over the last few years, balance training has been further diversified into traditional, perturbation-based, and multitask balance training approaches (see e.g., Granacher et al., 2011). A growing body of literature deals with the specific neurophysiological effects of balance training. Balance training may be able to reduce coactivation of antagonist muscles, to shorten onset latency of muscle activation, to augment reflex activity, to increase maximal and explosive force production capacity, increase the length of recovery steps subsequent to external perturbations (see Granacher et al., 2011). On a functional level, gait speed and step length, with or without external perturbations, have been reported to be increased, while step time variability seems to be reduced. Performance in clinical tests like Berg Balance Scale (BBS) and Timed Up and Go (TUG) test appears to be improved. This improvement was backed up by electrophysiological correlates, such as, the reduction of the Hoffmann reflex (Granacher et al., 2011; Nagai et al., 2012). However, it is unclear as yet, how balance training affects physiological subsystems of postural control, such as, use of sensory input, central processing, and motor output.

In the current study, we aimed to assess the main subsystems of elderly people's altered postural control with a focus on their sensitivity to balance training using parameter identification techniques based on brain network model simulations. We expected that balance training could change elderly subjects' postural control so that it resembles postural control of younger subjects, similar to a "juvenescence."

### METHODS

Forty elderly subjects between 65 and 80 years who lived independently in the community, were randomly allocated either into a balance training or into a control group that did not receive balance training. Allocation followed a matched-pair protocol on the basis of age and sex (see **Figure 1**). Each subject was examined by a senior consultant neurologist in order to identify sensory deficits or neurodegenerative diseases. In addition, we asked for the amount of physical activity, fear of falling and number of falls during the last 3 years prior to the study (for questionnaire, see Supplementary Material). As part of the neurological examination, vestibular function was specifically tested using Frenzel goggles on a turning chair (vestibuloocular reflex, VOR). Proprioceptive function was evaluated by testing position sense and by measuring vibration sense with a tuning fork. Elderly subjects with relevant sensory deficits were excluded from the study. Moreover, subjects suffering from any other acute or chronic disease that may interact with the postural control were excluded. Finally, 35 elderly subjects [73 ± 3.3 years (mean age ± SD)] contributed to the study.

In order to identify age-related changes in elderly subjects' postural control, we used younger subjects' data [n = 35, 37 ± 11.2 years (mean age ± SD)] generated with a similar experimental set in our laboratory, as a reference group. The study was approved by the ethics committee of the University of Freiburg and performed according to the ethical standards of the Declaration of Helsinki. Written informed consent was obtained from all subjects prior to study participation.

For evaluating postural control we used a dynamic posturography approach. Subjects were standing on a custombuilt motion platform (**Figure 2B**) with eyes open (eo) and with eyes closed (ec). Spontaneous sway was recorded with the platform fixed while postural reactions were measured during continuous platform perturbations. Posturographic assessments were composed of 20 trials divided into two sessions: During the first 10 trials subjects were told to close their eyes while the other 10 trials were carried out with eyes open. The first and last trial of each ten-trial sequence (eyes closed or eyes open) was a 'spontaneous sway' trial. The other eight trials were conducted while the platform tilted. Each trial took 1 min. Breaks of about 10 s were taken between trials, according to the subject's needs. Subjects were told to stand comfortably in an upright position. They were asked not to talk.

Spontaneous sway was quantified by center-of-pressure (COP) sway paths detected with the help of a force transducing platform (Kistler platform type 9286, Winterthur, Switzerland). Extracted measures consisted of sway amplitude (Root Mean Square, RMS), sway velocity (Mean Velocity, MV), and the frequency content of sway (Mean Frequency, MF). Postural reactions were measured on a tilting platform. The tilts consisted of platform rotations in the sagittal plane with the axis running through subjects' ankle joints. Platform tilts were designed as pseudorandom stimuli (PRTS, pseudorandom ternary sequence, **Figure 2A**) with two peak angular displacements (0.5 and 1◦ ) and analyzed at 11 frequencies (0.05, 0.15, 0.3, 0.4, 0.55, 0.7, 0.9, 1.1, 1.35, 1.75, and 2.2 Hz).

Angular excursions of the platform and the body (hipto-ankle, shoulder-to-hip) in space were quantified with an optoelectronic device using markers attached to shoulder, hip, and a rigid bar on the platform (Optotrak 3020, Waterloo, Canada). Each marker contained three light-emitting diodes (LEDs). 3-D LED positions were used to calculate marker movements (**Figures 2A,B**). Kistler <sup>R</sup> and Optotrak <sup>R</sup> output as well as the stimulus signals were sampled at 100 Hz using an analog-digital converter and stored on a PC via LabView <sup>R</sup> (National Instruments, Austin, Texas, USA) for offline analysis. Data was analyzed using custom-made software programmed in MATLAB <sup>R</sup> (The MathWorks Inc., Natick, MA, USA).

The relationship between the postural reactions and platform stimuli were represented by "transfer functions" in the frequency domain. Transfer functions were calculated using discrete Fourier transforms. From transfer functions, GAIN, PHASE, and Coherence values were extracted as a function of stimulus frequencies. GAIN represents the size of the postural reaction, i.e., lower body or upper body response in terms of angular excursion, as a function of stimulus size (platform angle). A

joints. Angular excursions of the platform, the upper body (UB) and the lower body (LB) in space were quantified with an optoelectronic device using markers attached

to shoulder, hip, and a rigid bar on the platform. GAIN of 1 would indicate a perfect match between body and

platform excursion. PHASE is related to the relative timing between postural reaction and stimulus. Negative PHASE values (PHASE lag) represent delays. Coherence is a measure for reproducibility of postural reactions across stimulus cycles. Coherence values of 1 signify perfectly reproduced postural reactions; zero would indicate no similarity between subsequent postural reactions.

Findings in the elderly were compared with data of a young reference group. In addition, data of the elderly group before (first assessment, A1) was compared to data after balance training (second assessment, A2). The second assessment was conducted between 4 and 10 days following the last training session.

In addition, the TUG and the Functional Reach Test (FRT) were assessed twice (A1 and A2) in elderly subjects. The TUG quantifies the time (in seconds), a subject needs to do the following motor task: standing up from a chair, walking 3 m straight, turning around, walking back, and sitting down. The FRT measures the maximum distance in centimeters before and after reaching the arm forward at shoulder level without losing balance (Enkelaar et al., 2013). Mean FRT was calculated across three attempts.

#### Parameter Identification

Transfer functions served as a basis for simulating postural control using well-established models of upright stance to extract relevant parameters (Peterka, 2002; Engelhart et al., 2014). The model includes a body defined by mass and height, a Neural Controller containing stiffness and damping, a feedback time delay, and a sensory feedback mechanism. A negative feedback loop links body excursions perceived by visual, vestibular, and proprioceptive channels to a corrective torque through a Neural Controller with proportional [P], derivative [D] and integral [I] contributions (PDI-controller, **Figure 3**). The external stimuli, i.e., anterior-posterior platform tilt angles, serve as an input of the model. Body sway, represented by the center of mass (COM) angle, is the model output. Since Neural Controller values depend on mass and height of the individual subjects (see Peterka, 2002; Cenciarini et al., 2010), these values are corrected by (mgh), which corresponds to the gravitational pull (body mass) × (gravitational constant) × (height of the COM from the ankle joint), leading to [P/mgh], [D/mgh], and [I/mgh]. Other parts of the model were: a lumped time delay [Td], representing the time interval between the postural reaction and the stimulus, and a sensory weighting mechanism. The sensory weighting mechanism specifies the reference frame for body orientation (space coordinates vs. platform coordinates), represented by [Wp]. The value [Wp] stands for the proprioceptive share of the sensory feedback. A value of 1 corresponds to 100 % proprioceptive control, i.e., stabilization in platform coordinates, a value of 0 relates to 0% proprioceptive control and 100% stabilization in space. Moreover, the model includes a biomechanics part that represents torque related to passive elasticity [Ppas] and damping [Dpas] of muscles and tendons, in

parallel to the active corrective torque, which is determined by the Neural Controller (**Figure 3**). With the help of an optimization procedure (fmincon/ Matlab, Mathworks), model simulations were fitted to the experimental transfer functions under different stimulus amplitudes and visual conditions. Similar to the model provided by Peterka (2002), we assume that all sensory feedback signals add up to a feedback gain of unity. For example, if the proprioceptive gain is 0.6 (60%) in the eyes-closed condition, the vestibular gain would be 0.4 (40%). In the eyes-open condition, vestibular and visual gains both contribute to the space reference. The strength of the visual feedback could be estimated by subtracting the gain of the space reference in the eyes-closed condition from the space reference in the eyes-open condition. Another assumption refers to the allowed range of all gains ([P/mgh], [D/mgh], [I/mgh], [Wp], [Ppas], [Dpas]), and time delay [Td], which were constrained to positive values. Goodnessof-fit measures, limitations of the model, and comparisons to simulations of datasets from other studies are provided as Supplementary Material.

#### Balance Training Program

The balance training group received a balance training twice a week over a period of 10 weeks (one session 60 min, 20 sessions in total). It was developed and conducted by professional instructors from the Institute for Sports and Sport Science (University of Freiburg) and the Institute for Exerciseand Occupational Medicine (Department of Internal Medicine, University Hospital Freiburg) on the basis of previous research (Granacher et al., 2010). The training group was divided into two smaller exercise groups which were supervised by two instructors in order to guarantee a small participant-to-instructor ratio (1 instructor vs. 5 subjects). Each session included 10 min warm-up and 10 min cool-down. Exercises were carried out under static (standing) and dynamic (walking) conditions and were modified by using either a stable or an unstable support surface (e.g., foam mats), by closing or opening the eyes and by performing the exercises in bipedal, semi-tandem, tandem and monopedal stance with an additional motor task like catching and throwing a ball. One set of exercises consisted of 4 periods of 20 s exercise and 40 s rest. The number of sets was increased over time. Many exercises were performed in pairs or as circuit training.

### Statistical Analysis

Statistical analyses were performed using Microsoft Excel and statistic programs (JMP <sup>R</sup> and Statview by SAS Institute Inc., Cary, NC, USA). Statistical significance of the difference between healthy young and elderly subjects before training was tested with the help of an analysis of variance (ANOVA). The between-subject variable was group (young, elderly). For spontaneous sway, the within-subject variables were: visual condition (eyes open, eyes closed), sway direction (mediolateral, anteroposterior), and body segment (COP, hip, shoulder). For the perturbed stance experiments, the within-subject variables were: visual condition, stimulus amplitude (0.5 and 1◦ ) and body segment (hip, shoulder). Intervention effects in the two groups of elderly subjects before (A1) and after (A2) balance training including FRT and TUG was tested by multivariate analyses of variance (MANOVA) with time (A1, A2) as an additional degree of freedom. Statistical significance was assumed at p ≤ 0.05. Moreover, the relationships between parameters related to platform measures and clinical test parameters were examined with a Pearson Correlation Test. A matrix of correlation coefficients was created, which illustrates the strength of linear relationships between each pair of parameters.

### RESULTS

#### Baseline Characteristics

Thirty-five healthy elderly [73 ± 3.3 years (mean age ± SD), 17 female, 18 male] and 35 young subjects [37 ± 11.2 years (mean age ± SD), 19 female, 16 male] were included in the analysis. None of the subjects reported any training or test-related injuries. For detailed information see **Tables 1**, **2**. Three subjects of the control group dropped out due to failure to attend the second assessment for personal reasons and illness. Two subjects of the training group dropped out during the training period due to personal reasons not associated with balance training. Both, the training and control groups were well balanced at baseline concerning age, sex, body mass, and physical activity (**Table 2**). In total, 11 subjects claimed to have fallen during the last 3 years prior to the study (training group: 6, control group: 5). The number of falls was similar in both groups (training group: 10 falls, control group: 7 falls). Reasons for falling were e.g., tripping or leisure time activities. Seven subjects reported fear of falling (training group: 4, control group: 3) which was always associated with particular situations such as, clear ice or standing on a ladder.

#### Spontaneous Sway

Root Mean Square (0.51 vs. 0.42 cm; F = 23.4, p < 0.001), MV (1.03 vs. 0.70 cm/s; F = 60.9, p < 0.001), and MF (0.49


TABLE 1 | Information about the young group.

#### TABLE 2 | Information about the training and control group.


*35 subjects; 19 female; 16 male. m, masculine; f, feminine; ys, years; kg, kilogram; m, meters; #, no data available.*

vs. 0.41 Hz; F = 20.5, p < 0.001) were significantly larger in elderly before training (A1) compared to younger subjects (see **Figure 4**). We found no significant interactions between age and visual condition (F = 0.8, p = 0.39), sway direction (mediolateral, anteroposterior; F = 1.8, p = 0.18), and body segments (F = 1.9, p = 0.16). None of the measures significantly interacted with balance training for the elderly (RMS: F = 2.6, p = 0.11, MV: F = 1.7, p = 0.20, MF: F = 0.05, p = 0.83).

#### Externally Perturbed Stance GAIN

In elderly subjects before training, GAIN was significantly larger (2.31; F = 553.7, p < 0.001) than in young subjects (1.77). Across the age groups, GAIN was significantly larger with eyes closed *35 subjects; 17 female; 18 male. m, masculine; f, feminine; control, control group; training, training group; ys, years; kg, kilogram; m, meters; h, hours.*

than with eyes open (eyes closed, ec: 2.36, eyes open, eo: 1.72; F = 766.7, p < 0.001). Stimulus amplitudes (0.5◦ : 2.24, 1◦ : 1.84; F = 307.4, p < 0.001), stimulus frequencies (F = 4954.3, p < 0.001), and body segments (hip: 1.60, shoulder: 2.48, F = 1482.5, p < 0.001) significantly influenced GAIN. Age group significantly interacted with frequency (F = 12.7, p < 0.001), with the most prominent GAIN difference between age groups in the lower frequency range (see GAIN plots in **Figure 5A1**). Moreover, we found a significant interaction between age group and body segments (F = 379.1, p < 0.001). This exemplifies that elderly

subjects' shoulder GAIN (2.98) was almost twice as large as hip GAIN (1.64), whereas, young subjects' shoulder GAIN (1.99) was 20% larger than hip GAIN (1.55, **Figure 5B1**). Lastly, age group did not significantly interact with visual condition (F = 3.3, p = 0.07) or stimulus amplitude (F = 0.7, p = 0.41).

GAIN as a function of time (A1/A2) significantly interacted with balance training (F = 25.4, p < 0.001, **Figure 5C1**). While GAIN of the training group significantly decreased from 2.36 to 2.23 (p < 0.05), GAIN of the control group slightly increased (A1: 2.26, A2: 2.31, p > 0.05). Frequency did not interact significantly with GAIN as a function of time (F = 0.4, p = 0.95). However, GAIN as a function of time significantly interacted with body segments (F = 5.2, p = 0.02): Whereas GAIN of the shoulder decreased over time (2.97 to 2.90), GAIN of the hip hardly changed as a function of time (1.64 to 1.65). The decrease in shoulder GAIN is an effect of balance training. Shoulder GAIN of the training group decreased from 3.0 to 2.8, whereas shoulder GAIN of the control group slightly increased from 2.9 to 3.0 (F = 17.9, p < 0.001). In both groups, GAIN of the hip was nearly equal as a function of time (training group A1: 1.70, A2: 1.71; control group A1: 1.57, A2: 1.59). There were no significant interactions between time and visual condition (F = 0.4, p = 0.54), and between time and stimulus amplitude (F = 0.08, p = 0.78).

#### Phase

PHASE, indicating the temporal relationship between response and stimulus, differed significantly between the age groups (young subjects: −127.27◦ , elderly subjects: −122.34◦ ; F = 8.9, p = 0.003). Across the age groups, PHASE was mainly determined by frequency, showing a PHASE lead in the low frequency range (F = 1035.9, p < 0.0001). In general, the significant interaction between age and frequency (F = 4.1, p < 0.001) showed the effect of age on PHASE as a function of frequency. The young group showed a moderate slope of PHASE as a function of stimulus frequencies, whereas the elderly group displayed a steeper relationship between PHASE and frequencies (see **Figure 5A2**). PHASE lag was found to be significantly smaller with eyes closed (−120.63◦ ) than with eyes open (−128.99◦ , F = 25.7, p < 0.001), significantly smaller at the hip (−101.07◦ ) than at the shoulder level (−148.54◦ , F = 828.6, p < 0.001) across all age groups. It did not significantly vary with different stimulus amplitudes (F = 0.009, p = 0.9). We found a significant interaction between age group and body segment (F = 45.2, p < 0.001) representing the fact that PHASE difference between shoulder and hip decreases with age (**Figure 5B2**). Age group did not significantly interact with visual condition (F = 1.6, p = 0.2) or stimulus amplitude (F = 0.6, p = 0.5).

PHASE lag as a function of time significantly interacted with balance training (F = 5.3, p = 0.02, **Figure 5C2**). Both, PHASE lag of the training group (−123.67◦ to −126.69◦ , p > 0.05) and PHASE lag of the control group (−120.94◦ to −131.70◦ , p < 0.05) increased as a function of time, with the increase being more pronounced in the control group. Time did not interact significantly with frequency (F = 0.1, p = 1.00). In addition, PHASE as a function of time significantly interacted with body segments (F = 22.1, p < 0.001). Whereas, PHASE lag of the hip was nearly stationary over time (−104.1 to −103.1), PHASE lag of the shoulder increased as a function of time (−140.5 to −155.3). However, we found no significant interaction between balance training and body segments as a function of time (F = 1.0, p = 0.3).

#### Coherence

In both, young and elderly subjects, coherence significantly depended on frequency (higher coherence with lower frequencies, F = 931.9, p < 0.001, **Figure 5A3**), stimulus amplitude (higher coherence with larger stimulus amplitude, F = 908.6, p < 0.001), on body segments (hip 0.49, shoulder 0.48; F = 7.5, p = 0.006), but not on visual condition (F = 0.6, p = 0.4). The coherence of the elderly group (0.52) was significantly higher than the coherence of the young group (0.46; F = 217.2, p < 0.001). There were significant interactions between age group and frequency (larger coherence differences between groups with lower frequencies, F = 5.8, p < 0.001),

amplitudes, body segments, and visual conditions. \* Statistically significant difference (*p* < 0.05). ◦Degree; Hz, Hertz. (B) GAIN, PHASE, and Coherence, interactions between age and body segments. GAIN (1), PHASE (2), and Coherence (3) of the two age groups across all stimulus amplitudes and visual conditions separated by body segments. \* Statistically significant difference (*p* < 0.05). ◦Degree; Hz, Hertz. (C) Influence of balance training on parameters of perturbed stance. GAIN (1) and PHASE (2) curves of the training and control group before (A1) and after (A2) training. \*Statistically significant difference (*p* < 0.05). ◦Degree.

between age group and body segment (larger shoulder than hip coherence in elderly, smaller shoulder than hip coherence in young subjects, F = 142.9, p < 0.001, **Figure 5B3**). Age group did not significantly interact with stimulus amplitude (F = 0.4, p = 0.5) and coherence did not significantly interact with balance training (F = 0.1, p = 0.77).

#### Model Parameters (see Figures 6A–C)

The integral gain, [I/mgh], was significantly higher in the young group (0.12 s−<sup>1</sup> ·rad−<sup>1</sup> ) than in the elderly group (0.10 s−<sup>1</sup> ·rad−<sup>1</sup> ; F = 35.0, p < 0.001). It was significantly higher with eyes open (0.12 s−<sup>1</sup> ·rad−<sup>1</sup> ) than with eyes closed (0.10 s−<sup>1</sup> ·rad−<sup>1</sup> ; F = 20.5, p < 0.001). In addition, [Td] was significantly larger in elderly (0.17 s) compared to young subjects (0.16 s; F = 19.4, p < 0.001). Moreover, [Wp] was significantly larger in elderly compared to young subjects (0.71 vs. 0.67; F = 5.8, p = 0.016, **Figure 6B**). It was significantly larger with eyes closed than with eyes open (0.79 vs. 0.58; F = 143.3, p < 0.001) and it was larger at a stimulus amplitude of 0.5◦ (0.74) than at 1◦ (0.64; F = 34.5, p < 0.001). Age group significantly interacted with visual condition (F = 4.3, p = 0.04). The difference of [Wp] between the eyes-open and eyes-closed condition was greater in young (0.24) than in elderly subjects (0.17). We found no significant interaction between age group and stimulus amplitude. The derivative gain, [D/mgh], was not significantly different between the age groups (elderly subjects: 0.376 s·rad−<sup>1</sup> , young subjects: 0.378 s·rad−<sup>1</sup> ; F = 0.06, p = 0.8). The proportional gain, [P/mgh], was significantly lower in elderly subjects (1.33 vs. 1.44 rad−<sup>1</sup> in young subjects; F = 10.7, p = 0.001) and significantly lower at a stimulus amplitude of 0.5◦ (1.35 vs. 1.43 rad−<sup>1</sup> at 1◦ ; F = 6.3, p = 0.013). Passive stiffness, [Ppas], and passive damping, [Dpas], were significantly larger in young subjects compared to elderly subjects ([Ppas], young: 89.4, elderly: 84.4; F = 15.1, p = 0.001; [Dpas], young: 60.3, elderly: 57.4; F = 8.5, p =0 .004). In general, [Ppas] und [Dpas] were larger with eyes open ([Ppas], eo: 92.9, ec: 80.9; F = 78.1, p < 0.001; [Dpas], eo: 61.6, ec 56.1; F = 32.5, p < 0.001). Visual condition and age group significantly interacted (F = 4.1, p = 0.042). The difference in [Ppas] between eyes-open and eyesclosed was greater in young (14.5) than in elderly subjects (10.7). Stimulus amplitude did not have a significant effect on [Ppas] (F = 2.6, p = 0.1).

Balance training did not have a significant effect on most model parameters ([I/mgh]: F = 0.004, p = 0.5, [P/mgh]: F = 0.8, p = 0.4, [D/mgh]: F = 1.3, p = 0.3, [Ppas]: F = 1.3, p = 0.3, [Dpas]: F = 0.3, p = 0.6, [Td]: F = 0.0, p = 1.0). However, the proprioceptive sensory weight, [Wp], changed significantly as a function of time (F = 4.0, p = 0.048, **Figure 6C**). [Wp] of the training group decreased from 0.73 to 0.70 (p < 0.05), [Wp] of the control group increased from 0.69 at the first assessment to 0.70 at the second assessment (p > 0.05).

#### Clinical Tests

The average reach distance of the training group increased significantly from 28.24 cm before to 32.08 cm after training (F = 7.4, p = 0.01). The reach distance of the control group decreased from 31.40 cm (A1) to 29.92 cm (A2) without being significant (F = 0.9, p = 0.4). Data of the TUG of the training group decreased during training but was not significant (A1: 8.48 s, A2: 8.34 s; F = 0.2, p = 0.7). Similar to the training group, there was no significant change in the TUG of the control group as a function of time (A1: 8.27 s, A2: 8.02 s; F = 0.2, p = 0.6).

#### Correlations

A correlation matrix was computed between measures (spontaneous sway and perturbed stance measures) and parameters that differed significantly between young and elderly subjects. Spontaneous sway measures RMS (r = 0.56, p = 0.0005) and MV (r = 0.42, p = 0.013), significantly correlated with GAIN. RMS also correlated with [Wp] (r = 0.36, p = 0.03). GAIN correlated with [Wp] (r = 0.37, p = 0.03) and [Td] (r = 0.45, p = 0.008, see **Figure 6D**).

## DISCUSSION

Here, the effect of balance training on postural control in elderly people was analyzed using a disturbance-related reactive motor approach. Postural control was assessed by spontaneous sway measures and measures of externally perturbed stance. Stimulus-response data were interpreted using a systems analysis approach (Engelhart et al., 2014; Pasma et al., 2014; Wiesmeier et al., 2015). We hypothesized that elderly subjects' postural control differed from that of young subjects, and that it was modified by balance training toward young subjects' postural control. In fact, elderly subjects displayed larger spontaneous sway amplitudes, velocities, and larger postural reactions than young subjects. Balance training reduced postural reaction sizes, which approached the range of values of young subjects. Using parameter identification techniques based on brain network model simulations, we found that balance training reduced overactive proprioceptive feedback and restored vestibular orientation in elderly. In the next paragraphs, we discuss the main findings sorted by the parameters analyzed, starting with age effects and followed by training effects, respectively.

Spontaneous sway was assessed using amplitude-related (RMS), velocity-related (MV), and frequency-related (MF) measures. All these measures have been reported to be higher in elderly than in young people (e.g., Prieto et al., 1996; Maurer and Peterka, 2005). In the present study, these differences were reproduced consistently.

While some authors reported effects of elderly's balance training on spontaneous sway (Judge, 2003; Hue et al., 2004; Nagy et al., 2007), we did not find significant effects. Some researchers interpreted smaller postural sway as improved balance (Judge, 2003; Hue et al., 2004). Others interpreted increased sway after balance training as an improved balance due to increased confidence (Nagy et al., 2007). As discussed in recent papers, different postural control deficits might lead to similar abnormalities in spontaneous sway measures reducing its usability for specific assessments of balance (Ghulyan et al., 2005; Wiesmeier et al., 2015).

Subjects' postural reactions as a function of external perturbations, i.e., anterior-posterior platform tilts, were characterized using GAIN and PHASE curves. Similar to reports in earlier papers, elderly subjects' postural reactions, i.e., GAIN values, were larger than in young subjects. This effect was more pronounced at the shoulder than at the hip level (Ghulyan et al., 2005; Wiesmeier et al., 2015). In other words, elderly subjects were dragged with the platform, while young subjects were more stable in space.

Balance training significantly reduced GAIN values toward the values of young subjects. The benefit of training amounted

FIGURE 6 | Model parameters and correlation matrix. (A) Model parameters of the two age groups ([P/mgh] in rad−<sup>1</sup> , [D/mgh] in s·rad−<sup>1</sup> , and [I/mgh] in s−<sup>1</sup> ·rad−<sup>1</sup> ); eo, eyes open; ec, eyes closed. \* Statistically significant difference (*p* < 0.05). (B) [Wp] (proprioceptive sensory weight) of the two age groups with respect to visual condition (1) and stimulus amplitude (2); eo, eyes open; ec, eyes closed. \* Statistically significant difference (*p* < 0.05). (C) Influence of balance training on [Wp]. [Wp] of the training group before (A1) and after (A2) balance training. (D) Correlation matrix of measures (spontaneous sway and perturbed stance measures) and parameters that differed significantly between young and elderly subjects. Only significant correlations are shown. RMS, Root Mean Square; MV, Mean Velocity; MF, Mean Frequency; [Wp], proprioceptive sensory weight; [Td], time delay.

to about 30% of the GAIN difference between elderly and young subjects. If we assume a linear relationship between deterioration of postural control and age, based on former studies (Era et al., 2006; Wiesmeier et al., 2015), the training effect could be extrapolated as a juvenescence of about 10 years given the average age difference of 36 years between the two age groups.

In general, larger postural reactions in elderly could be due to an intensified use of ankle proprioception and a reduced use of vestibular information. Vestibular information would be used to stabilize the body in space. Other reasons for large postural reactions might include general muscle weakness, or an increased time delay between stimulus and response. In order to separate the different possible subsystems responsible for increased postural reactions, we applied a systems analysis approach based on a well-known postural control model (Engelhart et al., 2014; Pasma et al., 2014).

Using model simulations, we were in fact able to identify larger contributions of proprioception to sensory feedback in elderly as compared to young subjects. The higher the proprioceptive feedback, the lower the contributions of space cues, i.e., the vestibular information, when eyes were closed, and vestibular and visual information with eyes open. With a decreased vestibular feedback, elderly people are closer to vestibular loss patients (see in Maurer et al., 2006) than to patients suffering from polyneuropathy (unpublished data). As we showed before (Maurer et al., 2006), vestibular loss patients, who are forced to rely 100% on proprioception, tend to fall on tilting platforms, signifying the problems with a pure proprioceptive strategy.

After training, the proprioceptive feedback was significantly reduced with both eyes open and eyes closed. The decrease of proprioceptive feedback in the eyes-closed condition can only be explained by an increase of the vestibular feedback. This indicates that elderly subjects learned to weigh vestibular information higher. Because the increase of the weight for space cues is similar in the eyes-open condition, we assume that this, again, is caused by an increase of the vestibular feedback. Interestingly, proprioceptive feedback was not only the most prominent difference between elderly and young subjects, but also the only parameter that was affected by training. The other parameters, that differed between young and elderly subjects (larger time delay [Td], smaller integral gain [I/mgh], smaller proportional gain [P/mgh], smaller passive stiffness factor [Ppas], and smaller passive damping factor [Dpas] in elderly), were not significantly affected by balance training. This could be due to the fact that the parameters not affected by training represent those physiological constituents of postural control that may be closely related to anatomical features of the subject (Peterka, 2002), such as, height of the COM, mass distribution, patterns of muscle recruitment, or nerve conduction time. While balance training could principally affect plastic central weighting processes, it is less likely that they directly influence anatomical constraints of the body.

Additional differences between young and elderly subjects' postural reactions were related to a more pronounced PHASE slope and a smaller PHASE lag (Ghulyan et al., 2005; Wiesmeier et al., 2015). More specifically, PHASE difference between shoulder and hip decreased with age, pointing to a different coordination of body segments. We interpreted this as a change in reactive balance strategy using more hip flexion and extension. This effect might be in accordance to a hypothesis presented by Kuo et al. (1998) that the use of hip flexion/extension may be enhanced in conditions where the support surface is not reliable, i.e., in unstable platform conditions. All these additional differences were not significantly affected by balance training.

Experimental results were compared with known clinical tests for the assessment of postural deficits, namely the FRT and the TUG TEST (Enkelaar et al., 2013). The FRT was significantly ameliorated by balance training. This is in accordance with the expected effects of balance training, as FRT clinically stands for the ability to balance. FRT values of our elderly group were similar to the ones reported by Duncan et al. (1990) who assessed 128 volunteers between 21 and 87 years. The TUG was improved after balance training (not significantly). The TUG scores of the training and control groups corresponded to scores reported by Nagy et al. (2007; 8.9–10.3 s) and Enkelaar et al. (2013; 9.3 s).

The correlation analysis of elderly's data before and after training revealed that the significant effect of balance training, represented by the larger postural reactions (GAIN) significantly correlated with the strength of proprioceptive feedback [Wp], with RMS, with MV, and with time delay [Td]. This correlation pattern indicates that there is a certain tendency that the elderly ameliorated spontaneous sway measures and time delay with training, which is correlated with the main training effect of the recovery of vestibular function.

## CONCLUSION

Balance training reduced elderly subjects' overactive proprioceptive feedback and enhanced vestibular orientation. The modified use of sensory information can be interpreted as a change in postural control strategies representing a higher level adaptive mechanism. Based on the assumption of a linear deterioration of postural control across the life span, the training effect can be extrapolated as a juvenescence of about 10 years. This is even more surprising, given the fact that the elderly subjects evaluated here were in a healthy and active state prior to the study. We hold that this study points to a considerable benefit of a continuous balance training in elderly, even without any sensorimotor deficits.

## ETHICS STATEMENT

This study was carried out in accordance with the recommendations of Ethics Committee of Freiburg University with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Ethics Committee of Freiburg University.

## AUTHOR CONTRIBUTIONS

IW and CM: substantial contributions to the conception and design of the study, data acquisition, analysis, and interpretation, drafted and revised manuscript, DD: substantial contributions to the conception and design of the study, acquisition, analysis, and interpretation of data, AW and JD: conception and design of the study, data acquisition, revised manuscript, UG, TM, CW, and AG: substantial contributions to the conception and design of the study, data acquisition and interpretation, revised manuscript.

#### FUNDING

CM was partially funded by a European Union FP7 grant (EMBALANCE: Grant Agreement no 610454), the Brainlinks-Braintools Cluster of Excellence funded by the

### REFERENCES


German Research foundation (DFG, grant no ADV139) and by DFG (MA 2543/3-1).

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnagi. 2017.00273/full#supplementary-material


motor control. Front. Aging Neurosci. 7:97. doi: 10.3389/fnagi.2015. 00097

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Wiesmeier, Dalin, Wehrle, Granacher, Muehlbauer, Dietterle, Weiller, Gollhofer and Maurer. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Open- and Closed-Skill Exercise Interventions Produce Different Neurocognitive Effects on Executive Functions in the Elderly: A 6-Month Randomized, Controlled Trial

Chia-Liang Tsai <sup>1</sup> \*, Chien-Yu Pan<sup>2</sup> , Fu-Chen Chen<sup>2</sup> and Yu-Ting Tseng1, 3

1 Institute of Physical Education, Health and Leisure Studies, National Cheng Kung University, Tainan, Taiwan, <sup>2</sup> Department of Physical Education, National Kaohsiung Normal University, Kaohsiung, Taiwan, <sup>3</sup> School of Kinesiology, University of Minnesota, Minneapolis, MN, United States

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Hartmut Heinrich, Universitätsklinikum Erlangen, Germany Hamid R. Sohrabi, Macquarie University, Australia István Czigler, Institute of Cognitive Neuroscience and Psychology (HAS), Hungary Dezhong Yao, University of Electronic Science and Technology of China, China Jason Michael Bruggemann, University of New South Wales, Australia Belinda Brown, Murdoch University, Australia

> \*Correspondence: Chia-Liang Tsai andytsai@mail.ncku.edu.tw

Received: 31 May 2017 Accepted: 29 August 2017 Published: 12 September 2017

#### Citation:

Tsai C-L, Pan C-Y, Chen F-C and Tseng Y-T (2017) Open- and Closed-Skill Exercise Interventions Produce Different Neurocognitive Effects on Executive Functions in the Elderly: A 6-Month Randomized, Controlled Trial. Front. Aging Neurosci. 9:294. doi: 10.3389/fnagi.2017.00294 This study aimed to explore the effects of open- and closed-skill exercise interventions on the neurocognitive performance of executive functions in the elderly. Sixty-four healthy elderly males were randomly assigned to either a closed-skill (bike riding or brisk walking/jogging, n = 22), open-skill (table tennis, n = 21), or control (n = 21) group. Various neuropsychological [e.g., accuracy rates (AR) and reaction time (RT)] and electrophysiological [e.g., event-related potential (ERP) P3 component] measures were assessed during a variant of the task-switching paradigm, as well as an N-back task at baseline and after either a 6-month exercise intervention or control period. The results showed that, when performing the task-switching paradigm, the two exercise groups relative to control group showed significantly faster RTs in the switch trials after the exercise intervention. However, the RT facilitation in the non-switch and switch trials post-exercise relative to pre-exercise only emerged in the open-skill group. In terms of the N-back task, the two exercise groups significantly increased ARs in the 1-back condition after the exercise intervention, and the beneficial AR effect on the 2-back condition only emerged in the closed-skill group. In addition, the two exercise groups exhibited significantly larger P3 amplitudes on the frontal-to-parietal cortex areas after the exercise intervention relative to the baseline when performing the two cognitive tasks. These neurocognitive results still remained unchanged even when the confounding factors (e.g., cardiorespiratory fitness, social participation, and BMI) were controlled for. The present study concluded that, although 6-month open- and closed-skill exercise interventions facilitate overall electrophysiological effects (i.e., increased ERP P3 amplitudes) on the frontal-to-parietal cortices in the elderly, the two exercise modes produced different levels of neuropsychologically beneficial effects on RTs of the task-switching paradigm (i.e., lessened RTs) and ARs of the N-back task (i.e., enhanced ARs). The distinctive neurocognitive changes induced by open- and closed-skill exercise have implications for task switching and working memory in elderly individuals, especially with such cognitive functioning impairments.

Keywords: cognition, behavior, event-related potential, exercise modes, elderly

### INTRODUCTION

Life expectancy has been increasing in developed countries, resulting in a rapid growth in the elderly population. Since aging is the main risk factor for neurodegenerative diseases (e.g., Alzheimer's disease), which affect a substantial and growing part of the global population (Rodríguez-Arellano et al., 2016), one important issue is how to counteract such neurocognitive declines in order to reduce the medical costs associated with geriatric care. Although a decrease in certain cognitive functions is an unavoidable part of normal aging, the degree to which this occurs varies within the healthy older population. Among an array of cognitive functions, the executive-control processes and brain areas that support them have been shown to undergo large age-related performance declines (West, 1996; Colcombe and Kramer, 2003; Anderson and McConnell, 2007), as seen in capacities such as working memory (Wingfield et al., 1988), visuospatial attention (Greenwood et al., 1993), and task switching (Friedman et al., 2008).

The vast majority of related studies have proposed that physical fitness and exercise are factors that strongly promote healthy cognitive aging (Lee et al., 2012; Kimura et al., 2013; Tsai et al., 2015). Improvements in physical fitness via exercise training are thus reflected in enhancements of a number of cognitive functions, such as processing speed, visuospatial function, and control processes (e.g., inhibition, planning, scheduling, and working memory) in older adults, with the largest effect sizes being on tests thought to depend more on the executive function (e.g., task switching, inhibitory control, and working memory; Diamond, 2013) (Hall et al., 2001; Voelcker-Rehage et al., 2010; Guiney and Machado, 2013; Tsai et al., 2015). However, a broad range of physical exercise types is possible, and different kinds of exercise seem to have specific effects on neurocognitive performance, due to the differences in the secretion of some biomarkers (e.g., brain-derived neurotrophic factor, insulin-like growth factor-1, and homocysteine) in the neurochemical system (Neeper et al., 1995; Liu-Ambrose et al., 2010; Cassilhas et al., 2012; Tsai et al., 2014b,c), and the differences in brain tissue volumes and activation patterns induced by different types of exercise (Luft et al., 2008; Park et al., 2008; Liu-Ambrose et al., 2010; Erickson et al., 2011; Tsai and Wang, 2015; Tsai et al., 2016). These earlier works seem to support the view that different types of physical exercise could affect the brain in different ways.

Different forms of physical exercise with different cognitive executive process loads and different motor-coordination skills have been reported to be strongly associated with improved neurocognitive performances (Voelcker-Rehage et al., 2011; Tsai et al., 2016). Therefore, the present study divided exercise into two main modes, open- and closed-skill (Di Russo et al., 2010; Dai et al., 2013; Tsai and Wang, 2015), since the former (e.g., table tennis and badminton) requires rich cognitive and executive loadings and different sets of motor-coordination skills to adapt to a unpredictable/changing environment and various opponents (van Praag et al., 2000; Artola et al., 2006; Di Russo et al., 2010), while the later (e.g., running and biking) is performed according to the individual's own pace in a stable and predictable environment (Di Russo et al., 2010). From the perspective of motor-coordination skill, Voelcker-Rehage et al. (2011) found that older adults experienced beneficial effects on executive control (assessed using the Flanker task) and perceptual speed (assessed using the Visual Search task) due to cardiorespiratory and coordination training. However, the two exercise modes produced different effects on speed and accuracy, with coordination training leading to improved accuracy rates (ARs) on executive control and perceptual speed, but cardiorespiratory training only leading to better ARs on executive control. In terms of reaction time (RT), only the perceptual speed task was significantly improved by coordination training. In addition, cardiorespiratory training increased activation of the sensorimotor network in the elderly, while coordination training elevated activation of the visual–spatial network (Voelcker-Rehage et al., 2011). In contrast, Hötting et al. (2012) found that, although significant increases in episodic memory learning scores were found for both the cycling and stretching/coordination groups as compared with the sedentary control group, cycling training had greater effects on the episodic memory recognition scores than the stretching/coordination training. They also found that the latter was particularly effective in improving selective attention as compared with the cycling training. This suggests a specific relation between particular types of exercise and cognitive functions, with the increase in memory functions being linked to an increase in cardiovascular fitness, whereas the increase in attention is more pronounced after stretching/coordination training. The two exercise modes (i.e., open- and closed-skills) thus seem to be capable of producing different effects on the various cognitive domains (Voss et al., 2010) and neural processes (Tsai and Wang, 2015; Tsai et al., 2016).

Although recent studies have explored the effects of openand closed-skill exercise on neuropsychological performance in disabled athletes (Di Russo et al., 2010) and the young adults (Wang et al., 2013a,b), and neurocognitive (neuropsychological and electrophysiological) performances in the elderly (Dai et al., 2013; Tsai and Wang, 2015; Tsai et al., 2016), the findings in the rather limited research literature remain somewhat ambiguous. More importantly, even though previous cross-sectional studies have demonstrated that regular participation in open- and closed-skill exercise has distinct benefits for neurocognitive performances (e.g., specific cost, RT, P3 amplitudes, and strength of inhibitory control) in the elderly when performing the task switching paradigm (Dai et al., 2013; Tsai and Wang, 2015) and visuospatial attention task (Tsai et al., 2016), the elderly subjects who showed more of the benefits of open-skill exercise on neurocognitive performance might have had some inherently better aspects of their executive control functions (e.g., visuospatial attention and task switching) as compared to their counterparts participating in the closed-skill exercise mode, and this may have induced them to adopt this kind of exercise (Snowden et al., 2011). Therefore, these cross-sectional studies cannot establish causality between exercise and cognitive aging, which is nonetheless required for more accurate and effective public health recommendations (Snowden et al., 2011; Miller et al., 2012), as well as to better explain the beneficial effects of the two exercise modes.

Two cognitive tasks, the task-switch paradigm and N-back task, were adopted in the current study to investigate the impacts of the various exercise-mode mechanisms responsible for specific kinds of executive-control functioning, since earlier works found that open- and closed-skill exercise could have different neurocognitive effects on different cognitive tasks executive functions (Di Russo et al., 2010; Dai et al., 2013; Tsai and Wang, 2015; Tsai et al., 2016), and, crucially, open-skill exercise (e.g., table tennis) affects more of the prefrontal cortex areas responsible for attention, task-switching and inhibition, while closed-skill exercise (e.g., jogging) works more on the hippocampus, which is important for memory (e.g., long-term memory and working memory; Axmacher et al., 2010; Burrel, 2015). Moreover, there are previous reports of age- and physical-activity-related impacts on the neuropsychological (e.g., AR and RT) and electrophysiological [e.g., event-related potential (ERP) P3 component] outcomes of the task-switching paradigm (Hillman et al., 2006; Friedman et al., 2008; Adrover-Roig and Barceló, 2010; Guiney and Machado, 2013) and N-back task (Voelcker-Rehage et al., 2010; Guiney and Machado, 2013; Saliasi et al., 2013) among elderly subjects.

ERP recordings made during the cognitive task performance permitted on-line measures of cognitive processes on the order of milliseconds, which cannot be obtained by neuropsychological performance alone (Tsai et al., 2014b). Given that the P3 activity has nonspecific qualities that are often associated with indexing stimulus evaluations and the intensity of the concomitant executive function processes (e.g., task-set updating processes and reconfiguration, updating working memory, integrating information into existing networks) (Kok, 2001; Kieffaber and Hetrick, 2005; Nicholson et al., 2005; Polich, 2007), the ERP component was used to illustrate the effects of the different exercise-mode interventions on executive cognitive functions in the elderly in the present study. With regard to the electrophysiological index, the P3 activity induced by the taskswitch paradigm represents the set of processes subsumed under the construct of the task-set reconfiguration (Kieffaber and Hetrick, 2005; Nicholson et al., 2005). ERP P3 represents the memory-related neural processing that is involved in categorizing incoming information and updating the context of the working memory (e.g., encoding, rehearsal, recognition, and retrieval) (Duncan-Johnson and Donchin, 1977; Donchin and Coles, 1988; Rugg, 1995).

To date there is a lack of intervention research on the impact of open- and closed-skill exercise modes on various forms of executive function (e.g., task switching and working memory) involved in cognitive aging in older adults. The main goal of this study was thus to clarify the distinctive effects of a 6-month open- and closed-skill exercise intervention on the neuropsychological (e.g., AR and RT) and electrophysiological (e.g., ERP P3 latency and amplitude) performances in older adults with a sedentary life-style when performing the taskswitching paradigm and N-back task, with rigorous controls on the confounding factors in neurocognitive performance [e.g., cardiorespiratory fitness, social participation, and body mass index (BMI), since these parameters could be changed to different extents after exercise] (Messier and Gagnon, 2009; Miller et al., 2012). The elderly subjects with regular openskill exercise participation in the literature general show better switch-related neurocognitive performances than those with only closed-skill experience (Dai et al., 2013; Tsai and Wang, 2015), and short-term closed-skill exercise intervention (e.g., resistance exercise) cannot improve task-switching performance (Kimura et al., 2010), while closed-skill exercise training with the goal of enhancing cardiorespiratory fitness can facilitate white matter integrity, increase the size of the hippocampus, and improve memory performance in the elderly (Erickson et al., 2011; Ruscheweyh et al., 2011; Voss et al., 2013; Maass et al., 2016) and have more benefits on memory functions than stretching/coordination training (Hötting et al., 2012). As such, we hypothesized the following: (1) that a 6-month open-skill exercise intervention (e.g., table tennis) in contrast to a closedexercise one (e.g., bike riding or brisk walking/jogging) would have more benefits for neurocognitive performance with regard to task-switching in the elderly; and (2) that closed-skill exercise would have more beneficial effects on the cognitive functioning involving the memory domains.

### MATERIALS AND METHODS

#### Participants

Sixty-four community-dwelling men were recruited with the use of an informative flyer and underwent screening by a standardized telephone interview, with subjects being eligible for inclusion in this study if they (1) were aged 60–80 years old and a non-smoker; (2) were living independently in their own home with a sedentary life-style; and (3) did not have a current medical condition for which exercise is contraindicated. The participants consisted solely of men because neurocognitive and endocrinological responses to exercise could be genderdependent (Baker et al., 2010). They then underwent a routine laboratory testing and medical examination, including blood pressure and heart rate measurements, electrocardiography, a standardized neurological and psychiatric examination, and a structured interview on previous medical history to ascertain whether they were free of a history of neurological disorders, brain injury, depressive symptoms [scores above 13 on the Beck Depression Inventory, 2nd edition (BDI-II)], and cognitive impairment [scores below 26 on the Mini Mental State Examination (MMSE)] (Ruscheweyh et al., 2011). The Edinburgh Handedness Inventory assessed all participants as right-handed (Oldfield, 1971). Written informed consent, as approved by the Institutional Ethics Committees in the organization within which the study was performed, was obtained from all the participants.

#### Procedures

The Consolidated Standards of Reporting Trials (CONSORT) flowchart outlining the number of participants from first contact to study completion is shown in **Figure 1**. The original cohort consisted of 79 participants. After the assessment of a physician specializing in geriatric care, two subjects were excluded due to a history of heart disease, five due to neurological disorders, musculoskeletal problems, or psychiatric illness (e.g., scores

above 13 on the BDI-II or below 26 on the MMSE), and three due to regular participation in open- and/or closed-skill exercise in the previous 3 months. The remaining 69 participants were randomly assigned to either an open-skill, closed-skill, or active control group by drawing an envelope with the treatment assignment enclosed after matching for age.

Before the exercise intervention, the participants visited the cognitive neurophysiology laboratory on two separate occasions. During the first session, each participant completed an informed consent form, the basic information form (e.g., a medical history and demographic questionnaire), and a handedness inventory. Two certified fitness instructors then completed all assessments of senior functional physical fitness for each participant. On a separate day within 1 week of the completion of the baseline evaluation, the participants performed two cognitive tasks (i.e., a task-switching paradigm and an N-back task) in a counter balanced order with concomitant electrophysiological recording (i.e., event-related potentials, ERPs).

Before the final exercise interventions were completed, one participant in the open-skill group, two in the closed-skill group, and two in the active control group, terminated their participation, leaving 64 participants in the final sample (openskill, n = 22; closed-skill, n = 21; active control, n = 21). The three groups did not significantly differ at baseline in any of the demographic characteristics, including years of formal education, body mass index, systolic and diastolic pressure, social participation, MMSE, BDI-II, and senior functional physical fitness (see **Table 1**). Within 1 week after completing a 24 week exercise intervention, the participants completed the same questionnaires and senior functional physical fitness assessment, and received the same neurocognitive measurements, as in the pre-exercise procedure. Compliance with the prescribed training protocol remained high throughout the study period (90 ± 2%).

## Training Protocol

#### Open-Skill Exercise Condition

The participants in the open-skill group were trained individually and regularly in a series of 40-min sessions that took place three times per week for 24 weeks, with the following structure: a warm-up, the main part of table tennis training, playing games of table tennis with the coach, and cooling down at the end. The training intervention was carried out in a sequence of increasing complexity. The table tennis training program was intended to improve the participants' general skills, and had seven main components over the whole training session: (a) footwork (e.g., ready position, one-, two-, and cross-step), (b) serving (including how to give sidespin, backspin, topspin, no-spin, and so on), (c) forehand and backhand driving, (d) forehand bouncing, backhand bouncing, and alternate bouncing, (e) smashing, (f) continuously hitting back a ball that was randomly delivered by


TABLE 1 | Demographic characteristics of the open-skill exercise, closed-skill and control groups before and after the exercise intervention.

BMI, body mass index; MMSE, Mini Mental State Examination; BDI-II, Beck Depression Inventory, 2nd edition; VO2max , maximal oxygen uptake. \*p < 0.05.

the ball-projection machine from fixed or random directions, and (g) comprehensive practice. Training for each skill began with a simple movement and then progressed to more complex variations. The assessor for the neurocognitive tests and the exercise leader/coach were blinded to group membership. A more detailed training manual is available on request from the authors.

#### Closed-Skill Exercise Condition

Participants in the closed-skill group attended three supervised exercise sessions per week for 24 consecutive weeks on a bicycle ergometer or a motor-driven treadmill (Medtrack ST55, Quinton Instrument Company, United States), with exercise intensity corresponding to 50–60% of the individual target heart rate reserve (HRR) during the first 2 weeks and 70–75% of the HRR for the remainder of the program. Each aerobic exercise session involved a 5-min warm-up period, followed by 30 min of continuous bike riding or brisk walking/jogging at an intensity that would maintain the heart rate within the assigned training range, and 5 min of cool-down. A Polar HR monitor (RX800CX, Finland) was used to monitor each participant's heart rate during the exercise.

#### Control Condition

Participants in the active control group attended a balance and stretching program led by a trained exercise leader three times a week for 24 weeks. Every class included a 5-min warm-up period, static stretching and balance training, and a 5-min cooldown period. Different stretching and balancing techniques used various equipment, such as balance boards and fitness balls, to maintain the participants' interest.

#### Physical Fitness Assessment

The Senior Functional Physical Fitness (SFPF) test (Rikli and Jones, 2012) is a battery of tests that was used to assess the participants' physical fitness in the current study. The participants first undertook 10 min' warm-up before the test and then completed the component tests in the designated order. Five items in the SFPF test were measured, as follows: (1) the Arm Curl test, which assesses arm muscle (specifically of the biceps) strength endurance, with the score being the number of repetitions in 30 s using the elbow of the dominant hand to flex and extend with a weight (men: 8 lb; women: 5 lb) through the complete range of motion; (2) the Chair Stand test, which measures lower body strength, based on the number of repetitions in 30 s using a full standing position from a chair and then returning to a fully seated position; (3) the 8-Foot Up-and-Go test, which evaluates agility and dynamic balance, using the number of seconds needed to get up from a seated position from a chair, walk eight feet, and return to a fully seated position on the chair; (4) the Back Scratch test, which assess upper body (shoulder) flexibility based on the number of centimeters being short of touching (minus score) or overlapping (plus score) between the third fingertip of each hand; (5) the Chair Sit-and-Reach test, which assesses the flexibility of the lower extremities, with the score being the distance in centimeters between the fingertips and toes; the number of centimeters short of reaching the toes (minus score) or reached beyond the toes (plus score). With regard to cardiorespiratory fitness, the Rockport Fitness Walking Test was used to estimate VO2max (Kline et al., 1987), in which the participants were required to walk one mile as quickly as possible, during which their heart rate was continuously recorded using a Polar HR monitor.

## Cognitive Tasks

#### Task-Switching Paradigm

The task switching paradigm employed in the present study has been shown to effectively assess the variations among elderly subjects who regularly participate in open- and closed-skill types of exercise (Tsai and Wang, 2015). The stimulus used in this test was a white digit (1–9, excluding 5) shown focally in the center of the screen against a black background. The same digit was not repeated in successive trials, and the digits were put into eight task blocks (blocks 1–2 and 7–8: homogeneous tasks; blocks 3– 6: heterogeneous tasks), with a short rest period in the middle of each. The homogenous (i.e., non-switch) blocks each included 56 trials. Within each homogenous block, the participants only responded if the focal digit was greater or less than 5 (e.g., blocks 1 and 7), or if it was odd or even (e.g., blocks 2 and 8). The heterogeneous (i.e., task-switching) blocks contained 224 trials, each with 20 switches. With the heterogeneous blocks, the participants started with one task (e.g., even/odd) and then switched to the other (e.g., more/less than 5), as signaled by a simultaneously presented rectangle drawn around the digit, after at least seven or no more than 13 intervening trials. The participants were asked to press one of two buttons on a small response box that they held in the right hand as quickly and accurately as possible. The digit was shown on the screen until the participants pressed the response button, and the following trial began 500ms after the RT response. The prompts "more less" or "even odd" appeared simultaneously with and below the digit during all trials, dependent on which was appropriate to the task. The homogenous and heterogeneous blocks were counterbalanced across participants. All the participants were given instructions about the tasks, and both single-task and taskswitch trials were practiced before the formal test until they understood the whole procedure.

#### N-back Task

A continuous stream of white letters (stimuli) with pseudorandom sequences of vowels and consonants was presented with 10% gray noise, embedded in a 50% random noise gray rectangular background patch, on a computer screen. Targets were defined according to the N-back design. Participants pressed a button with the index finger of their right hands as soon as a target appeared, and no motor response was needed for non-target trials. Stimuli were presented for a duration of 500 ms, followed by a 3 s inter-trial-interval during which a dot helped participants maintain fixation. The cognitive task consisted of three different working-memory-load conditions: (1) a simple detection (control) condition with sequential letters or background patches without any letters being presented, during which the participants responded when the background patches without letters appeared (memory-free condition); (2) the 1-back condition, with the target being any letter identical to the one immediately preceding it (moderately demanding); (3) the 2-back condition, with the target being any letter that was identical to the one presented two trials back (highly demanding). Before the formal test, the participants were given the task instructions and initially practiced a brief version of the task, consisting of two blocks of 45 trials each (one block of moderately demanding and one of highly demanding trials). Following this practice, the participants completed nine blocks of trials with 120 trials in each (three blocks per condition). Nine blocks were tested following the sequence: blocks 1, and 8–9: the control condition, blocks 2, and 6–7: the 1-back condition, and blocks 3–5: the 2-back condition. No feedback was provided during this period.

### ERPs Recording and Analysis

Continuous electroencephalographic (EEG) signals were acquired from 18 electrodes (F7, F3, Fz, F4, F8, C3, Cz, C4, T5, T3, T4, T6, P3, Pz, P4, O1, Oz, and O2) placed using a 10/20 extended Quik-Cap system (Compumedics Neuroscan, Inc., El Paso, TX). Horizontal and vertical electrooculogram (EOG) activity for eye movements was monitored bipolarly with ocular electrodes placed on the supero-lateral right canthus and infero-lateral to the left eye. A ground electrode was placed on the mid-forehead on the Quik-Cap. References were placed at vertex by default, but were subsequently re-referenced off-line to averaged mastoids. The values of interelectrode impedance were kept at 5 K for all electrodes. The raw EEG signal was recorded with an A/D rate of 500 Hz/channel, a band-pass filter of 0.1–50 Hz, and a 60-Hz notch filter using an on-line amplifier. These data were written continuously to hard disk for off-line analysis using Neuroscan Scan 4.3 analysis software (Compumedics Neuroscan, Inc., El Paso, USA).

Trials with a response error or EEG artifacts (e.g., VEOG, HEOG, and electromyogram) exceeding peak-to-peak deflections over 100µV were discarded before averaging. ERPs were extracted off-line and averaged in epochs starting 200 ms prior to stimulus activity onset, and lasting for 1,000 ms of post-stimulus activity. Since the ERP P3 component is widely examined among studies that independently and simultaneously investigate the externalizing spectrum and executive functioning (Baskin-Sommers et al., 2014), it was the focus of the current work. The effects of the two cognitive tasks on the P3 component in the elderly were clearly visible in the frontal-parietal regions of the scalp in the current study (see also Kieffaber and Hetrick, 2005; Friedman et al., 2008; Baskin-Sommers et al., 2014). Three electrodes (Fz, Cz, Pz) were thus analyzed in the present work (Themanson et al., 2006; Tusch et al., 2016). P3 mean amplitudes were calculated for the time-window between 300 and 600 ms post stimulus.

#### Data Processing and Statistical Analyses

One-way analysis of variance (ANOVA) was used to examine the homogeneity of the demographic backgrounds of the participants at the baseline in the open-skill, closed-skill, and control groups. A two-tailed paired t-test was used to analyze the differences within the three groups at baseline and at 24 weeks.

Different trials/conditions in the two cognitive tasks were subjected to neuropsychological (i.e., AR and RT) and electrophysiological (i.e., P3 amplitude) statistical analyses. Only the RT and ERP data corresponding to correct answers were averaged according to the task trial/condition. With regard to the task-switching paradigm, two switch costs were determined by the RTs performance: (1) general-switch cost, which was obtained by subtracting the mean RT between homogeneous trials during homogeneous blocks and non-switch trials during heterogeneous blocks; (2) specific-switch cost, which was calculated by subtracting the mean RT between non-switch trials and switch trials during heterogeneous blocks.

#### Task Switching Paradigm

With regard to the neuropsychological performance, the AR and RT were separately submitted to a 3 (Group: open-skill vs. closed-skill vs. control) × 2 (Time: pre-exercise vs. post-exercise) × 3 [Trial: homogeneous (during homogeneous blocks) vs. non-switch vs. switch (during heterogeneous blocks)] mixed repeated measures analysis of variance (RM–ANOVA). In terms of the electrophysiological performance, the P3 latency and amplitude from ERP recordings were submitted to a 3 (Group: open-skill vs. closed-skill vs. control) × 2 (Time: pre-exercise vs. post-exercise) × 3 (Trial: homogeneous vs. non-switch vs. switch) × 3 (Electrode: Fz vs. Cz vs. Pz) RM–ANOVA.

#### N-back Task

With regard to the neuropsychological performance, the AR and RT were separately submitted to a 3 (Group: open-skill vs. closedskill vs. control) × 2 (Time: pre-exercise vs. post-exercise) × 3 (Condition: 0-back vs. 1-back vs. 2-back) RM–ANOVA. In terms of the electrophysiological performance, the P3 latency and amplitude from ERP recordings were submitted to a 3 (Group: open-skill vs. closed-skill vs. control) × 2 (Time: pre-exercise vs. post-exercise) × 3 (Condition: 0-back vs. 1-back vs. 2-back) × 3 (Electrode: Fz vs. Cz vs. Pz) RM–ANOVA.

Bonferroni post-hoc analyses were performed when there was a significant difference. Simple main effects were determined following observation of any significant interaction effects. Because cardiorespiratory fitness, social participation, and BMI are confounding factors on cognitive performance (Messier and Gagnon, 2009; Miller et al., 2012), the neuropsychological and electrophysiological performances of the three groups after the exercise intervention were assessed separately using an analysis of covariance (ANCOVA) to examine the effects on the neurocognitive performance in relation to different types of intervention, with the three post-exercise factors as a covariate. The Greenhouse–Geisser (G–G) correction adjusted the significance levels of the F ratios whenever RM–ANOVA detected a major violation of the sphericity assumption. Partial Eta squared (η<sup>p</sup> 2 ) was used to calculate effect sizes for significant main effects and interactions, with the following criteria used to determine the magnitude of mean effect size: 0.01–0.059 indicated a small effect size; 0.06–0.139, a medium effect size; and >0.14, a large effect size. A p-value less than 0.05 for the differences between the mean values is considered statistically significant.

#### RESULTS

#### Demographic Characteristics

**Table 1** presents an overview of the pre- and post-exercise characteristics of the participants. At baseline there were no significant differences (all ps > 0.05) at the group level with regard to age, height, weight, BMI, systolic and diastolic pressure in the three groups (i.e., open-skill, closed-skill, and control). Other confounding factors (e.g., the years of education, social participation, memory depth, global cognitive function, depression, and cardiorespiratory fitness) in relation to cognition and the tests of functional fitness also did not achieve significant difference in the three groups (all ps > 0.05) across the three groups at baseline. Paired t-tests showed that after the exercise intervention the open-skill group had significantly improved scores for the social participation, arm curl, chair stand, and 8 foot up-and-go items, and approached significance for the level of cardiorespiratory fitness (VO2max); and the results for the closed-skill group showed that the values of weight and BMI decreased significantly, the level of cardiorespiratory fitness was significantly enhanced, and the performance of memory depth approached significance.

### Accuracy Rate (AR)

#### Task-Switching Paradigm

As shown in **Figure 2**, RM–ANOVA performed on the ARs only highlighted the main effects of Time [F(1, 61) = 10.36, p = 0.002, η<sup>p</sup> <sup>2</sup> = 0.15] and Trial [F(2, 122) = 35.57, p < 0.001, ηp <sup>2</sup> = 0.37], with a higher AR post- (90.96%) than pre-exercise (88.53%) and with the following trial gradient across the three groups: homogeneous (93.52%) > non-switch (88.96%) > switch (86.75%).

#### N-back Task

RM–ANOVA performed on the ARs for the N-back task revealed the main effects of Time [F(1, 61) = 20.45, p < 0.001, η<sup>p</sup> 2 = 0.25] and Condition [F(2, 122) = 129.89, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.68], with a higher AR for post- (90.61%) than pre-exercise (88.43%), and with the following condition gradient: 0-back (98.01%) > 1-back (90.17%) > 2-back (80.37%). The interactions between Time × Group [F(2, 61) = 4.74, p = 0.012, η<sup>p</sup> <sup>2</sup> = 0.14], Time × Condition [F(2, 122) = 10.51, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.15], and Time × Group × Condition [F(4, 122) = 2.61, p = 0.012, η<sup>p</sup> <sup>2</sup> = 0.39] were also significant. Post-hoc analysis for the Time × Group × Condition interaction showed that (1) there were no significant differences at any condition among the three groups at baseline and after the exercise intervention (all ps > 0.05); and (2) when post- compared to pre-exercise, the open- [F(1, 21) = 4.44, p = 0.047] and closed-skill [F(1, 20) = 21.45, p < 0.001] groups exhibited significantly higher ARs in the 1-back condition, and only the closed-skill [F(1, 20) = 36.59, p < 0.001] group exhibited a significantly higher AR in the 2-back condition. In addition, the significant post- and pre-exercise differences in ARs among the three groups only emerged in the 2-back condition, with a greater improvement after the exercise intervention for the closed-skill group than the open-skill and control groups (closedskill vs. open-skill: p = 0.002; closed-skill vs. control: p =

0.001). Even when the post-exercise cardiorespiratory fitness, social participation, and BMI were controlled for, the results of ANCOVA on the ARs in the 2-back condition still indicated a significant difference in the three groups [F(2, 58) = 9.90, p < 0.001], with post-hoc analysis indicating that closed-skill group performed significantly better than the open-skill (closed-skill vs. open-skill: p < 0.001) and control (closed-skill vs. control: p < 0.001) groups after the exercise intervention.

## Reaction Time (RT)

#### Task-Switching Paradigm

As illustrated in **Figure 2**, RM–ANOVA conducted on mean RTs revealed the main effects of Time [F(1, 61) = 10.22, p = 0.002, ηp <sup>2</sup> = 0.14] and Trial [F(2, 122) = 490.16, p < 0.001, η<sup>p</sup> 2 = 0.89], suggesting that RTs were faster post- (999.2 ms) rather than pre-exercise (1,056.1 ms), and that RTs followed the trial gradient: homogeneous (587.4 ms) < non-switch (1,136.6 ms) < switch (1,358.9 ms). The interactions of Time × Group [F(2, 61) = 7.34, p = 0.001, η<sup>p</sup> <sup>2</sup> = 0.19], Time × Condition [F(2, 122) = 5.55, p = 0.005, η<sup>p</sup> <sup>2</sup> = 0.08], and Time × Condition × Group [F(4, 122) = 3.69, p = 0.007, η<sup>p</sup> <sup>2</sup> = 0.11] were also significant. Post-hoc analysis for the Time × Condition × Group interaction showed that (1) the open- and closed-skill groups showed significantly faster responses than the control group in the switch trials after the exercise intervention [F(2, 61) = 10.31, p < 0.001; open-skill vs. control: p < 0.001; closed-skill vs. control: p = 0.027], and (2) when post- compared to pre-exercise, the open-skill group responded faster in the non-switch [F(1, 21) = 24.80, p < 0.001] and switch [F(1, 21) = 27.15, p < 0.001] trials. Even when the postexercise cardiorespiratory fitness, social participation, and BMI were controlled for, the results of ANCOVA on the RTs in switch trials still indicated a significant difference in the three groups [F(2, 58) = 7.70, p = 0.001], with post-hoc analysis indicating that the two exercise groups showed significantly faster responses than the control group (open-skill vs. control: p < 0.001; closedskill vs. control: p = 0.027) after the exercise intervention.

In terms of RT switch costs, RM-ANOVA for the generalswitch cost revealed the main effect of Time [F(1, 61) = 15.01, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.20], indicating that the general-switch cost was smaller post- (501.9 ms) rather than pre-exercise (596.5 ms) across the three groups. RM-ANOVA for the specific-switch cost revealed that the interaction of Time × Group [F(2, 61) = 3.35, p = 0.042, η<sup>p</sup> <sup>2</sup> = 0.10] was significant. Post-hoc analysis showed that only the value of the specific-switch cost approached significance [F(1, 21) = 4.33, p = 0.050] post-exercise (149.2 ± 163.0 ms) relative to pre-exercise (227.5 ± 208.1 ms) in the open-skill group.

#### N-back Task

RM–ANOVA conducted on mean RTs for the N-back task only revealed a main effect of Condition [F(2, 92) = 210.94, p < 0.001, ηp <sup>2</sup> = 0.82], with the following gradient: 0-back (586.8 ms) < 1-back (693.5 ms) < 2-back (812.4 ms).

#### P3 Latency

#### Task-Switching Paradigm

As shown in **Figures 3**, **5**, no significant difference was observed with regard to any main effect or interaction in the P3 latency. These results indicate that the P3 latencies did not show obvious changes between three groups when performing the

task-switching paradigm throughout the exercise intervention stage.

#### N-back Task

No significant difference was observed with regard to any main effect or interaction in the P3 latency. These results indicate that the P3 latencies did not show obvious changes when performing the N-back task between three groups throughout the exercise intervention stage.

### P3 Amplitude

#### Task-Switching Paradigm

As shown in **Figures 3**, **5**, RM–ANOVA performed on the P3 amplitudes of the task-switching paradigm showed main effects of Time [F(1, 61) = 30.30, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.33], Condition [F(2, 122) = 39.35, p < 0.001, η 2 <sup>p</sup>= 0.39] and Electrode [F(2, 122) = 15.99, p < 0.001, η 2 <sup>p</sup> = 0.21], with the post-hoc analyses indicating that the P3 amplitudes were larger post- (4.89µV) rather than pre-exercise (3.82µV); the P3 amplitudes in the three conditions had the following gradient: homogeneous (5.85µV) > non-switching (3.99µV) > switching (3.22µV); and significantly greater amplitudes at the Fz (4.73µV) and Cz (4.57µV) electrodes were found as compared to the Pz (3.77µV) electrode. The interactions of Time × Group [F(2, 61) = 11.47, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.27] and Condition × Electrode [F(4, 244) = 13.96, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.19] were also significant. Post-hoc analysis for the Time × Group interaction showed that (1) there were no significant differences among the three groups at baseline [F(2, 61) = 0.04; p = 0.961]; (2) there were significant differences among the three groups after the exercise intervention [F(2, 61) = 6.37; p = 0.003], with post-hoc analysis showing that both open- (5.79µV) and closed-skill (5.38µV) groups showed significantly larger P3 amplitudes as compared to the control group (3.50µV) after the intervention (open-skill vs. control: p = 0.004; closedskill vs. control: p = 0.025); and (3) both the open- and closedskill groups exhibited significantly larger P3 amplitudes [openskill: F(1, 21) = 32.99, p < 0.001; closed-skill: F(1, 20) = 23.34, p < 0.001] across all three conditions and electrodes after the intervention relative to baseline. Even when the post-exercise cardiorespiratory fitness, social participation, and BMI were controlled for, the results of ANCOVA on the averaged P3 amplitudes in the three conditions and three electrodes still indicated significant post-exercise group differences [F(2, 58) = 8.07, p = 0.001], with post-hoc analysis indicating that the two exercise groups showed significantly larger P3 amplitudes than the control group (open-skill vs. control: p < 0.001; closed-skill vs. control: p = 0.003) after the intervention.

#### N-back Task

As illustrated in **Figures 4**, **6**, RM–ANOVA performed on the P3 amplitudes of the N-back task showed the main effects of Time [F(1, 61) = 13.42, p = 0.001, η<sup>p</sup> <sup>2</sup> = 0.18] and Condition [F(2, 122) = 9.85, p < 0.001, η<sup>p</sup> <sup>2</sup> = 0.14], with the post-hoc analyses indicating the P3 amplitudes were larger post- (3.57µV) rather than preexercise (2.27µV); and significantly greater amplitudes at the 0-back (3.31µV) and 1-back (3.27µV) conditions were found as compared to the 2-back (2.19µV) condition. The interaction of Time × Group [F(2, 61) = 3.89, p = 0.026, η<sup>p</sup> <sup>2</sup> = 0.11] was also significant, with post-hoc analysis indicating that (1) there were no significant differences among the three groups at baseline [F(2, 61) = 0.02; p = 0.985]; (2) there were significant differences among the three groups after the exercise intervention [F(2, 61) = 6.09; p = 0.004], with post-hoc analysis showing that both open- (4.02µV) and closed-skill (4.52µV) groups showed significantly

the N-back task at baseline (dashed line) and after exercise intervention (solid line) in the open-skill (red), closed-skill (green), and control groups (blue).

FIGURE 5 | P3 amplitudes (mean ± SE) of three electrodes (Fz, Cz, and Pz) for three conditions (homogeneous, non-switch, and switch) during the task-switching paradigm in the open-skill, closed-skill, and control groups before and after the exercise intervention.

FIGURE 6 | P3 amplitudes (mean ± SE) of three electrodes (Fz, Cz, and Pz) for three conditions (0-, 1-, and 2-back) during the N-back task in the open-skill, closed-skill, and control groups before and after the exercise intervention.

larger P3 amplitudes as compared to the control group (2.17µV) after the intervention (open-skill vs. control: p = 0.032; closedskill vs. control: p = 0.005); and (3) both the open- and closed-skill groups exhibited significantly larger P3 amplitudes [open-skill: F(1, 21) = 7.59, p = 0.012; closed-skill: F(1, 20) = 16.83, p = 0.001] across all three conditions and electrodes after the intervention relative to baseline. Even when the postexercise cardiorespiratory fitness, social participation, and BMI were controlled for, the results of ANCOVA on the averaged P3 amplitudes in the three conditions and three electrodes still indicated significant post-exercise group differences [F(2, 58) = 5.56, p = 0.006], with post-hoc analysis indicating that the two exercise group showed significantly larger P3 amplitudes than the control group (open-skill vs. control: p = 0.016; closed-skill vs. control: p = 0.002).

### DISCUSSION

### Main Findings

The present study aimed to explore the effects of 6-month open- (e.g., table tennis) and closed-skill (e.g., bike riding or brisk walking/jogging) exercise interventions on neurocognitive performance in the elderly when performing the task-switching paradigm and N-back task. The findings showed that, with regard to the neuropsychological performance, after participation in the two types of exercise the elderly participants did not increase their ARs in the task-switching paradigm, but showed significantly faster responses than the control group in the switch trials of the heterogeneous condition after the exercise intervention. In addition, compared to the baseline, regular participation in openskill exercise for 6 months could effectively enhance RTs in the elderly when performing the non-switch and switch trials of the heterogeneous condition, and the specific-switch cost approached significance post- relative to pre-exercise in the openskill group. In terms of the N-back task, although performing the open- and closed-skill exercise for 6 months did not improve RTs in the elderly when performing such a cognitive task, the ARs in the 1-back condition were significantly enhanced after the exercise intervention in both exercise groups, and the beneficial effects on the 2-back condition only emerged in the closedskill group. In terms of electrophysiological performance, both open- and closed-skill groups exhibited significantly larger P3 amplitudes across conditions and electrodes after the exercise intervention relative to baseline when performing not only the task-switching paradigm, but also the N-back task. These post-exercise neurocognitive benefits still remained even when the confounding factors (e.g., cardiorespiratory fitness, social participation, and BMI) were controlled for.

### Neuropsychological Performances

In the present study, the older adults participating in the openskill types of exercise (e.g., table tennis) not only had to follow the rules of the game, but also switch strategies and select relevant sensory information when encountering the various skill levels of the other players within a constantly changing environment (Di Russo et al., 2010). In addition, they needed to continually adapt or switch to more suitable movements/responses to initiate appropriate actions or inhibit inappropriate ones based on the opponent's actions. The capabilities of motoric and cognitive switching are thus facilitated during this type of exercise. Indeed, with regard to neuropsychological performance in the taskswitching paradigm, although no differences in ARs were found among the three groups after the exercise intervention in the present study, only the open-skill group showed faster responses in the non-switch and switch trials of the heterogeneous condition and a lower specific-switch cost after participation in 6-month table tennis exercise training. This result is in line with the previous findings of cross-sectional research (Tsai and Wang, 2015). Since task switching involves stimulus perception and identification, attentional reallocation, task-set updating, response conflict detection, and monitoring processing (Friedman et al., 2008), the elderly subjects participating in the open-skill exercise modes could show greater cognitive flexibility at switching from one task to another. However, both the open and closed-skill groups showed significantly faster responses when compared to the control group in the switch trials of the heterogeneous condition after the exercise intervention, partly supporting the findings of earlier studies which showed that, relative to the control group, the older adults who regularly participated in physical exercise or regular participation in openor closed-skill exercise displayed a generalized reduction in the time efficiency of the central processing of cognitive functions when performing a task switching paradigm (Hillman et al., 2006; Themanson et al., 2006; Dai et al., 2013; Tsai and Wang, 2015). In addition, it is worth pointing out that, after engaging in 6-month open- or closed-skill exercise, no improvement in the generalswitch cost was found in the older adults in the present study, suggesting that the two exercise modes could not facilitate the process of selecting between and coordinating the two competing tasks (Friedman et al., 2008).

Although the older adults participating in the closedskill types of exercise (e.g., bike riding and jogging) in the present study stayed in a predictable and stable environment to perform the related exercise at their own pace (Di Russo et al., 2010), and thus they had a lower cognitive load than seen with the open-skill exercise, their cardiorespiratory fitness was significantly enhanced. This was because, compared to the open-skill exercise, repeatedly performing similar movements coupled with continuously higher HR maintenance could much more effectively improve cardiorespiratory fitness. Previous studies have demonstrated that physical exercise interventions aimed at increasing cardiorespiratory fitness are associated with improvements in the neuropsychological (e.g., response accuracy) and electrophysiological (e.g., P3 and CNV components) performances of working memory in preadolescent children (Pesce et al., 2009; Kamijo et al., 2011; Tsai et al., 2014a) and the elderly (Voss et al., 2013). The potential mechanisms could be that such an exercise mode can increase cerebral blood flow (Seifert and Secher, 2011), cerebral structure (Colcombe et al., 2003), and brain-derived neutrophic factors (Seifert et al., 2010), and regulate hippocampal neurogenesis and synaptic plasticity (Erickson et al., 2009). Given that previous studies have linked cardiorespiratory fitness to enhanced memory in the elderly (Kramer et al., 2001; Voss et al., 2013), and aerobic exercise is associated with increased hippocampal size and function (Erickson et al., 2011), it is not surprising that the closed-skill group in the present study participating in the exercise type that had the greatest impact on cardiorespiratory fitness showed improved ARs on the N-back task involving moderate and high working-memory demands, which require continual processing and updating (Bopp and Verhaeghen, 2005), although there were still no significantly different ARs among the three groups post-exercise. Nevertheless, it should be noted that the open-skill group also exhibited significantly higher ARs on the N-back task involving a moderate working-memory demand (i.e., the 1-back condition) after the exercise intervention, suggesting that the potentially neuropsychological benefit derived from the open-skill exercise on the working memory should not be negated in the elderly. One possible explanation for this is that the effect of improved cardiorespiratory fitness approached significance in the elderly participating in the table tennis intervention in the present study. The results on the relation between the N-back task and exercise modes in the present study were partly in line with Hötting et al.'s (2012) findings that cycling training improved learning and recognition scores in an episodic memory test in middle-aged adults, while stretching/coordination training only improved the learning score, and such beneficial effects could be attributed in an increase in cardiovascular fitness. However, it is worth noting that, when the cofounding factors (also including cardiorespiratory fitness) were controlled for in the present study, the beneficial effect on the AR in the 2-back condition still remained for the closed-skill group relative to the open-skill and control groups. Therefore, the potential mechanisms of the effects (e.g., the increases in the hippocampal size/function and the cerebral blood flow) of the closed-skill exercise on the working memory in the elderly are worth exploring in the future (Erickson et al., 2011).

#### Electrophysiological Indices

P3 latencies did not show obvious changes after 6-month exercise intervention in the two exercise groups, suggesting that the perceptual/central processing could not be facilitated by the open- and closed-skill exercise in the elderly when performing the two cognitive tasks. In line with a previous study reporting a larger P3 amplitude for the active rather than for the sedentary elderly when performing the task switching paradigm (Hillman et al., 2006), the older adults participating in the 6-month openand closed-skill exercises programs in the present study could effectively increase their P3 amplitudes and so have larger P3 amplitudes across all conditions relative to the control group, suggesting that the two physical exercise modes could facilitate the attentional set that makes it possible to better evaluate the stimulus in either of the two tasks. However, the openand closed-skill groups exhibited similar benefits on the neural processes at work in processing the current task-switching paradigm. These results are somewhat inconsistent with those of a previous cross-sectional study (Tsai and Wang, 2015), in which the older adults regularly participating in open-skill exercise (e.g., table tennis and badminton) showed a significantly larger P3 amplitude in the switch condition when performing the task-switching paradigm compared to their counterparts participating in closed-skill exercise. The lack of consistency in these results may be attributable to the inherently better taskswitching capacity that may encourage some older adults to choose an open-skill exercise mode (Snowden et al., 2011; Tsai and Wang, 2015). The current 6-month exercise intervention study seems to clarify that both open- and closed-skill exercise modes could produce similar electrophysiological benefits across all conditions when older adults perform the task-switching paradigm. However, even though lower P3-and-RT correlation was found in the older individuals (Pfefferbaum et al., 1980), the distinctive effects of open-skill exercise on neuropsychological performance (i.e., better exercise-training-induced effects on specific-switch cost and RTs in the heterogeneous conditions) in the elderly, as mentioned above, cannot be ignored.

Similarly, relative to the open-skill exercise, although a closedskilled exercise intervention could have more neuropsychological benefits (i.e., significantly increasing ARs under the high load condition) on the elderly when performing the N-back task, the effects of greater P3 amplitudes were comparable in the two exercise groups in the present study. This suggests that not only closed- but also open-skill exercise could facilitate the memoryrelated neural processing which is involved in categorizing incoming information and updating the context of the working memory (e.g., encoding, rehearsal, recognition, and retrieval), due to the greater efficiency by which cognitive resources are allocated (Duncan-Johnson and Donchin, 1977; Donchin and Coles, 1988; Rugg, 1995). Although there were no significant between-group AR differences post-exercise among the three groups, the increase in exercise-induced P3 amplitudes observed in the two exercise groups after the 6-month interventions, as compared with the figures seen before training, could reveal that they allocated more resources for target classification and evaluation, which might result in higher ARs in the working memory task since larger overall P3 amplitudes during the Nback could be associated with better task performance in older adults (Tusch et al., 2016). However, the P3 amplitude could be influenced by a greater latency jitter of P3 in the high than in the low memory condition (Kok, 2001), and regular exercise could change P3 latency in the older adults (see review, Hillman et al., 2003), the potential response jitter thus needs to be clarified in further investigations, since different conditions in the two cognitive tasks are involved in different cognitive loads in the present study.

It is worth noting that there was no significant interaction of Time × Group× Electrode in the present study, suggesting that 6-month open- and closed-skill exercise interventions induced similar electrophysiological effects (i.e., increased P3 amplitudes) from the frontal to parietal cortices. However, previous studies have suggested that different physical exercises could affect the brain in different ways (Erickson et al., 2011; Voelcker-Rehage et al., 2011; Burrel, 2015). For example, Erickson et al. (2011) found that 1-year of aerobic exercise training could effectively increase the size of the anterior hippocampus in the elderly, but not the caudate nucleus and thalamus volumes, accompanied by improved memory function. Moreover, such effects were not shown in the individuals performing a stretching and toning program. Voelcker-Rehage et al.'s (2011) longitudinal research reported that although the older adults participating in cardiorespiratory training aimed at enhancing cardiorespiratory fitness or coordination training to increase fine- and gross-motor body coordination could improve their executive functioning and perceptual speed, although with different effects on speed and accuracy, the two types of exercise had different impacts on neural activity, with an increased activation of the sensorimotor network and less prefrontal activation in the cardiovasculartraining group, and increased activation in the visual-spatial network (e.g., right inferior frontal gyrus, superior parietal cortex, thalamus, and caudate body) in the coordinationtraining group. In addition, animal studies showed that regular cardiovascular training in rats did not increase the number of synapses, but could increase the density of capillaries (Black et al., 1990) and shorten the diffusion distance from the blood vessels in the molecular layer of the paramedian lobule (Isaacs et al., 1992). In contrast, regular complex motor-skill training in the rats did not increase the density of capillaries, but could substantially increase the number of synapses per Purkinje neuron and blood vessels, thus maintaining the diffusion distance (Isaacs et al., 1992). These findings from both human and animal studies suggest that different types of exercise intervention (e.g., openvs. closed-skill) could produce distinct training effects on the brain tissues and neural activations. However, in the present study the effects of exercise interventions on neural activity in the frontal-to-parietal cortices seem to be comprehensively covered. In fact, in terms of aging, cognition, and brain function, the phenomenon of dedifferentiation, which characterizes a simple marker of cognitive decline, is often found in the elderly. That is, the regions of the brain that are recruited to perform a variety of cognitive tasks are less specific among older rather than younger adults (Cabeza, 2001), possibly due to additional cortical areas being recruited to compensate for losses in neural efficiency in the former. Therefore, while increased prefrontal activation to compensate for processing impairment, particularly in posterior areas, is a consequence of age-related structural and functional declines in various brain regions (Greenwood, 2007), the findings of the present study suggest that both open- and closed-skill exercise modes could not only facilitate anterior cortical processing efficiency, but also compensate for neural processing impairment in the posterior cortical areas due to cognitive aging. In addition, whether the improvement in cognitive function that results from the open- and closedskill exercise would produce more distinctive effects in different age groups is one area for future works to address. It is likely that longitudinal assessments of the effects of open- and closed exercise interventions on electrophysiological performance in young adults would help clarify the benefits of different types of exercise with regard to cognitive functions.

Colcombe and Kramer's (2003) meta-analytic study suggested that aerobic fitness training increases cognitive performance by 0.5 SD on average, regardless of the exercise training method and types of cognitive task, especially in the executive-control processes. Therefore, although the distinct benefits of openand closed-skill exercises on both working memory and task switching performance were found in the present research, cardiorespiratory fitness could play an important role and have a beneficial influence on these two types of cognitive functions in older adults (Netz et al., 2011; Voelcker-Rehage et al., 2011; Wang et al., 2016). Additionally, in light of evidence that depression, education, social stimulation, and BMI could also mediate the exercise-cognition association in the elderly (Miller et al., 2012; Ronan et al., 2016; Tomioka et al., 2016), adequate controls to take into account the confounding factors that participants in the intervention groups are impacted by, in addition to physical exercise, need to be considered. Indeed, in the present study the two groups showed different levels of improvements in cardiorespiratory fitness, social participation, and BMI after participation in either open- or closed-skilled exercise. However, when including these improved confounding factors as covariates in the analysis of the improved neurocognitive performance after the exercise intervention, the difference among the three groups remained significant, showing that both types of exercise modes do indeed lead to improved neurocognitive performance.

#### Strengths and Weaknesses

Although a number of confounding factors which could mediate the exercise-cognition association were rigorously controlled for in the current study, some potential limitations of this work need to be addressed. First, since changes in the brain and neurocognitive performance are not always proportional to each other, a decrease in brain size and plasticity that results in cognitive changes is associated with normal aging (Peters, 2006). Further MRI/fMRI studies to explore the changes in the sizes/densities of the brain tissues and in patterns of brain activation would be helpful to understand the complex relationship between different exercise modes and neurocognitive performance, and to determine the exact mechanism of cognitive enhancement in the elderly. Second, some exercise items that do not aim to increase cardiorespiratory fitness are also included in the category of closed-skill exercise, such as resistance exercise and yoga. It would thus be informative in the future to compare the neurocognitive performances of elderly adults on different cognitive tasks using these closed-skill exercise items, in order to more clearly explain the effects of the closed-skill exercise intervention on executive functioning, especially in the memory domains, in this population. Lastly, the ERP recordings in the present study were referred to a linked mastoid reference, and this is not an ideal zero reference, and thus the task-related effect could have been impacted by this, and a potential bias produced (Yao, 2001). However, the adopted reference did not change the topography map of the P300 components (Yang et al., 2017), and the P3 amplitudes decreased significantly with increasing cognitive loads (Kok, 2001) in the present study. These findings suggest that the electrophysiological findings from such a reference site are still reliable. However, further EEG-fMRI studies aiming to explore the effects of exercise interventions might consider applying the Reference Electrode Standardization Technique (REST) (Yao, 2001; Yao et al., 2005) as the reference method for ERP data (Yang et al., 2017).

#### CONCLUSIONS

Extending earlier cross-sectional studies on aging that relied on volunteer participants, which could inevitably include some selection bias with an overrepresentation of individuals with inherently higher executive functioning, the present study of a 6-month exercise intervention confirmed recent cross-sectional results showing that open- and closed-skill exercise modes relate differently to various forms of executive functioning (e.g., task switching and working memory) in relation to cognitive aging in older adults. However, the current findings more clearly revealed that both open- and closed-skill exercise could effectively enhance overall brain cortical activity. These beneficial effects of the two exercise interventions on the neuropsychological and electrophysiological performances in the elderly remained unchanged after statistical adjustment for improved cardiorespiratory fitness, social participation, and BMI. Although exercise is a simple and healthy lifestyle factor that has

#### REFERENCES


been proposed to be protective against neurocognitive declines during aging, possibly even retarding or reversing age-associated degeneration in the brain, different exercise modes seem to have different effects on various forms of executive function in the elderly.

#### AUTHOR CONTRIBUTIONS

CT designed the study, wrote the protocol, and the first draft of the manuscript. CP analyzed the data. FC and YT helped collect data.

#### ACKNOWLEDGMENTS

This research was supported by a grant from the Ministry of Science and Technology in Taiwan (MOST 102-2628-H-006-003- MY3). The authors are also grateful to the participants who gave their precious time to facilitate the work reported here.


players and swimmers with sedentary controls. J. Mot. Behav. 45, 55–63. doi: 10.1080/00222895.2012.740522


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Tsai, Pan, Chen and Tseng. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Acute Stress Affects the Expression of Hippocampal Mu Oscillations in an Age-Dependent Manner

Samir Takillah1,2,3,4,5 \* † , Jérémie Naudé1† , Steve Didienne1† , Claude Sebban2,3 , Brigitte Decros 2,3 , Esther Schenker <sup>6</sup> , Michael Spedding<sup>7</sup> , Alexandre Mourot <sup>1</sup> , Jean Mariani 2,3‡ and Philippe Faure<sup>1</sup> \* ‡

<sup>1</sup>Team Neurophysiology and Behavior, Institut de Biologie Paris Seine (IBPS), UMR 8246 Neuroscience Paris Seine (NPS), Sorbonne Universités, Université Pierre et Marie Curie (UPMC), CNRS, INSERM, U1130, Paris, France, <sup>2</sup>Team Brain Development, Repair and Ageing, Institut de Biologie Paris Seine (IBPS), UMR 8256 Biological Adaptation and Ageing (B2A), Sorbonne Universités, Université Pierre et Marie Curie (UPMC), CNRS, Paris, France, <sup>3</sup>APHP Hôpital Charles Foix, DHU Fast, Institut de la Longévité, Ivry-sur-Seine, France, <sup>4</sup>Département Neurosciences et Contraintes Opérationnelles, Institut de Recherche Biomédicale des Armées (IRBA), Unité Fatigue et Vigilance, Brétigny-sur-Orge, France, <sup>5</sup>EA7330 VIFASOM, Université Paris Descartes, Paris, France, <sup>6</sup>Neuroscience Drug Discovery Unit, Institut de Recherches Servier, Croissy-sur-Seine, France, <sup>7</sup>Spedding Research Solutions SARL, Le Vésinet, France

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Imre Vida, Charité Universitätsmedizin Berlin, Germany William Griffith, Texas A&M University, United States

#### \*Correspondence:

Samir Takillah samir.takillah@gmail.com Philippe Faure phfaure@gmail.com

†These authors have contributed equally to this work. ‡These authors have jointly directed this work.

Received: 29 May 2017 Accepted: 29 August 2017 Published: 21 September 2017

#### Citation:

Takillah S, Naudé J, Didienne S, Sebban C, Decros B, Schenker E, Spedding M, Mourot A, Mariani J and Faure P (2017) Acute Stress Affects the Expression of Hippocampal Mu Oscillations in an Age-Dependent Manner. Front. Aging Neurosci. 9:295. doi: 10.3389/fnagi.2017.00295 Anxiolytic drugs are widely used in the elderly, a population particularly sensitive to stress. Stress, aging and anxiolytics all affect low-frequency oscillations in the hippocampus and prefrontal cortex (PFC) independently, but the interactions between these factors remain unclear. Here, we compared the effects of stress (elevated platform, EP) and anxiolytics (diazepam, DZP) on extracellular field potentials (EFP) in the PFC, parietal cortex and hippocampus (dorsal and ventral parts) of adult (8 months) and aged (18 months) Wistar rats. A potential source of confusion in the experimental studies in rodents comes from locomotion-related theta (6–12 Hz) oscillations, which may overshadow the direct effects of anxiety on low-frequency and especially on the high-amplitude oscillations in the Mu range (7–12 Hz), related to arousal. Animals were restrained to avoid any confound and isolate the direct effects of stress from theta oscillations related to stress-induced locomotion. We identified transient, high-amplitude oscillations in the 7–12 Hz range ("Mu-bursts") in the PFC, parietal cortex and only in the dorsal part of hippocampus. At rest, aged rats displayed more Mu-bursts than adults. Stress acted differently on Mu-bursts depending on age: it increases vs. decreases burst, in adult and aged animals, respectively. In contrast DZP (1 mg/kg) acted the same way in stressed adult and age animal: it decreased the occurrence of Mu-bursts, as well as their co-occurrence. This is consistent with DZP acting as a positive allosteric modulator of GABA<sup>A</sup> receptors, which globally potentiates inhibition and has anxiolytic effects. Overall, the effect of benzodiazepines on stressed animals was to restore Mu burst activity in adults but to strongly diminish them in aged rats. This work suggests Mu-bursts as a neural marker to study the impact of stress and DZP on age.

Keywords: aging, stress, hippocampus, Mu-rhythm, synchronized oscillation

## INTRODUCTION

Stress is a set of physiological responses triggered by an aversive situation (Kim and Diamond, 2002). It is generally associated with anxiety disorder (Pêgo et al., 2008; Bessa et al., 2009), a state characterized by ''hypervigilance'' (i.e., a high level of arousal) and sustained alertness for potential threats (Sylvers et al., 2011; Adhikari, 2014; Tovote et al., 2015). Stress also promotes avoidance and is often associated with fear generalization (Duvarci et al., 2009; Davis et al., 2010). A key point is that reaction to stress is strongly age-dependent, with elderly people enduring stressful situations more frequently and reacting to pressure more profoundly (Prenderville et al., 2015). In particular, aging may induce sustained stress reactions (Wikinski et al., 2001; Leite-Almeida et al., 2009; Pietrelli et al., 2012). The neurological consequences of stress and age appear furthermore strikingly similar: both are associated with alterations of neuronal plasticity and increased risk of brain disorders (Morrison and Baxter, 2012; Prenderville et al., 2015). These similarities suggest that age itself may act as a stressor factor (Buechel et al., 2014). This link between age and stress is highlighted by an altered brain plasticity in elderly after exposure to new-onset stress (Morrison and Baxter, 2012; Lindenberger, 2014; Prenderville et al., 2015), and that aged individuals often cope with stressful situations (Barrientos et al., 2012; Buechel et al., 2014).

A proper understanding of the interactions between age and stress is crucial when considering the wide use of anxiolytic drugs such as benzodiazepines in the elderly (Gleason et al., 1998; Kirby et al., 1999). Benzodiazepines like diazepam (DZP) have a number of clinically approved uses (reduction of sleep latency, muscle relaxation, anxiolysis. . .) but also have unwanted side effects, in particular a decreased alertness, anterograde amnesia, dependence and addiction (Tan et al., 2011). Benzodiazepines influence behavioral activity and, accordingly, neural oscillations in cortical circuits. DZP is a positive allosteric modulator of the GABA<sup>A</sup> receptor that acts by potentiating the natural ligand GABA (Tan et al., 2011). At the synaptic level, DZP enhances the amplitude and duration of inhibitory postsynaptic events, and thus increases phasic inhibition (Scheffzük et al., 2013). At the network level, this potentiation of inhibition results in characteristic alterations of rhythmic activity patterns (Dimpfel et al., 1988; van Lier et al., 2004; Botta et al., 2015).

The organized activity of neural networks, as reflected in multi-neuronal, extracellular fields potential (EFP) recordings, frequently presents a rhythmic quality. In humans, the central 7–12 Hz rhythm, also called Mu-rhythm in the sensorimotor/parietal area, reflects an idling state (Gastaut et al., 1965). This oscillatory index, characterized by bursts of oscillations of high amplitude, has been mostly observed in somatosensory cortex and is known to be modulated by attention (Wiest and Nicolelis, 2003; Fontanini and Katz, 2005; Tort et al., 2010; Coll et al., 2017). However, other studies found Mu oscillations in many fronto-parietal regions (Sakata et al., 2005; Marini et al., 2008; Tort et al., 2010) and even in the cerebellum (Hartmann and Bower, 1998). In rats, the role of 7–12 Hz cortical rhythm remains a topic of intense debate (Nicolelis et al., 1995; Nicolelis and Fanselow, 2002; Wiest and Nicolelis, 2003; Shaw, 2004, 2007; Fontanini and Katz, 2005). It has been proposed to represent a dynamical filter for detecting weak or novel tactile stimuli (Wiest and Nicolelis, 2003) or a withdrawal state (i.e., with internally-directed attention; Fontanini and Katz, 2005). This rhythm is associated with whisker twitching (WT), during which rats stand still and twitch their whiskers in small-amplitude movements, inducing an increase of sensitivity to weak sensory signals (Nicolelis et al., 1995; Fanselow et al., 2001). However, it remains unknown whether stressful situations can switch the vigilance state towards such quiet alertness, reflected by an increased occurrence of the 7–12 Hz oscillations.

In this study, we recorded EFP in the dorsal and ventral hippocampus (d/v-HPC), prefrontal cortex (PFC) and parietal associative cortex (PAR, formerly called sensorimotor cortex in the somatosensory system) in adult and aged rats, at rest (control) and on an elevated platform (EP; stress condition), with systemic injections of saline or DZP, in order to assess the interactions between stress, age and anxiolytics on alertness-related cortical rhythms. We show that stress increased the 7–12 Hz rhythms of the PFC, PAR and HPC in adult rats but that, inversely, it decreased these rhythms in aged rats. Furthermore, we reveal an interaction between DZP, age and stress that may bear important implications for the anxiolytic effects of DZP in the elderly.

### MATERIALS AND METHODS

### Animals Care, Housing Conditions and Ethics Statement

All experiments were approved by the Ethic committee CAPSUD/N◦ 26 (Ministère de l'Enseignement Supérieur et de la Recherche, France) and conducted in agreement with institutional guidelines and in compliance with national and European laws and policies (Project no. 01272.01). Experiments were performed on 17 adults (8 months) and 10 aged (strictly speaking, late middle-aged rats, 18 months (Prenderville et al., 2015), but referred to here as ''aged'') male Wistar rats from Janvier Laboratories. The animals were singly housed, in a 12 h light/dark cycle and temperature-controlled room (22 ± 2 ◦C) with food and water available ad libitum.

### In Vivo Electrophysiological Recordings

Rats were anesthetized with ketamine and xylazine and placed in a stereotaxic frame. Anesthesia was maintained with inhalation of a mixture of isoflurane 3% and oxygen. Bipolar stainless steel electrodes were chronically implanted bilaterally in each rat into the infralimbic/prelimbic of PFC, the PAR, the CA1 dHPC and the CA1 vHPC. Monopolar ground electrodes were laid over the cortical layer of the cerebellum and the olfactory bulb. Electrodes were connected to an electrode interface board (QuickClip Connect EIB-16-QC-H, Neuralynx) and dental acrylic was used to fix them to the skull during the surgery. Six bipolar electrodes (the distance between the recording tips and the reference tips were 0.7 mm for the PFC or PAR and 0.5 mm for the different part of the HPC) were implanted through burr holes targeting the following coordinates from Bregma: depth over the cortex 3.8 mm, AP +3 mm, ML ± 0.8 mm for the PFC; depth over the cortex 0.7 mm, AP −4 mm, ML ± 4 mm for PAR; depth 2.8 mm, AP −3.6 mm, ML ± 2.2 mm for the dHPC; depth over the cortex 5.3 mm, AP −6.3 mm, ML ± 5.6 mm for the vHPC. The recording tips were located in the deep layers and the local reference tips at the surface of the corresponding cortices.

Finally, to reduce electrical noise, two grounds (monopolar electrodes) were implanted over the cortex at the following coordinates from Bregma, AP +6.7 mm, ML ± 1 mm for the olfactory bulb; AP −11 mm, ML ± 1 mm for the cerebellum (Paxinos and Watson, 2006).

After surgery, an antiseptic (Povidone-iodine solution) and a local anesthetic (lidocaine ointment) were applied in all areas where the scalp had been incised. Animals were permitted to recover until regaining pre-surgery body weight.

### Protocol Design

Experimental setting is based on previous studies (Sebban et al., 1999a,b, 2002). EFP obtained from the dHPC generally exhibit prominent theta-frequency oscillations. Two types of hippocampal theta activity were described in the rat. One type was termed atropine-sensitive theta, since it was abolished by the administration of atropine. Atropine-sensitive theta occurred during immobility in rodents in the normal state. The other type of theta was termed atropine-resistant, since it was not sensitive to treatment with atropine but was abolished by locomotors activities or anesthetics. Atropine-sensitive theta became known as type II (immobility-related) theta. Atropineresistant theta became known as type I theta, since it occurred during Type I (voluntary) motor behaviors, such as walking, rearing and postural adjustments. These oscillations can be modulated by stress, in the case of type II theta, related to cognition (Hsiao et al., 2012) but also by other behavioral variables, in particular by locomotion, in the case of type I theta (Vanderwolf, 1969; Buzsáki, 2002). In the theta frequency range, Mu oscillations are thought to reflect the vigilance state (Kramis et al., 1975; Sakata et al., 2005; Popa et al., 2010) but can be affected by sensory inputs (Fanselow et al., 2001; Tort et al., 2010; Aitake et al., 2011; Fries, 2015). We thus aimed at eliminating potential confounds related to type I theta and sensory influences on Mu oscillations. To minimize the EFP modulations induced by spontaneous locomotor activity (type I theta), rats were restrained in a resting-state environment box and were gradually accustomed to be restricted in their movements, 10 min the first day and with an additional 10 min every day (D1: 10 min, D2: 20 min, D3: 30 min. . .), until the recording time was reached. This procedure required approximatively 10–14 days for each animal to remain quiet. Furthermore, a cold light source of 100 lux was applied at a distance of 10 cm in front of the rat's nose to keep the animal still, with the head up (the animal voluntarily kept the head up due to the light source) and wide-open eyes (Sebban et al., 1999a,b). During recordings, rats were isolated into a large, electrically and acoustically insulated chamber, in a specific recording room, to eliminate as far as possible to any sensory input that might affect the Mu oscillations.

Recordings were obtained using a Digital Lynx SX (Neuralynx). Sixty minutes baseline EFP recordings were obtained while the animals remained relatively still. The effect of stress was evaluated 1 day later by placing rats on an EP (Xu et al., 1998; Rocher et al., 2004) of small size in the same experimental room and condition. To avoid any bias linked to circadian variation of EFP (Sebban et al., 2002) both recording sessions took place exactly at the same hour of the day. Adult and aged rats were examined simultaneously excepted for five additional adult rats. We checked that the restraining procedure do not produce stress symptoms by itself: no attempts to escape or notable stress reactions were observed (i.e., defecation, urination, freezing) at rest at the end of the habituation procedure, contrarily to the stress condition. Throughout all the recording, rats showed quick reactions when probed by slight sound stimulation, by turning their head toward the sound. We did not observe any difference in the propensity to detect the sound, nor in the reaction times between adult and aged rats, but no quantitative comparison of sensory functions between adult and aged rats were performed.

#### Neurophysiological Data Analysis

Data were analyzed using Matlab (Matworksr) built-in and custom-written codes. All EFP: (i) were acquired at 1000 Hz and offline band-pass filtered between 0.1 Hz and 100 Hz with a zero-phase shift filter function (zero-phase digital filtering filtfilt function); (ii) detrended using local linear regression (locdetrend function from the Chronux toolbox; Bokil et al., 2010): window size 1 s, overlap 0.5 s) to remove slow drifts; and (iii) notch-filtered (iirnotch function), with the notch located at 50 Hz to remove any possible power line noise. EFP signal was expressed in z-score units. The z-score normalization used the mean and the standard deviation from the baseline (entire rest session) of each electrode. Multitaper spectrogram method from the Chronux toolbox (Bokil et al., 2010) with time-bandwidth product of 5 and 10 slepian sequences of orthogonal data tapers was used to calculate power spectral density (PSD) of the EFP data, using a window size of 5 s, with 2 s overlap. PSD was averaged over two similar brain regions (right and left hemisphere) in each animal, for each frequency and time bin. The multitaper coherogram method was used to calculate the coherence (normalized spectral covariance) between the EFP from two structures with time-bandwidth product of 30 and 60 slepian sequences of orthogonal data tapers, using a windows size of 30 s without overlap. The signal was bandpass-filtered to extract Mu-oscillations by applying a 7–12 Hz finite impulse response (FIR) bandpass with zero-phase shift filter function (filtfilt function).

Instantaneous amplitude and phase from the EFP were obtained using a continuous Morlet wavelet transform, with matcher filter construct parameters: center frequency = 1 and bandwidth = 2, for the 0.1–30 Hz range. Wavelet coherence was computed by smoothing the product of the two relevant wavelet transforms over time (window for time smoothing = 0.6 s) and over scale (pseudo-frequency) steps (window for scale smoothing = 3 Hz).

We measured the phase locking value (as an index for synchrony) between EFP in the PFC and dHPC from the wavelet coherence, using the distribution of the phase differences between EFPs (Lachaux et al., 2002). This measurement is a normalized index of the stability of phase shifts which varies between 0 (random distribution, no phase synchrony) and 1 (perfect phase synchrony locking; Le Van Quyen et al., 2001).

The international classification of the borders between the different frequency bands was arbitrarily drawn (Delta, 0.5–4 Hz; Theta 4–8 Hz; Alpha, 8–12 Hz; Beta, 12–30 Hz; Gamma >30 Hz). In the freely moving rodent, hippocampal theta should be designated theta-alpha, according to the committee's recommendation, since theta varies between 6–7 Hz and 12 Hz (Vanderwolf, 1969; Winson, 1978; Bland, 1986; Buzsáki, 2002; Yamamoto et al., 2014). Hence, we did not separate alpha from theta and considered the whole 7–12 Hz range for statistical testing. A great variety of rhythms in the same 7–12 Hz range have been described in the thalamocortical system, including Mu rhythm (8–12 Hz, sensorimotor system), together with alpha waves (8–12 Hz, visual system), tau rhythms (8–12 Hz, auditory system) or sleep spindles (10–20 Hz). Hence, in rodents, Mu and Theta oscillations share the same frequency component. However, they differ in their voltage intensity i.e., Mu exhibiting higher amplitude of oscillations. These particularly large amplitude oscillations makes sometimes called high-voltage spindle or spike-and-wave discharge (Robinson and Gilmore, 1980; Inoue et al., 1990; Shaw, 2007). We thus separated Mu oscillations from type II Theta rhythm by an amplitude threshold, with the highest voltage events classified as Mu and residual activity considered as theta, with the following procedure. We measured the amplitude in the 7–12 Hz range from the area under the curve (AUC; trapz function) of the complex Morlet wavelet transform in this range. Finally, Mu-bursts were extracted by: (i) smoothing the filtered power of 7–12 Hz wavelet transform with a Kalman filter; and (ii) using a double threshold (for the beginning and end of a burst) and a persistence greater than 3 s, i.e., Mu-burst started when the EFP was above the upper thresholds for more than 3 s, and ended after switching below the lowest threshold for a duration greater than 3 s.

### Drugs Preparation and Pharmacological Protocol

DZP (Sigma-Aldrich), a standard anxiolytic in humans and rodents (van Lier et al., 2004; Scheffzük et al., 2013), was prepared in a 10% 2-hydroxypropyl-β-cyclodextrin solution and saline. The same solvent was used as vehicle in control experiments. DZP was injected intra-peritoneally at a single dose of 1 mg/kg, which is known to exert an anxiolytic effect. Higher doses were not tested as they may induce sedation (Wikinski et al., 2001; van Lier et al., 2004). Recordings started immediately after DZP administration. Five days before experiments, rats were daily prepared for intra-peritoneal administration by exerting a light pressure on the body with a syringe. The effect of DZP vs. vehicle administration was evaluated in the two conditions, i.e., at rest and under stress (in the EP), for 145 min. Moreover, to avoid the bias linked to circadian variation of EFP, both recording sessions took place exactly at the same hour. The rats were randomly assigned to a given treatment according to a withinsubject ''Latin square''. A free week was imposed between two interventions.

### Statistical Analysis

All datasets were tested for normality using Shapiro-Wilk and Lilliefors tests. For statistical comparison, three bands of the PSD were analyzed for each structure: 0.1–4 Hz (Delta), 7–12 Hz (Mu) and 12–30 Hz (Beta). No statistical methods were used to predetermine sample sizes, but our sample sizes are similar to those reported in previous publications.

For single comparisons, paired-sample T-tests (normally distributed data) or non-parametric Wilcoxon's signed rank tests (non-Gaussian distribution or for small samples) were used to compare PSD and coherence estimates of the Mu-rhythm from the same animals. Two-sample T-tests or nonparametric Wilcoxon's rank-sum tests were used to compare PSD and coherence estimates between adult and aged groups.

For multiple comparisons of normally distributed data, we used one-way ANOVA (e.g., adults/aged rats) or two-way ANOVA (e.g., with frequency bands and stress/rest as factors). For data with non-Gaussian distribution or for small samples, non-parametric tests were used: Kruskal-Wallis test (instead of one-way ANOVA) and Friedman test (instead of two-way ANOVA). Post hoc tests were performed to identify which frequency bands differed in the spectral analysis and which groups differed in the wavelet-transform analysis. We did not compare different frequency bands from different conditions (e.g., theta in adults with delta in aged rats). We used respectively the stepwise algorithms Holm-Bonferroni to correct family-wise error rate (i.e., potential interferences during multiple comparisons) by ordering the p-values and adjusting the significant level α and Tukey's honest significant difference (HSD) criterion.

Results are expressed as mean ± standard error of the mean (SEM) and were listed in table. SEM intervals were calculated through a jackknife method (Bokil et al., 2010). The level of statistical significance was set at 5% for all tests (two-sided). Significance levels are shown in figures with one to three asterisks ( <sup>∗</sup>p < 0.05, ∗∗p < 0.005, ∗∗∗p < 0.0005).

### RESULTS

#### Electrophysiological Signatures of Stress

Electrophysiological signatures of stress-induced activities in the PFC and HPC were evaluated by comparing PSD of EFP obtained for adult rats in a state of quiet wakefulness at rest,


5 September 2017 | Volume 9 | Article 295

(2) α/2 < 0.025, (3) α < 0.05.


Significant levels are presented with one to three asterisks (

∗P < 0.05, ∗∗P < 0.005, ∗∗∗P <

0.0005).


(2) α/2 < 0.025, (3) α < 0.05.

or in a stressful condition when animals were placed on an EP (**Figures 1A,B**). Restraining rats allowed to avoid theta oscillations related to locomotion (type I theta, prominent in the dHPC, see ''Materials and Methods'' Section) and thus to correctly evaluate the impact of acute stress on EFP frequency content, and in particular type II theta and Mu oscillations, which are both related to arousal and vigilance (Kramis et al., 1975; Shaw, 2004; Tort et al., 2010; Sobolewski et al., 2011; Wells et al., 2013).

Comparison of EFP content in specific frequency bands (i.e., delta (0.1–4 Hz), Mu (7–12 Hz) and beta (12–30 Hz), see **Table 1**) showed a significant increase in the Mu and Beta ranges of the dHPC EFP (n = 13, χ <sup>2</sup> = 4.36; <sup>∗</sup>P = 0.0369 Friedman's test followed by Wilcoxon's signed rank-test ∗∗Pmu = 7.3242e-4, and paired sample T-test <sup>∗</sup>Pbeta = 0.0188, **Figure 1C**, middle). In contrast, the PSD of vHPC (n = 12) was globally reduced following the stress procedure, in particular in the Mu range (F(1,66) = 7.04; <sup>∗</sup>P = 0.01 Two-way ANOVA test followed by Wilcoxon's signed-rank test ∗∗Pmu = 4.8828e-4, **Figure 1C**, right).

These differences in EFP co-occurred with a modification of the functional connectivity between brain areas. This functional connectivity was evaluated using pairwise coherences (i.e., co-modulation in amplitude and phase-shift stability between two structures) between the PFC, dHPC and vHPC (**Figure 1D**). At rest, coherence in the delta/Mu ranges was relatively higher between the ventral and dorsal HPC than between the PFC and either part of the hippocampus (**Figure 1D**). The stress protocol induced only a clear increase in coherence between the PFC and dHPC in the Mu and beta ranges (**Figure 1D** left (χ <sup>2</sup> = 4.61; <sup>∗</sup>P = 0.0318 Friedman's test followed by paired T-test: <sup>∗</sup>Pmu = 0.0076 and Wilcoxon's signed-rank test: <sup>∗</sup>Pbeta = 0.0085). Eliminating potential locomotion-related effects thus revealed that stress increased both the amplitude of the Mu rhythm in the dHPC and its coherence with mPFC in adult rats.

### Stress-Induced Modifications of Mu Rhythms Are Composed of Mu-Bursts of Oscillations, Associated with Whisker Twitching and Alertness

We next investigated whether the stress-induced increase in the dHPC Mu rhythm was related to a state of alertness, and thus could reflect an effect of stress on vigilance. Indeed, the time-dependent spectrogram of the dHPC EFP revealed transient bursts of oscillations in the Mu range (**Figure 2A**), thereafter called Mu-bursts, which were observed in every rat. While they rarely occurred in adult rats at rest (**Figure 2A**, left), their occurrence increased under the stress condition (**Figure 2A**, right). Mu-bursts were almost systematically associated with an exploratory behavior of ''WT'' (**Figure 2B**), i.e., an alert state where rats are still, keep their eyes open, and twitch their whiskers in rhythmic, smallamplitude movements (Fanselow et al., 2001; Sobolewski et al., 2011). This contrasted with the usual behavioral pattern at rest, where rats moved their head left and right, without rhythmic whisker movements. A careful examination of the time-resolved power (of dHPC EFP) obtained from the Morlet wavelet transforms revealed that Mu-bursts were composed of two main frequency contents: one between 7 Hz and 12 Hz, and one at higher frequencies possibly reflecting a ''biological harmonic'' (**Figure 2C**). PSD obtained from wavelets transforms of the dHPC EFP in the 7–12 Hz range at rest followed a unimodal distribution (**Figure 2D**, black curve), as commonly found throughout cortices (Roberts et al., 2015). This type of distribution indicates that synchronous events, i.e., high-amplitude oscillations, were irregularly interspersed with smaller-sized events. Nevertheless, in the stressful situation, another peak appeared in this distribution for large PSD values, while the rest of the distribution was unchanged (**Figure 2D**, blue). Hence Mu-bursts did not reflect an overall increase in oscillation amplitude, which would have resulted in a rightward shift of the distribution. Rather, they constituted discrete events that were clearly distinct from baseline oscillations and that co-occurred with WT. Furthermore, analysis of the AUC of the wavelet PSD confirmed an increase in the 7–12 Hz oscillations under stress condition (∗∗∗p = 5e-4, Wilcoxon' sign rank test **Figure 2E**). Overall the increase in the number of Mu-bursts recapitulated the increase in dHPC Mu band induced by stress.

### Stress-Induced Mu-Bursts Co-Occur in the dHPC and PFC

We next assessed whether the Mu-bursts observed in the HPC were correlated with similar activity in the PFC. Mu-burst were indeed detected both in the PFC, dHPC (**Figure 3A**, leftmiddle) and in the parietal cortex (**Figure 4**), but not in the vHPC. Coherence between the dHPC and PFC was maximal during the Mu-bursts (**Figure 3A**, right), which may explain why coherence increases during stress (see **Figure 1D**, left). Individual detection of Mu-bursts (see ''Materials and Methods'' Section, **Figure 3B**) indicated that they occurred more often in the PFC than in the dHPC, both at rest and under stress (**Figure 3C**, top). At rest, about half (56% ± 16, mean ± SEM) of the dHPC Mu-bursts appeared concomitantly in the PFC, while one fourth (26% ± 10, mean ± SEM) of the PFC Mu-burst were concomitantly detected in the dHPC (**Figure 3C**, top left). Hence, Mu-bursts could occur independently in these two structures. In the stress condition, the total number of Mu-bursts increased (**Figure 3C**, top right) in the dHPC (∗p = 0.0156, Wilcoxon's signed rank test) but not in the PFC (p = 0.8389, Wilcoxon's signed rank test). Moreover, co-occuring Mu-bursts were detected in the PFC first and then in the dHPC, both at rest and under stress (∗∗p = 0.0016 and ∗∗∗p = 8e-15, respectively), with a median delay that was shorter at rest (0.20 s, **Figure 3D**, left) than under stress (0.32 s, **Figure 3D**, right). Hence, stress increased the occurrence of dHPC Mu-bursts, especially after the initiation of a PFC Mu-burst. We also detected Mu-burst in the PAR. These events could also be observed independently from the ones in the two neighboring structures (PAR and dHPC, **Figure 4**).

At a finer timescale (**Figure 3D**, top), phase shifts (derived from wavelet coherence) among simultaneous bursts appeared nearly constant during Mu-bursts (**Figure 3D**, bottom),

FIGURE 1 | (A) Stress protocol: 60 min at rest, followed 1 day later by 60 min under stress on an elevated platform (EP). A cold light source (100 lux) was applied at a distance of 10 cm in front of the rat's nose to keep the eyes of the animal wide open and its head held up. (B) Representative traces of the Z-scored extracellular fields potential (EFP) simultaneously recorded from the same animal in the prefrontal cortex (PFC), dorsal hippocampus (dHPC) and ventral hippocampus (vHPC) at rest. Raw traces are plotted in gray and filtered (7–12 Hz range) traces are overlaid in black. (C) Spectral analysis of the EFP recorded at rest for each structure (black) and under acute stress (red: PFC n = 13, blue: dHPC n = 13, purple: vHPC n = 12). The top right insert represents the averaged relative change, expressed in percentage of variation. Horizontal dashed line at zero indicates no change. Data are presented as mean ± standard error of the mean (SEM) and shaded area indicates SEM. (D) Coherence for PFC-dHPC, PFC-vHPC and dHPC-vHPC at rest (black) and under stress (orange). In top right insert, averaged relative change expressed in percentage of variation. Horizontal dashed line at zero indicates no change. Data are presented as mean ± SEM and shaded area indicates SEM.

consistent with a value of coherence around one, indicating a strong stability of the phase shift and a high covariation in amplitude. The dHPC PSD amplitude in the Mu-range was low in the absence of Mu-bursts, regardless of the stress or rest condition (**Figure 3E**, top). As expected, amplitude in the 7–12 Hz range increased strikingly during co-occurring Mu-bursts, an effect that was less pronounced when we focused on Mu-bursts occurring in single structures, e.g., only in the PFC or dHPC (Kruskal-Wallis test χ <sup>2</sup> = 31.64; ∗∗∗Prest = 6.238e-7 and χ <sup>2</sup> = 31.61; ∗∗∗Pstress = 6.309e-7, **Table 2** and **Figure 3E**, top). This profile was similar at rest or in stress condition. More intriguingly, phase-locking was also higher during Mu-bursts in one of the two structure (whatever the context) than in non-bursting episodes, suggesting synchronization processes between the PFC and dHPC even during ''subthreshold'' oscillations in the Mu range (Kruskal-Wallis test χ <sup>2</sup> = 23.2 ∗∗∗Prest = 3.661 e-5 and χ <sup>2</sup> = 23.65; ∗∗∗Pstress = 2.961 e-5, **Table 2** and **Figure 3E**, down). These results indicate that Mu-bursts can be generated independently in the PFC and dHPC, while being highly synchronized, and that stress affected the occurrence of Mu-bursts rather than the fine temporal relations between them. Acute stress increased the occurrence of Mu-bursts in both the PFC and dHPC, but impacted the total PSD in the dHPC only. Hence stress can affect the generation of Mu-bursts independently from the basal amplitude and phase of background oscillations in each structure.

#### Interactions of Age and Stress on PFC and HPC Oscillations

We then characterized the modifications of EFP spectral properties upon aging (late middle aged rats of 18 months, henceforth called ''aged rats'', n = 10 for PFC and dHPC recordings, n = 5 for vHPC) and after stress exposure. Strikingly, at rest, aged rats exhibited Mu-bursts in the PFC and in the dHPC (**Figures 5A,B**), reminiscent of the Mu-bursts induced by stress in adults (**Figures 1**, **2**), but no Mu-bursts in the vHPC (**Figure 5B**, see below for analysis), as in adults. Moreover, significant differences were observed: (i) between the average spectral properties of adults and aged rats whatever the structure but without specific frequency range (**Figure 5C** and **Table 3**); and (ii) in hippocampalprefrontal synchrony in the PFC-dHPC and PFC-vHPC. Post hoc test showed that this difference did not implicate

( ∗∗∗Prest/

∗∗∗Pstress). See Table 2 which indicated statistical difference between pairs.

any specific band after statistical correction (**Figure 5C**, bottom).

Stress in aged animal decreased the PSD amplitude in all brain structures (**Table 1** and **Figure 5C**, insert), contrasting with adult rats where only the vHPC was affected (**Figure 1C**). This decrease was prominent in the dHPC whatever the band (∗Pdelta = 0.0078 Wilcoxon's signed rank test; <sup>∗</sup>Pmu = 0.0291 Paired sample t-test; <sup>∗</sup>Pbeta = 0.0118 Paired sample t-test). This might reflect a different reactivity of the vigilance state to stress in aged rats compared to adults. Furthermore, stress decreased dramatically the coherence between the PFC and dHPC in aged rats, for all frequency ranges taken separately (**Figure 5D**, insert right). This clearly contrasted with the stressinduced increase in coherence observed in adults (**Figure 1D**), and provides evidence for stress impacting cortical activity of aged and adult rats in an opposite fashion. Finally, significant difference was also found in PFC-vHPC coherence without incrimination of a specific frequency band, while coherence between the two parts of the hippocampus was not affected by stress.

We thus characterized the Mu-burst activity to assess the effect of stress in aged rats. Both at rest and under stress (**Figure 6A**), the vast majority of Mu-bursts co-occurred in the two structures PFC and dHPC (**Figure 6B**, top). Under stress, the total number of Mu-bursts decreased in the dHPC ( <sup>∗</sup>p = 0.0313 Wilcoxon signed rank test) and in the PFC ( <sup>∗</sup>p = 0.0298, paired sample T-test). However, at rest there was no significant delay between the PFC and dHPC bursts on average (**Figure 6B**, bottom left), while under stress, bursts were detected in the dHPC first (**Figure 6B**, bottom right). Hence temporal relations between Mu-bursts in the dHPC and PFC were inverted in adult and aged rats under stress. Finally, at rest, time-dependent spectrogram analysis suggested a high occurrence of Mu-bursts in aged rats in dHPC, similar to adult rats under stress (**Figure 2A**, right). The AUC from wavelet analysis confirmed a reduction under stress of dHPC

neighboring structures. Rest: 0.53 s median delay, paired sample T-test ∗∗∗P < 0.001 stress: 0.48 s median delay, Wilcoxon signed rank test ∗∗∗P < 0.001. Mu-bursts in aged rats (**Figure 6C**, ∗∗Paged = 0.0057 Paired-

represents the median lag. Note that all Mu-bursts detected in PAR co-occurred in the dHPC. Note that the Mu-burst are observed independently in the two

sample T-test). A similar decrease was observed in dHPC-PFC coherence (**Figure 6D**, ∗∗Paged = 0.0075 paired sample T-test). Overall, both the occurrence of Mu-bursts in the PFC and dHPC, and the synchrony between these structures, were higher at rest in aged rats when compared to adults, and were differentially affected by stress (i.e., increased in adult vs. decreased in aged rats).

Paradoxically, compared to the adult group, dHPC PSD remained low only in the stress condition, during occurring (in a single structure) and co-occurring (PFC and dHPC) Mu-bursts (Kruskal-Wallis test χ <sup>2</sup> = 15.16, ∗∗Prest = 0.0017 and χ <sup>2</sup> = 5.41, Pstress = 0.1444, **Table 2** and **Figure 6E**, top). Interestingly at rest, phase-locking remained significantly higher exclusively when Mu-bursts appeared at the same time in both structures. Lastly, phase locking appeared to be significantly higher when these events were detected together, or only in the PFC, under stress (One-way ANOVA F(3,23) = 7.66, ∗∗Prest = 0.001 and F(3,22) = 9.57, ∗∗∗Pstress = 0.0003, **Table 2** and **Figure 6E**, bottom).

These results appear fully consistent with the notions that aging is itself a stress factor (Morrison and Baxter, 2012; Lindenberger, 2014; Prenderville et al., 2015), and that aged individuals differently cope with stressful situations (Barrientos et al., 2012; Buechel et al., 2014).

#### Effect of the Anxiolytic Diazepam on PFC and HPC Oscillations

Finally, we assessed how DZP, a widely-used anxiolytic, affects stress- and age-related changes on the dHPC-PFC coherence (145 min, n = 5 for each group; **Figures 7A,B**). In the stress condition, for each group, DZP decreased the coherence between the PFC and dHPC in the Mu-range, compared to saline (**Figure 7B**: left, Adults <sup>∗</sup>Pmu = 0.0204, right, Aged ∗∗∗Pmu = 0.0009). Time-dependent spectrograms of dHPC suggested that DZP abolished the increase in stress-related Mu-bursts (**Figure 7C**). In the adults group, DZP partially abolished the effects of stress on the coherence of hippocampal and prefrontal field potentials (**Figure 7D**, left). Contrarily, in aged rats, Mu-rhythms were reduced in the stress-DZP condition compared to the stress-vehicle (SV) condition and rest-vehicle (RV) condition (**Figure 7**). Furthermore, the effect of DZP on the Mu-rhythms seemed to be specific to stress: there was no changes in control non-stressed rats treated with DZP (**Figure 8**). Overall, these results suggest an efficient effect of DZP on adults (DZP abolishes stress effects), and an additive effect of

each structure and each group. Adults (black) Aged (color); red: PFC n = 10, blue: dHPC n = 9, purple: vHPC n = 5). Differences were observed between the average spectral properties of adults and aged rats at rest, whatever the structure (PFC: χ <sup>2</sup> = 57.19; ∗∗∗P < 0.001 Kruskal-Wallis test; dHPC: F(2,63) = 132.45; ∗∗∗P < 0.001 One-way ANOVA test; vHPC Kruskal-Wallis test χ <sup>2</sup> = 41.61 ∗∗∗P ≤ 0.001) and without specific frequency range. Inserts corresponded at the average relative changes under stress, expressed in percentage of variation. Stress in aged animal decreased the PSD amplitude in all brain structures (Friedman's test: (Continued)

#### FIGURE 5 | Continued

χ <sup>2</sup> = 5.15, <sup>∗</sup>P < 0.05 for the PFC; χ <sup>2</sup> = 15.61, ∗∗∗P ≤ 0.001 for the dHPC; χ <sup>2</sup> = 5.53, <sup>∗</sup>P < 0.05 for the vHPC) whatever the band after correction (see text) (D) Coherence for PFC-dHPC, PFC-vHPC and dHPC-vHPC at rest for each group (black: Adults; gold: Aged; PFC-dHPC n = 10, PFC-vHPC n = 5, dHPC-vHPC n = 5). Hippocampal-prefrontal synchrony at rest was significantly different between aged and adults rats, in the PFC-dHPC and PFC-vHPC (Kruskal-Wallis test: χ <sup>2</sup> = 14.81 ∗∗∗PCOHPFC-dHPC <sup>&</sup>lt; 0.001; χ <sup>2</sup> = 24.26, ∗∗∗PCOHPFC-vHPC <sup>&</sup>lt; 0.001). Post hoc (Holm's Bonferroni) test showed that this difference did not implicate any specific band. Top-right insert: relative change after stress, expressed in percentage of variation. Horizontal dashed line at zero indicates no change. Shaded area indicated SEM. Stress decreased dramatically the coherence between PFC and dHPC in aged rats, for all frequency ranges taken separately (χ <sup>2</sup> = 5.76, <sup>∗</sup>P < 0.05 Friedman's test followed by post hoc tests (Holm's Bonferroni), <sup>∗</sup>Pdelta = 0.0166 paired-sample t-test; <sup>∗</sup>Pmu = 0.0098 Wilcoxon's signed rank test; \*Pbeta = 0.0137 Wilcoxon's signed rank test). Significant difference was also found in PFC-vHPC coherence without incrimination of a specific frequency band (χ <sup>2</sup> = 4.98, <sup>∗</sup>P < 0.05 Friedman's test), but not in dHPC-vHPC (F(1,24) = 0.02, P = 0.8766 Two-way ANOVA; ns p > 0.05; <sup>∗</sup>p < 0.05; ∗∗p < 0.01, ∗∗∗p < 0.001).

DZP and stress in aged rats. These results are summarized in the average coherograms from the pharmacological protocol (**Figure 9**).

### DISCUSSION

### Electrophysiological Markers of Stress in Immobile Rats

There is an ongoing debate on the implication of the different parts of the hippocampus in response to stress (Fanselow and Dong, 2010; Bannerman et al., 2014). While the vHPC is known to be directly implicated in anxiety-related processes through direct connections with the amygdala and bed nucleus of stria terminalis (Adhikari, 2014; Adhikari et al., 2015; Padilla-Coreano et al., 2016), the dHPC is believed to exert a role in contextual fear learning only (Bannerman et al., 2004; Fanselow and Dong, 2010). In most studies in rodents, analysis of hippocampal EFP focused on theta (4–12 Hz) oscillations, which in the dorsal part reveal prominent movement-dependent theta-rhythms (Buzsáki, 2002). Theta rhythms in the dHPC are generally of two types: type I theta, which is related to movement and is generated by the entorhinal cortex; and type II theta, which relates to alert immobility, arousal and anxiety and is generated by the medial septum and diagonal band of Broca (Vanderwolf, 1969; Kramis et al., 1975; Wells et al., 2013). Here we used a setup where rats could not move, enabling us to record type II theta and Mu rhythm, while avoiding contamination by type I theta, and we showed that dHPC rhythms were in fact modified by acute stress.

This is to our best knowledge the first report providing evidence that the dHPC PSD in adult rats significantly increased in the 7–12 Hz band under stress. Interestingly, these changes were exclusively caused by Mu-bursts rather than due to type II theta oscillation. These results were not observed in other studies, most probably because animals were free to move, e.g., in an elevated plus maze or an open field (Adhikari et al., 2010; Jacinto et al., 2013). In our experimental paradigm, changes in dHPC rhythms may be explained by animals being immobile (no theta I) or displaying a form of resignation to the long restraining time, with no escape possible (Balleine and Curthoys, 1991). Whatever the reason, these PSD increases in the 7–12 Hz range can be explained by fear experienced by rats when subjected to the EP, or by memorization of the stress context (but see ''Interpretation of Mu Burst Events'' Section for alternative interpretations).

Studies suggest that anxiety-like behaviors decrease, together with the activity of the vHPC circuit, when the environment becomes familiar. Likewise, we observed that vHPC PSD was globally desynchronized, probably due to a long exposure of the same environment, which is consistent with previous studies (Jacinto et al., 2013). Indeed, the environment in which the experiments took place was familiar (following habituation) to all groups of rats.

We found that, at rest, coherence in the 7–12 Hz range was very high between the two parts of the hippocampus (vHPC and dHPC). Coherence was also high between the PFC and the hippocampus but, unexpectedly, significantly higher with the dorsal than with the ventral part. Synchronizations were significantly increased by stress, yet only between the PFC and dHPC. These results are somewhat surprising considering the monosynaptic connections between the PFC and vHPC and the role of these structures in in anxiety (Verwer et al., 1997; Parent et al., 2010). Nevertheless, a strong coherence between the PFC and the dHPC is consistent with their anatomical relationship, which includes not only polysynaptic connections but also monosynaptic drive from the dorsal anterior cingulate cortex to the CA1/CA3 subfield (Rajasethupathy et al., 2015). Coherence analysis reflect functional cell assemblies, e.g., related groups of cells in distant brain structures with synchronized discharge to encode and store information (Battaglia et al., 2011). Hence, our results can be explained by a propagation of activity in the 7–12 Hz range from dHPC neurons, a crucial structure for the fast encoding of initial fear information, to the PFC, a structure with larger storage capacity, but slower learning, resulting into the consolidation of the fear memory.

#### Modulation of Mu-Bursts by Age, Stress and Benzodiazepine

We detected in the PFC and dHPC (but not the vHPC) transient bursts of activity consisting in large amplitude oscillations in the 7–12 Hz frequency range, which we called Mubursts, and that seem associated with WT. These events were modulated in the dHPC by multiple factors, including age, stress and benzodiazepines. At rest, we observed a striking effect of the animal's age, with more Mu-bursts in aged rats compared to adults, which is in agreement with previous studies (Aporti et al., 1986; Buzsáki et al., 1988; Ambrosini et al., 1997). This increased occurrence of Mu-bursts with age suggests these one may arise from the pathway specifically alters with aging. An important finding is that the 7–12 Hz rhythm of the dHPC was the frequency range the most impacted by stress, which suggests these events may be used as a biomarker for stress. Yet, stress acted on Mu-bursts in

relative to PFC. Red line corresponds to the zero-lag and the purple line represents the median lag. Under stress condition, the total number of Mu-bursts decreased in the dHPC (∗p < 0.05, Wilcoxon signed rank test) and in the PFC (∗p < 0.05, paired sample T-test). At rest, there was on average no significant delay (median delay = −0.004 s) between PFC and dHPC bursts, while under stress, bursts were detected first in the dHPC (median delay = −0.0781 s, Wilcoxon signed rank test, ∗∗∗p < 0.001). (C,D) AUC computed from the wavelet transform (left) and the wavelet coherence (right) from the 7–12 Hz range. In aged group, both decreased significantly under stress (black: rest; color: stress). Significant differences were found for AUC (Wilcoxson signed rank test n = 13 ∗∗∗Padults < 0.001 and paired sample T-test n = 9 ∗∗Paged < 0.001) and for coherence (paired sample T-test ∗∗Padults < 0.01 and ∗∗Paged < 0.01). (E) Top: square of the absolute value of the wavelet transform when no Mu-bursts occur ("No"), when Mu-bursts occur in both the PFC and dHPC ("both") or only in one of the two structures ("only PFC" and "only dHPC) in two condition (R: rest; S: stress). dHPC PSD remained low in the absence of Mu-burst and during of occurring and co-occurring Mu-bursts only in the rest condition (Kruskal-Wallis test χ <sup>2</sup> = 15.12, ∗∗Prest < 0.05 and χ <sup>2</sup> = 5.41, Pstress = 0.1444) Bottom: phase-locking value still remained significantly higher during co-occurring of the Mu-bursts at rest, and during PFC-occurring only (One-way ANOVA F(3,23) = 7.66, ∗∗Prest < 0.01 and F(3,22) = 9.57 ∗∗∗Pstress < 0.001). See table which indicated statistical difference between pairs.

opposite fashion, i.e., increase vs. decrease, in adult and aged animals, respectively. Aged rats mays exhibit a hypersecreting HPA axis with increased corticotroprin release, and such glucocorticoid signaling might result in an exaggerated stress response (Buechel et al., 2014; Barrientos et al., 2015). This paradoxical result may alternatively be explained by the fact that, at rest, animals already exhibited different levels of Mu-burst activity.

In addition, we show that DZP decreased the occurrence of Mu-bursts, together with their co-occurrence, across all age in the stress condition. This is consistent with DZP acting as a positive allosteric modulator of GABA<sup>A</sup> receptors, hence globally potentiating inhibition and inducing anxiolytic effects. However, because stress differently affect adult vs. aged animal, DZP overall reverted Mu-bursts occurrence and coherence in adult animals but almost abolished them in aged rats. Aging is associated with an altered composition in α1 and especially α5 subunits of GABA<sup>A</sup> receptors (Yu et al., 2006; Schmidt et al., 2010). The GABAergic inhibition is less active in enhancing benzodiazepine binding in older animals, potentially due to the loss of functional GABA<sup>A</sup> subunits (Calderini et al., 1981; Hoekzema et al., 2012). However, an increase in benzodiazepine binding sites was observed in aged rats, mainly in the hippocampus, striatum and cerebellum (Calderini et al., 1981). Hence, starting from

FIGURE 7 | (A) Stress protocol (same as Figure 1) with an acute i.p injection of Diazepam (DZP; 1 mg/kg) vs. vehicle, in adult rats (n = 5) and aged rats (n = 5). (B) Coherence for PFC-dHPC in stress condition, under vehicle (orange) and under DZP (orange red). Top right insert: averaged relative change in coherence, expressed in percentage of variation. Horizontal dashed line at zero indicates no change. Shaded area indicates SEM. A significant decrease was found in the Mu band both for adult (left panel) and aged rats (right panel; <sup>∗</sup>Padults < 0.01; ∗∗∗Paged < 0.001, paired sample T-test). (C) Spectrogram of dHPC under stress for each group after a vehicle or DZP injection. Note the decrease of Mu-bursts under stress condition after an i.p injection of DZP. (D) PFC-dHPC coherence computed from the wavelet transform in the Mu range in three conditions (RV, rest vehicle; SV, stress vehicle; SD, Stress DZP (1 mg/kg)). Adults: COHPFC-dHPC: Kruskal-Wallis test χ <sup>2</sup> = 1.63, P = 0.4431 for the PFC-dHPC coherence. Aged: Kruskal-Wallis test: χ <sup>2</sup> = 6.02, <sup>∗</sup>P < 0.05, followed by paired-sample t-test and Holm-Bonferroni correction <sup>∗</sup>PRV-SD = 0.0071, <sup>∗</sup>PRV-SV = 0.0102, <sup>∗</sup>PSV-SD = 0.0146 for the PFC-dHPC coherence.

a reduced inhibitory drive, acute administration of DZP may be more efficient in enhancing GABA<sup>A</sup> function in old rats (Reeves and Schweizer, 1983). Our results suggest a definite effect of age on stress response and DZP administration. How this relates to alterations in oscillatory activity will be the focus of further work.

#### Interpretation of the Mu-Burst Events

Mu-bursts, like spindles, have been traditionally observed in the cortex. Mu-bursts are associated with bursting in thalamic neurons and are believed to support distal communication between cortical areas and with the hippocampus (Fanselow et al., 2001). Although EFP reflect the activity of large groups of synapses, allowing identification of synchronous oscillatory activity within and across the brain areas, the exact anatomical origins of EFP must be tempered as voltage fluctuations can originate from volume conduction of distal signals. Our findings suggest that this was not the case here. First, we used a local reference (bipolar electrode) that minimizes electrical transfer (because common distal signals are subtracted), and provides the ''intrinsic'' EFP of the structure. Second, these bursts were found in adult rats in different combinations: only in the PFC, only in the dHPC, or in both structures. It is thus unlikely that dHPC bursts originate from the neocortex. Third, we also recorded from the parietal cortex, in the adult group and found that Mu-bursts could also be observed independently in two close neighboring structures (parietal and dHPC **Figure 4**). Therefore, our work, together with a previous study in the cerebellum (Hartmann and Bower, 1998), suggest that a much larger network of somatosensory structures (i.e., rather than the sole somatosensory cortex) may be flexibly involved in Mu-burst generation. The functional role of Mu-bursts is still a subject of debate in the literature: it has been hypothesized to reflect either a pathological (i.e., absence epilepsy, Inoue et al., 1990; Shaw, 2004, 2007) or physiological state (e.g., alertness or idling; Fanselow and Nicolelis, 1999; Fontanini and Katz, 2005). Even though we cannot definitely discard the hypothesis of an epileptic phenomenon, we believe that the effects of stress, age and anxiolytic we have observed on the 7–12 Hz bursts are of physiological background. First, similar ''Mu rhythms'' occur in 10%–30% of normal human subjects at rest (Nicolelis et al., 1995; Fontanini and Katz, 2005; Sakata et al., 2005; Tort et al., 2010) and have also been observed in cats (Guido and Weyand, 1995; Reinagel et al., 1999), guinea pigs (Edeline et al., 2000), rabbits (Swadlow and Gusev, 2002) and monkeys during periods of sensory processing (Ramcharan et al., 2000).

These data suggest a functionally important and conserved physiological phenomenon. Second, we and others have observed that rats respond rapidly to stimuli during periods where Mu-bursts are detected (Vergnes et al., 1982; Fanselow et al., 2001) and these prominent oscillatory activities are invariably suppressed by movement, but not affected by eye opening (Buzsaki et al., 1988). This suggests that Mu-bursts do not reflect an epileptic state, which would be associated with an impaired sensory detection, but rather a ''hyper alert'' state of vigilance (Fanselow et al., 2001; Sobolewski et al., 2011) or, alternatively, an idling state during quiet immobility (Fontanini and Katz, 2005). Finally, Mu bursts have been associated with sensory-motor processing, which may provide an alternative interpretation of the Mu-burst modifications with age we observed. Animals were indeed isolated from any sensory inputs that might affect Mu oscillations, by placing them in an acoustically insolated chamber. Yet we cannot totally exclude the possibility that Mu oscillations reflected the sensorymotor processing of WT, rather than the level of alertness associated with WT. Along the same line, alterations of Mu oscillations in aged rats may have been caused by sensory impairment associated with aging. We did not quantify sensory function per se, even if aged rats were able to quickly detect sound.

Given that a high occurrence of Mu rhythms has been observed in humans during mind wandering (Braboszcz and Delorme, 2011; Kerr et al., 2013), we propose Mu-bursts in rats could correspond to a similar state of internal attention (Corballis, 2013a,b). Mind-wandering consists in disengaging from goal-oriented interactions with the external environment, with attention being directed inwardly to selfgenerated, stimulus-independent and task-unrelated thoughts. It is plausible that stress and age favor this mind state together with a disengagement from the environment (Killingsworth and Gilbert, 2010; Forster et al., 2015). Nonetheless, our study puts forward that stress-induced theta in the HPC and PFC is composed of Mu-bursts related to arousal. This provides an interesting electrophysiological framework to study the neurobiology of anxiety and anxiolytics, especially in the elderly.

## AUTHOR CONTRIBUTIONS

CS, ES, MS, PF and JM designed the study. ST and BD performed the manufacture of bipolar electrodes and surgery. ST, BD and SD realized the experiment. ST, SD, JN and PF analyzed the data. JM provided the animal facility, the experimental sites and the electrophysiological equipment. ST, JN, AM and PF wrote the manuscript with inputs from SD and JM.

## FUNDING

This work was supported by the Ministère de l'Education nationale de l'Enseignement supérieur et de la Recherche, the region of Ile de France, Centre National de la Recherche Scientifique (CNRS), the University Pierre et Marie Curie, the Agence Nationale pour la Recherche (ANR Programme Blanc 2012 for PF), the foundation for Medical Research (FRM, Equipe FRM DEQ2013326488 to PF), the Fédération pour la Recherche sur le Cerveau (FRC et les rotariens de France, ''espoir en tête'' 2012 to PF) and the IMI Newmeds grant (to ST). The laboratories of PF and JM are part of the École des Neurosciences de Paris Ile-de-France RTRA network. PF and JM are members of the Laboratory of Excellence, LabEx Bio-Psy, and of respectively the DHU Pepsy and the DHU Fast.

#### REFERENCES


### ACKNOWLEDGMENTS

We thank Marie-Louise Dongelmans and Sebastien Valverde for their critical reading of the article, Amélie Cougny and Vincent Grosjean for help with experiments during their internship, Thomas Watson and Lu Zhang for their critical interpretations of the Mu-bursts and the discussion of wavelet coherence analysis, Soizic Jezequel, Laura Legouestre, Grégoire Mauny and Laurent Poitier, for animal care and the animal facility of Université Pierre et Marie Curie based in the Hôpital Charles Foix.


of AMPA-, α1- and 5-HT2A-receptors. Br. J. Pharmacol. 135, 65–78. doi: 10.1038/sj.bjp.0704451


**Conflict of Interest Statement**: MS and ES were employees of Servier during a part of this work.

The other authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Takillah, Naudé, Didienne, Sebban, Decros, Schenker, Spedding, Mourot, Mariani and Faure. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Cognitive Training Enhances Auditory Attention Efficiency in Older Adults

Jennifer L. O'Brien<sup>1</sup> \*, Jennifer J. Lister<sup>2</sup> , Bernadette A. Fausto<sup>3</sup> , Gregory K. Clifton<sup>2</sup> and Jerri D. Edwards<sup>4</sup>

<sup>1</sup> Department of Psychology, University of South Florida St. Petersburg, St. Petersburg, FL, United States, <sup>2</sup> Department of Communication Sciences and Disorders, University of South Florida, Tampa, FL, United States, <sup>3</sup> School of Aging Studies, University of South Florida, Tampa, FL, United States, <sup>4</sup> Department of Psychiatry and Behavioral Neurosciences, College of Medicine, University of South Florida, Tampa, FL, United States

Auditory cognitive training (ACT) improves attention in older adults; however, the underlying neurophysiological mechanisms are still unknown. The present study examined the effects of ACT on the P3b event-related potential reflecting attention allocation (amplitude) and speed of processing (latency) during stimulus categorization and the P1-N1-P2 complex reflecting perceptual processing (amplitude and latency). Participants completed an auditory oddball task before and after 10 weeks of ACT (n = 9) or a no contact control period (n = 15). Parietal P3b amplitudes to oddball stimuli decreased at post-test in the trained group as compared to those in the control group, and frontal P3b amplitudes show a similar trend, potentially reflecting more efficient attentional allocation after ACT. No advantages for the ACT group were evident for auditory perceptual processing or speed of processing in this small sample. Our results provide preliminary evidence that ACT may enhance the efficiency of attention allocation, which may account for the positive impact of ACT on the everyday functioning of older adults.

#### Edited by:

Ana B. Vivas, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Jingwen Niu, Temple University, United States Paloma Mari-Beffa, Bangor University, United Kingdom

#### \*Correspondence:

Jennifer L. O'Brien jenobrien@usf.edu

Received: 29 September 2016 Accepted: 19 September 2017 Published: 04 October 2017

#### Citation:

O'Brien JL, Lister JJ, Fausto BA, Clifton GK and Edwards JD (2017) Cognitive Training Enhances Auditory Attention Efficiency in Older Adults. Front. Aging Neurosci. 9:322. doi: 10.3389/fnagi.2017.00322 Keywords: aging, cognitive training, auditory cognition, attention, event-related potentials

## INTRODUCTION

Hearing loss is a distressing impairment that becomes increasingly prevalent with age, affecting 30% of adults aged 65 to 74 and almost 50% of older adults over the age of 75 (National Institute on Deafness and Other Communication Disorders [NIDCD], 2010). Hearing loss causes speech perception difficulties (Humes et al., 2012) and has been linked to subsequent cognitive impairment (Lin, 2011; Lin et al., 2011a,b,c, 2013; Harrison Bush et al., 2015) as well as increased social isolation (Mick et al., 2014), reduced quality of life (Ciorba et al., 2012), increased risk for depression (Li et al., 2014), and reduced engagement in independent activities of daily living (Dalton et al., 2003). Because hearing loss results in effortful auditory processing in normal aging (Pichora-Fuller et al., 2016), remediation of auditory processing may improve cognition and enhance quality of life in older adults. The current study investigates the efficacy of auditory cognitive training (ACT) to improve underlying neurophysiological mechanisms of auditory perception and cognition.

Incipient, age-related hearing loss has been conventionally corrected with hearing aids. Despite improved audibility provided by current hearing aid technology, hearing aids alone often do not compensate for decreased ability to comprehend meaningful auditory stimuli among background noise. Speech perception, for instance, depends not only on intact auditory function but also intact

cognitive function including the ability to attend to relevant stimuli, to process incoming stimuli, and to maintain new information all while actively manipulating temporarily stored stimuli. These cognitive functions all appear to decline in normal aging (for reviews, see Akeroyd, 2008; Arlinger et al., 2009). Thus, while hearing aids can enhance audibility, concomitant and subsequent age-related cognitive decline may persist, impairing older adults' ability to process meaningful sound in challenging listening situations.

There is growing evidence that cognitive training programs can ameliorate or minimize age-related sensory and cognitive decline through neuroplastic change (Smith et al., 2009). However, the neural mechanisms underlying these changes are still relatively unknown (although see O'Brien et al., 2013, 2015). The purpose of the current study is to investigate the neural mechanisms underlying cognitive gains following auditory-based cognitive training.

The current study employs computerized ACT in an attempt to improve auditory perception and cognition (attention, speed of processing) in older adults. ACT is a process-based training targeting certain neural circuits via perceptual information processing. Process-based training is hypothesized to lead to transfer to other tasks that engage the same or overlapping neural circuit(s) regardless of whether the other tasks were specifically trained (Jonides, 2004). ACT has a positive effect on behavior reflecting auditory perception, memory, attention, and speed of processing (Mahncke et al., 2006; Smith et al., 2009; Anderson et al., 2013) and there is some evidence for far transfer (Strenziok et al., 2014). However, the underlying mechanisms of these gains remain relatively unexplored. Most of the extant literature has measured efficacy of cognitive training behaviorally through neuropsychological and psychometric tests. Behavioral measures reflect combined effort stemming from several stages of processing (i.e., sensory, cognitive, and motor) and performance is influenced by outside factors such as motivation and physical function, making it difficult to draw conclusions about the neurophysiological processes underlying behavioral changes.

An alternate approach used in the current study is measuring event-related potentials (ERPs), which are averaged signals from electroencephalogram (EEG) time-locked to a perceptual and/or cognitive event. ERPs are reflective of ongoing brain activity and are particularly sensitive to the timing of mental processes (on the order of ms), such that early perceptual activity can be distinguished from post-perceptual cognitive processes (Luck and Kappenman, 2012). Of particular interest in the current study is the P1-N1-P2 complex and the P3b component. The P1-N1-P2 complex is a fronto-centrally occurring series of ERP components measuring the physiological response of the auditory cortex to a stimulus, reflective of the neural detection of sound (for a review, see Key et al., 2005). The effect of age on these components is unclear. Data from numerous studies show both increases and decreases to amplitudes and latencies with age depending on methodological differences such as the attentional demands of the task (for a review, see Ceponien ˇ e et al., ˙ 2008).

The P3b component is a posterior-parietal component thought to reflect the attentional resources allocated for categorization of a target (Donchin, 1981; Pfefferbaum et al., 1984; Kok, 2001). It is sensitive to target probability, with unexpected or deviant stimuli eliciting a larger P3b than stimuli occurring with a high probability. Effects of aging on the P3b show a more frontally distributed, attenuated P3b amplitude and longer P3b latency (Anderer et al., 1996; Polich, 1997). This is consistent with recent theories of cognitive aging (Davis et al., 2008; Park and Reuter-Lorenz, 2009) indicating that prefrontal cortex (PFC) processing is recruited to counteract sensory, cognitive, and physical brain changes associated with normal aging. O'Brien et al. (2013) demonstrated an increase in older adults' parietal P3b amplitude following process-based training in the visual modality compared to no-contact controls. Supporting evidence shows that P3b latency decreases after visual training are associated with better Useful Field of View performance (O'Brien et al., 2015).

In the present study, we predicted that ACT would result in improved attention consistent with previous evidence (Smith et al., 2009) and reflected cortically by a change in parietal P3b amplitude. Speed of processing improvements would be reflected in changes to parietal P3b latency. We also predicted that a frontal P3b would be present for all participants at baseline and would change following training. Also, if ACT impacts early cortical perception of auditory stimuli, we predicted a change to one or more components in the P1-N1-P2 complex in the form of amplitude or latency shifts, or a combination of both. Behavioral measures of cognition (Cognitive Self-Report Questionnaire, CSRQ) and speed of processing (Time-compressed speech, TCS) were included as corresponding evidence of neurophysiological changes. Changes in neurophysiological measures were predicted to correspond with changes in behavioral measures as a result of training.

### MATERIALS AND METHODS

#### Participants

Twenty-four experimentally naïve healthy older adult subjects (17 female, mean age = 70.88 years, mean years of education = 15.29) participated in exchange for cognitive training (see **Table 1** for demographic information). Participants were recruited from a list compiled of older adults who contacted the lab in response to a newspaper article or an ad placed in



PTA, pure tone average. Count in brackets.

local media. This study was carried out in accordance with the recommendations of the University of South Florida institutional review board with written informed consent from all subjects in accordance with the Declaration of Helsinki. The protocol was approved by the University of South Florida institutional review board.

#### Inclusion and Exclusion Criteria

Participants were required to: have a Mini-Mental State Examination (Folstein et al., 1975) score of 24 or greater (no severe cognitive impairment or dementia), have no self-reported neurological disorders, major strokes, or head injuries, have sufficient hearing (pure tone hearing thresholds <70 dB HL at 1000 and 2000 Hz in both ears), be a native English speaker, be available and willing to commit to the time and travel requirements of the study (maximum 22 visits), not be concurrently enrolled in another cognitive or training-related study, and not have previously completed a cognitive training program before participating.

#### Group Assignment

Training-eligible participants were randomly assigned to computer-based ACT using Brain Fitness© (n = 9) or a no-contact control group (n = 15). During recruitment, participants were informed that they would be receiving cognitive training either immediately after baseline testing or subsequent to a second testing session 10 weeks after their baseline session. Chi-square analysis showed no significant differences between groups based on sex, p = 0.668. Independent samples t-tests revealed no significant differences between the groups in age or education, ts < 1.

#### Measures

#### Audiometric Testing

A standard comprehensive audiometric evaluation was completed (American Speech-Language-Hearing Association, 2005) for both ears at octave frequencies between 250 and 8000 Hz to determine sufficient hearing to discern testing and training stimuli (pure tone hearing thresholds < 70 dB HL at 1000 and 2000 Hz in both ears). Testing was completed in a single-walled sound-treated booth suitable for audiometric testing. A three-frequency pure tone average (PTA) was calculated for each ear for using thresholds measured at 500, 1000, and 2000 Hz. PTAs lower than 25 dB are considered within normal hearing limits, 26–40 dB constitutes mild hearing loss, and moderate hearing loss at 41–55 dB. Participants in the trained group had PTAs ranging from 0 to 27 dB, significantly lower than those in the control group, ps < 0.025, which ranged from 5 to 55 dB (see **Table 1** for more descriptives).

#### Auditory Oddball Task

Frequent pure tone stimuli were presented (80% of the time) at 1000 Hz; oddball pure tone stimuli were presented (20% of the time) at 1500 Hz. Participants indicated the presence of an oddball stimulus by pressing a key on a computer keyboard. The task contained 8 blocks of 60 trials each (12 oddballs made up 20% and 48 frequents made up the remaining 80% of the stimuli presented) for a total of 480 stimuli (96 oddballs, 384 frequents) for each stimulus condition. The stimuli were 60 ms in duration and were presented at 80 dB SPL and the same wav file was used on each presentation. The task took approximately 15 min to complete.

#### Cognitive Self-Report Questionnaire (CSRQ)

The CSRQ is a 25-item self-report questionnaire comprising statements about an individual's self-reported perceptions of hearing, cognition, and mood (Spina et al., 2006). Participants are asked to rate each statement (e.g., "I have had trouble hearing conversations on the telephone"; "I have felt I have a good memory"; "I have been in a bad mood" on a 5-point Likert scale from 1 "Almost Always" to "Hardly Ever." The sum of all 25 items is calculated for a total score, with higher scores indicating more cognitive difficulties. The CSRQ has been reported to have excellent internal consistency (α = 0.91) and good 2-month test–retest reliability (r = 0.85) and has been used as a pre-and-post cognitive training tool to examine cognitive training effects on hearing, cognition, and mood in older adults (Spina et al., 2006).

#### Time-Compressed Speech (TCS)

The TCS is a low redundancy measure of auditory processing speed in which speech is digitally accelerated (compressed) to resemble fast speech (Beasley et al., 1980). For the current study, the TCS stimuli comprised the Northwestern University Number 6 words. Fifty words spoken by a female were presented binaurally at a 65% compression rate. Immediately after each word presentation, the participant was asked to repeat the word, even if he or she was unsure of the answer. TCS performance was defined as the percent of words correctly recognized with higher scores indicating better performance. Performance typically decreases with age (Sticht and Gray, 1969; Gordon and Fitzgibbons, 2001). The TCS is a routinely used clinical measure for auditory processing deficits and has been previously validated in older adults (Letowski and Poch, 1996).

#### Procedure

Participants completed a screening visit to determine eligibility for the study and a baseline assessment of behavioral tasks (detailed above). EEG was recorded at baseline during performance of an auditory oddball task (detailed above). After baseline assessment, participants in the cognitive training group worked on computerized training exercises with the goal of completing a minimum of 16 training hours. Participants completed the auditory training program Brain Fitness (Posit Science). The program consists of six adaptive auditory exercises that are aimed at enhancing speed and accuracy of auditory processing. **Table 2** describes each exercise. The exercises are designed to simulate realistic listening contexts in a progressive fashion, moving from simple to complex auditory stimuli across exercises. Within each exercise, the stimuli become less discriminable and duration of stimulus presentation decreases as performance improves.

#### TABLE 2 | Brain fitness exercises.

fnagi-09-00322 September 30, 2017 Time: 16:0 # 4


Training sessions were 60 min in duration, 2 days per week, for up to 10 weeks, based on prior study protocols of cognitive training (e.g., Edwards et al., 2002, 2005; Ball et al., 2007). On average, participants completed training in 62 days (Min = 56, Max = 70, SD = 6.01), missing at most 1 week between sessions. During each training session, individuals were required to take at least one 5-min break, and were allowed to take additional breaks as necessary. Based on prior findings that the interval between sessions could vary without affecting efficacy (Vance et al., 2007), participants could skip training days if necessary, although frequent or extended missing of sessions was discouraged. Participants were supervised by a trainer in a group computer lab setting. The trainer was present to ensure on-task participation for the full session, as well as to clarify task instructions and handle any technical difficulties if necessary. On average, participants completed 18.78 h of training (Min = 14, Max = 20, SD = 1.99). Following training, participants repeated the same behavioral and auditory oddball tasks as completed at baseline.

Participants in the no-contact control group completed a second testing session 10 weeks following their baseline assessment, during which they repeated the same behavioral and auditory oddball tasks as completed at baseline. They were then invited to complete 10 weeks of training. We chose a no-contact control because previous research of cognitive training has revealed no differences between no-contact and social- and computer-activity control conditions (Wadley et al., 2006) on behavioral outcome measures.

#### EEG Recording

Continuous EEG activity was recorded from 64 Ag/AgCl electrodes at standard 10/20 locations using NeuroscanTM (SCAN version 4.3.1) with a SynAmps2 amplifier, with a vertex midline electrode position halfway between Cz and CPz as reference. For five trained and five control participants, continuous EEG activity was recorded using the NuAmps (NuAmp, Neuroscan, Inc., El Paso, TX, United States) single-ended, 40-channel amplifier according to the NuAmps International 10–20 electrode system using a Quikcap with sintered Ag/AgCl electrodes, and a continuous acquisition system (Scan 4.3 Acquisition Software, Neuroscan, Inc.). A right mastoid electrode was used as a reference. For all participants, four additional electrodes were placed on the outer canthus of each eye and on the supra and infraorbital ridges of the left eye to monitor eye movement and blink activity. Data was sampled at 1000 Hz with a 100 Hz low pass filter (time constant: DC). Electrode impedances were kept below 5 k for most electrodes.

The experiment took place in a dimly lit, sound-attenuating booth. A Pentium 4 PC running E-Prime 1.1 (Schneider et al., 2002) recorded behavioral data and presented auditory stimuli. Responses were registered using a keyboard.

#### Data Analysis and Predictions

Continuous EEG was high-pass filtered at a corner frequency of.1 Hz and low-pass filtered at a corner frequency of 30 Hz with a squared Butterworth zero-phase filter (12 dB/octave roll-off). Ocular artifact from eye movement and blinks were corrected for each subject by extracting the electroocular signals from the EEG. EEG for frequent and oddball trials was separated into epochs of 1000 ms (−200 ms before trial onset to 1200 ms after) and baseline corrected (−100 to 0 ms). Epochs in which EEG amplitude exceeded criteria of ±100 µV were rejected prior to averaging (M = 6%, MAX = 17%). Data were then averaged separately for each stimulus type (frequent, oddball). Data for the P3b component were re-referenced to averaged mastoids. Mean amplitudes and peak latencies were measured at parietal electrode site Pz and frontal electrode FCz for frequent and oddball stimuli in a 250 – 750 ms post-stimulus time window. Data for the P1-N1-P2 complex were re-referenced to a global reference. Mean amplitudes and peak latencies were measured at FCz for frequent stimuli in a 45 – 95 ms post-stimulus time window for P1, 105 – 155 ms post-stimulus time window for N1, and 225 – 275 ms post-stimulus time window for P2.

For each component, an Analysis of Variance (ANOVA) (or independent t-test where applicable) was first used to compare the two conditions at baseline, and repeated-measures ANOVA was used to examine training effects. All tests were two-sided and had an alpha level of.05. P3b analyses at Pz included within-participant factors of Testing Session and Stimulus Type, and the between-participants factor of Group. Effect sizes were calculated using omega squared (ω 2 ). A significant Testing Session × Stimulus Type × Group interaction for P3b amplitude was expected to support the hypothesis that attentional allocation is enhanced post-training and the same interaction for P3b latency was expected to support the hypothesis that speed of processing is enhanced post-training. Frontal P3b analyses at FCz were conducted in the same manner with the same expected results. P1-N1-P2 analyses for frequent stimuli included the within-participant factors of Testing Session and the betweenparticipants factor of Group. For all the above analyses, follow-up ANOVAs and t-tests were conducted to examine any significant effects within each subgroup.

Behavioral data from the oddball task were analyzed with repeated measures ANOVAs including within-participant factors of Testing Session and the between-participants factor of Group. The auditory oddball task was designed to be very easy to ensure high accuracy rates to preserve as many trials as possible for the electrophysiological analyses. Therefore, behavior from the oddball task was not predicted to change following training.

TABLE 3 | Mean amplitudes and peak latencies for P3b and P1-N1-P2 at baseline and post-test per group.


a, electrode site Pz; b, electrode site FCz. Standard deviations in parentheses; †Oddball stimuli; ‡Frequent stimuli.

Due to the small sample size, the sample was underpowered to be able to detect behavioral changes in the two auditory behavioral measures (CSRQ and TCS). However, prior research (Smith et al., 2009) has documented the positive effect of ACT on the CSRQ.

Finally, Pearson correlations were conducted to determine the relationship between any significant neurophysiological training gains and changes in auditory behavior measures. Alpha was set to 0.05. Mean amplitudes and peak latencies for all components are reported in **Table 3**. Behavioral scores for the CSRQ and TCS are reported in **Table 4**. Effects relevant to the proposed hypotheses are summarized below and all main effects and interactions are reported in **Tables 5**–**9**.

#### RESULTS

#### Auditory Oddball Task Accuracy

Behavioral performance, accuracy for detecting the occurrence of oddballs, was above 99% (SDs below 2%) for both groups at both time points. ANOVA of behavioral accuracy revealed that performance did not significantly differ by group or by time, Fs < 1. Reaction times to oddball stimuli also did not significantly vary by group or time point, ps > 0.200.

#### Parietal P3b Amplitude

**Figure 1** illustrates the grand average ERP waveforms of the oddball and frequent stimuli for the trained and control groups at both testing points at electrode Pz. ANOVA of initial P3b

TABLE 4 | Mean behavioral scores for Cognitive Self-Report Questionnaire and time-compressed speech at baseline and post-test per group.


CSRQ, Cognitive Self-Report Questionnaire; TCS, time-compressed speech at 65% compression; <sup>∗</sup>Score for one training group participant is missing (n = 8). Standard deviations in parentheses.

amplitudes at baseline revealed that they did not significantly differ between groups, F(1,22) = 1.17, p = 0.292, ω <sup>2</sup> = 0.007. When comparing P3b amplitudes across time, the hypothesized interaction between Stimulus Type, Testing Session, and Group was marginally significant, F(1,22) = 4.01, p = 0.058, ω <sup>2</sup> = 0.007; there was also a significant interaction of Testing Session and Group F(1,22) = 14.58, p < 0.001,ω <sup>2</sup> = 0.025. Follow-up analysis of the trained group showed a significant interaction of Stimulus Type and Testing Session, F(1,8) = 11.03, p = 0.011, ω <sup>2</sup> = 0.036, reflecting a significant decrease in P3b amplitude to the oddball stimulus from baseline (M = 5.14, SD = 3.96) to post-test (M = 3.12, SD = 3.64), t(8) = 5.60, p = 0.001. There were no significant changes in amplitude across testing sessions after 10 weeks of no contact for the control group, F < 1.

Given the difference between the trained and control groups' hearing PTAs at baseline, it is possible that the pattern of results were associated with this difference. To test this, we removed six participants from the control group who had average PTAs greater than 2.5 SDs from that of the trained group to conduct sensitivity analyses. Initial P3b amplitudes at baseline did not significantly differ between groups, F(1,16) = 1.32, p = 0.268, ω <sup>2</sup> = 0.017. When comparing P3b amplitudes across time, the hypothesized interaction between Stimulus Type, Testing Session, and Group was now not significant, F(1,16) = 2.68, p = 0.121, ω <sup>2</sup> = 0.005; however, there was still a significant interaction of Testing Session and Group F(1,22) = 13.20, p = 0.002,ω <sup>2</sup> = 0.030. Critically, there remained no significant changes in amplitude across testing sessions for the control group, F < 1.

#### Frontal P3b Amplitude

**Figure 2** illustrates the grand average ERP waveforms of the oddball and frequent stimuli for the trained and control groups at both testing points at electrode FCz. ANOVA of initial P3b amplitudes at baseline revealed that they did not significantly differ between groups, F < 1. When comparing P3b amplitudes across time, the hypothesized interaction between Stimulus Type, Testing Session, and Group was not significant, F(1,22) = 1.60, p = 0.220, ω <sup>2</sup> = 0.002. However, there was a significant interaction of Testing Session and Stimulus, F(1,22) = 5.46,

#### TABLE 5 | Results of P3b amplitude at Pz repeated measures ANOVA.


#### TABLE 6 | Results of P3b amplitude at FCz repeated measures ANOVA.


TABLE 7 | Results of P3b latency at Pz and FCz repeated measures ANOVA.


#### TABLE 8 | Results of P1-N1-P2 Amplitudes at FCz Repeated Measures ANOVA.


#### TABLE 9 | Results of P1-N1-P2 latencies at FCz repeated measures ANOVA.


p = 0.029, ω <sup>2</sup> = 0.016 and a marginally significant interaction of Testing Session and Group, F(1,22) = 3.42, p = 0.078, ω <sup>2</sup> = 0.008.

Given these effects and the a priori hypothesis of frontal shifts of P3b with aging, we conducted follow-up analyses to determine if significant P3b amplitude decreases occurred frontally as well for trained participants. Follow-up analysis of the trained group showed a significant interaction of Stimulus Type and Testing Session, F(1,8) = 7.38, p = 0.026, ω <sup>2</sup> = 0.040, indeed reflecting a decrease in P3b amplitude to the oddball stimulus from baseline (M = 1.42, SD = 4.93) to post-test (M = −0.84 SD = 4.95), t(8) = 3.28, p = 0.011. There were no significant changes in amplitude across testing sessions after 10 weeks of no contact for the control group, F < 1.

#### P3b and Frontal P3b Latencies

ANOVA of initial P3b latencies at baseline revealed that they did not significantly differ between groups, F < 1. When comparing P3b latencies across time, there was a significant main effect of

Testing Session, F(1,22) = 5.63, p = 0.027, ω <sup>2</sup> = 0.026. P3b latencies decreased from baseline (M = 447.85, SD = 132.63) to post-testing (M = 405.23, SD = 121.31), with no significant interaction of Group or Stimulus Type, Fs < 1. There were no noteworthy significant latency effects at the frontal P3b.

#### P1-N1-P2 Amplitude and Latencies

baseline (gray line) and post-test (black line) at electrode Pz.

**Figure 3** illustrates the grand average ERP waveforms of the frequent stimuli for the trained and control groups at both testing points. Comparison of initial P1-N1-P2 amplitudes and latencies at baseline revealed that neither significantly differed between groups, ps > 0.187. When comparing P1-N1-P2 amplitudes across time, there were no significant effects of time, group, or interactions for any of the three components, ps > 0.281. This was also the case for P1-N1-P2 latencies, ps > 0.256.

#### ERP Correlations with Behavioral Data

To determine whether significant training effects on the P3b component corresponded with positive auditory behavioral

gains, we first conducted a Pearson correlation between posterior P3b amplitude differences and differences self-reported perceptions of cognition (CSRQ) from baseline to post-test. A significant positive correlation was found between scores on the CSRQ and P3b amplitudes, with larger P3b amplitudes for the individuals with more self-reported auditory cognitive difficulties, r = 0.53, p = 0.010. We also conducted a Person correlation between posterior P3b latency differences and differences in our behavioral measure of auditory processing speed (TCS) from baseline to post-test, which was not significant, p = 0.501.

### DISCUSSION

The goal of the current study was to elucidate the role of ACT on attentional and perceptual processing as measured by ERPs to help determine the underlying mechanisms of training gains and transfer. We provide electrophysiological evidence showing that engaging in process-based ACT can enhance attentional mechanisms in older adults.

#### Training-Related Effect on P3b Amplitude

P3b amplitude is considered to reflect the attentional resources involved in comparing a significant or relevant event with an internal representation to categorize as a match or mismatch (Kok, 2001). Rare events, as a mismatch to internal representations, typically capture more attentional resources and result in larger P3b amplitudes than expected frequent events. Consistent with this, oddball stimuli (rare events) did elicit a larger P3b than frequent stimuli for both groups at both testing time points in the current study. However, oddball stimuli elicited a P3b following ACT that was smaller compared to pre-training amplitudes, while oddball amplitudes showed no change for controls across time. There was no significant difference between groups in baseline P3b amplitudes for oddball stimuli, but visual inspection shows a descriptively larger baseline amplitude for the trained group compared to the control group (see **Table 3**, **Figure 1**). Further study is needed to determine how baseline differences in cognitive abilities such as attention may impact neurophysiological measures after training.

While decreases in P3b amplitude are usually interpreted as diminished attentional resources, it is also possible that decreased P3b amplitude reflects diminished allocation due to the need for fewer attentional resources to categorize a stimulus. This interpretation suggests that ACT leads to an efficiency of attentional resource allocation and/or interaction with working memory updating. There is some prior evidence to support this hypothesis. Using meditation training, which promotes a broader attentional state, Slagter et al. (2007, 2009) showed decreased P3b amplitude to the first target in an attentional blink task following training compared to untrained controls. The attentional blink task requires attention to a rapid stream of stimuli and the subsequent report of two embedded target stimuli. Ability to report the second target in the stream is typically reduced if it appears within 500 ms after the first target ("attentional blink," Raymond et al., 1992). A reduction in P3b amplitude to the first target suggests that training improved efficiency of attentional engagement such that resources were not solely devoted to the first target and were instead balanced between the two. In support of this, behavioral results showed a parallel reduction in the attentional blink following training. Mindfulness training has also been shown to reduce the P3b amplitude to incongruent words in a Stroop task possibly reflecting more efficient allocation of attentional resources during stimulus processing (Moore et al., 2012).

In a sample of young adults, Ben-David et al. (2010) showed a reduction in the late positive complex (LPC), a parietal response of which P3b is considered to be a subcomponent (Rushby et al., 2005), following an hour of auditory perceptual training of speech sounds. This was interpreted to reflect improvement in stimulus categorization and perceptual processes and possibly improvement in memory updating. Interestingly, Ben-David et al. (2010) showed the opposite effect of speech sound practice on the LPC, an increased LPC amplitude following training. They cite differences in experimental design between the two studies as a possible cause of amplitude reversal. Similar to this, O'Brien et al. (2015) reported larger P3b amplitudes following visual cognitive training compared to controls. In addition to being in a different modality, the oddball task used in their study was significantly more complex involving locating a perceptually different (tilted) target among an array of identical (horizontal) distractors. The possibility that visual and auditory process-based training impacts P3b amplitude differently needs to be further explored.

P3b amplitude was also significantly correlated with participants' self-reported perceptions of their own auditory cognition (CSRQ). Participants who reported more cognitive difficulties had larger P3b amplitudes compared to those with fewer cognitive difficulties. ACT has previously been shown to result in CSRQ self-reported improvements by participants (Smith et al., 2009), suggesting that the neurophysiological changes occurring due to training may be behaviorally

significant. Taken together, these findings show that a reduction in P3b amplitude can reflect efficiency in attentional resource allocation, such as flexibility in engaging and disengaging from relevant target stimuli. Attentional resource allocation following ACT is likely more efficient, resulting in fewer resources needed to categorize an oddball stimulus.

### Training-Related Effect on Frontal P3b Amplitude

Auditory cognitive training also resulted in a marginally significant decrease in frontal P3b amplitude compared to untrained controls. P3b activity shifts to a more frontal distribution with age and this shift has been interpreted as reflecting frontal lobe activity (for meta-analysis studies supporting this, see Friedman et al., 1997). Friedman and colleagues (Friedman et al., 1993; Friedman and Simpson, 1994; Fabiani and Friedman, 1995) report that older adults have two scalp distributions in response to auditory oddball stimuli – one frontal and the other parietal – suggesting that older adults activate frontal lobe processes to help encode these stimuli. Frontal areas are often activated in target detection tasks, corresponding with P3b generation (for reviews, see Polich, 2003; Fjell et al., 2007). As previously described, current theories of cognitive aging propose that PFC processing is recruited for cognitive tasks to compensate for parietal network functions that decline with age. Reduction in amplitude of the frontal P3b following training suggests that ACT potentially reduces the demand for PFC recruitment during attention processing.

#### P3b Latency Changes

P3b latency did not show any change based on training as hypothesized but instead showed an effect of testing. Participants showed faster processing for both stimuli types during their post-test regardless of whether they were in the training or the control group. This is consistent with evidence that P3b latency reliability in an auditory oddball task is low in older adults both within a session (r = 0.07–0.24) and from 1 year to the next (r = 0.14–0.40) (Walhovd and Fjell, 2002). In addition, the correlation between P3b latency and behavioral auditory processing speed (TCS) was not significant.

Auditory cognitive training is designed to target speed of processing and has previously been shown to improve auditory processing speed in a sample of healthy older adults (Anderson et al., 2013) as well as in a sample of older adults with heart failure (Athilingam et al., 2015). It is possible that P3b latency as a measure of processing speed in the current study did not have enough power to detect an effect given the inconsistencies likely occurring within subjects. However, it is also possible that ACT primarily enhances allocation of attention rather than speed of processing.

#### No P1-N1-P2 Changes

Auditory cognitive training did not impact auditory perceptual processing in the current study, as measured by P1-N1-P2 amplitude and latency. Enhancements in subcortical neural timing and speech perception following ACT have previously been observed using evoked potentials representing auditory brainstem responses to a speech syllable in noise (Anderson et al., 2013). It is possible that changes to early auditory perception following ACT are only measurable subcortically and that our later ERP measure was insensitive to these changes. The stimuli used in the current study – pure tones in quiet – were likely much easier to perceive than the speech in noise condition used in Anderson et al. (2013) study and it is therefore plausible that the insensitivity of our measure is due to a ceiling effect.

#### Limitations

A significant limitation to this study is the small sample size. Randomized controlled studies involving cognitive training with older adults across a significant period of time are resourceintensive and prone to high levels of attrition for multiple reasons (e.g., voluntary withdrawal of consent, no longer meeting inclusion/exclusion criteria during course of study, poor adherence to training regimen). A large-scale, multi-site trial investigating the cognitive impact of ACT in older adults has been conducted (Smith et al., 2009), but this did not include any ERP measures to determine underlying mechanisms of training gains. To further investigate these underlying mechanisms, a study using neurophysiological measures similar in scope to Smith et al. (2009), utilizing intent-to-treat analysis modeling for attrition needs to be conducted.

Study demographics (primarily female, Caucasian, and well educated) also limit the interpretation of the current findings. Given the proposal that ACT targets speed of processing, and the current finding that ACT benefits attention and not speed of processing, multiple converging measures of these two functions should be included in future studies to clarify their role in training gains. Finally, the impact of ACT on functional outcomes as well as the long-term maintenance of training gains need to be investigated.

### CONCLUSION

The present finding of decreased P3b amplitudes following ACT reinforces the hypothesis that there is plasticity in the attentional control systems of older adults. Control over attentional resource allocation is vulnerable to age-related decline, but is shown here to be ameliorated by ACT. In light of previous findings demonstrating that portions of this training program result in improved cognition and transfer of gains to functional tasks (e.g., Smith et al., 2009; Anderson et al., 2013; Strenziok et al., 2014), our study is the first to provide preliminary neurophysiological evidence that ACT may particularly be enhancing the efficiency of attention allocation, which may account for the positive impact of ACT on the everyday functioning of older adults.

#### AUTHOR CONTRIBUTIONS

JO, JL, GC, and JE contributed to the conception and design of the work; JO and GC contributed to the acquisition of the data, JO, JL, BF, GC, and JE contributed to the analysis and interpretation of the data, drafted the work, JO, JL, BF, and JE revised the work critically for important intellectual content, JO, JL, BF, GC, and JE gave final approval of the version to be published and agree to be accountable for all aspects of the work.

#### REFERENCES

fnagi-09-00322 September 30, 2017 Time: 16:0 # 10


### ACKNOWLEDGMENTS

Special thanks to Courtney Matthews and Amanda Brandino for assistance in data collection.



**Conflict of Interest Statement:** From June to August 2008, JE worked as a limited consultant to Posit Science, who currently markets the auditory cognitive training program used in this study. JE currently serves on the data safety and monitoring board of NIH grants awarded to employees of Posit Science. JE worked as a consultant to Wilson, Sonsini, Goodrich, and Rosati across an approximate 3 month period between May and August of 2015.

The other authors declare that the research was conducted in the absence of any other commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 O'Brien, Lister, Fausto, Clifton and Edwards. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Enhancing Innovation and Underlying Neural Mechanisms Via Cognitive Training in Healthy Older Adults

Sandra B. Chapman<sup>1</sup> \* † , Jeffrey S. Spence1† , Sina Aslan1,2 and Molly W. Keebler <sup>1</sup>

<sup>1</sup>Department of Behavioral and Brain Sciences, Center for BrainHealth, The University of Texas at Dallas, Dallas, TX, United States, <sup>2</sup>Advance MRI, LLC, Frisco, TX, United States

Non-invasive interventions, such as cognitive training (CT) and physical exercise, are gaining momentum as ways to augment both cognitive and brain function throughout life. One of the most fundamental yet little studied aspects of human cognition is innovative thinking, especially in older adults. In this study, we utilize a measure of innovative cognition that examines both the quantity and quality of abstracted interpretations. This randomized pilot trial in cognitively normal adults (56–75 years) compared the effect of cognitive reasoning training (SMART) on innovative cognition as measured by Multiple Interpretations Measure (MIM). We also examined brain changes in relation to MIM using two MRI-based measurement of arterial spin labeling (ASL) to measure cerebral blood flow (CBF) and functional connectivity MRI (fcMRI) to measure default mode and central executive network (CEN) synchrony at rest. Participants (N = 58) were randomized to the CT, physical exercise (physical training, PT) or control (CN) group where CT and PT groups received training for 3 h/week over 12 weeks. They were assessed at baseline-, mid- and post-training using innovative cognition and MRI measures. First, the CT group showed significant gains pre- to post-training on the innovation measure whereas the physical exercise and control groups failed to show significant gains. Next, the CT group showed increased CBF in medial orbitofrontal cortex (mOFC) and bilateral posterior cingulate cortex (PCC), two nodes within the Default Mode Network (DMN) compared to physical exercise and control groups. Last, significant correlations were found between innovation performance and connectivity of two major networks: CEN (positive correlation) and DMN (negative correlation). These results support the view that both the CEN and DMN are important for enhancement of innovative cognition. We propose that neural mechanisms in healthy older adults can be modified through reasoning training to better subserve enhanced innovative cognition.

Keywords: innovation, cognitive training, aging, creativity, CBF, functional connectivity, reasoning training, randomized trial

#### INTRODUCTION

Innovative cognition is widely recognized as a vital capacity, undergirding adaptive and flexible thinking. This cognitive domain is of interest in older adults due to its centrality to human cognition, intellect, decision-making, life achievement, resilience and psychological well-being (McFadden and Basting, 2010; Li et al., 2015; Beaty et al., 2016; Heilman, 2016; Palmiero et al., 2016; Saggar et al., 2016). Innovative thinking may be a pivotal cognitive capacity and brain function

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Daria Antonenko, Charité Universitätsmedizin Berlin, Germany Hans-Peter Müller, University of Ulm, Germany

> \*Correspondence: Sandra B. Chapman schapman@utdallas.edu

†These authors have contributed equally to this work.

Received: 09 March 2017 Accepted: 14 September 2017 Published: 09 October 2017

#### Citation:

Chapman SB, Spence JS, Aslan S and Keebler MW (2017) Enhancing Innovation and Underlying Neural Mechanisms Via Cognitive Training in Healthy Older Adults. Front. Aging Neurosci. 9:314. doi: 10.3389/fnagi.2017.00314 allowing one to respond effectively to challenging and constantly changing life demands (Saggar et al., 2016). Cognitive neuroscientists are becoming increasingly interested in elucidating the domain of innovative thinking, its neurobiological underpinnings; and whether this important capability can be enhanced (Fink et al., 2015). Thus, the present study offers one of the first pilot trials: (1) to examine whether innovative cognition can be improved as well as; and (2) to elucidate associated neural changes following cognitive or physical exercise training in healthy older adults.

Innovative thinking purportedly declines even before young adulthood (Kim, 2011) and may worsen with increasing age. Most aging evidence has focused largely on insidious cognitive declines in areas such as executive function, cognitive control and memory as well as losses in both structural and functional aspects of brain systems (Raz et al., 1997; Kennedy and Raz, 2009; Lu et al., 2011). Declines in these domains reportedly accumulate with increasing age even in the absence of frank dementia. The sparse evidence that does exist about age-related changes in innovative cognition is equivocal. Some evidence suggests that innovative thinking may follow the same degradation pattern as other executive functions and memory with a peak in early adulthood followed by accumulating declines starting as young as 30 s to 40 s (Alpaugh and Birren, 1977; McCrae et al., 1987; Reese et al., 2001). Other accounts have challenged this age-related loss pattern, showing preserved innovative cognition with aging (Roskos-Ewoldsen et al., 2008; Greenwood and Parasuraman, 2010). Li et al. (2015) have shown that real life success as reflected in publication productivity in university professors is related to maintaining innovative cognition with increasing age. Other researchers have shown that divergent thinking, one facet of innovative thinking, stabilizes in middle-age and is preserved across the lifespan (Palmiero, 2015) especially when controlling for processing speed (Elgamal et al., 2011).

With regard to age-related decline in brain function, significant changes occur in nodes across two brain networks linked to innovative thinking, namely, the central executive network (CEN) and the Default Mode Network (DMN; Beaty et al., 2016). Specifically, age-related declines are identified on measures of brain function including: (a) reductions in cerebral blood flow (CBF) as measured by arterial spin labeling MRI (ASL MRI) across brain regions (Lu et al., 2011); and (b) reduced functional connectivity in these specific regions (Sambataro et al., 2010; Hafkemeijer et al., 2012; Geerligs et al., 2015). With regard to how brain networks are linked to innovative thinking, the findings are inconsistent. Beaty et al. (2016) reports an inverse correlation between CEN and DMN that is associated with higher performance on innovation (Green et al., 2015; Beaty et al., 2016); whereas Takeuchi et al. (2012) reported increased connectivity between these regions in relation to innovation. Most participants in prior studies were college students. Therefore, it is unclear how the neural and cognitive findings generalize to healthy older adults or to older adults in response to training.

Whether or not innovative cognition can be improved in older adults remains an important issue to address. Clinical trials provide evidence that the neuroplasticity of the aging brain may indeed be harnessed to leverage a perspective shift towards one that refuses to accept the well-documented, insidious age-related loss as a definite outcome of the aging process (Chapman and Mudar, 2014; Rebok et al., 2014). Specifically, research findings reveal that a significant degree of age-related cognitive and brain losses can be halted, reversed or even inoculated against through the building of cognitive and brain reserves to stave off subsequent decline (Mahncke et al., 2006a; Anguera et al., 2013; Rebok et al., 2014; Chapman et al., 2015, 2016; Hohenfeld et al., 2017). Among a variety of opportunities to modify age-related losses, two non-pharmacological intervention-types have been shown to enhance cognition and neural systems, specifically cognitive training (CT) protocols (Levine et al., 2000; Mahncke et al., 2006b; Chapman and Mudar, 2014; Chapman et al., 2015, 2016) and physical exercise regimens (Kramer et al., 1999; Chapman et al., 2013, 2016). We previously reported that reasoning training (SMART©) improved performance on cognitive control measures of complex abstraction and working memory; whereas aerobic exercise improved immediate and delayed memory (Chapman and Mudar, 2014; Chapman et al., 2015, 2016). Linked to these cognitive gains, we also identified corresponding significant increases in resting CBF (Chapman et al., 2016). However, whether the cognitive (SMART©) protocol can improve innovative cognition and neural mechanisms has yet to be investigated in aging populations.

We extend our prior work by addressing whether the CT can also improve innovative cognition, influence brain systems and show correspondence between enhanced innovative cognition and brain changes in the same group of participants. The specific aims of this randomized pilot study were: (a) to determine whether innovative cognition in older adults can be improved through cognitive reasoning training; (b) to compare CBF changes following CT compared to physical training (PT) and wait-list controls; and (c) to elucidate brain mechanisms related to improved innovative thinking in cognitively normal adults (56–75 years of age). We set out to test three questions: would the CT affect: (1) innovative thinking as measured by the Multiple Interpretations Measure (MIM); (2) brain plasticity as measured by resting state CBF; and (3) correspondence between enhanced innovation performance and brain connectivity changes in two prominent brain networks, i.e., DMN and CEN.

### MATERIALS AND METHODS

### Participants

A total of 140 subjects were screened in a multi-stage screening process comprising online, telephone and in-person questionnaires as well as a physical examination to ensure good health, see Supplementary Figure S1 for the consort chart. Participants were adults between the ages of 56 and 75 years, right-handed native English speakers, with at least a high school diploma, no history of psychiatric or neurological conditions, no history of medication changes or surgery entailing general anesthesia within 3 months, and no more than 20 min of aerobic activity, twice per week. The online questionnaire was followed



Two-sample t-test was conducted to assess potential baseline differences. IQ, Intelligence Quotient; MoCA, Montreal Cognitive Assessment; TICS-M, Telephone Interview of Cognitive Status-Modified; BDI, Beckman Depression Inventory; BMI, Body Mass Index; VO<sup>2</sup> Max, maximal oxygen consumption; MIM, Multiple Interpretations Measure to measure innovation; pCASL, pseudo-Continuous Arterial Spin Labeling; fcMRI, functional connectivity MRI.

up by a telephone interview to answer questions about the study, verify online responses, and screen for cognitive status using Telephone Interview for Cognitive Status-M ≥ 28. The third stage comprised of an in-person Intelligence Quotient (IQ) using Wechsler Abbreviated Scale of Intelligence (WASI) ≥ 80, mood screen using Beck Depression Inventory (BDI) ≤ 14 and cognitive status screen using Montreal Cognitive Assessment (MoCA) ≥ 26. Finally, in the fourth stage, a physician examined each participant's physical ability to comply with the study's exercise requirements through an in-person physical assessment of height, weight, waist circumference, Body Mass Index (BMI) < 40, hypertension screen, basic blood test and graded stress test. Specifically, participants underwent a maximal oxygen consumption (relative VO<sup>2</sup> max: mL/kg/min) exercise stress test to assess maximal exercise capacity as well as blood pressure/ECG responses and rating of perceived exertion (RPE) according to the Borg scale, range: 6–20 (Borg, 1990). A repeat of this rigorous assessment was carried out at all three time points during and following the training.

This study was carried out in accordance with the recommendations of Institutional Review Boards (IRB) of University of Texas Southwestern Medical Center, University of Texas at Dallas and The Cooper Institute. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The participants were then randomized using a block randomization schedule stratified by gender into one of three groups: (CT, n = 19), (PT, n = 19) and Wait-listed control (CN, n = 20). All participants in the PT and CT groups were required to complete at least 90% of the training sessions over the 3-month period. No significant differences in age, gender, estimated IQ, MoCA, Telephone Interview of Cognitive Status-Modified (TICS-M) were noted between groups (p > 0.05), as shown in **Table 1**. This study was registered at Clinical Trials.gov, NCT#00977418.

#### Cognitive Training Program

The CT program used in this study is an evidence-based, manualized program focused on enhancing top-down executive functioning: Strategic Memory Advanced Reasoning Training or SMART© (Chapman and Mudar, 2014; Chapman et al., 2015, 2016). For treatment fidelity, the sessions for all participants in the CT group were led by the same clinician, whose three-stage training process included reviewing literature on the program, observing other trained clinicians and leading non-study SMART© training groups under the supervision of a trained clinician. The SMART© training sessions were comprised of 12 1-h in-person small group (n ≤ 5) sessions held once a week for 12 weeks. Additionally, each participant was assigned two 1-h pencil and paper assignments to complete at home each week for a total of 24 h of solo work, for a total of 36 h over the course of the study. In addition to completing the independent assignments, each participant kept a log of the assignments, which included the total amount of time spent and task completion.

SMART© trained and provided practice of three metacognitive strategies for each of the complex cognitive functions of Strategic Attention, Integrated Reasoning and Innovation. As stated in Chapman et al. (2016), Distinct Benefits of Cognitive vs. PT, Strategic Attention is the ability to filter important information from less relevant data which is routinely necessary in life to efficiently manage time and cognitive resources by prioritizing daily goal setting, blocking distractions, intentionally single tasking, and scheduling regular mental breaks during the day. Integrated Reasoning teaches individuals to synthesize information at deeper levels of interpretation by abstracting the essence or extracting key goals for tasks. Strategies for Integrated Reasoning exert cognitive control to ''zoom in'' on the important details or steps to a goal, then rapidly ''zooming out'' to synthesize, and abstract big picture ideas/goals, followed by ''zooming deep and wide'' to construct generalized application of derived ideas, interpretations, or goals-completed. It is a skill that allows one to make informed decisions and solve problems in dynamic and demanding environments. The strategies of Innovation encourage fluid and flexible thinking, perspective taking and problem solving. Innovation focuses on flexibly updating ideas and perspectives and continually seeking ways to improve everyday tasks. These three core strategies were trained in the first 3 weeks of in-person group meetings so that participants could understand the basics of SMART©. The remainder of the training hours, participants practiced generating synthesized ideas and relevant application of the strategies to everyday life. Trainees received feedback from the trainer not only relative to performance on in-session group interactions regarding complex cognitive activities but also regarding their responses to applied activities. SMART trains individuals to approach challenging cognitive tasks with a brain prepared to think deeply, to continually synthesize data encountered daily (e.g., movies, medical information, speeches) and to practice innovative thinking by generating diverse interpretations, solutions and perspectives.

#### Physical Training Program

The PT program used in the study, similar to the CT program, was comprised of three 1-h exercise sessions per week for 12 weeks. Every exercise session of aerobic activity occurred under supervision of trained personnel, an exercise physiologist and a nurse practitioner, with alternate use of treadmill walking and stationary cycling. By monitoring participants every 5 min, the supervising trainers ensured that they maintained 50%–75% of their VO<sup>2</sup> max during the individual sessions. Sessions were structured to include 5-min warm-up and cool-down periods with specified slower speeds and 50 min at the rate necessary to maintain the required VO<sup>2</sup> max For a complete description of both training protocols employed, interested readers are encouraged to reference ''Distinct brain and behavioral benefits of cognitive vs. PT: a randomized trial in aging adults'' (Chapman et al., 2013).

#### Multiple Interpretations Measure (MIM)

A shortcoming of assessment batteries for innovative cognition is the limited ability to measure novelty and relevance of ideas in responses that typify naturally occurring cognitive activities and challenges. Commonly used measures to characterize innovation include a variety of divergent thinking tasks such as Guilford's Alternative Use Task (e.g., list as many different uses of cardboard boxes, a brick, pencil, etc.), other verbal fluency tasks (e.g., list as many words as possible that begin with the letter ''d'' or exclude the letter ''k''), ideational fluency like some of the prompts present in the Torrance Test of Creative Thinking (e.g., ask as many questions as possible regarding a provided image or object, list as many consequences as possible for a given image, list as many improvements as possible for a toy, and as many consequences as possible for impossible scenarios like people no longer needing sleep; Guilford, 1967; Wallach, 1968; Kaufman and Sternberg, 2006; Runco and Acar, 2012; Runco and Jaeger, 2012). Whereas these measures may be informative, these are less common cognitive challenges faced in everyday life in older adults and may lack ecological validity.

For the present study, we utilized The MIM, a subtest of Test of Strategic Learning (TOSL). This test is comprised of three expository texts about an historical person who is unknown but has generalizable life experiences from which distinct high-level themes may be gleaned, e.g., the meaning of success, self-actualization, courage, strength during moments of adversity, etc. One of the three versions was randomly administered at each assessment time point: pre- (T1), mid- (T2) and post-training (T3) periods. The primary scale of the Test of Strategic Learning measures cognitive control of complex abstraction as represented through the ability to understand and synthesize the overall meaning in a synopsis of text, much like you would in the abstract of an article or a synopsis of a movie. The Multiple Interpretation Measure subtest measures an individual's ability to generate multiple interpretations of the expository text, a task motivated by the work of Kaufman and Sternberg (2006). Specifically, participants are asked to construct as many high-level interpretations as possible that can be drawn from the expository texts but which are not explicitly stated.

In this way, the MIM subtest taps the ability to combine presented information with their world knowledge in a multitude of ways. This subtest that solicits multiple interpretations represents a real-life task, similar to what a person could encounter in everyday life when they express a range of ideas and/or solutions. For instance, when engaged in conversation, there are an infinite number of possible interpretations for a movie, political speech, medical scenario dilemma, or future financial advice. These interpretations are self-generated ideas, which are not explicitly conveyed in the texts. Instead, individuals use cognitive control processes to create abstracted responses. Abstracted interpretations require the individual to decipher meanings expressed in the immediate context and combine these meanings with their own experiences and world knowledge to construct plausible and relevant responses.

For scoring purposes, every interpretation was first rated along two dimensions: either high-quality (HQ) or Othertype. Responses were coded as HQ when they were judged to convey generalized/abstracted ideas that showed an ability to combine the meanings from the text within the context of more generalized real world knowledge. In short, HQ responses were those that represented a depth of understanding and synthesis of meaning whereas Other-type responses tended to represent more of a reiteration of literal facts or obvious ideas from the text. To exemplify, one of the texts of the MIM describes the life of a man who was not considered a success during his lifetime in terms of predominant societal measures, who nonetheless in retrospect made incredible contributions to humankind. A specific example of a HQ interpretation is, ''Often the perspective of time can redeem a person's ideas and ideals''. Or ''Empathy can impact the lives of many by creating societal change''. Example of Other-type responses would be ''He had a lot of jobs and failed at them all'' or ''He never seemed to be satisfied with his choices''. The first examples clearly represent synthesized statements that generalize beyond what is explicitly stated in the text whereas the latter responses relate only to the meaning as conveyed in the text.

Three clinicians, trained in the scoring method for the measure, utilized a coding manual with sample responses to make response judgments. Three raters scored the responses separately and were blinded to the participant's group membership and time interval of test, i.e., whether they were scoring T1, T2 or T3. Disagreements on scores were resolved by consensus. Changes in participants' innovative responses over time were determined by comparing the number of HQ interpretations by training group (i.e., T1 to T2 and/or T3).

### MRI Experiment

MRI investigations were performed on a 3 Tesla MR system (Philips Medical System, Best, Netherlands). A body coil was used for radiofrequency (RF) transmission and an 8-channel head coil with parallel imaging capability was used for signal reception. We used different MRI techniques to investigate changes at rest: a pseudo-Continuous Arterial Spin Labeling (pCASL) sequence was used to measure CBF, functional connectivity MRI (fcMRI) was used to assess functional connectivity of the brain. Additionally, a high-resolution T1-weighted image was acquired as an anatomical reference. The details of imaging parameters and their processing techniques are provided below.

Imaging parameters for pCASL experiments were: single-shot gradient-echo EPI, field-of-view (FOV) = 240 × 240, matrix = 80 × 80, voxel size = 3 × 3 mm<sup>2</sup> , 27 slices acquired in ascending order, slice thickness = 5 mm, no gap between slices, labeling duration = 1650 ms, post-labeling delay = 1525 ms, time interval between consecutive slice acquisitions = 35.5 ms, TR/TE = 4020/14 ms, SENSE factor 2.5, number of controls/labels = 30 pairs, RF duration = 0.5 ms, pause between RF pulses = 0.5 ms, labeling pulse flip angle = 18◦ , bandwidth = 2.7 kHz, echo train length = 35, and scan duration 4.5 min. The sequence parameters for fcMRI were FOV = 220 × 220, matrix = 64 × 64, slice thickness = 4 mm, no gap between slices, voxel size = 3.44 × 3.44 × 4 mm<sup>3</sup> , 36 axial slices, TR/TE = 2000/30 ms, flip angle = 70◦ , 120 image volumes, and scan duration = 4 min. The high-resolution T1-weighted image parameters were magnetization prepared rapid acquisition of gradient-echo (MPRAGE) sequence, TR/TE = 8.3/3.8 ms, shot interval = 2100 ms, inversion time = 1100 ms, flip angle = 12◦ , 160 sagittal slices, voxel size = 1 × 1 × 1 mm<sup>3</sup> , FOV = 256 × 256 × 160 mm<sup>3</sup> , and duration 4 min.

pCASL image series were realigned to the first volume for motion correction (SPM5's realign function, University College London, UK). An in-house MATLAB (Mathworks, Natick, MA, USA) program was used to calculate the difference between averaged control and label images. Then, the difference image was corrected for imaging slice delay time to yield CBF-weight image, which was normalized to the Brain template from Montreal Neurological Institute (MNI). This procedure was carried out using a nonlinear elastic registration algorithm, Hierarchical Attribute Matching Mechanism for Elastic Registration (HAMMER, University of Pennsylvania, PA, USA). The HAMMER algorithm detects and corrects for region-specific brain atrophy which is commonly seen in elderly subjects. Last, the absolute CBF was estimated by using Alsop and Detre's equation in the units of mL blood/min/100 g of brain tissue (Aslan et al., 2010).

For voxel-based analyses (VBA), the individual CBF maps were spatially smoothed (with full-width half-maximum (FWHM) of 4 mm) to account for small differences in sulci/gyri location across subjects. For cluster extent inference, we used 3dClustsim in AFNI (NIMH Scientific and Statistical Computing Core, Bethesda, MD, USA), which controls falsepositive activation clusters over the set of all activation clusters throughout the whole-brain volume. We refer to this procedure in the ''Results'' Section as family-wise error correction (FWE corrected). For cluster inference, we tested the volume of clusters, which is conditional on two criteria: smoothness of the voxel map and cluster-defining threshold. We estimated the smoothness to be 9.3 mm FWHM (inherent smoothness plus additional smoothness applied—described above) and set the cluster-defining threshold to the 99.5th percentile of t-statistic distribution. Then, the minimum cluster size of 98 voxels (784 mm<sup>3</sup> ) yielded an FWE-corrected significance level of 0.05.

Functional connectivity images were analyzed by using AFNI (NIMH Scientific and Statistical Computing Core, Bethesda, MD, USA). The dataset was preprocessed with slice timing correction, motion correction (realignment), removal of the linear trend, smoothing by a Gaussian filter with a FWHM of 6 mm and band-pass filtering (0.01–0.1 Hz) to keep appropriate frequency fluctuations. Next, images were spatially normalized to MNI template. In the DMN analysis, we correlated the time series of the orbitofrontal cortex (OFC) and posterior cingulate cortex (PCC) regions (an ROI analysis since the regions were significant in the CBF analysis and are part of DMN). Then, a Pearson correlation was conducted between the time series of PCC and OFC; followed by a z-transformed using Fisher's transformation. In the CEN analysis of functional connectivity, the preprocessed images were analyzed using a seed-based approach by choosing bilateral dorsolateral prefrontal [± 45 + 16 + 45] cortices based on MNI coordinates (Chapman et al., 2015). The cross-correlation coefficient between these seed voxels and all other voxels was calculated to generate a correlation map. Next, the correlation maps were converted to a z-transformed using Fisher's transformation. Last, an ROI analysis was performed based on two CEN regions: dorsolateral prefrontal cortex (DLPFC; composed of BA 9 and 46) and inferior parietal cortex (IPC). The functional ROIs were defined as follows: first, each region's anatomical region was defined based on Talairach Daemon database in AFNI. Then, a functional ROI was defined by choosing the top 200 voxels at each time point (i.e., T1, T2 and T3) and the intersection (i.e., common voxels) of the masks was calculated (Chapman et al., 2015). The CEN Z-Score was calculated by averaging the values of the all four nodes of CEN: L/R DLPFC and L/R IPC.

#### Statistical Analyses

All tests were t-statistic contrasts of parameter estimates from a linear mixed model. The computations were implemented in the R computing language<sup>1</sup> . We modeled HQ innovations as additive effects of training type (CT, CN, control and PT), time of assessment (T1—baseline, T2—mid-training and T3—post-training), and the interaction between type of training and assessment period in a standard linear mixed effects model framework. Two variance components—one due to variability across subjects, and one due to variability in the same subject over time—were included to account for the different

<sup>1</sup>http://cran.r-project.org

Chapman et al. Enhancing Innovation Induces Neural Plasticity

levels of variability and estimated by restricted maximum likelihood. We were primarily interested in how the groups differed across the training sessions. Thus, we hypothesized that the CT group would show a larger positive change in mean number of HQ innovations by T2 and/or T3, relative to the control and PT groups. This hypothesis led to the following one-sided t-statistic contrasts of means from the linear mixed effects model: (1) time contrasts (T23 − T1) for each group, where we define (T23 − T1) = (T2 + T3)/2 − T1 as the ''sustained change'' following training; (2) interaction contrasts (T23 − T1)CT − (T23 − T1)CN and (T23 − T1)CT − (T23 − T1)PT. One additional interaction contrast was also tested as (T23 − T1)CT − (T23 − T1)CN/PT, where CN/PT denotes the average of the two control groups. These six contrasts were tested as single degree-of-freedom t-tests from the linear mixed effects model without multiple comparisons adjustments.

We modeled CBF similarly. That is, in the VBA, we used the same linear mixed effects model for voxel-level CBF as noted above: training type (CT, CN, PT), assessment period (T1, T2, T3), and their interaction. Our hypothesis for CBF was also similar to that of HQ innovations—we hypothesized that the CT group would show a larger positive change in mean CBF by T2 and/or T3, relative to the control and PT groups. For this hypothesis, we tested only the single interaction contrast (T23 − T1)CT − (T23 − T1)CN/PT. We did not, however, hypothesize specific regions of the brain in which we expected this CBF relationship. Therefore, as noted above in the description of our VBA analysis, we FWE corrected through AFNI's cluster-extent inference.

Last, we used separate linear models to assess the HQ innovations/DMN connectivity relationship and the HQ innovations/CEN connectivity relationship. In the first linear model the dependent variable was (T23 − T1) for HQ innovations, and the independent variable was (T23 − T1) for DMN connectivity on a z-transformed scale (described above). In the second linear model, based on the CEN findings of Chapman et al. (2015), the independent variable was ''transient change'' (T2 − T1) for CEN connectivity on a z-scale, and the dependent variable was, similarly, (T2 − T1) for HQ innovations. For both models, training type (CT/CN/PT) was also included as an additive term, as well as the interaction of training type with the independent variable DMN or CEN connectivity change, respectively. From these models, we calculated regression coefficients for each group and tested their respective differences from zero as t-statistics. Additionally, we tested the single degree-of-freedom interaction contrast BCT − BCN/PT, where B denotes the estimated regression coefficient. In the first model our primary hypothesis was that the functional relationship would be restricted to the CT group relative to controls and the PT group, yielding a significant interaction test, but without a directional hypothesis. In the second model, however, our hypothesis was that the HQ innovations/CEN connectivity relationship was positive and restricted to the CT group. This directional hypothesis was also based on the CEN findings of Chapman et al. (2015).

#### RESULTS

All control (CN, n = 20), physical training (PT, n = 19) and cognitive training (CT, n = 19) participants completed the neurocognitive assessments at each time point. However, several participants in the control (CN), CT and physical exercise (PT) groups were not included in the analysis due to incomplete MRI time points, gross movement of >3 mm, and >3◦ and/or artifacts. All physical exercise and CT participants were required to complete at least 90% of training sessions over the 3-month training period, which means they completed 32 h or more of the 36 h of training. One baseline-only measurement for the neurocognitive assessment of one CT participant was removed because it had been scored incorrectly. No participant was excluded due to missing too many sessions to meet the 90% criterion.

#### Neurocognitive Analysis

**Figure 1** displays mean HQ innovations for each group and each assessment period, and **Table 2** lists all the relevant contrasts of interest from the linear mixed model. The CT group shows a significant mean ''sustained increase'' in number of HQ innovations from T1 to T23 (t<sup>109</sup> = 2.23, p = 0.014); whereas the same contrast for the control and PT groups were not significant (t<sup>109</sup> = 0.44, p = 0.33 and t<sup>109</sup> = −0.08, p = 0.53, respectively). Comparing the sustained increase for the CT group with the comparable change in the control group and the PT group (i.e., from T1 to T23), we found that the sustained increase for the CT group is marginally larger than the control group (t<sup>109</sup> = 1.30, p = 0.098) and, similarly, marginally larger than the PT group (t<sup>109</sup> = 1.63, p = 0.053). Averaging the sustained change over the two control groups (i.e., CN and PT), we find that the sustained increase for the CT group is also marginally larger than the average change of the controls and PT groups (t<sup>109</sup> = 1.69, p = 0.047).

FIGURE 1 | High-quality (HQ) innovation results. Mean number of HQ innovations across three assessment periods T1, T2 and T3 (baseline, mid-training, and post-training, respectively). The cognitive training (CT) group shows a sustained improvement in mean HQ innovations (T23 − T1), while the controls (CN) and physical training (PT) groups do not. See text and Table 2 for tests of the relevant contrasts. Error bars indicate 95% least significant intervals for contrasts T23 − T1 by group.


TABLE 2 | HQ innovation results.

Contrasts from the linear mixed model for HQ innovations for cognitive training (CT), control (CN) and physical exercise (PT) groups.

#### CBF Analysis

**Figure 2** shows the results of the interaction contrast described above for the VBA of the CBF maps. That is, we tested whether the sustained increase for the CT was greater than that of the average change between the control (CN) and physical exercise (PT) groups. A significantly larger increase in blood flow was observed at T23 in bilateral medial orbital frontal cortex (mOFC) and bilateral PCC of the CT group compared to the PT/control group, shown in **Figure 2**. Both mOFC and PCC are major nodes of DMN (Fox et al., 2005). **Table 3** summarizes these findings for cluster-level inference as well as descriptive statistics for peak voxel within cluster. Cluster volumes larger than 784 mm<sup>3</sup> (FWE alpha level of 0.05) yield FWE p-values less than 0.05. Our observed cluster volumes for PCC and mOFC are 3792 and 992 mm<sup>3</sup> , respectively.

#### Neurocognitive and Regional Connectivity Relationship

**Figure 3A** shows a scatterplot of the relationship between the sustained change in functional connectivity of the DMN and the sustained change in High Quality innovation scores, coded separately for each group. **Table 4** displays the regression statistics from the linear model. An inverse relationship was found for the CT group (t<sup>36</sup> = −4.57, p < 0.001); whereas the control (CN) and physical exercise (PT) groups showed no significant relationship (CN: t<sup>36</sup> = 1.15, p = 0.25; PT: t<sup>36</sup> = −0.81, p = 0.43). Furthermore, the inverse relationship for the CT group was significant relative to the control (CN) and physical exercise (PT) groups (interaction test in **Table 4**: t<sup>36</sup> = −4.36, p < 0.001).

**Figure 3B** shows a scatterplot of the relationship between the transient change in functional connectivity of the CEN and the transient change in HQ innovation scores, coded separately for

FIGURE 2 | Regional cerebral blood flow (CBF) results. Voxel-based analysis for the interaction contrast described in text, superimposed on an average CBF map of all participants. Both cluster volumes k = 3792 mm<sup>3</sup> for posterior cingulate cortex (PCC) and k = 992 mm<sup>3</sup> for medial orbitofrontal cortex (mOFC) are significant at an family-wise error correction (FWE) alpha level of 0.05 (k = 784 mm<sup>3</sup> ).

TABLE 3 | Regional CBF results.


Regions that showed significant cerebral blood flow (CBF) increase at rest in cognitive training (CT) compared to control (CN) and physical training (PT) groups. The coordinates depict the peak of clusters.

FIGURE 3 | HQ innovation changes in relation to changes in connectivity of default mode network (DMN) and central executive network (CEN). (A) Scatterplot of the sustained change (T23 − T1) in HQ innovation scores against the sustained change (T23 − T1) in DMN connectivity z-scores. The CT group shows a significant negative relationship, while the controls (CN) and PT groups do not show significant relationships between behavior and connectivity. (B) Scatterplot of transient change (T2 − T1) in HQ innovation scores against the transient change (T2 − T1) in CEN connectivity z-scores (outlier removed, see text). The CT group shows a significant positive association, while the controls (CN) and PT groups show no significant relationship. Table 4 displays the regression statistics from both linear models. each group. One subject in the CT group has been removed based on outlier diagnostics (studentized residual = −4.16; outlier test—Bonferroni p-value = 0.010), see Supplementary Figure S2. A positive association was found for the CT group (t<sup>37</sup> = 1.837, p = 0.037); whereas the control (CN) and physical exercise (PT) groups showed no significant relationship (CN: t<sup>37</sup> = −0.239, p = 0.594; PT: t<sup>37</sup> = −0.483, p = 0.684). Furthermore, the positive relationship for the CT group was significant relative to the control (CN) and physical exercise (PT) groups (interaction test in **Table 4**: t<sup>37</sup> = 1.81, p = 0.039). In **Table 4**, we display the regression statistics from the linear model for ∆HQ innovation as a function of ∆DMN and ∆CEN, respectively.

#### DISCUSSION

This randomized pilot study evaluated whether innovative cognition was improved in a group of older adults (56–75 years) in response to CT vs. physical exercise training (PT) or a wait-list control group (CN). In previously published research, we showed that CT improved cognitive control on measures of complex abstraction and working memory; whereas physical exercise enhanced immediate and delayed memory (Chapman and Mudar, 2014; Chapman et al., 2015, 2016). In the current article, we used the same cohorts, but compared a distinct measure from that previously reported, to examine this new question as to whether the CT protocol would also improve innovative cognition. The outcome measure was a novel cognitive innovation task intrinsically related to real life demands, i.e., being able to formulate multiple interpretations for a lengthy expository text.

Our preliminary results can be summarized as three key findings. First, we found that the CT group showed significant gains in high quality innovation performance. In contrast, neither the exercise nor the control group showed significant changes in innovation performance over time. Second, we identified mechanisms related to training-induced brain changes, namely increases in CBF within the CT group only. The CT


Regression statistics from linear model for ∆HQ innovation as a function of (A) ∆ DMN connectivity and (B) ∆ CEN connectivity (∗outlier removed, studentized residual = −4.16, Bonferroni p-value = 0.010) for the cognitive training (CT), control (CN) and physical training (PT) groups.

group showed significant change from baseline bilaterally in the mOFC and the PCC, major nodes in the DMN. Lastly, we found significant associations between changes in high quality Innovation scores and the connectivity of two major neural networks, the CEN and the DMN using resting state fcMRI in the CT group. Specifically, individuals in the CT group with high quality innovation scores showed increased connectivity in CEN nodes (a positive correlation) as contrasted with decreased connectivity in DMN nodes (a negative correlation) on resting state fcMRI.

Overall, the findings support a potential to harness latent innovative thinking capacity and neuroplasticity in a cognitively normal older adult population (56–75 years) with a short-term cognitive reasoning training protocol, namely SMART©. These results add to growing data showing that older adults benefit from different forms of CTs (Mahncke et al., 2006b; Ball et al., 2010; Greenwood and Parasuraman, 2010; Anguera et al., 2013; Hohenfeld et al., 2017). The current study is one of the first known studies to show gains in innovative cognition and corresponding neural networks linked to reasoning training in older adults. Taken together with prior research showing enhanced neurocognitive effects with reasoning training (Chapman et al., 2013, 2016), the present findings support the potential for such training to have broad-based benefits manifested not only on measures of cognitive control but now these results also implicate a potential to improve innovative cognition in middle-age to older adults. This promise of improved innovative cognition capacity in cognitively normal adults warrants further validation in a larger study.

Our evidence suggests that the CT (SMART©) may be deployed to induce an experience-driven neuroplasticity in cognitively normal older adults. This enhanced innovative cognition performance had a direct association with gains in the CEN regions' connectivity but an inverse association with the DMN regions' connectivity in the CT group. The advantageous patterns of connectivity within the CEN and DMN are reinforced by previous evidence linking such a dynamic relation to innovative cognition (Greicius et al., 2003; Beaty et al., 2016). Further evidence that the neural changes reflect positive brain reorganization with CT is supported by the distinct pattern for the CT group only with no significant innovation or neural changes for the physical exercise (PT) and control (CN) groups. Thus, we propose that the change in connectivity of CEN and DMN following reasoning training may represent a redesigned ''healthier neural mechanism'' in older adults that is able to better subserve enhanced innovative cognition. Specifically, continued research toward this effort would help determine if reasoning training builds a more resilient system to counteract failure between two major networks; which previously has been linked to inefficient cognitive performance in healthy and compromised brains (Bonnelle et al., 2011, 2012). This interacting neural pattern between networks is consistent with the claim by Beaty et al. (2016) that innovative thinking engages dynamic interactions of large-scale brain networks, especially the CEN and DMN.

The nature of this complex and dynamic interaction of the CEN and DMN in relation to innovative thinking is equivocal. Jung et al. (2013), concluded in their review article that both increased and decreased brain ''fidelity'' across major brain networks was linked to creative innovation. In contrast, other studies report the opposite inverse innovationconnectivity relation between the two networks in relation to elevated divergent thinking performance, an aspect of innovative cognition (Takeuchi et al., 2012; Benedek et al., 2014; Mayseless et al., 2015; Beaty et al., 2016). Despite the disparity in directionality of CEN and DMN in support of innovative cognition, the consensus supports a dynamic interplay between the two functionally distinct but complementary networks (Jung et al., 2013). We propose that the complex operations of innovative thinking are not isolated to single neural hubs, but rather are supported through the involvement of at least two brain networks of DMN and CEN.

A number of factors may contribute to this seemingly disparate pattern across studies such as: (1) the nature of the innovative paradigm; (2) resting-state vs. task-induced studies; (3) age of participants; and (4) single time point measurement vs. longitudinal measurement in response to an intervention. First, the measure of innovation that we used in the present study is distinct from those used in prior work. Our innovative task taps top-down processes, drawing upon controlled retrieval of information, combining and integrating the selected ideas with world knowledge to generate and create a multitude of abstract interpretations. Second, different mechanisms are tested when comparing resting-state vs. task-induced brain imaging. The majority of studies that have shown increased DMN with higher creativity were task-induced studies whereas ours was resting-state. Third, previous innovation-connectivity patterns were identified in younger adults (ages 19–36 years), which may not necessarily be comparable to an older group (ages 56–75 years). Last, we were interested in neural changes following a 12-week CT protocol whereas many of the prior findings examined a single time point, with a few exceptions involving young adults (Fink et al., 2015; Saggar et al., 2016). We conclude that the functional changes in two neural networks relative to innovative cognition following training leaves a footprint in the resting state networks to better support enhanced innovative cognition in the aging brain.

This pilot study must be interpreted in the context of a number of limitations. First, the present task lacks the degree of validation of prior tasks used to measure divergent thinking, namely tasks which prompt for as many alternative uses of an object, (i.e., a ''tissue'') as designed decades ago by Guilford (1967). Whereas we recognize this is a limitation and are in the process of establishing its validity; we propose that the task of deriving multiple interpretations for commonly encountered information may be a practical, functional task that is related to higher-order cognitive capacities that may have ecological validity. Other limitations include small sample size and lack of follow-up after training ended to shed light on the persistence of the gains. We were able to address whether this particular sample enhanced their performance from baseline. However, we were not able to evaluate whether individuals regained lost capacity or perhaps may be able to maintain and mitigate declining innovative cognition in the ensuing years. Addressing these latter issues requires longitudinal studies and quite possibly proactive interventions along the way to test whether declining abilities can be strengthened at life stages where decline emerges. Another possibility to consider for subsequent research is whether this older adult group achieved a level of performance that was superior to how they would have performed as a younger version of themselves. Some evidence suggests that the older mind may be able to take advantage of prior experience to engage in innovative cognition.

#### CONCLUSION

The objective of the present study was to examine the effects of CT on innovative cognition in older adults. This study revealed that reasoning training via SMART improved innovative cognition which correlated to the key nodes of CEN and DMN networks. In sum, the current findings suggest that short-term and cost effective interventions, such as CT, may be beneficial in enhancing cognitive capacities and supporting neural mechanisms in healthy older adults. The new finding related to improved innovative cognition in healthy older adults is heartening; given innovative thinking is one of the most valued assets and fruitful outputs of the human mind across the lifespan (Kaufman and Sternberg, 2006; Palmiero et al., 2016). The potential to strengthen innovative cognition may tap into a positive and valuable resource of the aging mind that could support an individual's ability to reinforce and retain an active mental lifestyle, engage in complex decision-making, intellect and psychological

#### REFERENCES


well-being with advancing age (Baltes et al., 1999; Kaufman and Sternberg, 2006). Much work needs to be done but this feasibility study motivates a continued push to harness the potential and reduce the gap of cognitive brain decline as we age.

#### AUTHOR CONTRIBUTIONS

SBC designed the study, interpreted the data and drafted the manuscript; JSS performed statistical analysis and drafted the manuscript; SA performed neuroimaging analysis, interpretation of neuroimaging data and reviewed the manuscript; MWK performed cognitive assessments and reviewed the manuscript.

#### FUNDING

This work was supported by a grant from the National Institute of Health (RC1-AG035954, R01-NS067015, R01-AG033106) and by grants from the T. Boone Pickens Foundation, the Lyda Hill Foundation and Dee Wyly Distinguished University Endowment. Funding to pay the Open Access publication charges for this article was provided by funds from the Dee Wyly Distinguished University Endowed Chair held by SBC.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi.20 17.00314/full#supplementary-material


study. Proc. Natl. Acad. Sci. U S A 103, 12523–12528. doi: 10.1073/pnas.06051 94103


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Chapman, Spence, Aslan and Keebler. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Cognitive Vulnerability in Aging May Be Modulated by Education and Reserve in Healthy People

María D. Roldán-Tapia\*, Rosa Cánovas, Irene León and Juan García-Garcia\*

Department of Psychology, University of Almería, Almería, Spain

Aging is related to a deterioration of cognitive performance and to multiple alterations in the brain. Even before the beginning of a noticeable cognitive decline, the framework which holds cognitive function experiences these alterations. From a systemvulnerability point of view of cognition, the deterioration associated with age would be the collection of repercussions during a life. Brain function and structure are modified in a multidimensional way, which could concern different aspects like structural integrity, functional activity, connectivity, or glucose metabolism. From this point of view, the effects of aging could affect the most brain systems and their functional activity. In this study, we analyze the functional development of three cognitive domains in relation to aging, educational level, and cognitive reserve (CR). A total of 172 healthy subjects were divided into two age groups (young and old), and completed a battery of classic neuropsychological tests. The tests were organized and analyzed according to three cognitive domains: working memory and flexibility, visuoconstructive functions, and declarative memory. Subjects also completed a questionnaire on CR. Results showed that the performance in all cognitive domains decreased with age. In particular, tests related to working memory, flexibility, and visuoconstructive abilities were influenced by age. Nevertheless, this effect was attenuated by effects of education, mainly in visuoconstructive domain. Surprisingly, visual as well as verbal memory tests were not affected either by aging, education, or CR. Brain plasticity plays a prominent role in the aging process, but, as other studies have shown, the plasticity mechanism is quite different in healthy vs. pathological brains. Moreover, this plasticity brain mechanism could be modulated by education and CR. Specially, cognitive domains as working memory, some executive functions and the visuoconstructive abilities seem to be modulated by education. Therefore, it seems to be crucial, to propose mechanisms of maintenance of a healthy and enriched brain, since it promotes auto-regulatory mechanisms of well-aging.

Keywords: well-aging, educational attainment, cognitive reserve, brain compensation, neuroplasticity, cognitive domains

## INTRODUCTION

As human life expectancy increases, maintaining cognitive function has become a target to pursue. It seems that a combination of educational and occupational fulfillment, recreational activities [all aspects included in the concept of cognitive reserve (CR)], and the biological process of aging itself could be the predictor of how to develop cognitive performance as we age.

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Dina Di Giacomo, University of L'Aquila, Italy Carlo Semenza, Università degli Studi di Padova, Italy Frederick Robert Carrick, Bedfordshire Centre for Mental Health Research in Association with University of Cambridge, United Kingdom

#### \*Correspondence:

María D. Roldán-Tapia mdroldan@ual.es Juan García-Garcia jgarciag@ual.es

Received: 13 March 2017 Accepted: 09 October 2017 Published: 24 October 2017

#### Citation:

Roldán-Tapia MD, Cánovas R, León I and García-Garcia J (2017) Cognitive Vulnerability in Aging May Be Modulated by Education and Reserve in Healthy People. Front. Aging Neurosci. 9:340. doi: 10.3389/fnagi.2017.00340

In recent decades there has been an increasing interest in finding out why some older adults are able to maintain adequate cognitive abilities despite age, while others show clear cognitive decline with advancing age (Wilson et al., 2002). The concept of CR accounts for individual differences in which people with higher CR are able to enlarge the performance on cognitive tasks by recruiting different brain systems and/or using alternative cognitive strategies compared to individuals with lower CR (Stern et al., 2003; Stern, 2009; Tucker and Stern, 2011; Meng and D'Arcy, 2012).

In recent decades it has been demonstrated that age do not affect cognitive functions in the same way nor do they evolve at the same rate throughout life. Processing speed is one of the cognitive abilities that most consistently decline with age (Holdnack et al., 2013; Hong et al., 2015; Tam et al., 2015), even when motor dexterity is statistically controlled (Ebaid et al., 2017). Detriment with age in the case of working memory and inhibition has been also largely demonstrated (Turner and Spreng, 2012; Calso et al., 2016). With advancing age inhibitory mechanisms become increasingly less efficient and no longer prevent irrelevant information from saturating working memory capacity (De Beni et al., 2007). Normal aging seems also to be characterized by a general reduction in cognitive flexibility, defined as attention switching and task shifting. In fact, it has been argued that correlations between cognitive flexibility and primary mental abilities were relatively stable across the adult lifespan, suggesting that individual differences in cognitive flexibility could be responsible for age-related changes in other cognitive abilities (Hülür et al., 2016).

Finally, the extant literature regarding declarative memory indicates that episodic and semantic memory measures are associated with differential aging patterns across lifespan (Nyberg et al., 2012). Thus, several cross-sectional studies have stablished an early onset for episodic memory decline following a linear pattern that begins around the age of 20 or 30 years and results in as much as 1 standard deviation (SD) unit below peak level performance by the age of 60 years and 2 SD units at age 80 years (Schaie, 1994; Nilsson et al., 1997; Verhaeghen and Salthouse, 1997; Park et al., 2002). Longitudinal studies, however, suggest that episode memory decline begins by the age of 60 years followed by an accelerated deterioration (Schaie, 1994; Zelinski and Burnight, 1997). Nevertherless, that discrepancy may be explained by the increase of the educational level in the last decades. On the other hand, semantic memory or knowledge is among the more stable memory systems across the adult lifespan (Park and Reuter-Lorenz, 2009). Young to young–old adults (from age 35 to age 65 years) maintain or even slightly improve their semantic performance, and deficits appear only at very late ages in this domain (Kaufmann, 2001).

Advancing knowledge of influence of CR and educational attainment on aging can be very hopeful since, although the brain will be modified by age, there are numerous opportunities to adapt, restrict, or improve cognitive changes, allowing older adults continue to function independently. Therefore, the purpose of this study is to examine the effect of education and CR in the process of healthy brain aging. We seek to discover whether aging follows similar patterns in different cognitive domains, and how these variables might influence the aging process.

#### MATERIALS AND METHODS

The study was approved by the Ethics Committee of the University of Almeria, and conducted in accordance with Helsinki declaration and Spanish legislation on personal data protection. Participation was voluntary and all subjects gave written consent.

#### Subjects

A sample of 140 subjects was recruited from social clubs, entertainment centers, and the University of Almeria's Center for Adult Education.

The sample was divided into two age groups according to the traditional Spanish retirement age (65 years): adults (aged 36–64 years) and elderly adults (≥65 years). None of them had a history of psychiatric or neurological disorders, drug consumption, or head injury that could potentially affect their cognitive performance. Besides, for all subjects older than 64 years a score of 27 or lower in the Mini-Examen Cognoscitivo [Mini-Mental State Examination (Lobo et al., 2002)] was a criterion of exclusion. Following these criteria, three men were excluded and four women (one of them with fibromyalgia diagnosis) under the suspicion of suffering Mild Cognitive Impairment. Additionally, the elderly adults completed the Barthel index (Mahoney and Barthel, 1965). **Table 1** shows the sociodemographic characteristics of the participants.

#### Methods

Subjects completed the cognitive reserve scale (CRS), developed by the authors, that measures the reserve throughout a person's lifetime by means of taking part in cognitively stimulating activities (reading, playing a musical instrument, collecting things, speak several languages or dialects, traveling, or play sports) (León-Estrada et al., 2017). Each item was completed several times about different age periods: the older the participant, the more times each item had to be completed. A Likert-type scale of 0–4 points was used and the total CRS score was the sum of the mean scores for each item (24 items). The

TABLE 1 | Participant demographics, cognitive reserve scale (CRS) scores (León et al., 2014), and descriptive scores for elderly adults.


CRS gave scores ranging from 0 to 96, with higher CRS scores indicating more frequent participation. The CRS is available upon request from the authors and additionally, it is available on the Internet: http://www2.ual.es/CognitiveReserveScale/thecognitive-reserve-scale-crs/.

#### Neuropsychological Assessment

A trained psychologist carried out the neuropsychological assessments in several cognitive domains: Working memory and flexibility, Digit Span subtest (backward) (Peña-Casanova et al., 2009c), the Stroop test (Peña-Casanova et al., 2009b), TMT-B (Peña-Casanova et al., 2009c), Controlled Oral Word Association Test (COWAT) (Benton and Hamsher, 1989), and the Corsi Block task (backward) (Peña-Casanova et al., 2009c), visuoconstructive abilities: matrix reasoning and Block Design subtests (Wechsler, 1993), and Rey–Osterrieth Complex Figure Test (ROCF) (quality of copy) (Peña-Casanova et al., 2009a), and declarative memory: Verbal Learning Spanish–Complutense Test TAVEC (Benedet and Alejandre, 1968), sum of the learning slope, short-term recall and delayed-memory, and ROCF short-term recall and delayed-memory. **Table 2** shows the main scores of neuropsychological tests.

In addition, subjects' IQ score was estimated by Vocabulary subtest (Wechsler, 1993). All the tests are translated and validated in Spanish and the raw scores of all tests were converted to standard scores adjusted for age and educational level (scale scores, z-scores, or percentile scores) following Spanish normative studies, except for COWAT, which was standardized following a normative study in an English population (Benton and Hamsher, 1989).

#### Statistical Analysis

In the present study, data were analyzed with multivariate analyses of covariance (MANCOVA). A total of three MANCOVAs were conducted separately to investigate the effect of age on the cognitive domains. Three dependent variables were used: the neuropsychological performance in the cognitive domains previously mentioned. Years of formal education and CR were entered as covariates for all statistical analyses to control for individual differences. Age was used as independent variable and divided into two age groups: adults (36–64 years) and elderly adults (≥65 years).

As statistical assumptions underlying the MANCOVAs were not fully met, the estimation of the parameters in the linear model was analyzed using the resampling method of simple bootstrapping with 1000 bootstrap samples. The bootstrap bias-corrected accelerated method was used as a corrective method. Analyses were carried out using statistical package IBM-SPSS 22.0 for Windows (IBM Corp., 2013).

Results with p < 0.05 were considered statistically significant. The effect size was obtained by using the partial eta squared (η 2 p ) and r of regression coefficient (B).

### RESULTS

In relation to working memory and cognitive flexibility [Digit Span subtest (backward), the Stroop test, TMT-B, COWAT, and the Corsi Block task (backward)], there were significant differences between adults and elderly adults (3Wilks = 0.874,

TABLE 2 | Multivariate analyses of covariance (MANCOVA) classified according to the three functional areas (anterior, posterior and temporal).


Years of formal education and cognitive reserve were entered as covariates for all statistical analyses to control for individual differences. Age was used as independent variable and divided into two age groups: adults (36–64 years) and elderly adults (≥65 years). TMT-B, trail making test part B; COWAT, controlled oral word association test; ROCF, Rey–Osterrieth complex figure test; TAVEC, Verbal Learning Spanish-Complutense Test; CR, cognitive reserve; NS, non-significant.

F5,<sup>130</sup> = 3.735; p = 0.003; η 2 <sup>p</sup> = 0.126). The global effect of the covariates (education and CR) was not statistically significant.

Additionally, estimation of the effect of the independent variables and covariates on the dependent variables showed significant relationships between the Stroop test and age group (B = −1.427; p = 0.001; r = 0.28), education (B = −0.098; p = 0.049; r = 0.16), and CR (B = 0.49; p = 0.007; r = 0.23), as well as between COWAT and age group (B = −13.70; p = 0.019; r = 0.20), and the Corsi Block task (backward) and age group (B = 0.98; p = 0.07; r = 0.15) (**Table 2**).

In the case of the visuoconstructive abilities [the Matrix Reasoning, Block Design, and ROCF (quality of copy)], there were statistically significant differences between adults and elderly adults (3Wilks = 0.847, F3,<sup>130</sup> = 7,817; p < 0.01; η 2 <sup>p</sup> = 0.15). The effect of education was significant (3Wilks = 0.722, F3,<sup>130</sup> = 16.645; p < 0.01; η 2 <sup>p</sup> = 0.278), but no significant effect from CR.

The estimation of the effect of the independent variables and covariates on the dependent variables using simple bootstrapping disclosed significant relationships between the Matrix Reasoning subtest and age group (B = −1.75; p = 0.001; r = 0.28) and education (B = 0.14; p = 0.007; r = 0.23); the Block Design subtest and age group (B = −2.45; p = 0.001; r = 0.28) and education (B = −0.187; p = 0.001; r = 0.28), and the ROCF (copy) and education (B = −0.231; p = 0.001; r = 0.28) (**Table 2**).

Regarding the memory tests (TAVEC sum, TAVEC short-term recall, TAVEC delayed recall, and ROCF short-term recall and long-term recall), there were no significant differences between adults and elderly adults. Besides, the effect of the two covariates (education and CR) was not statistically significant.

Finally, estimation of the effect of the independent variables and covariates on the dependent variables using simple bootstrapping revealed significant relationships between TAVECsum and CR (B = 0.017; p = 0.045; r = 0.17), TAVEC short-term recall and CR (B = 0.021; p = 0.014; r = 0.21), TAVEC delayed recall and CR (B = 0.019; p = 0.036; r = 0.18), and ROCF delayed recall and education (B = −0.149; p = 0.005; r = 0.24) (**Table 2**).

#### DISCUSSION

In the present study, a functional analysis of brain has been performed using neuropsychological tests in two agedifferentiated populations. The aim was to assess the evolution of cognitive changes across the life cycle and their relation to educational level and CR.

Our results, generally speaking, show a large influence of aging on inhibition, flexibility, working memory, and visuoperceptive functions, although there is also an important contribution of education in these domains. However, surprisingly, the same pattern was not found in declarative memory tasks where the effect of education and CR but not of aging could be verified. It seems clear that at least in a healthy population, an enriched environment allows us to face the tasks of flexibility, updating, monitoring, or memory as we get old.

Given our results in cognitive domains related to the frontal region, the data in the flexibility, inhibition, and spatial working memory tests have shown an influence of aging, while in inhibition tasks there is also a significant influence of education and reservation.

In the case of frontal areas, it is now well-established that normal aging is associated with cognitive decline (Craik and Salthouse, 2000; Rypma and D'Esposito, 2000; Kennedy et al., 2009), particularly in tasks involving executive functions (Salthouse et al., 2003; Podell et al., 2012). However, all executive functions do not decline in the same way with advanced age (Collette and Salmon, 2014). A previous study from our group (Roldán-Tapia et al., 2012) reported that the "performance of functions related to the dorsolateral prefrontal cortex (PFC) (verbal fluency, behavioral spontaneity, reasoning, divided and complex attention, and working memory) was associated with aging." Or, for example, regarding to shifting abilities, older people would show difficulties to maintain and to manipulate two mental plans but not to alternate between both of them (Kray et al., 2004). However, these results are not consistent with Schaie (1958, 1994) and Stawski et al. (2013) works, that show a stable performed in flexibility during lifespan (although it is true that they did an intra-subjects analyses).

In addition, numerous researches have shown how CR and education exert a protective effect on the decline associated with age in executive functions (Ardila et al., 2000; Opdebeeck et al., 2016).

On the other hand, the effect of aging on the visuoconstructive domains has also been demonstrated earlier, mainly in face recognition, mental rotation, and visuospatial abilities (Adduri and Marotta, 2009; Iachini et al., 2009; Daniel and Bentin, 2012). It seems that the occipital decrement and sensory deficits are consistent with perceptual processing declines as a function of aging (Zhuravleva et al., 2014).

The relationship between cognitive decline that occurs in the elderly, and age-related changes that take place in brain morphology and functioning, is still being documented.

Virtually no area of the brain is fully spared from the effects of aging, although certain brain systems seem to be particularly vulnerable to aging effects, which takes place earlier and in greater degree (Raz, 2000). From a system-vulnerability approach, the deterioration associated to age would be the collection of repercussions during a life. Brain structure and function are modified in a multidimensional way, which could affect different aspects such as structural integrity, functional activity, and connectivity, as a result of continuous interactions between endogenous and exogenous factors (Khachaturian, 2011; Jagust, 2013). The areas found to be most vulnerable to normal age changes are the PFC and the medial temporal cortex (Raz, 2000, 2004; Buckner, 2004; Hedden and Gabrieli, 2004).

But perhaps the most relevant fact is the influence of education on complex visuoconstructive tasks such as spatial reasoning, visuoconstruction, and the grafomotor integration of a complex design. Our data are in line with previous studies such as that of Ardila et al. (2000) which points out the importance of education for the execution of cognitive processes, especially in older people and even seems to play a relevant role in elderly "illiterates."

The results of the present study are also supported by data from other studies. For example, Kim et al. (2015) demonstrated

a protective effect of education on the cortical thinning in cognitive normal older individuals. According to the authors, this protective effect could be achieved by increasing resistance to structural loss from aging.

In memory domains, our results point to the idea of certain stability across different age periods. This result is not new; other studies have supported the idea of maintenance of different types of memory (semantic memory) in healthy subjects (Park and Reuter-Lorenz, 2009). For example, the perirhinal cortex appears to undergo little age-related atrophy (Insausti et al., 1998). Also, performance in older adults seem be associated with the recruitment of additional brain regions in the medial temporal lobe and in frontal regions (Cabeza, 2001; Maguire and Frith, 2003; Giovanello and Schacter, 2012). It is no less true, however, that a group of studies have shown hippocampal degeneration (Raz, 2004; Raz and Rodrigue, 2006) or a reduction in BOLD signal in the medial temporal lobe (Maguire and Frith, 2003).

While in the case of the working memory, flexibility, and visuoconstructive domains, we do not find any relevant influence from CR in the aging process, the same cannot be said for mnesic processes. Structural MRI studies in healthy aging consistently reported positive associations between CR and increased gray and white matter volumes in associative frontal and temporoparietal cortices, as well as reduced mean diffusivity in the bilateral hippocampus (Akbaraly et al., 2009; Marioni et al., 2012). Declarative memory is greatly influenced by education, primarily, and also by CR, as well in short-term memory and delayed recall, as using visual and verbal stimuli. Valenzuela et al. (2008) found that hippocampal decline occurs more slowly in normal aging when performing cognitive challenging life.

It has also been shown that greater CR contributes to delay or attenuate pathological changes such as those that occur in Alzheimer's disease (Carnero-Pardo and Del Ser, 2007), vascular injury (Dufouil et al., 2003; Elkins et al., 2006), Parkinson's disease (Glatt et al., 1996), traumatic brain injury (Kesler et al., 2003), neuropsychiatric disorders (Barnett et al., 2006), and multiple sclerosis (Sumowski et al., 2009); greater CR may even prevent accumulation of amyloid plaque (Landau et al., 2012). Observations using cerebral blood flow (CBF) have found that Alzheimer's disease patients with higher education have a lower resting rCBF (Stern et al., 1992).

In individuals with AD and mild cognitive impairment, higher education and occupation (as proxies of CR) are correlated to more severe hypometabolism in temporoparietal areas (memory and visual abilities) and in the precuneus (Garibotto et al., 2008). Moreover, they are associated with an increased metabolism in the dorsolateral PFC (working memory and flexibility), suggesting a compensatory mechanism against ADrelated cerebral neurodegeneration (Grady et al., 1994). A review paper calculated that a higher CR reduced the chances of suffer dementia by 46% (Valenzuela and Sachdev, 2005). However, even higher CR cannot maintain functions when pathology turns very severe.

Could it be that education affects cognitive abilities (processing speed, working memory, verbal fluency, or verbal episodic memory) differently?

Maybe, in demanding memory or complex visuoperceptive tasks, the influence of variables such as reserve and education is more marked in the case of healthy aging or even more visible in the case of pathological aging than in other cognitive domains. The domain of memory could be more sensitive to the effect of training, education, and lifestyle.

Our results show that, in healthy people, if we keep level of education and CR constant, the only variable that can influence the neuropsychological outcome in flexibility, working memory, and visuoconstructive domain, seems to be aging. Nevertheless, our data are collected from subjects who were healthy and not very advanced in age. It may be that the possible effect of CR depends on the baseline of normal versus pathological brain conditions. From this point of view Hayden et al. (2011) defined "reserve" as "the difference between cognitive performance as predicted by an individual's brain pathology and that individual's observed cognitive performance." Thus, people whose measured cognitive performance is better than predicted by their pathology have high reserve, whereas those who perform worse than predicted have low reserve.

Regarding the age of our study participants, a 15-year longitudinal study among elderly Catholic clergy members who were participating in the religious Orders (healthy aging) showed a typical profile of extraordinarily slow deterioration of cognitive functions, which last for decades. These findings suggest that cognitive changes associated with aging may be minimal when the process is not associated with a neurodegenerative disease (Christensen et al., 1994; Hayden et al., 2011; Reed et al., 2011).

Both our results and those from morphological and cognitive studies lead us to the idea that "brain aging" is an interactive and synergistic process in which several variables play an important role.

For example, the health condition versus pathology, the genetic load, and the influence of an enriched environment (CR and education). A good method to further investigate this interactive process is the use of brain mapping approaches to complete information obtained through the neuropsychological assessment of cognitive domains (Agis and Hillis, 2016) such as the ones described in the present study (working memory, flexibility, visuoconstructive functions, and declarative memory).

Additionally, the new developments of neuropsychological brain mapping batteries seem to contribute to enhance clinical diagnosis, pre-surgical mapping, and follow-up (Karakas et al., 2013). Hence, advances in the knowledge of neural substrates involved in cognitive tests and revealed by the interpretation of human brain mapping studies represent a strong resource to increase the comprehension of brain and how age and different levels of education or CR might affect cognitive vulnerability.

#### Limitations

Limitations of the current study are the size of the sample and the lack of inclusion of people with presumably low reserve (without access to education or social club). The present sample seem likely to have a higher than average CR, but the CRS includes a high variety of different activities that are not limited exclusively to the centers where the participants were recruited. Besides, in the elderly adults, the SD related to the years of formal education

reflects variability among them (from just 5–15 years of formal education). However, future studies should include an older sample, as well as specific experimental tasks, to help discern the real influence of CR in healthy aging populations.

As a final conclusion, regarding the importance of CR in the domain of memory, its limited effect on fluency, divided attention, interference, spatial reasoning, and visuospatial tasks may be due to the brain's own compensation mechanism, which under healthy conditions, makes it independent of education or CR. It is the effect of CR could be quite different depending of the brain status: pathological or healthy.

#### REFERENCES


#### AUTHOR CONTRIBUTIONS

Conceived and designed the experiments: MR-T, JG-G, and IL. Performed the experiments: IL. Analyzed the data: JG-G and IL. Wrote the paper: RC, MR-T, and JG-G.

#### ACKNOWLEDGMENT

This work was supported by the Ministry of Culture, Education and Sport, Government of Spain, under Senior Mobility Program Grant (PRX16/00362).


perceptual priming and learning. Neuropsychology 23, 475–491. doi: 10.1037/ a0015377


letter and number sequencing, trail making test, and symbol digit modalities test. Arch. Clin. Neuropsychol. 24, 321–341. doi: 10.1093/arclin/acp038


Verhaeghen, P., and Salthouse, T. A. (1997). Meta-analyses of age-cognition relations in adulthood: estimates of linear and nonlinear age effects and structural models. Psychol. Bull. 122, 231–249. doi: 10.1037//0033-2909.122. 3.231

Wechsler, D. (1993). Escala de Inteligencia de Wechsler Para Adultos. Madrid: TEA.


attention in the very old. J. Clin. Neurophysiol. 31, 261–271. doi: 10.1097/WNP. 0000000000000056

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Roldán-Tapia, Cánovas, León and García-Garcia. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# "Cerebellar Challenge" for Older Adults: Evaluation of a Home-Based Internet Intervention

Zoe Gallant and Roderick I. Nicolson\* †

Department of Psychology, University of Sheffield, Sheffield, United Kingdom

There is converging evidence that maintenance of function in the multiple connectivity networks involving the cerebellum is a key requirement for healthy aging. The present study evaluated the effectiveness of a home-based, internet-administered "cerebellar challenge" intervention designed to create progressive challenges to vestibular function, multi-tasking, and dynamic coordination. Participants (n = 98, mean age 68.2, SD 6.6) were randomly allocated to either intervention (the cerebellar challenge training for 10 weeks) or no intervention. All participants undertook an initial series of pre-tests, and then an identical set of post-tests following the intervention period. The test battery comprised five suites of tests designed to evaluate cognitive-sensori-motor-affective functions, including Physical Coordination, Memory, Language Dexterity, Fluid Thinking and Affect. The intervention group showed significant pre- to post improvements in 9 of the 18 tests, whereas the controls improved significantly on one only. Furthermore, the intervention group showed significantly greater improvement than the controls on the "Physical Coordination" suite of tests, with evidence also of differential improvement on the Delayed Picture Recall test. Frequency of intervention use correlated significantly with the improvement in balance and in peg-moving speed. It is concluded that an internet-based cerebellar challenge programme for older adults can lead to benefits in balance, coordination and declarative memory. Limitations and directions for further research are outlined.

Keywords: declarative memory, cerebellum, hippocampus, sensorimotor, balance, vestibular stimulation, functional networks

### INTRODUCTION

The brain and body form a complex, self-regulating system capable of coping with a range of environmental and cognitive challenges, together with the pervasive, age-related progressive impairment in function of many system components. In this article we develop the perspective that the functional networks involving the cerebellum represent a significant part of the degradation in aging. We then briefly review the many interventions that have proved efficacious with older adults, noting the current consensus that multi-component systems designed to maintain a progressive challenge appear to have greater effect than single component systems. On theoretical grounds we argue that interventions designed around ''cerebellar challenge'', combining coordinative exercise with cerebellar stimulation, should prove particularly effective. We finish by presenting an evaluation of an internet-based cerebellar challenge system, Zing, in terms of its effectiveness compared with a life-as-usual control group.

#### Edited by:

Panagiotis D. Bamidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Richard Camicioli, University of Alberta, Canada Manousos A. Klados, Aston University, Birmingham, United Kingdom

#### \*Correspondence:

Roderick I. Nicolson rod.nicolson@edgehill.ac.uk

#### †Present address:

Roderick I. Nicolson, Department of Psychology, Edge Hill University, Ormskirk, United Kingdom

Received: 02 August 2016 Accepted: 27 September 2017 Published: 27 October 2017

#### Citation:

Gallant Z and Nicolson RI (2017) "Cerebellar Challenge" for Older Adults: Evaluation of a Home-Based Internet Intervention. Front. Aging Neurosci. 9:332. doi: 10.3389/fnagi.2017.00332

Traditional approaches to the causes of cognitive decline with aging considered primarily the frontal lobes (Jackson, 1958; Dempster, 1992; Greenwood, 2000). Over the past three decades there has been an explosion of research on all aspects of aging. Early this century extensive research was undertaken on changes in brain structure with aging (Raz et al., 2005), genetics (Deary et al., 2004; Erraji-Benchekroun et al., 2005), together with risk factors including increased white matter (Bartzokis, 2004; Head et al., 2004; Westlye et al., 2010; Sexton et al., 2014); excess homocysteine (Schafer et al., 2005) and reductions in dopamine and acetylcholamine neurotransmitters (Castner and Goldman-Rakic, 2004; Sarter and Bruno, 2004; Erixon-Lindroth et al., 2005).

Following these discoveries, arguably the greatest recent development has been the change of emphasis from these individual components and processes of the aging brain to consideration of the brain as a whole system. A major recent development in cognitive neuroscience has been the development of techniques for determining functional connectivity (Greicius et al., 2003; Fox et al., 2005; Buckner et al., 2008), and the consequent identification of a range of intrinsic networks (Yeo et al., 2011). The approach has great potential for characterizing the connectivity problems that affect brain function. A recent review (Bamidis et al., 2014) highlights the key role of connectivity changes in brain aging, and its implications for assessment and intervention.

It is notable that the cerebellum is also involved in seven of the major intrinsic networks (Buckner et al., 2011; Bernard et al., 2012; Kipping et al., 2013). It is therefore particularly interesting that circuits involving the cerebellum are strongly affected by age (Seidler et al., 2010; Balsters et al., 2013; Bernard et al., 2013; Humes et al., 2013; Bernard and Seidler, 2014; Koppelmans et al., 2015). Furthermore, it appears that the pattern of cerebellar degeneration with age in healthy adults is analogous to that shown by cerebellar patients (Hulst et al., 2015).

It is long established that there are major declines with age in sensory function (Humes et al., 2013; Wayne and Johnsrude, 2015; Roberts and Allen, 2016), in motor function (Seidler et al., 2010) and proprioceptive function (Goble et al., 2009). The cerebellum is centrally involved in sensorimotor processing (Chadderton et al., 2004, 2014; Ramakrishnan et al., 2016) and the involvement of the cerebellum in cognitive function is now fully established (Balsters et al., 2013; Mariën et al., 2014), as are direct, two-way links between the cerebellum and not only motor cortex but also prefrontal and posterior parietal cortex and the basal ganglia (Strick et al., 2009; Bostan et al., 2013).

Taken together, these results converge on the hypothesis (Bernard and Seidler, 2014) that the cerebellum—given its pervasive connectivity, its involvement in multiple sensory, cognitive and motor circuits; and its central role in adapting to internal changes—may be a critical component in the system degradation with age. This new conceptualization offers the promise that interventions designed to maintain or enhance cerebellar function may alleviate the affects of aging on sensorimotor-cognitive performance.

There are many successful interventions for alleviating age-related decline. A recent review (Ballesteros et al., 2015) focused on three modes of intervention: physical activity, computerized cognitive training and social enhancement and concluded that although single domain interventions were effective the simultaneous training of both cognitive and physical domains offers a greater potential on daily life functioning. One of the key problems identified by the authors was the issue of how to combine different interventions and how to evaluate their effectiveness. The systems approach to healthy aging provides a theoretical perspective on this issue, suggesting that if a major cause if impairment is functional loss in the intrinsic connectivity networks, the optimal intervention should target function in the network as a whole, rather than individual components thereof.

Computerized cognitive training (CCT) approaches, using computer programs to boost core cognitive capabilities such as working memory, speed of processing and visual attention have proved highly effective in some studies, but less so in others. Systematic reviews of brain training programmes with older adults (Gross et al., 2012; Kueider et al., 2012) concluded that computerized training is an effective, less labor intensive alternative to cognitive training. In contrast, a recent analysis (Lampit et al., 2014) concluded that the overall effect size of CCT vs. control was small and statistically significant for nonverbal memory, verbal memory, working memory, processing speed, and visuospatial skills but not for executive functions and attention. A meta-analysis for younger groups (Melby-Lervåg and Hulme, 2013) concluded that WM programs produced reliable short-term improvements in WM skills but that the effects were ''short-term, specific training effects that do not generalize''.

One of the clear limitations, from a systems view, both of CCT and of direct brain stimulation, is that the intervention is artificial, and isolated from the physical or mental activities involved in normal system functionality. There is strong evidence that natural activities, such as exercise, can improve not only physical fitness but also mental fitness, and even stimulate the growth of new brain neurons and connections (Hillman et al., 2008; Höetting and Röeder, 2013; Kirk-Sanchez and McGough, 2014). An innovative approach, the Long-Lasting Memories intervention, which combines both exercise and CCT approaches (''exergaming'') was shown to have beneficial effects for healthy older adults and for those with Mild Cognitive Impairment (MCI; González-Palau et al., 2014).

A recent discovery has been the differential effects of cardiovascular, high intensity, exercise and ''co-ordinative exercise'' such as balance training or tai-chi. There is strong evidence that exercise can potentiate the brain for new learning, with coordinative balance exercises leading to neural growth in the hippocampus—a core structure for explicit learning and memory (Niemann et al., 2014)—and also in the cerebellarcortical loop (Burciu et al., 2013)—a core network for implicit learning and coordination. A further study (Nascimento et al., 2014) concluded that multimodal physical exercise was effective in reducing pro-inflammatory cytokines and in improving brainderived neurotrophic factor (BDNF) peripheral levels, with positive reflexes on cognition in elderly individuals with MCI.

Recent studies of the effects of exercise on rat brains (Kellermann et al., 2012; Abel and Rissman, 2013) reveal strong effects on epigenetic changes and changes in the cerebellar Purkinje cells following a rat vestibular training exercise (Lee et al., 2015). There is also evidence that BDNF is expressed in the cerebellum following environmental enrichment for rats (Angelucci et al., 2009; Vazquez-Sanroman et al., 2013). Of particular interest, there is evidence (though sparse) that Quadrato exercise (like Tai Chi) led to increased creativity and changes in gray matter and white matter in the cerebellum (Ben-Soussan et al., 2015).

There have also been detailed neuroimaging studies of interventions for special groups. Daily clinic-based balance training for 2 weeks in cerebellar patients and age-matched healthy controls (Burciu et al., 2013) led to enhanced balance performance in the patients, with associated increased gray matter volume in the dorsal premotor cortex and within the cerebellum for both groups. A 6 week balance-training study with Parkinson's patients and healthy controls (Sehm et al., 2014) led to improved balance which was maintained for the following year, together with increased gray matter in the hippocampus for the controls and in several brain regions for the patients.

Of particular interest regarding functional connectivity, two recent studies with older adults with MCI have established functional connectivity changes following 8 week interventions. Klados et al. (2016) established that the Long Lasting Memories intervention led to increased beta-band EEG activity (reflecting increased bilateral connections in the occipital, parietal, temporal and prefrontal regions) after the intervention. Chirles et al. (2017) undertook a ''walking exercise'' intervention, and established that following the intervention the MCI group showed increased connectivity in 10 regions spanning frontal, parietal, temporal and insular lobes, together with the cerebellum.

In summary, current neuroimaging and behavioral research appears to be converging to a view that: (i) a systems approach to aging is the most promising framework for understanding the degradation in multiple functions with age; (ii) there is extensive evidence that the cerebellum is one of the key structures affected, and the multiple intrinsic connectivity networks linking the cerebellum with other brain and body structures may well mediate many of the actual deficits shown; (iii) ''single system'' interventions can be effective, but generally it is better to have multiple domain interventions; (iv) a range of interventions, from coordinative exercise to direct vestibular stimulation are likely to have beneficial effects on cerebellar function.

The above considerations informed the design of the current study. We wished to evaluate the effectiveness of a novel internet-based ''vestibular stimulation'' intervention, the Zing intervention<sup>1</sup> . This intervention was originally developed to tune up the coordination abilities of top sporting performers, using a series of graded exercises designed specifically to improve three performance dimensions: sensorimotor coordination, eye movement control and dual tasking. However, extensive feedback had suggested that the programme was valuable for many average performers. Consequently the system was embedded in an internet-based ''game'' format designed to challenge and stimulate the user to keep improving their performance. Zing Performance offer a number of courses specifically tailored to each individual user, with applications in sporting areas, organizational development, in education.

The Zing system involves a series of graded activities on three dimensions—dynamic activity (patterned movement sequences), focus activity (developing the ability both to concentrate and to ''dual task'') and stability activity (coordinative balance). Underpinning the approach is the technique of vestibular stimulation. Rather than cardiovascular exercise, which is designed to have energetic use of highly practised routines, or even coordinative balance such as tai-chi, which does involve learning new actions, vestibular activities are designed to cause abnormal input for the vestibular system, for example by requiring the user to put their head on one side while undertaking tasks. This presents the vestibular system, and the cerebellum, with an immediate challenge, requiring activation of many circuits to cope with the ensuing proprioceptive feedback.

A typical course lasts 6 months and is composed of daily physical activities and digital video games. An example of a low level focus activity (at the time of the study) is given in **Figure 1**. A video is also available for each activity.

The Zing platform therefore provides a user-orientated, motivating framework for delivering a cost-effective cerebellar challenge intervention that satisfies the criteria that have emerged for multimodal, challenging interventions in older adults.

We undertook the study to investigate whether internet-based approaches can indeed be an effective and popular method for older adults, and designed an 8 week intervention. It is important to highlight that although this is a Randomized Control Trial (RCT) study, in that there was a control condition and allocation to condition was random, it is not a full RCT, for which an active intervention condition, matched in time and form to the Zing intervention, would need to be used to counter placebo-type

FIGURE 1 | Sample screen from the Zing intervention. This is an example of a screen from the focus strand, at a low level and aimed at developing dual tasking ability.

<sup>1</sup>www.Zing.com

effects. Our view is that this non-equivalent–control RCT (NEC-RCT) design is appropriate for a user-centered trial that has the underlying question ''If participants undertake the intervention, will it help them, and, if so, in what ways?'' We are not investigating the theoretical issue—is intervention A more effective than intervention B, and if so, why? Each design has its strengths and weaknesses. For the purpose of evaluating whether a low-cost, home-based intervention might be beneficial compared with life-as-usual, the NEC-RCT design is the appropriate one.

A limitation of many previous intervention studies is that the set of tests used from pre-intervention to post-intervention focus on a limited range of performance measures. As noted above, there is reason to expect that a cerebellar challenge intervention might lead to changes in both the sensorimotor domain and in the cognitive domain. There is also longstanding evidence that the cerebellum is involved in emotional processing (Schmahmann and Sherman, 1997), with emerging evidence regarding its involvement in processing emotional salience (Styliadis et al., 2015; Adamaszek et al., 2017). There is also evidence that the resting state networks involving the cerebellum are associated with differences in crystallized intelligence (Pezoulas et al., 2017). Consequently we designed a battery of simple tasks designed to probe sensorimotor, performance, cognitive performance, emotional state and nonverbal reasoning.

The design allows the following hypotheses to be evaluated.

Hypothesis 1. Improvements in balance and sensorimotor coordination. This is the primary applied hypothesis, directly related to attempting to boost balance performance and thus decrease the danger of falling. One in three people of 65 fall at least once per year, with the incidence rising to one half of those over 80 years old (Todd and Skelton, 2004). Falls are a major cost to elderly people and to national health services, estimated to account for 21% of the Dutch health service costs for injuries (Hartholt et al., 2011). Hypothesis 1 states that Zing training will lead to significant improvements in sensorimotor coordination especially balance: (a) for each individual compared with their pre-training; (b) that the intervention group will improve significantly more than a control, no intervention, group.

Hypothesis 2. ''Hippocampal'' improvements. This hypothesis is derived from the research showing the benefits of coordinative balance training for hippocampal function. Hypothesis 2 states that Zing training will lead to significant improvements in ''declarative memory'' performance: (a) for each individual compared with their pre-training; (b) that the intervention group will improve significantly more than a control, no intervention, group.

Hypothesis 3. Improvement Specificity. Despite the emerging evidence of cerebellar involvement in affective processing, we would expect any such changes to be of secondary importance in terms of affective state. Consequently, although Zing training may lead to significant ''transfer'' to other areas, including language, affect and fluid reasoning, any such effects will be minor compared with the specific improvements in hippocampal and sensorimotor skills.

### MATERIALS AND METHODS

#### Participants

Ninety-eight volunteers (30 male, 68 female) aged 50–85 (mean 68.2, SD 6.6) were recruited through advertisements in local newspapers, churches and social groups. An advert also went out on the University of the Third Age Sheffield website. Participants were all without a known diagnosis of dementia. The ethics committee of the Department of Psychology, University of Sheffield, approved the study. Participants gave fully informed prior consent. They were also informed that their information would be anonymised and kept securely. They were also informed that they could withdraw from the study at any time without needing to give any reason. All participants were healthy older adults.

#### Design

The aim of this study was to test the effectiveness of vestibular stimulation on physical and mental function. Therefore, a repeated measures design was used. Participants were asked to complete a baseline set of tests at the University of Sheffield Department of Psychology before taking part in the 8 week Zing intervention at home. They were then asked to return to the department for a repeat of the baseline tests.

#### Test Battery

The same tests were used both pre and post-test. While there may be some practice effects here, it would be expected that this would affect both groups equally, and therefore any relative difference in the intervention group's performance is likely to be attributable to the exercises.

We wished to evaluate changes in all the core physical, mental and affective domains, using simple but normed tests where possible. We based the battery on the Dyslexia Adult Screening Test (Fawcett and Nicolson, 1998), which covers the majority of the necessary tests in a 30 min package. We constructed a battery of 14 tests, divided into five suites. Suite 1 was for Physical Coordination and comprised the DAST balance test, the two Purdue pegboard (Tiffin and Asher, 1948) tests (Peg Moving and Peg Assembly), and the DAST writing (copying) test. Suite 2 investigated memory. It included two tests of working memory, the DAST backwards digit span test and the South Yorkshire Ageing Study (Tarmey, 2012) Spatial Memory test which determines spatial memory span for non-verbalizable pictures presented in one of eight locations. There was one declarative memory test, the South Yorkshire Ageing Study (Tarmey, 2012) Picture Memory test which assesses recall for a set of 20 pictures of common objects, presented sequentially for 1 s, including both immediate recall and delayed recall after 20 min. The Language Suite comprised the DAST Rapid Naming, Phonological Processing, Reading, Nonsense Passage and Spelling tests. The Fluid Reasoning suite comprised the DAST Nonverbal Reasoning, Semantic Fluency and Verbal Fluency tests. Finally two tests of affect were administered: the Beck Depression Inventory (BDI; Beck et al., 1996) and the Authentic Happiness Inventory (Seligman, 2002).

#### Intervention Training

Participants were required to undertake a minimum of 8 weeks and maximum of 10 weeks balance and sensorimotor coordination training using an online set of activities. These were provided by Zing Performance Ltd., and were designed specifically to stimulate brain regions involved in coordinative balance. Initially, participants had to undergo an assessment to determine their strengths and weaknesses. After this a 30 day programme was set for them, specifically designed to target their biggest needs. Participants were required to do two exercises a day, before rating how difficult they found that particular activity. A screen shot is shown in **Figure 1** below. Each week, three exercises were assigned, with two of the three appearing each day. After 30 days, participants were reassessed before continuing onto unit two. It should be noted that a full Zing 360 session programme is designed for 6 months, with two sessions per day. Consequently this study is very much shorter than intended by the Zing designers.

#### RESULTS

Data were converted to standard scores (mean 100, SD 15) to allow direct comparison across tasks. Where possible, population norms and standard deviations were used to normalize test scores. The population norm for age 55+ was used for all participants, irrespective of age, to represent absolute performance rather than age-adjusted performance.

#### Effect Sizes

In order to facilitate comparison of the improvements (or otherwise) in performance from pre-test to post-test, effect sizes were calculated using the formula ES = (post-test − pre-test) / SD (all groups on pre-test), which is a form of Cohen, 1988 applied to change analysis. No change would result in an effect size of 0, whereas a score of +1.0 indicates a change of one standard deviation unit. Cohen (1988) suggests that effect sizes of 0.8, 0.5 and 0.2 be labeled large, medium and small, respectively.

Effect sizes for the two groups are shown in **Figure 2**. Tests have been grouped in order of the hypotheses. The Physical Coordination suite—postural stability, peg moving, peg assembly and handwriting speed are on the left, then the Memory Suite—immediate picture recall, delayed picture recall, immediate spatial memory and immediate verbal memory. Group 3 include the affect measures—the BDI (reverse scored such that higher means less depressed) and the Authentic Happiness Index. The remaining tests are the Language Suite and the Fluid Thinking Suite but were not predicted to be affected by the intervention.

#### Correlations with Zing Usage

Next correlational analyses were undertaken utilizing data collected automatically on ''compliance'' for the Zing group. Of the 53 participants allocated to the Zing group, 38 completed at least 40 sessions, as requested, but the differential uptake allowed us to investigate the effects both of frequency of Zing use (sessions per week) and the duration (number of weeks). For the frequency of Zing use significant correlations were found for

peg movement (r = 0.31, p < 0.05), and for postural stability (r = 0.303, p < 0.05). A significant correlation with number of weeks of the intervention occurred only for nonverbal reasoning (r = 0.288, p < 0.05).

#### Within-Group Statistical Tests

Inferential statistical tests were then undertaken for the 16 tests within the five suites of tests. First repeated measures multivariate analyses of variance were undertaken for each suite separately on the data for pre-test and post-test for each test within the suite.

For the control group, none of the set of MANOVAs approached significance. In fact the only individual comparison to reach the uncorrected 0.05 significance level was for peg moving (F(1,43) = 6.26, p = 0.016).

For the Zing group, the MANOVA analyses of the change from pre-test to post-test were highly significant for the suites for Physical Coordination, for Declarative Memory, for Language, and for Fluid Thinking (F(1,47) = 22.95, p < 0.001; F(1,52) = 10.71, p = 0.002; F(1,52) = 15.99, p < 0.001; F(1,52) = 5.72, p = 0.020 respectively), whereas there was no difference for the Affect suite. It is not sensible to undertake a Bonferroni correction for multiple comparison when all comparisons are significant in the same direction (Moran, 2003), and consequently uncorrected probabilities are reported. The changes for Balance, Peg Assembly and Peg Movement were significant (F(1,50) = 14.07, p < 0.001; F(1,51) = 5.53, p = 0.023; F(1,50) = 4.10, p = 0.048 respectively). The improvements for Delayed Picture Recall, Immediate Picture Recall and Memory Span were also significant (F(1,52) = 14.44, p < 0.001; F(1,52) = 15.41, p < 0.001; F(1,52) = 4.20, p = 0.046 respectively). Two of the improvements for nonsense passage reading, 1 min reading, rapid naming and spelling were significant [F(1,52) = 3.72, p = 0.059; F(1,52) = 6.28, p = 0.015; F(1,52) = 4.73, p = 0.034; F(1,52) = 3.28, p = 0.076 respectively). The improvement for verbal fluency was also significant (F(1,52) = 5.13, p = 0.028).

#### Between-Group Statistical Tests

Finally, in the most stringent test of the changes, a series of multivariate 2-factor analyses of variance was undertaken, with the independent groups factor being the group (Zing vs. Control) and the repeated measure being time-of-test (pre-test vs. post-test. Manovas were undertaken separately for each of the five suites (see **Table 1**). For the MANOVA entry, only the key statistic, the interaction term between time of test (pre vs. post) and Group is reported. A significant interaction would typically indicate that the Intervention led to a significant difference between groups at post-test whereas performance at pre-test was equivalent.

It may be seen that the only suite returning a significant MANOVA result was the Physical Coordination suite. For each of the four tests a univariate two factor mixed measures analysis of variance was undertaken, with the within-group variable being time-of-test (pre-intervention vs. post-intervention) and the between-group variable being group (intervention vs. control). Significant (uncorrected) interactions—all reflecting greater improvement for the intervention group than the control group—were obtained for postural stability and for peg assembly. By contrast, there were no differences for peg moving speed or writing speed.

The MANOVA results for the other four suites of tests were not close to significance. Uncorrected significant differences were obtained for Delayed Picture Memory and for Nonsense Passage Reading.

### Correlations with Age

Finally, correlations with age were calculated. Significant correlations were found for performance on the majority of tests, with correlations between age and each dependent variable in descending order being −0.47 (Nonverbal reasoning), −0.40 (peg assembly), −0.38 (immediate picture memory), −0.35 (immediate picture memory), −0.34 (writing), −0.28 (spatial

TABLE 1 | Multivariate and univariate analyses of variance for the variables of


memory), −0.27 (semantic fluency), −0.26 (postural stability) and −0.26 (spelling). Correlations between age and the amount of improvement for the Zing group were also calculated. Few correlations were significant, with only peg assembly (−0.37) being more extreme than −0.25.

### DISCUSSION

The primary issue addressed by this study was whether a home-based cerebellar challenge internet-administered intervention was feasible for use with older adults and, if used, whether it would result in better balance, and hence reduce danger of falling (Hypothesis 1). A secondary, theoretical issue, was whether the intervention might also improve cognitive functions previously found to be improved by coordinative balance training (Hypothesis 2).

A set of five ''suites'' of tests was applied before and after the intervention, allowing comparison with a ''life as usual'' control group. Comparing individual performances across the intervention period, the control group performance remained roughly constant, with no significant change for 17 of the 18 tests. By contrast, the intervention participants showed significant improvement in their scores for 9 of the 18 tests administered, with only the tests of the Affect suite showing no significant multivariate improvement.

Furthermore, a series of two factor multivariate analyses of variance revealed that the intervention group improved significantly more than the control group on the Physical Coordination suite, but not on the memory suite, the affect suite, the language suite or the fluency suite.

Hypothesis 1 is therefore clearly supported. Not surprisingly—but crucial for applied purposes—the intervention group did improve significantly on balance compared both with their own pre-intervention performance and with the control group's change in balance over the period of the study. There was also transfer of this training effect to manual dexterity (as indicated by the ''peg assembly'' task).

Clear support is also provided for Hypothesis 2, though the effect is masked by the inclusion of tests of working memory and declarative memory within the memory suite. It is clear from the effect sizes and the between-groups anova data that there was a significant benefit for the Zing group for the tests of declarative memory (especially the key task of delayed picture recall, which is more akin to a realistic memory use task) but not for the tests of working memory. It is therefore legitimate to infer that, consistent with the literature on the benefits of balance training on hippocampal function, there was transfer of this benefit to declarative memory in the delayed picture recall condition (Hypothesis 2).

The specificity of the differential benefit findings to the hypotheses suggests strongly that the changes are not a practice, placebo or Hawthorne effect (Hypothesis 3). Given the null effect on affect, we examined the individual scores on the BDI in order to investigate whether participants more at risk of depression showed differential effects. Of the four participants in the intervention group classifiable as at least mildly depressed, one showed marked improvement. However, the 2 participants in the non-intervention group initially showing mild depression also showed marked improvement. We conclude that, at least in this group of heathy older adults, the results do not suggest that there is a direct effect of the cerebellar challenge intervention on affective state.

Of the 52 participants selected for the intervention condition, 38 (73%) completed the requested 40 sessions over 8 weeks. Analyses of the ''dose effect'' (that is, correlations of performance improvement with number of intervention sessions for the intervention group) revealed significant correlations with intervention frequency for peg movement, reading and balance, with a significant correlation with intervention duration for nonverbal reasoning.

In terms of the participants' response to the intervention, it should be noted that the Zing platform was a prototype version, not yet publicly available and in the process of substantial development and improvement. A sizeable minority of the Zing participants reported difficulties in accessing the system initially, though subsequently problems were relatively small.

It is important to acknowledge the limitations of this study. First, the population sampled was by no means random, involving respondents to a circular. They should therefore be seen as relatively high functioning and with good self efficacy. Furthermore, they represented a spread of ages, with 2 under 55, 27 aged 55–64, 55 aged 65–74 and 15 aged 75 and over. There was a considerable imbalance between the sexes, with 30 male and 68 female. All were living at home in reasonable health. All of these factors reduce the strength of inferences that can be made regarding generalization to the full population of community based older adults, and highlight the need for further research.

#### CONCLUSION

Prior research has established that home-based balance exercises are among the most cost-effective methods of improving balance ability and hence reducing falls in older adults. Recent developments in cognitive neuroscience have revealed that ''coordinative'' balance training is likely to have beneficial effects not only on physical coordination but also on hippocampal function. Our theoretical analyses suggested that a multicomponent, cerebellar challenge intervention should prove highly effective, combining the effectiveness of coordinative exercise with that of direct cerebellar stimulation, and therefore improving function in the intrinsic connectivity networks involving the cerebellum. The study design did not include brain imaging, and therefore it is not possible to assess directly any underlying neural changes. Furthermore the study design does

#### REFERENCES


not permit comparison of the Zing approach with a suitable active control. Nonetheless, the results were encouraging.

The present study is unique in two ways: first we investigated a highly cost-effective internet-based ''cerebellar challenge'' intervention, ''Zing''. Second we investigated physical coordination, mental coordination, language, fluid thinking and affect using a specially developed battery of tests. Significant benefits (comparing initial performance with post-intervention performance) were found for the intervention group on the majority of the tests, excluding only those for affect. Furthermore, significantly greater improvements were found for the intervention group (compared with the control group) for balance, for physical coordination and for declarative memory retrieval.

Further research, including research using an active control intervention, would be needed to pinpoint the theoretical causes of the improvements obtained. Nonetheless, given the minimal cost and considerable ease of access of the intervention, it provides a promising approach to improving the overall cerebellar-related function, protecting against subsequent balance problems, and may also benefit declarative memory in older adults.

### AUTHOR CONTRIBUTIONS

Both authors contributed to all aspects of the empirical work, the data analysis and the article writing. ZG had a stronger focus on the empirical work, and RIN had a stronger focus on design and theoretical interpretation.

#### FUNDING

The research was undertaken as part of ZG's doctoral research at the University of Sheffield. This was supported by a fees-only award from the University of Sheffield. The research was undertaken in 2014/15. In 2016 Nicolson was appointed to the scientific advisory board of Zing Performance Outreach, a non-profit charitable organization.

#### ACKNOWLEDGMENTS

We gratefully acknowledge the support of Samantha Critchley, Mark Manser and Gareth Dore of Zing Performance in helping to resolve any difficulties encountered by the participants. We acknowledge with thanks the contributions made in the pre-test and post-test assessments by Caroline Carta, Penny Jackson, Phil Roughsedge and Laura Scott.


**Conflict of Interest Statement**: Zing Performance Inc. provided free registration on the Zing Programme for the intervention participants and £30 in acknowledgment for those completing the intervention. Neither author received financial support from Zing Performance Ltd.

Copyright © 2017 Gallant and Nicolson. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Interactions of Insula Subdivisions-Based Networks with Default-Mode and Central-Executive Networks in Mild Cognitive Impairment

#### Ganesh B. Chand<sup>1</sup> \*, Junjie Wu<sup>2</sup> , Ihab Hajjar1,3† and Deqiang Qiu2,4†

<sup>1</sup> Department of Medicine, Emory University School of Medicine, Atlanta, GA, United States, <sup>2</sup> Department of Radiology and Imaging Sciences, Emory University School of Medicine, Atlanta, GA, United States, <sup>3</sup> Department of Neurology, Emory Alzheimer's Disease Research Center, Emory University School of Medicine, Atlanta, GA, United States, <sup>4</sup> Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, GA, United States

#### Edited by:

Ana B. Vivas, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Rui Li, Institute of Psychology (CAS), China Kaiming Li, Sichuan University, China

#### \*Correspondence:

Ganesh B. Chand ganesh.chand@emory.edu; ganeshchand@gmail.com

†These authors have contributed equally to this work.

Received: 25 August 2017 Accepted: 25 October 2017 Published: 09 November 2017

#### Citation:

Chand GB, Wu J, Hajjar I and Qiu D (2017) Interactions of Insula Subdivisions-Based Networks with Default-Mode and Central-Executive Networks in Mild Cognitive Impairment. Front. Aging Neurosci. 9:367. doi: 10.3389/fnagi.2017.00367 Interactions between the brain networks and subnetworks are crucial for active and resting cognitive states. Whether a subnetwork can restore the adequate function of the parent network whenever a disease state affects the parent network is unclear. Investigations suggest that the control of the anterior insula-based network (AIN) over the default-mode network (DMN) and central-executive network (CEN) is decreased in individuals with mild cognitive impairment (MCI). Here, we hypothesized that the posterior insula-based network (PIN) attempts to compensate for this decrease. To test this, we compared a group of MCI and normal cognitive individuals. A dynamical causal modeling method has been employed to investigate the dynamic network controls/modulations. We used the resting state functional MRI data, and assessed the interactions of the AIN and of the PIN, respectively, over the DMN and CEN. We found that the greater control of AIN than that of DMN (Wilcoxon rank sum: Z = 1.987; p = 0.047) and CEN (Z = 3.076; p = 0.002) in normal group and the lower (impaired) control of AIN than that of CEN (Z = 8.602; p = 7.816 × 10−18). We further revealed that the PIN control was significantly higher than that of DMN (Z = 6.608; p = 3.888 × 10−11) and CEN (Z = 6.429; p = 1.278 × 10−10) in MCI group where the AIN was impaired, but that control was significantly lower than of DMN (Z = 5.285; p = 1.254 × 10−<sup>7</sup> ) and CEN (Z = 5.404; p = 6.513 × 10−<sup>8</sup> ) in normal group. Finally, the global cognitive test score assessed using Montreal cognitive assessment and the network modulations were correlated (Spearman's correlation: r = 0.47; p = 3.76 × 10−<sup>5</sup> and r = −0.43; p = 1.97 × 10−<sup>4</sup> ). These findings might suggest the flexible functional profiles of AIN and PIN in normal aging and MCI.

Keywords: central-executive network, dynamical causal modeling, default mode network, insula subdivisions, insula-based network

## INTRODUCTION

fnagi-09-00367 November 7, 2017 Time: 16:47 # 2

Normal cognitive function involves an effective coordination between functionally associated brain regions or network(s) (Fox et al., 2005; Power et al., 2011). Two core brain networks, namely the default-mode network (DMN)—consisting of the posterior cingulate and ventromedial prefrontal cortices, and the centralexecutive network (CEN)—consisting of the posterior parietal and dorsolateral prefrontal cortices, exhibit anti-correlated network activities (Fox et al., 2005; Bressler and Menon, 2010; Chen et al., 2013) with the DMN being more active during internally directed actions while the CEN being more active primarily during externally directed actions (Bressler and Menon, 2010; Uddin, 2015). Recent evidence consistently suggests that this anti-correlation pattern is modulated by the anterior insulabased network (AIN), which primarily comprises of the anterior insula and the dorsal anterior cingulate cortex, in both young and elderly people with normal cognition (Sridharan et al., 2008; Chand and Dhamala, 2016a; Wu et al., 2016). We have recently found that the modulation effect of this AIN over the DMN and CEN is impaired in individuals with mild cognitive impairment (MCI) (Chand et al., 2017a,b). Furthermore, recent studies highlight that the insula subdivisions—the anterior insula and the posterior insula—exhibit overlapping profiles/activities that could flexibly involve in a wide range of cognitive processes, especially in restoring the coginitive functions (Starr et al., 2009; Segerdahl et al., 2015; Nomi et al., 2016; Namkung et al., 2017). However, as the modulation ability of the AIN declines in MCI, whether such control feature shifts over to the posterior insulabased network (PIN)—network that mainly comprises of the posterior insula and the sensorimotor areas (Deen et al., 2011; Nomi et al., 2016)—has not been previously investigated.

Dynamic interaction analysis between the intrinsic networks has emerged as a potentially valuable approach that may reveal the underlying neural processes in health and disease. Recent functional MRI studies suggest that the AIN is responsible for switching the activation and deactivation between the DMN and CEN in cognitively normal people, and these studies suggest that this control ability maintains individual's active and passive cognitive states (Menon, 2011; Goulden et al., 2014; Wu et al., 2016). The control functionality of the AIN is reasoned to be carried out with the aid of a unique anatomical cytoarchitectural feature of its key regions—the anterior insula and the dorsal anterior cingulate cortex (Sridharan et al., 2008; Chand and Dhamala, 2016b). Specifically, these regions are anatomically connected (Bonnelle et al., 2012; Jilka et al., 2014) and consist of a special type of neurons named von Economo neurons that are thought to facilitate rapid relays of information from the AIN to the other brain regions such as DMN and CEN (Allman et al., 2005, 2010; Watson et al., 2006; Sridharan et al., 2008). On the other hand, the PIN encompasses the posterior insula and sensorimotor areas, specifically temporal and posterior cingulate regions, and thus are primarily involved in sensorimotor processes (Cauda et al., 2011; Nomi et al., 2016). Structural connectivity analysis consistently demonstrates that posterior insula has direct white matter connections with the parietal and posterior temporal regions (Cerliani et al., 2012; Cloutman et al., 2012; Dennis et al., 2014). Alternation in the AIN activity has been recently reported in the diseases, including autism, frontotemporal dementia, schizophrenia, and MCI or Alzheimer's disease (Menon, 2011; Uddin, 2015; Chand et al., 2017a,b). Specifically, when the AIN modulation over the DMN and CEN is declined or impaired in MCI, whether the PIN tends to take over this control feature has not been formerly examined.

In the present study, we therefore seek to examine the differential modulation activities of the AIN and of the PIN, respectively, over the DMN and CEN in MCI people and compare with a group of healthy controls. We hypothesized that the control ability of the AIN over the DMN and CEN is disrupted in the MCI group but this control is retained in the healthy normal group. As the AIN preserves this control in the normal group, we further hypothesized that the PIN does not take such control in the normal group, but does tend to take over that control feature in the MCI where the AIN is impaired. We also hypothesized that the global cognitive test score correlates with the modulating probability of the AIN and of the PIN, respectively. To test our hypotheses, we analyzed resting state functional MRI data collected on with a sample of older adults with normal cognition and with the MCI, then applied dynamical causal modeling (DCM), and compared the network interactions between two groups. We also assessed the association between network modulation probability with cognitive performance within the same sample.

#### MATERIALS AND METHODS

### Subjects

This study was carried out in accordance with the recommendations of "Institutional Review Board (IRB) of Emory University" with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. This study protocol was reviewed and approved by IRB of Emory University. MRI scans were performed on 53 MCI subjects. The MCI subject inclusion criteria were: age ≥ 55 years, hypertension defined by systolic blood pressure ≥ 140 mm Hg or diastolic blood pressure ≥ 90 mm Hg, and MCI assessed based on previously defined criteria (Chao et al., 2009; Pa et al., 2009): Montreal cognitive assessment (MoCA) ≤ 26, cognitive performance at the 10th percentile or below on at least one of four screening tests trail marking test B, Stroop interference, digit span forward and digit span backward, verbal fluency and abstraction—and minimal functional limitation test as reflected by the functional assessment questionnaire ≤ 7. The subject exclusion criteria were: systolic blood pressure > 200 mm Hg or diastolic blood pressure > 110 mm Hg, renal disease or hyperkalemia, active medical or psychiatric problems, uncontrolled congestive heart failure (shortness of breath at rest or evidence of pulmonary edema on exam), history of stroke in the past 3 years, ineligibility for MRI (metal implants or cardiac pacemaker), inability to complete cognitive test and MRI scan, women of childbearing potential, and diagnosis of dementia (self-reported or care-giver

reported). In MCI group, mean age was 66.9 years (SD: 8.1), mean education was 15 years (SD: 2.4), 60% were African–Americans, 52.8% were women, mean systolic blood pressure 150.7 mm of Hg (SD: 21.3), and mean diastolic blood pressure 90.9 mm of Hg (SD: 13.5). MRI data were included from 20 normal older adults. The normal control subject inclusion criteria were age ≥ 50 years, MoCA ≥ 26, clinical dementia rating score of 0, and normal logical memory subscale defined as ≥11 for 16 or more years of education, ≥9 for 8–15 years of education, and ≥6 for less than 7 years of education. The exclusion criteria were history of stroke in the past 3 years, ineligibility for MRI (metal implants or cardiac pacemaker), inability to complete cognitive test and MRI scan, and diagnosis of dementia. In cognitively normal group, mean age was 65.8 years (SD: 8.8), mean education was 16 years (SD: 2.9), 20% were African–Americans, 70% were women, mean systolic blood pressure 128.8 mm of Hg (SD: 23.1), and mean diastolic blood pressure 71.7 mm of Hg (SD: 11.7), and eight subjects (out of 20) had hypertension. The MCI and normal control groups were not statistically different for age, sex, and education, but were different for systolic blood pressure, diastolic blood pressure, and MoCA score as shown in **Table 1**.

#### Image Acquisition

Siemens 3T Trio scanner was used for MRI data acquisition at Center for Systems Imaging of Emory University. Anatomical 3D images were acquired using T1-weighted magnetization prepared rapid gradient echo (MPRAGE) sagittal with the repetition time (TR) = 2300 ms, echo time (TE) = 2.89 ms, inversion time (TI) = 800 ms, flip angle (FA) = 8 ◦ , resolution = 256 × 256 matrix, slices = 176, and thickness = 1 mm. Resting state blood oxygenation level dependent (BOLD)-fMRI images were acquired axially using an echo-planar imaging sequence with the TR = 2500 ms, TE = 27 ms, FA = 90◦ , field of view = 22 cm, resolution = 74 × 74 matrix, number of slices = 48, thickness = 3 mm and bandwidth = 2598 Hz/Pixel. The subjects were asked to hold still, keep their eyes open, and think nothing during scan time.

TABLE 1 | Mean (standard deviation) of mild cognitive impairment (MCI) group and cognitively normal (NC) group regarding subjects' age, education, sex, race, systolic and diastolic blood pressures (BP), and Montreal cognitive assessment (MoCA) (p-value represents MCI vs. NC comparison using Wilcoxon rank sum test or chi-square test and p < 0.05 is considered statistically significant difference between two groups).


#### Image Preprocessing

MRI images were preprocessed using SPM12 (Wellcome Trust Centre for Neuroimaging, London, United Kingdom<sup>1</sup> ). The preprocessing included slice-timing correction, motion correction, co-registration to individual anatomical image, normalization to Montreal Neurological Institute template, and spatial smoothing of normalized images using a 6 mm isotropic Gaussian kernel. Independent component analysis (ICA) was carried out on the preprocessed data.

#### Independent Component Analysis

Independent component analysis is a promising technique for the functional brain activities. A spatially constrained ICA (Lin et al., 2010) has been proposed for the study of specific brain areas or networks. In this work, we used the templates of DMN, AIN, PIN, and CEN from previous study (Shirer et al., 2012) in Group ICA of fMRI Toolbox (GIFT<sup>2</sup> ) and computed ICA component of each network. Prior studies suggest that ICA component of each network/mask is more accurate than the average or first eigenvariate of network template/mask (Smith et al., 2011; Craddock et al., 2012; Shirer et al., 2012; Goulden et al., 2014). We first ran ICA analysis separately for the normal controls and MCI. We subsequently implemented a DCM on the ICA-components of networks. For cross-validation purpose, we also ran ICA analysis combinely for the normal controls and MCI and then implemented DCM.

#### Dynamical Causal Modeling

Dynamical causal modeling (Friston et al., 2003) infers the statistical measure of directed functional connectivity between brain areas or networks. DCM bases on Bayesian model selection and compares the user defined models with the measured data (Stephan et al., 2010). DCM has recently been implemented in resting state fMRI (Daunizeau et al., 2012; Friston et al., 2014).

In model construction, DCM models were designed with full intrinsic connections between the networks and the modulations were taken to represent the models. In DCM analysis, model 1 represents non-linear modulation of DMN on both connections between AIN (or PIN) and CEN. Similarly, model 2 represents non-linear modulation of AIN (or PIN) on the connections between DMN and CEN, and model 3 represents non-linear modulation of CEN on the connections between AIN (or PIN) and DMN. We performed both fixed effect and random effect Bayesian model selection methods. In brief, a fixed effect considers that the optimal model is homogeneous across subjects and provides the group log-evidence that measures the balance between fit and complexity of models and quantifies the relative goodness of models. On the other hand, a random effect accounts for heterogeneity of model structure across subjects and provides the posterior model probability, which measures how likely a specific model generated the data of randomly selected subject, and the posterior exceedance probability that measures how one model is more likely than any other model (Stephan et al., 2010). DCM analysis was performed by using

<sup>1</sup>www.fil.ion.ucl.ac.uk/spm/software/spm12

<sup>2</sup>http://mialab.mrn.org/software/gift

respectively).

SPM12 (Wellcome Trust Centre for Neuroimaging, London, United Kingdom<sup>1</sup> ).

### Statistical Analysis

Network modulation probabilities were compared between cognitively normal group and MCI group using Wilcoxon rank sum test. Correlation analysis was performed between the global neuropsychological test score assessed by MoCA and the modulation probability of AIN and/or PIN to the DMN and CEN using Spearman's correlation. Matlab software (Natick, MA, United States<sup>3</sup> ) was used to analyze the data.

## RESULTS

### Constrained ICA

**Figures 1**, **2** show the results of constrained ICA of DMN, CEN, AIN, and PIN for the normal control group and the MCI group, respectively.

<sup>3</sup>https://www.mathworks.com

### DCM Model Comparisons

**Figure 3** shows the fixed effect results for normal controls and MCI expressed in terms of log-evidence and posterior probability. Fixed effect method for the normal control showed that a control feature by AIN (model 2: first column) has higher probability compared with that of control feature by DMN (model 1: first column) and by CEN (model 3: first column). Fixed effect method further demonstrated that the control feature of AIN (model 2: second column) no longer has dominant probability in the MCI. We further investigated the control feature of PIN in the normal control (model 2: third column) and in the MCI (model 2: fourth column). In the normal control, we found that the model 2 (control feature of PIN) does not have dominant switching probability. In the MCI, we uncovered that the model 2 (control feature of PIN) came in play and possessed the higher probability compared with that of models 1 and 3. DCM analysis was also carried out by running ICA together in normal controls and MCI (**Figure 4**). We found the similar patterns of higher model probability. The AIN modulation has higher probability in the normal controls, but not in the MCI (model 2 in the first and second columns).

FIGURE 2 | The t-value maps of (A) DMN, (B) CEN, (C) AIN, and (D) PIN from the constrained ICA overlaid on mean BOLD images in mild cognitive impairment (MCI) (L and R indicate the left and right hemispheres, respectively).

Moreover, the PIN modulation did not have higher probability in the normal controls, but had higher probability than other models in the MCI (model 2 in the third and fourth columns).

We further performed the random effect analysis. **Figure 5** displays the random effect results for the normal control and the MCI expressed in terms of expected and exceedance probabilities. In the normal control, we found that a control feature of AIN (model 2: first column) has higher probability compared with that of control feature of DMN (model 1: first column) and of CEN (model 3: first column). In the MCI, we further revealed that the control feature of AIN (model 2: second column) no longer has dominant probability. We also examined that the control feature of PIN in the normal control (model 2: third column) and the MCI (model 2: fourth column). We found that control feature of PIN (model 2) does not have dominant probability in the normal control, but it has higher probability than other models in the MCI group. The modulation probability from PI to AI increased in MCI (Supplementary Figure S1).

We compared the model probability within the group and between the groups using Wilcoxon rank sum test. In normal control group, we found that modulation by AIN (model 2 in left column **Figure 6**) had statistically higher probability than model 1 (p = 0.046; Z = 1.987) and model 3 (p = 0.002; Z = 3.076), but there was no statistical difference between model 1 and model 3 (p = 0.531; Z = 0.627). In MCI group, AIN did not have higher probability, but model 3 had significantly higher probability than model 1 (p = 8.539 × 10−19; Z = 8.853) and model 2 (p = 7.816 × 10−18; Z = 8.602) and model 2 had higher probability than model 1 (p = 1.911 × 10−<sup>4</sup> ; Z = 3.731). Each model probability showed statistical difference between normal control and MCI groups: model 1 (p = 8.606 × 10−11; Z = 6.489), model 2 (p = 8.257 × 10−<sup>9</sup> ; Z = 5.763), and model 3 (p = 4.089 × 10−10; Z = 6.251). On the other hand, in normal control group we found that modulation by PIN (model 2 in right column) did not have higher probability. Model 3 had significantly higher probability than model 1 (p = 7.616 × 10−<sup>8</sup> ; Z = 5.376) and model 2 (p = 6.513 × 10−<sup>8</sup> ; Z = 5.404) and model 1 had higher probability than model 2 (p = 1.254 × 10−<sup>7</sup> ; Z = 5.285). In MCI group, our analysis revealed PIN modulations (model 2 in right column) had statistically higher probability than model 1 (p = 3.888 × 10−11; Z = 6.608) and model 3 (p = 1.278 × 10−10; Z = 6.429),

but there was no statistical difference between model 1 and model 3 (p = 0.197; Z = 1.289). In PIN modulations, there was statistical difference between normal control and MCI groups in model 2 (p = 5.216 × 10−10; Z = 6.212) and in model 3 (p = 1.557 × 10−<sup>9</sup> ; Z = 6.038), but not in model 1 (p = 0.985; Z = 0.019).

### Association between Network Interactions and Cognitive Scores

We studied the association between the network modulation probability and the cognitive scores. We found statistically significant correlation between the MoCA and the modulation probability of AIN (Spearman's correlation: r = 0.47;

respectively. The fourth and third columns display that the modulations by PIN had a higher probability in the MCI but not in the NC, respectively.

p = 3.76 × 10−<sup>5</sup> ) and of PIN (r = −0.43; p = 1.97 × 10−<sup>4</sup> ), respectively, as shown in **Figure 7**.

The modulation probability significantly associated with the delayed recall memory function and visuospatial-executive function "Subscores of MOCA". Delayed recall memory test score associated with the modulation probability of AIN (Spearman's correlation: r = 0.30; p = 0.010), and with the modulation probability of PIN (r = −0.299; p = 0.012). Visuospatialexecutive test score correlated with the modulation probability of AIN (r = 0.343; p = 0.004), and with the modulation probability of PIN (r = −0.304; p = 0.011) (**Figure 8**). Trail marking test B score showed weak correlation with the modulation probability of AIN (p = 0.173) and of PIN (p = 0.058).

### DISCUSSION

Here, we evaluated the switching/modulation effects of insula subdivisions-based networks—AIN and PIN—on the DMN and CEN in MCI group in comparison to a group of normal controls. The AIN was found to exert modulation effects on the DMN and CEN in control group, consistent with former studies (Sridharan et al., 2008; Chand and Dhamala, 2016a; Wu et al., 2016). However, this modulation effect of the AIN was impaired in MCI group (Chand et al., 2017a). Furthermore, the PIN did not provide modulation effects on the DMN and CEN in normal group, and in contrast the PIN took over some control feature of the AIN in MCI group (the AIN was impaired in MCI). Finally, the global cognitive test scores were correlated with the modulating probability of the AIN and of the PIN.

Previous investigations suggest that control feature of AIN in cognitively normal group might be carried out with the help of Von Economo neurons (Allman et al., 2005, 2010; Watson et al., 2006). Specifically, those studies report that Von Economo neurons are present abundantly in anterior insula and dorsal anterior cingulate cortex nodes of AIN, but

(r = –0.43; p = 1.97 × 10−<sup>4</sup> ).

there are no reports of their presence in posterior insula and sensorimotor area of PIN. Literature shows that the anterior insula of AIN is functionally connected to the networks responsible for adaptive behavior (Seeley et al., 2007) and to the fronto-parietal control network (Vincent et al., 2008). Anterior insula also has a direct white matter connections to other key brain nodes and lobes such as dorsal anterior cingulate cortex (Jilka et al., 2014), inferior-parietal lobe (Uddin et al., 2010), and the temporo-parietal junction (Kucyi et al., 2012). Thus, the anterior insula involves in a wide range of cognitive processes, including reorienting the attention (Ullsperger et al., 2010) and switching between cognitive resources (Uddin and Menon, 2009). The activity in the dorsal anterior cingulate cortex of AIN is crucial in monitoring the conflict, switching between cognitive states in association with anterior insula during harder decision-making tasks, and implement behavioral changes (Egner, 2009; Chand and Dhamala, 2016a). The control signal of AIN might be carried out by the neural bases mentioned above. The PIN encompasses the posterior insula and sensorimotor areas, specifically temporal and posterior cingulate regions, and is suggested to involve in interoceptive and/or sensorimotor processes (Cauda et al., 2011; Nomi et al., 2016). The posterior insula has well-developed functional connections with the auditory cortex and has been consistently reported in auditory processing, supporting the findings that it is mainly a sensory region (Cauda et al., 2011). Structural connectivity analysis consistently demonstrates that posterior insula has direct white matter connections with the parietal and posterior temporal regions, and anterior temporal regions to a lesser extent (Cerliani et al., 2012; Cloutman et al., 2012; Dennis et al., 2014). The posterior insula and the middle insula are consistently reported to exhibit overlapping cognitive functions (Deen et al., 2011). Emerging studies report that the insula subdivisions exhibit the unique and overlapping profiles in a wide range of cognitive processes and argue that such overlapping functional profiles might be helpful in restoring cognitive functions (Starr et al., 2009; Segerdahl et al., 2015; Nomi et al., 2016; Namkung et al., 2017). The control ability achieved by the PIN in MCI in the present study might thus support the putative roles of overlapping functional activities of the insula divisions.

Literature reports that the AIN atypically engaged in disease, including autism, schizophrenia, fronto-temporal dementia, and Alzheimer's disease (Menon, 2015; Uddin, 2015; Chand et al., 2017a,b). A large body of MCI and/or Alzheimer's disease investigations repeatedly suggest that the DMN activity decreases (Greicius et al., 2004; Greicius and Kimmel, 2012; Brier et al., 2014), but the role of CEN activity has been conflicted with the progression of disease (Diener et al., 2012). The CEN, especially its dorsolateral prefrontal cortex node, abundantly connects with visual, somatosensory, and auditory areas, and therefore might possess the crucial role in a wide range of cognitive functions, including goal-orientated actions (Petrides and Pandya, 1999; Miller and Cohen, 2001; Chand et al., 2016; Chand and Dhamala, 2017). The functional role of CEN activity has been inconsistently reported in disease (Diener et al., 2012). Whether the dorsolateral prefrontal cortex—a key node of CEN—hyperactive or hypoactive in disease has remained conflicting in those studies. In our case, we observed the higher probability of CEN modulation with the AIN and DMN in MCI. On the other hand, the modulation probability of CEN during interactions with the PIN and DMN was smaller in MCI than in normal control group. The alterations of CEN modulations thus remain unclear. Prior studies and our findings together suggest that, as AIN control is disrupted in MCI individuals, the PIN might come up to take over the control features, and this control might possibly decline when MCI changes to dementia

visuospatial-executive function subscore of MoCA (r = –0.304; p = 0.011).

or Alzheimer's disease. A detailed description of this decline mechanism can be explored in the future by including the data from individuals with dementia or Alzheimer's disease.

In summary, we evaluated the patterns of connectivity of the PIN and/or AIN over the DMN and CEN in MCI people and compared with a group of cognitively normal people. We revealed that the PIN took control over DMN and CEN in MCI group where the control activity of AIN was impaired. These findings provide important implications about the underlying flexible functional profiles of insula subdivision-based brain networks and their interactions in normal cognition and MCI.

#### DISCLOSURE STATEMENT

All authors have approved the manuscript and agree with submission to this journal. We have read and have abided by the statement of ethical standards for manuscripts submission.

## AUTHOR CONTRIBUTIONS

Designed the experiment: IH and DQ. Performed the experiment: GC, JW, DQ, and IH. Analyzed the data: GC, DQ, and IH. Wrote the paper: GC, DQ, and IH. Participated in the discussion and provided the comments: GC, JW, DQ, and IH.

#### FUNDING

NIA/NIH grants RF1AG051633 and R01AG042127 supported to IH. NIH grants AG25688, AG42127, AG49752, AG51633 supported to DQ.

#### ACKNOWLEDGMENT

fnagi-09-00367 November 7, 2017 Time: 16:47 # 10

The authors would like to thank the members of our team and the volunteers for their participation in the present study.

#### REFERENCES


#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi. 2017.00367/full#supplementary-material



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Chand, Wu, Hajjar and Qiu. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Linking Inter-Individual Variability in Functional Brain Connectivity to Cognitive Ability in Elderly Individuals

Rui Li1,2, Shufei Yin<sup>3</sup> , Xinyi Zhu1,2, Weicong Ren1,4, Jing Yu1,5, Pengyun Wang1,2 , Zhiwei Zheng1,2, Ya-Nan Niu1,2, Xin Huang1,2 and Juan Li1,2,6,7 \*

<sup>1</sup> CAS Key Laboratory of Mental Health, Institute of Psychology, Beijing, China, <sup>2</sup> Department of Psychology, University of Chinese Academy of Sciences, Beijing, China, <sup>3</sup> Department of Psychology, Faculty of Education, Hubei University, Wuhan, China, <sup>4</sup> Department of Education, Hebei Normal University, Shijiazhuang, China, <sup>5</sup> Faculty of Psychology, Southwest University, Chongqing, China, <sup>6</sup> Magnetic Resonance Imaging Research Center, Institute of Psychology, Chinese Academy of Sciences, Beijing, China, <sup>7</sup> State Key Laboratory of Brain and Cognitive Science, Institute of Biophysics, Chinese Academy of Sciences, Beijing, China

Increasing evidence suggests that functional brain connectivity is an important determinant of cognitive aging. However, the fundamental concept of inter-individual variations in functional connectivity in older individuals is not yet completely understood. It is essential to evaluate the extent to which inter-individual variability in connectivity impacts cognitive performance at an older age. In the current study, we aimed to characterize individual variability of functional connectivity in the elderly and to examine its significance to individual cognition. We mapped inter-individual variability of functional connectivity by analyzing whole-brain functional connectivity magnetic resonance imaging data obtained from a large sample of cognitively normal older adults. Our results demonstrated a gradual increase in variability in primary regions of the visual, sensorimotor, and auditory networks to specific subcortical structures, particularly the hippocampal formation, and the prefrontal and parietal cortices, which largely constitute the default mode and fronto-parietal networks, to the cerebellum. Further, the inter-individual variability of the functional connectivity correlated significantly with the degree of cognitive relevance. Regions with greater connectivity variability demonstrated more connections that correlated with cognitive performance. These results also underscored the crucial function of the long-range and inter-network connections in individual cognition. Thus, individual connectivity–cognition variability mapping findings may provide important information for future research on cognitive aging and neurocognitive diseases.

Keywords: individual variability, functional connectivity, cognitive aging, fMRI, brain networks

### INTRODUCTION

There is a marked heterogeneity in cognitive functioning during late adulthood and old age (Hedden and Gabrieli, 2004; Lustig et al., 2009; Nyberg et al., 2012). Some older people may display rapid cognitive decline or develop Alzheimer's disease (AD), whereas others may continue to exhibit a superior level of cognitive functioning. One of the main contributions to this heterogeneity originates from the variability of the brain (Hedden and Gabrieli, 2004; Reuter-Lorenz and Lustig, 2005; Bishop et al., 2010; Grady, 2012; Tomasi and Volkow, 2012;

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Xu Lei, Southwest University, China Jiaojian Wang, University of Electronic Science and Technology of China, China

#### \*Correspondence:

Juan Li lijuan@psych.ac.cn

Received: 26 July 2017 Accepted: 09 November 2017 Published: 21 November 2017

#### Citation:

Li R, Yin S, Zhu X, Ren W, Yu J, Wang P, Zheng Z, Niu Y-N, Huang X and Li J (2017) Linking Inter-Individual Variability in Functional Brain Connectivity to Cognitive Ability in Elderly Individuals. Front. Aging Neurosci. 9:385. doi: 10.3389/fnagi.2017.00385

**401**

Karama et al., 2014), particularly in regard to functional connectivity (Burke and Barnes, 2006; Andrews-Hanna et al., 2007; Bishop et al., 2010; Grady, 2012; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013; Fornito et al., 2015).

Previous studies demonstrated that preserved functional integration between distributed brain regions supports proficient cognitive function, while functional disruption of the interregional neural communication results in cognitive decline and AD (Andrews-Hanna et al., 2007; Eyler et al., 2011; Grady, 2012; Nyberg et al., 2012; Dennis and Thompson, 2014). Much of this evidence comes from direct comparisons of functional connectivity between groups that are pre-defined by neuropsychological questionnaires or clinical classifications of mental states. For instance, elderly individuals who performed better on a verbal fluency test demonstrated stronger connections between the precuneus and prefrontal regions compared to that in individuals with lower verbal fluency test performance (Yin et al., 2015). Similarly, patients with AD exhibited disrupted functional connectivity in the default mode and several frontoparietal attention networks, compared to that of healthy elderly individuals (Wang et al., 2007, 2015; Buckner et al., 2009; Myers et al., 2014). These "group differences" provide substantial insights into the brain connectivity correlates of cognitive aging. However, a fundamental issue regarding how functional brain connectivity itself differs among older individuals remains to be elucidated. Although many group-based investigations usually included individual-level results, the "individual difference" in functional connectivity remains largely uninvestigated. For example, Betzel et al. (2014) demonstrated the trajectory of individual functional connectivity of resting state networks with age (Betzel et al., 2014). Similarly, there are studies that have largely demonstrated individual-level correlations between functional connectivity and cognitive performance in normal elderly people (Andrews-Hanna et al., 2007; Sala-Llonch et al., 2014; Yin et al., 2015) and patients (Wang et al., 2015). However, it is still not clear how inter-individual variability in functional connectivity can vary in different brain regions and to what extent the inter-individual variability in connectivity impacts cognitive performance at an older age.

An important reason for the bias toward group differences is that traditional task-based neuroimaging studies are limited in their ability to systematically quantify individual brain function differences, given the diverse nature of the tasks used in different studies. Resting-state functional connectivity magnetic resonance imaging (fcMRI) that measures the intrinsic temporal synchronization of the blood oxygen level-dependent (BOLD) signals has been developed to delineate the neural functional architecture in human participants who are not engaged in any specific task. Similar to genomic and phenomic approaches, fcMRI is recognized as a remarkably powerful tool to understand individual variation in brain functioning (Mohr and Nagel, 2010; Buckner, 2013; Mueller et al., 2013; Zatorre, 2013). Mueller et al. (2013) recently measured individual differences of the restingstate connectivity of the cortical regions in 25 healthy adults. The authors reported higher variability in the association cortex and lower variability in the unimodal cortices. Similarly, Gao et al. (2014) examined the inter-individual variability of functional connectivity during infancy (Gao et al., 2014). However, to our knowledge, there have been no studies to date regarding the distribution of the inter-individual differences in functional connectivity in the brains of elderly individuals.

In the current study, we aimed to investigate two major issues as follows: (1) we sought to delineate the inter-individual variability map of functional brain connectivity during old age. The fcMRI data from 108 healthy older adults were collected during resting-state conditions. The brain was divided into 116 regions of interest (ROIs), including cortical, subcortical, and cerebellar regions, using the automated anatomical labeling (AAL) procedure (Tzourio-Mazoyer et al., 2002). The variation of the individual-to-individual functional connectivity in each ROI of these older adults were then estimated and used to generate the brain variability map. Further, to facilitate the inspection of the brain distribution for the inter-individual variability, the variability was compared in 6 distinct brain systems, including the default mode, fronto-parietal, visual, sensorimotor and auditory, subcortical, and cerebellar networks (Ferrarini et al., 2009; He et al., 2009); and 2) we then linked the inter-individual variability of the connectivity to cognitive function in the elderly. A battery of standardized neuropsychological tests was employed to assess the cognitive function of the older participants. The connectivity–cognition association was first examined by calculating the correlations between each region's connectivity and the cognitive test performance. This allowed us to determine whether the regions that had correlations between connectivity and cognitive ability were concentrated in the areas with large inter-individual variability for functional connectivity. Then, we defined a cognitive relevance index that was calculated as the number of cognition-correlated connections to quantify the role of each region's functional connectivity in cognitive functioning. To examine the cognitive significance of the distribution of interindividual variability of functional connectivity, a correlation between the value of inter-individual variability and the degree of cognitive relevance was computed across all ROIs. This allowed us to further determine whether regions with larger inter-individual variability in the brain connectivity would play a more important role in cognitive performance of the elderly. Recent studies have suggested that the long-range and internetwork regional connections function critically in cognitive processing and cognitive aging (Tomasi and Volkow, 2012; Park and Friston, 2013; Fjell et al., 2015). Therefore, to better describe the relationship between variability in connectivity to cognitive significance, we also investigated whether this relationship was more specific to the long-range and inter-network connections.

### MATERIALS AND METHODS

#### Participants

A total of 108 cognitively normal, older volunteers (70.3 ± 5.7 years; range: 60–80 years of age; 50 men and 58 women) were recruited from communities near the Institute of Psychology-Chinese Academy of Sciences. All participants met the following inclusion criteria: age ≥60 years; a score ≥ 21 on the Beijing Version of the Montreal Cognitive Assessment

(Yu et al., 2012); a score ≤ 16 on the Activities of Daily Living (Lawton and Brody, 1969); right-handed; and free of stroke, heart disease, diabetes mellitus, neurological and psychiatric disorders, and traumatic brain injury. The images were collected under resting-state conditions using a 3.0-T Siemens Trio scanner (Erlangen, Germany), located at the Beijing MRI Center for Brain Research. Functional imaging consisted of 33 T2<sup>∗</sup> -weighted echo-planar image (EPI) slices (time repetition (TR) = 2000 ms, time echo (TE) = 30 ms, flip angle = 90◦ , field of view (FOV) = 200 mm × 200 mm, thickness = 3.0 mm, gap = 0.6 mm, acquisition matrix = 64 × 64, and in-plane resolution = 3.125 × 3.125). We collected 200 functional volumes for each participant. T1-weighted anatomical images were collected using a magnetization-prepared rapid gradient echo (MPRAGE) sequence (176 slices, acquisition matrix = 256 × 256, voxel size = 1 mm × 1 mm × 1 mm, TR = 1900 ms, TE = 2.2 ms, and flip angle = 9 ◦ ) for co-registration with the functional images. Of the total number of participants, 85 participants completed a battery of neuropsychological assessments, which included the Digit Forward Span (DFS) and Digit Backward Span (DBS) (Gong, 1992), the Paired Associative Learning Test (PALT) (Xu and Wu, 1986), the Trail Making Test (TMT) Parts A and B (Reitan, 1986), and the Verbal Fluency Test (VFT) (Rosenberg et al., 1984).

Five participants were excluded due to poor image quality or gross structural abnormalities. Six participants were excluded because of excessive head movements (more than 2.0 mm maximum translation or 2.0◦ rotation) during the scan. Nine participants were excluded because of bad registration quality during the visual inspection for the normalization. Thus, the final statistical analysis included fMRI data from 88 older adults (70.2 ± 5.6 years; range: 60–80 years of age; 40 men and 48 women). Of these, 76 individuals (70.7 ± 5.5 years; range: 60–80 years of age; 35 men and 41 women) completed the neuropsychological assessments and provided behavioral data.

The institutional review board of the Institute of Psychology of Chinese Academy of Sciences approved the current study. All participants provided written informed consent prior to their participation in the experiments.

#### Image Preprocessing

Data pre-processing was performed using the Statistical Parametric Mapping program<sup>1</sup> (SPM8) and the Data Processing Assistant for Resting-State fMRI<sup>2</sup> (DPARSF). This included the following: removal of the first 5 volumes, corrections for the intra-volume acquisition time differences between the slices using the Sinc interpolation, corrections for the intervolume geometrical displacement due to head motion using a 6-parameter (rigid body) spatial transformation, a normalization to the standard Montreal Neurological Institute (MNI) space (resampling voxel size, 3 mm × 3 mm × 3 mm) using the DARTEL approach (Ashburner, 2007), spatial smoothing with a 4-mm full width at a half maximum Gaussian kernel to decrease the spatial noise, and de-trending and temporal band-pass filtering (0.01–0.08 Hz) to reduce the effects of low-frequency drifts and high-frequency physiological noise (Lowe et al., 1998). To remove the head motions for each participant, we performed a nuisance regression of the head motion, using a Friston 24-parameter model (6 head motion parameters, 6 head motion parameters one time point before, and the 12 corresponding squared items) (Friston et al., 1996) with scrubbing (Satterthwaite et al., 2013; Yan et al., 2013a,b; Power et al., 2014). We calculated the mean framewise displacement (FD), which was derived using the Jenkinson's relative root mean square (RMS) algorithm (Jenkinson et al., 2002). This was used as a covariate in the group analyses of the connectivity–cognition correlations to further control for any residual effects of head movement (Yan et al., 2013a,b; Power et al., 2014). In addition, we performed a nuisance regression of the global signal (the average voxel signal within the SPM apriori mask (brainmask.nii) thresholded at 50%, and the white matter and cerebrospinal fluid signals, which were calculated by averaging the voxel signals within the SPM apriori masks (white.nii and csf.nii, respectively) thresholded at 99%. The residual volumes were retained for use in the following functional connectivity analysis.

### Measuring the Inter-Individual Variability of Functional Connectivity

To create the regions for the functional connectivity analyses, we parcellated the brain into 116 ROIs, including 90 cerebral regions and 26 cerebellar regions, based on the AAL atlas (Tzourio-Mazoyer et al., 2002). To ensure that only the gray matter voxels within the AAL ROIs were included in the analyses, these ROIs were multiplied by the SPM's gray matter mask, which was thresholded at 20%, to further remove white matter, cerebrospinal fluid, and other non-brain tissue voxels. The mean time series of each ROI was calculated. Pearson's linear correlation coefficients (r values) were computed between each ROI pair of the averaged time series and subsequently transformed to Fisher z values, which yielded a 116 × 116 correlation matrix for each participant. For a given AAL ROI R<sup>i</sup> (i = 1, 2, . . . 116), the functional connectivity of the participant, S<sup>m</sup> (m = 1, 2, . . . 88), was denoted as a 1 × 115 correlation coefficient vector, FC(Sm)<sup>i</sup> , in which each element corresponded to its correlation with each of the remaining 115 regions. To quantify the inter-individual variability at R<sup>i</sup> , the interindividual similarity, FCS<sup>i</sup> was first calculated as the mean (E) of the correlation values between any two functional connectivity vectors of the 88 older participants:

$$F\text{CS}\_i = E[corr(FC(\mathbb{S}\_m)\_i, FC(\mathbb{S}\_n)\_i)],$$

where m, n = 1, 2, . . . 88, and m6=n.

The inverted similarity (1–FCSi) was thus defined as the interindividual variability (FCVi) of the functional connectivity at R<sup>i</sup> (Mueller et al., 2013). This calculation was repeated for all R<sup>i</sup> ROIs to derive the spatial distribution of the inter-individual variability of the functional connectivity across the entire brain.

Further, we investigated the inter-individual variability for distinct functional systems in the older participants. Previous

<sup>1</sup>http://www.fil.ion.ucl.ac.uk/spm

<sup>2</sup>http://www.rfmri.org

functional connectome analyses of the brain architecture indicated the existence of a hierarchical modularity, which is typically represented as intrinsic functional networks (Ferrarini et al., 2009; He et al., 2009; Park and Friston, 2013; Turk-Browne, 2013). Here, we associated the 90 cerebrum regions with five networks, including the sensorimotor and auditory network, visual network, fronto-parietal network related to attention and executive function, default-mode network, and the subcortical network (He et al., 2009), and another 26 regions to the cerebellar network. The inter-individual variability values were averaged across the regions from the same functional network. A oneway analysis of variance (ANOVA) with network as a factor (six networks) followed by post hoc pair-wise comparisons were performed to investigate the differences in the inter-individual variability between the different functional networks (Bonferroni corrected for 15 comparisons, threshold at 0.05/15 ≈ 0.0033).

### Linking Inter-Individual Functional Variability to Cognitive Ability

First, we calculated the correlations between functional connectivity and cognitive ability. Individual cognitive performance was assessed using four functional domains, including working memory (indexed by the average z-score of DFS and DBS), episodic memory (the z-score of PALT), executive function (inverted z-score of TMT B-A), and vocabulary (the z-score of VFT). In addition, the composite average z-score on all tests was considered a measure of individual global cognitive function. Correlation analyses between each functional domain and the global measure and connectivity of all ROI pairs were performed in a subset of participants (n = 76). With an emphasis on the overall trend of the connectivity–cognition relationship, we used a liberal threshold of p < 0.01 to map the correlation patterns between the cognitive measures and interregional connectivity of all ROI pairs. Age, sex, education level, and the mean head motion FD were considered covariates during the connectivity–cognition correlation analyses. In addition, to further describe the relationship between individual cognition levels to the connectome measures, the number of long-range (Euclidean distance >75 mm between the centroids of the connected regions in stereotactic space), short-range (Euclidean distance ≤ 75 mm) (Achard et al., 2006; Liang et al., 2013), intranetwork (connections within the 6 networks mentioned above), and inter-network (connections between the six networks) connections that were significantly related to each cognitive measure were calculated.

Then, to quantify the significance of the functional connectivity of each region with individual cognitive ability in elderly individuals, a cognitive relevance index was defined. It was measured as the number of connections (including the total connections, long-/short-range connections, and inter- /intra-network connections, respectively) that was significantly correlated with all cognitive variables at each ROI.

Finally, to evaluate the cognitive significance of interindividual variability in connectivity, we examined the correlation between the values of inter-individual variability and the values of cognitive relevance across all the AAL ROIs (p < 0.05). We were interested in examining whether a larger inter-individual variability in the brain connectivity would be more cognitively relevant.

### Evaluating Potential Confounding Factors

First, global signal regression (GSR) is a controversial step that may significantly affect the results and conclusions. Recent studies have suggested that GSR can decrease dependence on motion, remove artifactual variance, and provide increased tissue sensitivity (Fox et al., 2009; Satterthwaite et al., 2013; Yan et al., 2013a; Power et al., 2014). However, other studies have demonstrated that GSR may introduce undesirable negative correlations (otherwise largely absent from the connectivity data) that alter inter-individual differences (Fox et al., 2009; Gotts et al., 2013; Saad et al., 2013). In view of these conflicting reports, we included the results without GSR (nGSR) as Supplementary Material for the present study.

Second, different AAL regions vary in regional noise and volume, both of which may potentially drive the inter-individual variability distribution. To rule out these possibilities, we calculated the temporal signal-to-noise ratio (SNR), which was measured as the average signal across time divided by standard deviation across time for each voxel, and averaged the SNR of voxels within each ROI. We also calculated the number of voxels for each ROI to index the volume of each AAL region. A correlation analysis (p < 0.05) between the SNR/volume and the inter-individual variability values of the ROIs was examined.

Third, although its size in relation to the entire brain is small, recent studies mapping the cerebellar topographical organization suggest that the cerebellum is functionally heterogeneous (Buckner, 2013). Therefore, cerebellar ROIs may be more prone to contain functionally diverse gray matter compared to that of other ROIs. To rule out this potential confound, a connectivity atlas of the cerebellum, which was adopted by a previous study (Buckner et al., 2011) with large data set (N = 1000) to calculate the functional connectivity of different cerebellar regions with neocortical network, was used to perform an additional analysis. We chose the 17-network parcellation atlas of the cerebellum. The voxels assigned to the same network were considered as one ROI; thus, the 17 ROIs from Buckner et al. (2011) were used to replace the 26 AAL cerebellar ROIs, and to recalculate the interindividual functional variability in the brain. This allowed us to rule out the possibility that high functional heterogeneity in the cerebellum may influence the variability estimation.

Finally, to further confirm the robustness of the result with regard to functional inter-individual variability in the elderly, we validated the result by analyzing an independent replication resting-state fMRI dataset (N = 49; 12 men and 37 women; 67.1 ± 4.8 years; range: 60–76 years of age). The data were acquired using a Philips Achieva 3.0-T MRI scanner (Philips Healthcare, Andover, MA) at the MRI Center of the First Hospital of Hebei Medical University of China. Functional images were collected using an EPI sequence with TR = 2000 ms, TE = 30 ms, flip angle = 90◦ , FOV = 200 mm × 200 mm, thickness = 3.6 mm, matrix = 112 × 112; in-plane resolution = 1.786 × 1.786, 33 axial slices, and 200 volumes. T1-weighted MPRAGE image was collected with the following parameters: 176 slices;

matrix = 256 × 256; voxel size = 1 mm × 1 mm × 1 mm; TR = 1900 ms; TE = 2.2 ms; flip angle = 9 ◦ . The individual variability of functional connectivity in this dataset was estimated using the same procedure as described above.

### RESULTS

### Inter-Individual Variability in Functional Brain Connectivity

The exploration of the whole-brain functional connectivity in 116 AAL regions indicated a highly uneven distribution pattern for inter-individual variability in the 88 older participants (Supplementary Figure 1). There was an overall tendency that the inter-individual functional variability increased from the primary areas to the subcortical structures and association cortex to the cerebellum across the whole brain. The mean variability in the cerebellum (0.72 ± 0.10) was significantly larger (twosample t-test, p < 0.0001) than that in the cerebral regions (0.59 ± 0.07). In the cerebrum (**Figure 1A**), the inter-individual difference in functional connectivity was higher in the frontal and parietal cortices; pre- and post-central gyri; anterior, middle and posterior cingulated gyri; parahippocampus; hippocampus; and amygdala and lower in the occipital, temporal, and other subcortical regions.

The analyses in the six specific functional systems (**Figure 1B**) further highlighted a gradual increase in the functional variability from the visual, subcortical, and sensorimotor and auditory networks to the default and fronto-parietal networks, and to the cerebellar network. The ANOVA revealed a significant main effect of network in the functional variability (p < 0.001). The post hoc comparisons demonstrated that the mean interindividual variability in the cerebellar network was significantly larger (p < 0.001) than that of each of the other five networks at a Bonferroni-corrected threshold of p = 0.0033 (0.05/15). The fronto-parietal network exhibited a trend toward a higher variability compared with the visual network (p < 0.01) and subcortical network (p < 0.05).

#### Connectivity–Cognition Correlation

**Figure 2** shows the Pearson correlations of the connectivity of all ROI pairs with individual global cognitive function and the four specific cognitive domains (p < 0.01, uncorrected). The largest number of connections from the superior and orbital prefrontal cortex and the cerebellum consistently correlated with individual scores in global cognition and in the four specific cognitive domains. Further, the functional connectivity of the following connections were related to the four cognitive measures: (1) from the middle, anterior, and posterior cingulate; hippocampus; parahippocampus; amygdala; and precentral gyrus for working memory (DFS and DBS); (2) from the middle temporal pole, middle temporal gyrus, postcentral gyrus, precuneus, thalamus, parahippocampus, hippocampus, anterior and posterior cingulate, and putamen for episodic memory (PALT); (3) from the middle temporal gyrus, middle temporal pole, anterior cingulate, inferior parietal lobule, and parahippocampus for executive function (TMT B-A); and (4) from the fusiform, supramarginal gyrus, angular gyrus, middle temporal gyrus, middle temporal pole, and hippocampus for vocabulary (VFT).

Long-range and inter-network connectivity accounted for a considerable proportion of connections that predicted individual cognition (**Figure 2B**). There were more long-range connections than short-range connections (56.7% vs. 43.3% in total), which correlated with both global measures and specific measures, except for the vocabulary score. Moreover, inter-network connections accounted for 87.3% of all the connections that correlated with the cognitive measures in the whole brain, and consistently preponderated over the intra-network connections when the six functional networks were separately analyzed (**Figure 2C**). We also summarized the total number of internetwork connections that correlated with the four specific cognitive measures for each network. Interestingly, we found that the six networks were in the same variability rank order, except for the subcortical network, which moved up to second place (**Figure 3**).

Finally, we calculated the cognitive relevance, which was indexed by the number of connections that were significantly correlated with the cognitive measures, for each AAL ROI. The distribution map for the cognitive relevance (**Figure 4A**) was similar to the inter-individual functional connectivity variability map (**Figure 1A**). The correlation analysis revealed that the value of the inter-individual functional variability was significantly correlated with the cognitive relevance across the 116 ROIs (Pearson correlation r = 0.29, p = 0.001; **Figure 4B**). Regions with higher inter-individual functional connectivity variability demonstrated more connections that correlated with cognitive performance. More interestingly, when examining the number of long-/short-range and inter-/intra-network connections, the value of the inter-individual variability significantly correlated with the degree of cognitive relevance for the long-range (Pearson correlation r = 0.32, p < 0.001; **Figure 4C**) and inter-network (Pearson correlation r = 0.30, p = 0.001; **Figure 4D**) connectivity across all ROIs. There was no significant correlation between the inter-individual variability and the short-range (Pearson correlation r = 0.10, p = 0.27) or intra-network (Pearson correlation r = 0.16, p = 0.09) connectivity cognitive relevance measures in the brains of elderly individuals.

### Impact of Potential Confounds

First, we re-calculated the functional inter-individual variability without removing the global signal in the preprocessing. Variability maps, estimated with (**Figure 1**) and without (Supplementary Figures 1, 2) GSR, demonstrated a highly similar pattern (Pearson correlation r = 0.92, p < 0.0001). The cerebellum retained the largest mean inter-individual variability compared to that of cerebral regions (two-sample t-test, p < 0.0001). The network-level variability also consistently demonstrated significant statistical difference for the functional variability among the networks (p < 0.001), with gradually increased variability occurring in the subcortical network, then the primary networks (i.e., visual, sensorimotor, and auditory networks), to the association networks (i.e., default and frontoparietal networks), and to the cerebellar network (Supplementary

Figure 2B). However, as expected, the GSR largely affected the connectivity–cognition correlations, such that the GSR preprocessing introduced more negative correlations (**Figure 2**) than the nGSR preprocessing (Supplementary Figure 3). As an overall trend, this was consistent with the GSR results regarding the inter-network connectivity, especially for the connections from the superior and orbital prefrontal cortex, hippocampus, and the cerebellum predominating individual cognitive ability. It is important to note that the retention of the global signal diminished the correlation between the longrange connections and cognition, with a larger proportion of the long-range connections only found in the global measure

connections, as well as the intra-network and inter-network connections that are correlated with each cognitive domain. (C) The bars show the total number of connections within each functional network (transparent bars) and the total number of connections with other networks (non-transparent bars) that are correlated with each cognitive domain.

thickness of the connections is proportional to the connectivity–cognition correlation coefficients. (B) The bars show the total number of short-range and long-range

and vocabulary score. In addition, in the nGSR condition, the relationship between the value of the inter-individual functional connectivity variability and the cognitive relevance across all ROIs disappeared (Pearson correlation r = –0.13, p = 0.17; Supplementary Figure 4).

Next, we calculated the correlation between the regional SNR/size and the inter-individual functional connectivity variability values across all ROIs, to exclude the possibility that the ranking of the regional inter-individual variability was primarily driven by potential noise and size effects. The rank of the inter-individual variability derived with GSR was not influenced by the regional noise or size (p > 0.05). The supplementary nGSR result of the inter-individual variability, however, correlated significantly with the regional SNR (Pearson correlation r = 0.34, p < 0.01).

Third, to rule out the possibility that high functional heterogeneity in the cerebellar ROIs influenced the variability estimation, we used the 17-network parcellation atlas of the cerebellum (Buckner et al., 2011) to replace the 26 cerebellar AAL ROIs, which allowed us to perform an additional analysis of the inter-individual functional connectivity variability. Consistent with our findings using the cerebellar AAL ROIs, the additional analysis demonstrated that 5 of the 17 cerebellar ROIs ranked highly for the inter-individual functional variability in the brain. The mean variability of the 17 cerebellar ROIs (0.66 ± 0.12) was significantly larger (two-sample t-test, p = 0.0001) than that of

the cerebral ROIs (0.58 ± 0.07). No significant differences were found for the inter-individual functional connectivity variability in the cerebellum between the two different atlases (two-sample t-test, p = 0.09).

cognition-related inter-network connections for each functional network.

Finally, the robust validation analysis in an independent dataset further confirmed the distribution of inter-individual functional connectivity variability in the brains of elderly people. The distribution patterns in both datasets were highly similar (Pearson correlation r = 0.61, p < 0.0001). Further, the cerebellum had maximal inter-individual variability (0.72 ± 0.11), and the cerebrum demonstrated gradually increased inter-individual variability from the visual (0.61 ± 0.06), subcortical (0.62 ± 0.05), and sensorimotor and auditory (0.63 ± 0.09) networks to the fronto-parietal (0.64 ± 0.09) and default (0.69 ± 0.09) networks.

### DISCUSSION

There is fairly extensive research regarding the relationship between changes in brain connectivity and a broad range of cognitive decline and neuropsychiatric symptoms in aging populations (Hedden and Gabrieli, 2004; Reuter-Lorenz and Lustig, 2005; Andrews-Hanna et al., 2007; Wang et al., 2007, 2015; Bishop et al., 2010; Grady, 2012; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013; Li et al., 2013, 2015; Fornito et al., 2015). Although these studies strongly supported the notion that brain connectivity is an important determinant of cognitive aging, the contribution of personto-person variation remained unclear. Thus, the present study

aimed to bridge the gap in knowledge of how individual variability of functional connectivity and the inter-individual differences affect the cognitive ability of elderly individuals. Our novel study systematically mapped the distribution of individual functional variability on a whole-brain scale, which facilitated understanding of how inter-individual variability differs between different brain areas in the older adults. Further, we demonstrated that the inter-individual variability mapping has important cognitive significance. These findings may thus contribute a valuable reference or evidence for future cognitive aging studies.

### Inter-Individual Functional Variability in Elderly Individuals

Functional connectivity in the cerebral cortex indicated that there was higher inter-individual variability in the frontal and parietal cortices, and the pre- and post-central gyri, while there was lower variation in the occipital and temporal regions in elderly individuals. The cortical variability generally aligned with the results from Mueller et al. (2013) who conducted a study of the inter-individual differences in cortical connectivity in 25 healthy adults. In the current study, we expanded this previous work to include a global analysis of the brain in a large sample of older adults. Our findings indicate that inter-individual variability was the largest in the cerebellum, followed by the association regions that largely constitute the fronto-parietal and default mode networks, as well as some subcortical regions, especially the hippocampal formation. Primary regions, including visual and sensorimotor networks, and other subcortical structures exhibited minimal variability among individuals.

It is not surprising that the functional connectivity in the prefrontal and parietal cortices and the relevant fronto-parietal and default mode networks demonstrated major individual variations in the cortex, because extensive evidence suggests these association regions and network connections are the selective targets of aging effects (Andrews-Hanna et al., 2007; Grady, 2012; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013). The cerebellum has not been substantially investigated in most aging studies. However, there is increasingly converging evidence to suggest that the cerebellum is connected to cerebral association regions, including the prefrontal and posterior parietal cortices, and subcortical structures, including the vestibular nuclei and basal ganglia. Therefore, the cerebellum can contribute to a wide variety of functional domains and neuropsychiatric diseases (Stoodley and Schmahmann, 2009; Bostan et al., 2013; Buckner, 2013). Wagner et al. (2017) recently observed that the cerebellar granule cells could encode reward expectation, suggesting that the cerebellum was involved in cognitive processing (Wagner et al., 2017). Further, in a recent review of multidisciplinary findings, Sokolov et al. (2017) suggested a cerebro-cerebellar loop to explain the involvement of the cerebellum in higher cognitive functions, including attention, language, memory, and social cognition (Sokolov et al., 2017). We identified that the largest inter-individual variability resides in the cerebellum, further indicating that it is a noteworthy region for future aging studies. Additional potential studies include the exploration of how the cerebellum is mediated by the prefrontal and parietal regions in the association functional networks, which would provide a better understanding of its role in aging.

Several potential causes may underlie the distribution of the inter-individual variability in the brain functional connectivity of older individuals. First, the hemodynamic MRI signal is triggered by the metabolic demands of neuronal activities (Heeger and Ress, 2002). The variability map of functional connectivity is consistent with the previous metabolic topography of normal aging, as investigated by positron emission tomography; this technique demonstrated covariant metabolic changes in the prefrontal cortex, lateral temporal and parietal cortices, cerebellum, and basal ganglia (Moeller et al., 1996; Chiaravalloti et al., 2014). Thus, we speculated that the inter-individual variability in the functional connectivity had a physiologically reasonable metabolic basis. Second, our findings may be, in part, a functional consequence of the individual heterogeneity in brain structure morphology that occurs with aging. MRI volumetric studies have demonstrated heterogenic aging patterns across structures regarding neuroanatomical volume loss (Jernigan et al., 2001; Walhovd et al., 2005). Jernigan et al. (2001) observed that the cerebellum exhibited the same striking degree of gray matter reduction with aging as the frontal lobes, and exhibited a more accelerated volume loss than the hippocampus (Jernigan et al., 2001). In addition, other anatomical profiles, such as its cortical folding, thickness, and white matter fiber tracts, may also contribute to the individual differences in the functional correlations (Kanai and Rees, 2011; Mueller et al., 2013; Karama et al., 2014). For example, diffusion tensor imaging of white fiber tracts demonstrated that the variability of aging effects was also regionally complex; this was indicated by a gradient increase in the white matter deficits from the posterior to anterior cortex segments, but also by a greater impairment in the cerebellum (Davis et al., 2009; Bennett et al., 2010). Third, the diverse dynamics and heterogeneous distributions of neurons, as well as the selective vulnerability of synapses and neurons during aging, may also promote individual differences in functional connectivity (Morrison and Hof, 1997; Zhao et al., 2008; Bishop et al., 2010; Urban and Tripathy, 2012; Mejias and Longtin, 2014). Although the cerebellum only accounts for approximately 10% of the total brain weight, it accounts for half of its neurons. Thus, the cerebellum would naturally exhibit more variations due to its densely packed neuronal assembly. Finally, genetic and plasticity factors play critical roles in the inter-individual variability in brain connectivity (Mueller et al., 2013; Toro et al., 2014). Genes determine the individual differences in the evolutionarily recent association cortex, specifically in the prefrontal region (Thompson et al., 2001), where the gene expression patterns exhibit substantially greater heterogeneity in middle-old aged populations (Lu et al., 2004; Bishop et al., 2010). Furthermore, the prefrontal cortex and cerebellum are the final structures to achieve maturity, but are also the first structures to undergo involution in later life (Wang and Zoghbi, 2001; Hogan et al., 2011). This protracted development and prolonged degeneration processes can continue to accumulate deeper and more complex interindividual variations via environment- and lifestyle-dependent neural plasticity.

### Cognitive Relevance of the Inter-Individual Connectivity Variability

fnagi-09-00385 November 18, 2017 Time: 15:47 # 10

The connectivity–cognition correlation suggested that the connections that are related to cognitive ability lay mainly in regions with large inter-individual connectivity differences, including the prefrontal cortex, hippocampal formation, inferior parietal gyrus, middle temporal pole, middle temporal gyrus, and cerebellar regions. The prefrontal, parietal, and temporal regions are the uppermost components of the fronto-parietal and default mode networks, which support high-level cognition. Numerous molecular and neuroimaging studies have repeatedly confirmed their role in cognitive aging, such as in memory, attention, and executive function decline (Hedden and Gabrieli, 2004; Burke and Barnes, 2006; Andrews-Hanna et al., 2007; Bishop et al., 2010; Grady, 2012; Nyberg et al., 2012; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013; Karama et al., 2014). The hippocampal formation is also particularly vulnerable to the aging process. Here, we demonstrated that the functional connectivity of the hippocampus and parahippocampus is correlated with all four cognitive measures, including working memory (DFS and DBS), episodic memory (PALT), executive function (TMT B-A), and vocabulary (VFT). Our results are consistent with a recent meta-analysis of 114 fMRI studies of older adults, which suggested a set of regions that are remarkably involved in cognitive aging; these included the frontal gyrus, parahippocampal gyrus, fusiform gyrus, precentral gyrus, and functional networks, especially the fronto-parietal and default networks (Li et al., 2015).

Long-range and inter-network connections appeared to dominate cognitive ability differences among older adults. Longrange connections are well-known for their key role in efficient brain-wide information processing and functional integration of diverse cognitive functions (Jbabdi et al., 2013; Park and Friston, 2013). Previous evidence demonstrated that the longrange connections in the default and fronto-parietal attention networks are selectively vulnerable to aging and are susceptible to early Alzheimer's disease, compared to that of the shortrange connections (Andrews-Hanna et al., 2007; Tomasi and Volkow, 2012; Li et al., 2013; Wang et al., 2013; Sala-Llonch et al., 2014). Recently, Fjell et al. (2015) found extensive changes in inter-network functional connectivity across multiple cortical networks that were related to a decline in episodic memory with aging (Fjell et al., 2015). It is also interesting to note that although the subcortical network ranked lower for the average interindividual variability, some specific regions with larger interindividual differences, including the hippocampal formation, thalamus, caudate, and amygdala, have considerable connections to regions in other networks that are involved in cognition. Therefore, this may suggest that this network, specifically some specific regions, needs to be considered as having a role in cognitive function through its interactions with other cortical and cerebellar networks.

Importantly, we found larger inter-individual variation of functional connectivity was significantly correlated with higher cognitive relevance, in terms of the number of cognitioncorrelated connections. This relationship suggested that the functional connectome was a major root of individual behavior differences. Moreover, we demonstrated that the correlation between the inter-individual variability and the cognitive relevance of functional connectivity was specific to longrange and inter-network connections. Given the role of longrange and inter-network connections in cognitive performance, this finding further indicated that regions and networks with large inter-individual variability deserve attention in future studies. Thus, these results provide a new perspective for understanding cognitive aging. Currently, most studies are conducted by first assigning participants to different groups, and then exploring differences in the averaged brain activity signals among the groups. In these studies, inter-individual differences in brain function are essentially neglected, or simplified to group differences (Mohr and Nagel, 2010), limiting the full understanding of cognitive aging. Here, the mapping of the inter-individual functional connectivity variability and its correlation with cognition suggested regions and connections, which are typically overlooked but important to cognitive aging studies. For example, the cerebellum showed the largest inter-individual variability and was correlated with diverse cognitive domains. In fact, several studies investigating the role of the cerebellum in aging has emerged. Increasing evidence has indicated that the cerebellum is involved in frontally based functional decline in elderly individuals (Sullivan and Pfefferbaum, 2006; Hogan et al., 2011; Bernard and Seidler, 2014). Future studies should investigate how the prefrontal cortex interacts with the cerebellum, subcortical areas, and other cortical regions to contribute to the inter-individual differences seen with aging. This would be particularly important to distinguish the connectivity–cognition associations that are specific to aging from the inherent general relationships across the lifespan. For example, previously the prefrontal cortex has been overwhelmingly emphasized in cognitive aging. However, a previous molecular genetic expression study (Erraji-Benchekroun et al., 2005) and a recent cortical thickness study (Karama et al., 2014) have stressed that the prefrontal cortex is in fact linked closely with diverse cognitive abilities throughout the human life-span. The mapping of inter-individual variability thus brings a new perspective to future studies that seek key areas affected by cognitive aging. It will also be exciting to investigate inter-individual connectivity variability and its cognitive importance over time to further understand interindividual differences in the trajectories of cognitive aging and specific diseases, such as AD.

#### Limitations

A few limitations of the present study must be noted. First, given the complex and controversial involvement of the GSR in fcMRI studies, we included results both with and without GSR. The GSR was expected to influence the results. The inter-individual variability rank estimated from data with GSR appeared to be more sensitive to the noise than the nGSR results, despite a similar inter-individual ROI variability ranking with both strategies. In addition, the GSR produced negative biased correlations between the individual connectivity and

cognitive measures and magnified the proportion of long-range connections that were correlated with performance, compared to that of the nGSR results. However, both suggested an important role for the inter-network connections in cognitive aging. The influence of the regression of global signal in the present result needs to be carefully considered. Second, we noted that as the present study was confined to the estimation of the distribution of inter-individual functional variability of older adults, intraindividual variations were not considered. Variations in the intra-individual functional connectivity may be caused by measurement instability due to technical noise or changes of mental and biological states (Mueller et al., 2013). A recent study conducted by Chen et al. (2015) depicted the pattern of intra-individual functional variability in the brain of young adults with the use of ten repeated fMRI measurements. Another factor is the temporal moment-to-moment variation within an individual's BOLD signal, which has also been suggested to have predictive significance in relation to cognitive function and various clinical conditions (Garrett et al., 2013). It is necessary for future studies to investigate the distribution characteristic of these intra-individual variations and examine their cognitive correlations with aging and to further investigate how these variations may interact with inter-individual variability. Third, we acknowledge that the connectivity–cognition correlation was not corrected for multiple comparisons. This is because the focus of this study was not to report which regional connections were significantly correlated with cognitive performance. We used a threshold of p < 0.01 (corresponding to r > 0.30) to define the cognitive relevance index for each region, which helped disclose an overall relationship between functional connectivity variability and cognitive association across all brain areas. The definition of "cognitive relevance index" in our study was similar to that of other fMRI connectivity measurements, such as "functional connectivity density" or "degree centrality," which is usually calculated as the number of correlated connections at a liberal correlation coefficient threshold (e.g., r = 0.25), without multiple corrections on the correlations between mass voxels (Buckner et al., 2009; Weng et al., 2016). Fourth, the AAL atlas we used to calculate functional connectivity was defined on the basis of anatomical features. Although the use of an alternative connectivity atlas of cerebellum did not change the cerebellar rank in inter-individual variability, the influence of the ROI definition from using the AAL cannot be fully excluded. As the division of brain regions, as well as their functional characteristics, remains controversial, future studies of data-driven parcellation of brain regions and networks would present more precise estimation of inter-individual variability in elderly individuals. Finally, the current study was focused on mapping a general profile of inter-individual variability in an older population. No attempt was made to examine

#### REFERENCES

Achard, S., Salvador, R., Whitcher, B., Suckling, J., and Bullmore, E. T. (2006). A resilient, low-frequency, small-world human brain functional network with highly connected association cortical hubs. J. Neurosci. 26, 63–72. doi: 10.1523/ jneurosci.3874-05.2006

factors, such as the age of participants, that influence interindividual variability. Thus, further studies are required to investigate the effect of age, as well as other environmental or genetic factors, that can influence individual functional variability.

### CONCLUSION

In the current study, we delineated a map of inter-individual variability in whole-brain functional connectivity for older adults. These results revealed gradually increased variability from the primary regions (including the visual, sensorimotor, and auditory networks), to specific subcortical structures, particularly the hippocampal formation, and the prefrontal and parietal cortices that largely constitute the default mode and fronto-parietal networks, and the cerebellum. The connectivity–cognition results further stressed a crucial function for long-range and internetwork connections in inter-individual cognitive performance. Moreover, the associations between inter-individual variability and the cognition relevance of functional connectivity provide a new perspective for investigating the mechanisms underlying cognitive aging and relevant diseases.

### AUTHOR CONTRIBUTIONS

Conceived and designed the experiments: RL, JL. Performed the experiments: SY, XZ, WR, JY, PW, ZZ, Y-NN, XH. Analyzed the data and wrote the paper: RL.

#### ACKNOWLEDGMENTS

This work was supported by the National Natural Science Foundation of China (31671157, 31470998, 61673374, 31271108, 31200847, 31600904), Beijing Municipal Science and Technology Commission (No. Z171100000117006), the Pioneer Initiative of the Chinese Academy of Sciences, Feature Institutes Program (TSS-2015-06), the National Key Research and Development Program of China (2016YFC1305904), and the CAS Key Laboratory of Mental Health, Institute of Psychology (KLMH2014ZK02, KLMH2014ZG03).

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi. 2017.00385/full#supplementary-material



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Li, Yin, Zhu, Ren, Yu, Wang, Zheng, Niu, Huang and Li. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnagi-09-00385 November 18, 2017 Time: 15:47 # 13

# Biosystems Study of the Molecular Networks Underlying Hippocampal Aging Progression and Anti-aging Treatment in Mice

Jiao Wang<sup>1</sup> , Qian Li<sup>1</sup> , Yanyan Kong<sup>2</sup> , Fangfang Zhou<sup>1</sup> , Jie Li<sup>1</sup> , Weihao Li<sup>1</sup> , Kai Wang<sup>3</sup> , Ting Wu<sup>4</sup> , Yihui Guan<sup>2</sup> , Jiang Xie<sup>5</sup> \* and Tieqiao Wen<sup>1</sup> \*

<sup>1</sup> Laboratory of Molecular Neural Biology, School of Life Sciences, Shanghai University, Shanghai, China, <sup>2</sup> Position Emission Computed Tomography Center, Huashan Hospital, Fudan University, Shanghai, China, <sup>3</sup> Shanghai Key Laboratory of Molecular Andrology, Institute of Biochemistry and Cell Biology, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai, China, <sup>4</sup> Shanghai Stem Cell Group, Shanghai, China, <sup>5</sup> School of Computer Engineering and Science, Shanghai University, Shanghai, China

Aging progression is a process that an individual encounters as they become older, and usually results from a series of normal physiological changes over time. The hippocampus, which contributes to the loss of spatial and episodic memory and learning in older people, is closely related to the detrimental effects of aging at the morphological and molecular levels. However, age-related genetic changes in hippocampal molecular mechanisms are not yet well-established. To provide additional insight into the aging process, differentially-expressed genes of 3- versus 24- and 29 month old mice were re-analyzed. The results revealed that a large number of immune and inflammatory response-related genes were up-regulated in the aged hippocampus, and membrane receptor-associated genes were down-regulated. The down-regulation of transmembrane receptors may indicate the weaker perception of environmental exposure in older people, since many transmembrane proteins participate in signal transduction. In addition, molecular interaction analysis of the up-regulated immune genes indicated that the hub gene, Ywhae, may play essential roles in immune and inflammatory responses during aging progression, as well as during hippocampal development. Our biological experiments confirmed the conserved roles of Ywhae and its partners between human and mouse. Furthermore, comparison of microarray data between advanced-age mice treated with human umbilical cord blood plasma protein and the phosphate-buffered saline control showed that the genes that contribute to the revitalization of advanced-age mice are different from the genes induced by aging. These results implied that the revitalization of advanced-age mice is not a simple reverse process of normal aging progression. Our data assigned novel roles of genes during aging progression and provided further theoretic evidence for future studies exploring the underlying mechanisms of aging and anti-aging-related disease therapy.

#### Keywords: aging, immune response, hippocampal development, umbilical cord blood, molecular interaction network

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Enrique Hernandez-Lemus, National Institute of Genomic Medicine, Mexico Peter Csermely, Semmelweis University, Hungary Athanasios Vasilopoulos, Northwestern University, United States

\*Correspondence:

Tieqiao Wen wtq@shu.edu.cn Jiang Xie jiangx@shu.edu.cn

Received: 24 August 2017 Accepted: 13 November 2017 Published: 06 December 2017

#### Citation:

Wang J, Li Q, Kong Y, Zhou F, Li J, Li W, Wang K, Wu T, Guan Y, Xie J and Wen T (2017) Biosystems Study of the Molecular Networks Underlying Hippocampal Aging Progression and Anti-aging Treatment in Mice. Front. Aging Neurosci. 9:393. doi: 10.3389/fnagi.2017.00393

**Abbreviations:** BioGRID, The Biological General Repository for Interaction Datasets; DAVID, The Database for Annotation, Visualization and Integrated Discovery; DEG, differentially-expressed gene; GFAP, glial fibrillary acidic protein; HSCs, hematopoietic stem cells; PBS, phosphate-buffered saline; qPCR, real-time quantitative polymerase chain reaction.

## INTRODUCTION

fnagi-09-00393 December 4, 2017 Time: 15:1 # 2

Aging is the process of an organism, particularly mammals, becoming older and characterized by a progressive loss of physiological integrity (Lopez-Otin et al., 2013). It is the result of a series of normal changes in gene regulatory networks over time, and is usually reflected by physiological changes (Lopez-Otin et al., 2013). The hippocampus aging, which contributes to the loss of spatial and episodic memory and learning in aging people, is closely related to the detrimental effects of aging at the morphological and molecular levels (Castellano et al., 2017).

The aged hippocampus may impair neural and cognitive function and lead to many neurological disorders such as Alzheimer's disease (Stilling et al., 2014; Castellano et al., 2017), aphasia, agnosia, and Parkinson's disease (Ghika, 2008). The aging hippocampus seems to suffer a generalized loss of synapses (Miller and O'Callaghan, 2003), which influence spatial memory and plasticity functions in multiple species (Moser and Moser, 1998; Hampson et al., 1999; Burke and Barnes, 2006). Moreover, a decrease in neuronal density (Mani et al., 1986; Keuker et al., 2003) and hippocampal volume (Driscoll et al., 2003) were found in the aging hippocampus, and these age-related losses occur in some degenerative conditions (West et al., 1994). Certain genes were predicted, for instance, a deubiquitinating enzyme- UCHL5 hub protein (Kikuchi et al., 2013), and neuropeptide argininevasopressin (4−8) (AVP4−8) (Xiong et al., 2000) may be potential candidates for treating age-related disease. The increasing elderly population has increased labor costs and financial burden for both families and governments, thus, delaying or even partially reversing the aging process is of great interest (Castellano et al., 2017). However, due to the limited knowledge regarding aging, there are still numerous difficulties with different treatments for degenerative diseases linked to aging.

The most common theories for aging include telomere attrition, endocrine disorders, and oxidative stress (Patel et al., 2015). It has been suggested that low-grade inflammation, which is characterized by increased levels of inflammatory cytokines in response to environmental signals, contributes to aging processes (Nikodemova et al., 2016). The chronic inflammation eventually initiates immune-senescence in both the immune and central nervous system, resulting in the functional decline of the immune system with age (Regulski, 2017). Recently, advances in knowledge regarding neuroinflammation and immunity to brain aging have been reviewed (Di Benedetto et al., 2017). Since a plethora of research has been conducted in the exploration of aging mechanisms, recent research has been dedicated to elucidating ways in which to cease aging (Loffredo et al., 2013; Katsimpardi et al., 2014; Sinha et al., 2014; Castellano et al., 2017). GDF11 has been identified as a circulating factor in young mice that declines with age, and subsequently, GDF11 has been shown to be a rejuvenating factor for skeletal muscle (Sinha et al., 2014) and the brain (Katsimpardi et al., 2014). Furthermore, recent research has shown that tissue inhibitor of metalloproteinases 2, an umbilical cord plasma protein, provides a reservoir for the neuroplasticity-promoting process (Di Benedetto et al., 2017). Despite these research efforts, the molecular mechanisms underlying the progression of biological aging have not yet been established, likely due to the complexity of chronic aging progressions.

In the present paper, we aim to explore the molecular interaction mechanisms underlying aging using a biological systems approach that integrates computational and experimental methods. We assume that DEGs between advanced-age and young mice likely contribute to aging progression via certain mechanisms. Investigation into the molecular interactions of the genes identified potential aging regulators (or hub genes of the network). The key candidate genes inferred from network analysis were validated using qPCR. Our present study provides a more comprehensive analysis of the mechanisms underlying the aging process, and has laid the foundation for the exploration of genetic changes involved in the promotion of rejuvenation from advanced-age to young individuals.

### MATERIALS AND METHODS

#### Mice

Specific pathogen free C57BL/6 wild-type mice were fed under a controlled temperature (23◦C) and kept on a 12 h light/dark cycle with food and water provided ad libitum. 3–4 mice were kept in individually ventilated cages. Littermate male mice (4 and 20 months old) were used, and three different tissue samples for RNA extraction were obtained per strain. All animal handling protocols were approved by the Animal Ethics Committee of Shanghai University.

### RNA Extraction and RT-PCR

Hippocampus was rapidly dissected in 0◦C PBS and placed in RNA extraction solution (Promega, United States). RNA extraction was carried out in accordance with the manufacturer's protocol. 2 µg total RNA and 4 µl 5X RT Master Mix (TaKaRa, Japan) were added to RNA-free water to a final reaction volume of 20 µl for cDNA synthesis under different temperature of 25◦C for 5 min, 37◦C for 30 min, 85◦C for 10 s, and 12◦C for 10 min. The target genes were quantified by real-time PCR using qPCR SYBR Green Master Mix (Yeasen, China). Each reaction contained 1 µl cDNA sample (100–200 ng/µl), 10 µl qPCR SYBR Green Master Mix, 0.8 µl (10 µM) designated primers, and RNA-free water to a final volume of 20 µl. The PCR conditions were as follows: 95◦C for 30 s, 55◦C for 20 s, 72◦C for 30 s, 40 thermal cycles. The mRNA expression was normalized to GAPDH. The primers of each gene were shown in **Table 1**.

### Identification of DEGs and Function Annotation

For comparison of gene expression in the hippocampus of 3-, 24-, and 29 month-old mice, the following methods were used. Raw RNA-Seq reads of the gene expression series GSE61915 were downloaded from the NCBI Gene Expression Omnibus database (Stilling et al., 2014). Low quality bases and 5<sup>0</sup> and 3<sup>0</sup> (Q20 ≤ 20) ends of each read were trimmed using a locally developed Perl script, and adapters were removed by trimmomatic (version 0.36)

#### TABLE 1 | Primers used in qPCR.

fnagi-09-00393 December 4, 2017 Time: 15:1 # 3


(Bolger et al., 2014). The clear reads were mapped against mouse genes (GRCm38.p5) using HISAT2 (version 2.0.5) (Kim et al., 2015). The read counts were generated by BEDTools (Quinlan, 2014) and expression values were calculated using the RPKM (Reads per Kilobase per Million mapped reads) method. DEGs were obtained using the fold change of ≥1.50 or ≤0.67, and a t-test p-value of ≤0.05 was also required.

For comparison of gene expression in the hippocampus of mice treated with PBS or human cord plasma protein, human young plasma, and human advanced-age plasma, the processed data from GSE75416 were used directly (Castellano et al., 2017). DEGs were obtained using the fold change of ≥1.2 or ≤0.83, as well as a t-test p-value of ≤0.05.

Enriched functions of significantly changed genes were generated using DAVID 6.8 (Huang da et al., 2009).

### Molecular Interaction Network Construction

Human and mouse molecular interactions were collected from the BioGRID (version 3.4.153) (Chatr-Aryamontri et al., 2017) and IntAct (release 2017/9/2) (Orchard et al., 2014) databases, respectively. The BioGRID and IntAct databases contain various experimentally validated interactions, such as: two-hybrid, affinity capture, co-localization, far-western blotting, proximity label-MS, phage display, coimmunoprecipitation, pull-down, x-ray crystallography, and cross-linking. We were especially interested in the genes that directly interacted with the DEGs of special functions. All the interactions were collected from published literature, and could be retrieved from the original paper via the PubMed IDs. Thus, these data are more reliable than other computationally predicted interactions. All interactions were represented by gene symbol pairs, the BioGRID and IntAct dataset were merged based on these gene symbol pairs. Only human–human or mouse–mouse interactions were retained. For molecular interaction network construction, the first neighbors (most of which represents more reliable and usually direct interactions) of the genes related to the immune response (**Figure 2A**) were selected, and the interactions were displayed using Cytoscape 3.5.1 (Shannon et al., 2003).

The gene functions in hippocampal development were collected from Uniprot (Boutet et al., 2016) according to their GO annotation. The hippocampal development (Supplementary Table S1)-associated network was constructed using the same method. A slight difference was that all genes that interacted with hippocampal development-associated genes were DEGs, while the immune response-related genes were DEGs in the above network.

The human and mouse networks were constructed.

#### Plasmid Construction

PsiRNA-YWHAE was constructed for the intracellular silence of YWHAE expression. The hairpin sequence containing YWHAE (YWHAE-for 5<sup>0</sup> - TCC CAG CTA ACA CTG GCG AGT CCA AGG TTT CAA GAG AAA ACC TTG GAC TCG CCA GTG TTA GCT T -3<sup>0</sup> and YWHAE-rev 5<sup>0</sup> - CAA AAA GCT AAC ACT GGC GAG TCC AAG GTT TTC TCT TGA AAC CTT GGA CTC GCC AGT GTT AGC T -3<sup>0</sup> ) was inserted into the psiRNA-hH1neo vector at two restriction enzyme sites (Bpi I). The negative control (psiRNA-control) with a random sequence (non-targeting siRNA) was also constructed in the same way (**Supplementary Figure S1**).

#### Cell Culture and Transfection

HEK293T, an inbreeding of HEK293 (Human embryonic kidney cells 293) cells, were cultured in DMEM (Invitrogen, United States) containing 10% fetal bovine serum (Invitrogen, United States), at 37◦C with 5% CO2. Cells were plated on 6-well plates at a density of 5 × 10<sup>5</sup> /ml, and the following day transfected with 4 µg psiRNA-YWHAE or psiRNA-control as a negative control and cultured for 36 h. Transfection was carried out using Lipofectamine2000 (Invitrogen, United States), according to the manufacturer's protocol.

#### Immunofluorescence

The human brain sections (provided by Huashan Hospital, Fudan University, Shanghai, China) were washed three times with PBS, treated with 0.2% Triton X-100 (diluted with PBS) for 30 min, and then blocked in 5% albumin from bovine serum for 60 min. Next, the sections were incubated with an YWHAE antibody at room temperature overnight. The following

day, the sections were washed three times with PBS (5 min per wash) and then incubated with a secondary antibody for 1.5 h. Finally, the sections were stained with 4<sup>0</sup> ,6-diamidino-2-phenylindole for 20 min. The expression of YWHAE was detected by confocal microscopy. The experiment protocols were approved by the Human Ethic Committee of Huashan Hospital (Fudan University, Shanghai, China), and Shanghai University Ethics Committee.

### RESULTS

### DEGs in the Hippocampus between Young and Advanced-Age Mice

To determine the difference at the molecular level during the hippocampal aging process, we compared the gene expression in the hippocampus between young (3 months) and advanced-age (24 and 29 months) mice (Stilling et al., 2014). In comparison with 3 month-age mice, 29 month-age mice showed a larger difference than that of 24 month-age mice, as can be seen from the greater discrete degree in the scatter plot (**Figure 1A**). These DEGs were subsequently classified into two categories, which included 510 up-regulated and 132 down-regulated genes (**Figure 1B**). Specifically, 378 and 271 genes were up-regulated by at least 1.5-fold (t-test p-value ≤ 0.05) in the 24- and 29 month-age mice respectively and 139 genes were consistently up-regulated (**Figure 1B**). 44 and 99 genes, respectively, were down-regulated by at least 1.5-fold, and 11 genes were collectively down-regulated (**Figure 1B**). The small overlap between the up and down regulated genes of the 24 and 29 months old mice is due to the variable gene expressions between these stages, indicating the aging is a dynamically changing process. It can be inferred that there are more genes up-regulated than those down-regulated. Moreover, the number of overlapping genes in the up-regulated group (139) is substantially larger than in the down-regulated group (11), illustrating that these common genes may play a general role during the process of hippocampal aging. Furthermore, DAVID (Huang da et al., 2009) functional annotation shows that the majority of the up-regulated genes are associated with the innate and inflammatory immune responses (**Figure 1C**). Interestingly, the immune responserelated functions were also found to have significant changes during aging (High, 2004; Weinberger et al., 2008; Balivada et al., 2017), indicating that immune response-related genes may play important roles during the hippocampal aging process.

### Immune Response-Associated Molecular Interaction Network during Hippocampal Aging

The above analysis demonstrates the importance of the immune response (**Figure 1C**) in hippocampal aging, however, the underlying mechanism of this genes network is still unknown. Thus, we investigated the molecular interactions (see section "Materials and Methods") of these genes in the BioGRID (Chatr-Aryamontri et al., 2017) and IntAct (Orchard et al., 2014) databases. The molecular interactions formed a network (**Figures 2A,B**) that included 97 nodes and 78 edges, with 24 immune response-related genes being up-regulated (indicated in red nodes) and interacted with its partners. Meanwhile, there are 12 single genes without interactions to others. Through interaction with these 24 genes, their partner genes (green nodes) may indirectly participate in the aging-related immune response. Notably, Ywhae is shown to be linked to many immune response-related genes (**Table 2**), forming a hub in the network (**Figure 2B**). Considering its connection to many immune response genes (**Figure 2B**), it can be speculated that Ywhae may not only have an influence in the nervous system (Baxter et al., 2002; Toyo-oka et al., 2003) but also in the immune response during the aging process.

We also investigated the interactions in human. The hub gene Ywhae of the above immune response associated molecular interaction network in mice was not observed in human (**Supplementary Figure S2**). However, this does not mean the immune response associated molecular interaction network is not conserved between human and mouse, since the interaction dataset of both human and mouse were incomplete, the overlapped interactions were limited.

### Membrane Receptors Are Down-Regulated during the Hippocampal Aging Process

In addition to the 139 up-regulated genes (**Figure 1B**), 11 down-regulated genes were also observed, as shown in **Figure 3**. The majority of these genes, including Ret, Sftpc, Gpr17, Tpsb2, and Glp1r are associated with disulfide bonds that are crucial to protein structure (Butera et al., 2014), it has been reported that disulfide bonds may influence the function of proteins in the blood (Butera et al., 2014). In addition to disulfide bonds, cell membrane and receptorassociated genes were also observed, including Ret, Lims2, Gpr17 and Glp1r. Membrane receptors embedded in the plasma membrane bind extracellular molecules and allow communication between the extracellular and intracellular space, meaning that environmental exposure may influence aging through transmembrane receptors (Bettio et al., 2017). As shown in **Figure 3**, Sftpc, for example, encodes the membrane protein surfactant-associated protein C (SP-C) (Glasser et al., 2013). It has been reported that Sftpc plays an important role in innate host defense, enhancing macrophage-mediated phagocytosis and clearing, and limiting inflammatory responses (Glasser et al., 2008), which are all important during aging (**Figure 1C**). Further, GPR17 is a marker for progenitor progression within the oligodendroglia lineage (Boda et al., 2011) and is expressed in all regions of the brain, indicating that downregulation of GPR17 in our network may interfere with normal brain function and lead to aging. In conclusion, these down-regulated genes take part in many different regulatory pathways together with the up-regulated genes. The current results demonstrate that the identified genes have important biological significance in hippocampal aging research.

using the DAVID platform.

### DEGs during Hippocampal Aging Are Closely Related to Hippocampal Development

It has been reported that a change in hippocampal volume and structure is observed with age-related decline in learning and memory (Verbitsky et al., 2004; Arias-Cavieres et al., 2017; Bettio et al., 2017; Moret-Tatay et al., 2017). Additionally, in the aging hippocampus, neurobiological alterations have been clearly observed including those that alter intracellular signaling and gene expression, and cause neuroinflammation response (Bettio et al., 2017). In recent years, studies regarding of the age-related hippocampus have focused on cognitive decline and impaired memory (Arias-Cavieres et al., 2017; Bettio et al., 2017; Moret-Tatay et al., 2017). Several studies revealed the DEGs in hippocampus between young and the aging mice linked to the cognitive deficient (Xiong et al., 2000; Cheng et al., 2007; Thomas et al., 2014), synaptic plasticity (Dempsey and Ali, 2014) and various dysregulated pathways related to immune and inflammatory response (Landel et al., 2016). In order to understand how significantly up- and down-regulated genes are involved in hippocampal function during aging, their interactions were investigated using public databases and established a network using the privous methods (von Mering et al., 2003; Liu and Zhao, 2004). Interestingly, most of these genes interact with Ywhae (**Figure 4**), which is the hub gene in the immune response network during hippocampal aging (**Figure 2B**). Nucleotide polymorphisms in Ywhae would lead to abnormal hippocampal development (Kido et al., 2014), implying that Ywhae maybe influence the function of hippocampus during aging through certain signaling pathway. As shown in **Figure 4**, YWHAE interacts with VIM, which is an integral regulator of cell adhesion (Dave and Bayless, 2014), neurite extension (Dave and Bayless, 2014), and myelination (Triolo et al., 2012), suggesting that Ywhae may be involved in the regulation of neural function through Vim during aging. Furthermore, GFAP, which is expressed in astrocytes of the central nervous system (Eng and Ghirnikar, 1994), is the partner of YWHAE (**Figure 4**) and is important for the brain to accommodate

FIGURE 2 | The molecular interaction network of up-regulated genes that are associated with the immune response in mice. (A) Heatmap of up-regulated immune response-related genes in the 3 age groups of mice (3M, n = 8; 24M, n = 6; 29M, n = 3). Red represents genes with high expression and blue represents those with low expression. (B) The molecular interaction network related to the immune response (red nodes) and their partner genes (green nodes; 61). A gray line represents an interaction between two nodes.

neural activities or changes during development (Kim et al., 2011). Moreover, GFAP also plays a crucial role in neuroinflammation of central nervous system injury (Otani et al., 2003), which is consistent with the speculation that Ywahe is involved in the immune response (**Figure 2B**) together with GFAP. In conclusion, the interaction between YWHAE and its partners is obviously important for hippocampal development in the nervous system, implying an important function of these genes in the aging process. Interestingly, YWHAE interacts with C4B, which is up-regulated in aging mice and may influence the autoimmune process in diseases (Paul et al., 2002) or modulate interstitial inflammation (Welch et al., 2001), participating in the immune response. Moreover, C4B is also associated with ITGB2, which regulates the lineage distribution of hematopoietic cells in the blood and bone marrow (Gomez and Doerschuk, 2010), implying that Ywhae may be involved in different signaling pathways contributing to the aging process. Essentially, aging is a slow and complicated process that is accompanied by numerous changes in gene expression, especially those related to hippocampal development. This also partially explains why these genes influence hippocampal development and impair memory and learning ability, accelerating aging.

TABLE 2 | The up-regulated immune response-related genes that interacted with Ywhae.


### Verification of Gene Expression in the Interaction Network by QPCR

In order to further verify the expression of the genes in our current research (**Figure 4**), a number of genes in the network were randomly chosen and their expression was evaluated by qPCR in advanced-age and young mice. Hippocampal tissues were extracted from 4- and 20 month-old male mice and the

FIGURE 3 | The analysis of down-regulated genes in the hippocampus of advanced-age mice. (A) Heatmap of the 11 down-regulated genes in mice of different ages. The red and blue color spectrum represents high and low expressed genes, respectively (3M, n = 8; 24M, n = 6; 29M, n = 3). (B) Function enrichment analysis of 11 down-regulated genes. The DAVID platform was used for analysis; the 11 down-regulated genes have 10 functions exhibited in the column diagram. The majority of the genes were enriched in disulfide bonds.

mRNA amplified and compared with that of 4 month-old mice as a control. As shown in **Figure 5**, the majority of hippocampal development genes (Dclk2, Rara, Hdac1, and Ezh2) were upregulated, with the exception of Ywhae. Moreover, Tgm1, Cd74, Rbm3, and C4b were also up-regulated compared with the young mice group, which is consistent with the network shown in **Figure 4**.

### Interactions between YWHAE and Its Partners Are Conserved between Humans and Mice

In order to further explore the expression pattern of YWHAE (**Figure 4**) in humans, the expression of YWHAE was evaluated in human brain (**Figure 6A**). It was found that compared with the young human brain (15- and 25-years-old), YWHAE is down-regulated in the older human brain (69-year-old). These results are consistent with those seen in mice (**Figures 5A,B**). Furthermore, to evaluate whether the interaction network of YWHAE in mice is conserved in humans, qPCR was carried out to detect the mRNA expression in HEK293T cells after silencing YWHAE. The psiRNA-YWHAE vector was constructed to silence it in HEK293T cells. As shown in **Figures 6B,C**, with the obvious down-regulated expression of YWHAE, the mRNA levels of DCLK2, RARA, HDAC1, CD74, and RBM3 were significantly increased in HEK293T cells, which is consistent with the results observed in mice (**Figure 5A**). However, due to the diversity of gene expression in different species, not all genes exhibit the same expression pattern (**Figure 5A**). For instance, there is no significant difference in the expression of EZH2, TGM1, or C4B at the mRNA level. In conclusion, the gene expression pattern in humans is consistent with that in mice, indicating that the interaction network of YWHAE is conserved between humans and mice.

### The Genes That Contribute to the Revitalization of Advanced-Age Mice Are Different from the Genes Which Are Induced by Aging

It has been reported that human umbilical cord plasma proteins can revitalize hippocampal function in advancedage mice (Castellano et al., 2017), and that tissue inhibitor of metalloproteinases 2, which is an umbilical cord plasma

protein, provides a reservoir for the neuroplasticity-promoting process. In order to gain more information regarding this process, microarray data from mice treated with human cord plasma (Castellano et al., 2017) were re-analyzed (**Figure 7**). In comparison with the PBS control, a total of 148 up-regulated and 190 down-regulated genes (up- or down-regulated by least 1.2-fold, p-value ≤ 0.05) were observed in the samples treated with human cord plasma (**Figure 7A**). Function enrichment analysis indicates that these genes exhibit functions including sensory perception, skeletal muscle cell differentiation, G proteincoupled receptor signaling, and detection of chemical, amino acid, hormone stimuli (**Figures 7B,C**). The hypothesis exists that changes in textural perception (Conroy et al., 2017) and loss of skeletal muscle stem cells with age may drive the degeneration of age-related neuromuscular junctions (Liu et al., 2017), indicating that genes with these functions (**Figures 7B,C**) may participate in aging progression or the revitalization progress. Interestingly, certain inflammatory- and immune response-related genes were also observed among those genes that were down-regulated (**Figure 7D**). However, no overlapping genes (**Figure 7E**) were found between those up-regulated during hippocampal aging (**Figure 2A**) and those down-regulated following treatment with human cord plasma (**Figure 7D**), indicating that mice adopt a different process to become younger after human umbilical cord plasma treatment. Our current data suggest that advancedage mice do indeed benefit from these DEGs; however, it is noteworthy that this appears not to be a reverse process from advanced-age to young mice following treatment with umbilical cord plasma. In other words, the genes which contribute to the revitalization of advanced-age mice are different from the genes which are induced by aging during the normal hippocampal aging process.

#### DISCUSSION

In the present study, we identified DEGs between young and advanced-age mice, and found that the majority of the up-regulated genes in DEGs were enriched in the immune response process (**Figure 1**). Moreover, membrane receptorassociated genes were down-regulated (**Figure 3**). The analysis of the molecular interaction network indicated that Ywhae plays important roles during hippocampal aging (**Figure 2B**). Subsequently, we found that the interactions between YWHAE and its partners are conserved between humans and mice (**Figures 5**, **6**). Furthermore, our results also showed that the genes that contribute to the revitalization of advanced-age mice are different from the genes induced by aging (**Figure 7**). Our data assigned novel roles for genes during hippocampal aging progression and provided further theoretical evidence for future studies exploring the underlying mechanisms of hippocampal aging and anti-aging-related disease therapy.

The progression of aging is complex. In order to explore the mystery of aging, many researchers have used microarrayrelated approaches to perform aging-related research (Bishop et al., 2010; Stilling et al., 2014). In addition, transcriptional profiles were focused on (Pawlowski et al., 2009; Bishop et al., 2010; Stilling et al., 2014) with different research purposes. In our current research, this common approach was also used to analyze DEGs associated with hippocampal aging progression. Some researchers have revealed that DEGs are closely linked to synaptic plasticity (Dempsey and Ali, 2014), cognitive impairment (Cheng et al., 2007; Park et al., 2015), age-associated spatial learning impairment (ASLI) (Uddin and Singh, 2013), degenerative diseases (Xiong et al., 2000; Kikuchi et al., 2013) and even various dysregulated pathways (Landel et al., 2016). In addition, it has been reported that the immune system is dysregulated in the aging brain (Patterson, 2015), which was consistent with our results that the majority of up-regulated genes were enriched in the immune response-related process (**Figure 1C**).

In order to build a reliable network, we collected interaction data from the BioGRID (Chatr-Aryamontri et al., 2017) and IntAct (Orchard et al., 2014) databases. These databases contain various experimentally validated interactions curated from published literature. Moreover, the networks we built were

mean ± SEM, n ≥ 3. <sup>∗</sup>p ≤ 0.05; ∗∗p ≤ 0.01; ∗∗∗p ≤ 0.001.

first-neighbor networks, since the first neighbors represent the most reliable interactions. The second-neighbor network may contain many genes and interactions that are directly linked to the first neighbors but not the genes that we are interested in (immune response or hippocampal development), thus, the center of the network may be shifted. However, we also constructed a second-neighbor network for mice based on the immune response-related genes (**Supplementary Figure S3**). We found that the network is huge, containing 5,283 genes and 19,641 interactions. Link degree analysis of each node in

the network indicated that Ywhae, Eed, Ywhaz, Ubc, Hsd3b4, Pgls, Hsd3b7, Ywhab, Hmga2, and Dlg4 were the top 10 genes with the most interactions with other genes. YWHAE still had the most links with other genes, which is the similar phenomenon as seen in **Figure 2B**. Most interestingly, YWHAE, YWHAZ, and YWHAB belong to the 14-3-3 protein family, which mediate signal transduction by binding to phosphoserinecontaining proteins (Conklin et al., 1995). These results implied that the 14-3-3 protein family has a potential role during the hippocampal aging process. We also checked the conversation of the hippocampus development associated network in human and mice (**Supplementary Figure S4** and **Figure 4**). In mice, Ywhae was linked to many up-regulated DEGs (**Figure 4**, green nodes), indicating Ywhae may be involved in the regulation of those genes (green nodes) during hippocampus development. However, the same situation was not observed in human (**Supplementary Figure S4**). Considering the incompletion of the interaction data deposited in both BioGRID and IntAct, this may be due to the largely un-delineated overlapping of the two datasets.

cord plasma.

Consistent with the previous studies (Stilling et al., 2014), the constructed network implied that the up-regulated DEGs between young and advanced-age mice (**Figure 2A**) interact with their partners to participate in the immune response and facilitate hippocampal aging progression (**Figure 2B**). As people become older, their immune systems become weaker and antibody responses are slow (Montecino-Rodriguez et al., 2013), which induces many dysfunctions (Weigle, 1989). This could explain why older people are more likely to get sick. However, due to the limitations of the database, there were 12 genes (Supplementary Table S2) that did not appear in this network, meaning that these genes showed no functional associations with the other genes. Nevertheless, this does not mean that these genes are not related to hippocampal aging. For instance, CXCL5, a CXC-type chemokine, has been reported to be secreted by aging prostate stroma, implying its important function in aging progression (Begley et al., 2008). As shown in **Figure 3**, five down-regulated genes are associated with disulfide bonds, and oxidation of protein sulfhydryl groups to disulfide groups occurs as a normal part of human aging (Takemoto, 1996), indicating that disulfide bonds play an important unclarified role in aging, and further research is needed to explore the underlying mechanisms. In addition, cell membrane and receptor were also enriched in **Figure 3B**. For instance, Ret is a crucial gene that plays an important role in the development and function of the nervous system (Gabreski et al., 2016), as well as driving HSCs survival. Ablation of Ret cannot only impair the number of HSCs but also influences its normal differentiation potential (Fonseca-Pereira et al., 2014). The downregulation of Ret (**Figure 3A**) implies that HSC function is impaired during aging. Most importantly, the enrichment of many down-regulated membrane proteins may also indicate weaker perception ability to environmental exposure of older people, since many membrane proteins participate in different signaling pathways that transduce extracellular information across the plasma membrane. In conclusion, this indicates that these genes, although enriched in different functions, may work together to promote hippocampal aging through regulating cell communication.

Our results also showed that no overlapping genes (**Figure 7E**) were found between those up-regulated during hippocampal aging (**Figure 2A**) and those down-regulated following treatment with human cord plasma (**Figure 7D**), indicating that the revitalization induced by the treatment of human umbilical cord blood plasma protein is not a simple reverse process of natural aging progression. We noticed that the mice strains from the other research studies are different. Immuno-deficient (NSG) mice were chosen for the injection experiment, since the mice can receive intravenous injection of human plasma without an adverse immune response (Castellano et al., 2017), while, C57BL/6J mice were used for detecting gene expression differences in the other research. This likely provides an explanation for the lack of overlap in gene expression. Another possibility for our results could be that the mice treated with PBS and human cord plasma was 14-months-old rather than 24- or 29-months-old. The distance in age of the mice may have led to differences in gene expression.

### CONCLUSION

The present study provides a more comprehensive analysis of aging and anti-aging treatment using a biological systems approach. Our data suggested that a large number of immune and inflammatory response-related genes were up-regulated in the aged hippocampus, and that membrane receptorassociated genes were down-regulated. Moreover, Ywhae in the predicted network may be assigned novel roles during hippocampal aging progression. The results also implied that the revitalization of advanced-age mice may be not a simple reverse process of normal aging progression. Considering the association of aging-related diseases such as Alzheimer's disease, hypertension, and cancer, our study provides further theoretic evidence for future studies exploring the underlying mechanisms of aging and anti-aging-related disease therapy.

#### AUTHOR CONTRIBUTIONS

JW and TW contributed to the design, analysis, and interpretation of data for the study. JW and QL drafted the work, interpreted the data and conducted the experiments. QL, FZ, JL, WL, and KW conducted the experiments. YK, TW, and YG collected data. JX and TQW finally approved the version to be published.

### FUNDING

This work was sponsored by the National Natural Science Foundation of China (Grant Number: 31500827), Young Eastern Scholar (Grant Number: QD2015033), the Natural Science Foundation of Shanghai (Grant Number: 14ZR1414400, 17ZR1409900), the National Natural Science Foundation of China (81571345; 81471162; 31601163), the Science and Technology Commission of Shanghai (Grant Number: 14JC1402400), and China Postdoctoral Science Foundation (Grant Number: 2016M601661).

## ACKNOWLEDGMENT

The authors would like to thank Dr. Natalie Ward (Medical College of Wisconsin, Wauwatosa, WI, United States) for editing this manuscript.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi.2017. 00393/full#supplementary-material

FIGURE S1 | The interfere fragment of YWHAE was successfully inserted into the vector of psiRNA-hH1neo. (A) The diagram of plasmids construction. The sequence of YWHAE was marked in yellow. (B) The plasmids of psiRNA-hH1neo and psiRNA-YWHAE were identified by 0.6% agarose gels. The arrow shows the recombinant plasmid. (C) The product of plasmids that were cut by single enzyme (Xba I) was identified by 0.6% agarose gels. The arrows shows the sequence was successfully inserted into psiRNA-hH1neo.

FIGURE S2 | The molecular interaction network related to the immune response in human. Red nodes represent the immune responds genes. A gray line represents an interaction between two nodes.

### REFERENCES

fnagi-09-00393 December 4, 2017 Time: 15:1 # 12


FIGURE S3 | The second neighbor molecular interaction network related to the immune response in mice.

FIGURE S4 | The human molecular interaction network involved in hippocampus development. Hippocampus development associate genes (37) are marked with red nodes. The 13 green nodes refer to the homologous DEGs in the mice.



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Wang, Li, Kong, Zhou, Li, Li, Wang, Wu, Guan, Xie and Wen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Resting State Default Mode Network Connectivity, Dual Task Performance, Gait Speed, and Postural Sway in Older Adults with Mild Cognitive Impairment

#### Rachel A. Crockett, Chun Liang Hsu, John R. Best and Teresa Liu-Ambrose\*

Aging, Mobility, and Cognitive Neuroscience Laboratory, Department of Physical Therapy, Djavad Mowafaghian Centre for Brain Health, University of British Columbia, Vancouver, BC, Canada

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Gennady Knyazev, Institute of Physiology and Basic Medicine, Russia Erika J. Wolf, National Center for PTSD at VA Boston Healthcare System; Boston University School of Medicine, United States Daqing Guo, University of Electronic Science and Technology of China, China Laura Marzetti, Università degli Studi "G. d'Annunzio" Chieti - Pescara, Italy

> \*Correspondence: Teresa Liu-Ambrose teresa.ambrose@ubc.ca

Received: 13 August 2017 Accepted: 08 December 2017

#### Published: 21 December 2017 Citation:

Crockett RA, Hsu CL, Best JR and Liu-Ambrose T (2017) Resting State Default Mode Network Connectivity, Dual Task Performance, Gait Speed, and Postural Sway in Older Adults with Mild Cognitive Impairment. Front. Aging Neurosci. 9:423. doi: 10.3389/fnagi.2017.00423 Aging is associated with an increased risk of falling. In particular, older adults with mild cognitive impairment (MCI) are more vulnerable to falling compared with their healthy counterparts. Major contributors to this increased falls risk include a decline in dual task performance, gait speed, and postural sway. Recent evidence highlights the potential influence of the default mode network (DMN), the frontoparietal network (FPN), and the supplementary motor area (SMA) on dual task performance, gait speed, and postural sway. The DMN is active during rest and deactivates during task-oriented processes, to maintain attention and stay on task. The FPN and SMA are involved in top-down attentional control, motor planning, and motor execution. The DMN shows less deactivation during task in older adults with MCI. This lack of deactivation is theorized to increase competition for resources between the DMN and task-related brain regions (e.g., the FPN and SMA), increasing distraction from the task and reducing task performance. However, no study has yet investigated the relationship between the between-network connectivity of the DMN with these regions and dual task walking, gait speed or postural sway. We hypothesized that greater functional connectivity both within the DMN and between DMN–FPN and DMN–SMA, will be associated with poorer performance during dual task walking, slower gait speed, and greater postural sway in older adults with MCI. Forty older adults with MCI were measured on a dual taskwalking paradigm, gait speed over a 4-m walk, and postural sway using a sway-meter. Greater within-DMN connectivity was significantly correlated with poorer dual task performance. Furthermore, greater inter-network connectivity between the DMN and SMA was significantly correlated with slower gait speed and greater postural sway on the eyes open floor sway task. Thus, greater resting state DMN functional connectivity may be an underlying neural mechanism for reduced dual task ability, slower gait speed, and greater postural sway, resulting in the increased risk of mobility disability and falling in older adults with MCI.

Keywords: functional connectivity, default mode network, dual task, gait speed, postural sway, mild cognitive impairment

## INTRODUCTION

fnagi-09-00423 December 19, 2017 Time: 16:18 # 2

Walking has long been considered an automated skill required for daily functioning and independent living. In reality, walking requires the complex neural coordination of visual, proprioceptive, and vestibular sensory incoming information (Beurskens and Bock, 2012). Furthermore, walking in everyday life is not performed as a single task since environmental demands require the ability to perform additional cognitive tasks at the same time (i.e., dual task), such as walking while talking (Yuan et al., 2015). Thus, walking depends on higherorder cognitive processes, known as executive functions (Yogev et al., 2008). Research consistently demonstrates that executive functions are important for successful dual task performance (Coppin et al., 2006).

The risk for developing impaired mobility (e.g., slow walking) increases with age (Verghese et al., 2009; Seidler et al., 2010). Falls are a significant consequence of impaired mobility (Fuller, 2000) with the majority of falls occurring during dual task conditions, such as walking while performing a secondary task (Tideiksaar, 1996). Thus, it is hypothesized that falls may not be a result of balance deficits in isolation, but the inability to effectively allocate attention to postural stability in dual task situations (Lajoie et al., 1996). This hypothesis is supported by observations such as older adults who stop walking while engaged in conversation (i.e., less able to dual task) are more likely to fall than those who continue walking (Lundin-Olsson et al., 1997). Recent evidence also highlights the importance of maintaining dual task ability in older adults beyond falls risk; Montero-Odasso et al. (2017) demonstrated that reduced dual task gait performance among older adults with mild cognitive impairment (MCI) was associated with progression to dementia.

MCI is considered the prodromal stage for dementia and is characterized by global brain atrophy and cognitive decline beyond normal aging but that does not impact daily living (Feldman and Jacova, 2005). Older adults with MCI are also found to have a greater decline in physical functioning (Aggarwal et al., 2006) and poorer performance under dual task conditions (Nascimbeni et al., 2015). Consequently, those with MCI are five times as likely to fall compared to cognitively intact older adults (Tinetti et al., 1988). Despite this, knowledge regarding the underlying neural correlates associated with dual task and gait performance in MCI is lacking. A better understanding of the neural basis for impaired dual task and mobility in this population can inform future strategies to prevent the progression of MCI and subsequently reduce the risk of mobility disability and falling.

Current evidence suggests that both the cognitive and motor impairments associated with MCI have a common neurobiological basis (Silbert et al., 2008; Callisaya et al., 2013). Of relevance, aging is characterized by disruptions in the functional connectivity of neural networks that support both cognitive and motor functions. Recent evidence highlights the potential involvement of the default mode network (DMN) and the frontoparietal network (FPN). The DMN is involved in autobiographical memory, memory consolidation, and selfreferential thought (Andrews-Hanna et al., 2007; Buckner et al., 2008). It is active during rest and deactivates during task-oriented processes, to maintain attention and stay on task (Raichle et al., 2001). However, the DMN shows less deactivation on task for older adults and those with MCI (Lustig et al., 2003; Mevel et al., 2011). Although findings are inconsistent, it is thought this lack of deactivation may be due to greater resting state DMN functional connectivity (Mevel et al., 2011). It is suggested that as the posterior components of the network are among the first brain regions to be affected by both age and MCI-related atrophy (Buckner et al., 2005; Choo et al., 2010), compensation occurs as connections between the DMN and frontal regions increase (Davis et al., 2008), leading to a net increase in resting state functional connectivity and a lack of deactivation on task. This lack of deactivation while on task is theorized to increase distraction and reduce cognitive performance (Gardini et al., 2014). With age, walking becomes more dependent on executive functioning (Coppin et al., 2006). Therefore a decline in cognitive performance is subsequently hypothesized to also have a detrimental impact on walking in older adults and those with MCI.

The FPN is involved in top-down attentional control and allocation of available neural resources to important cognitive processes (Corbetta, 1998; Scolari et al., 2015), as well as in motor planning and motor execution (Ptak et al., 2017). It has also been associated with dual task walking performance. Compared to normal walking, performance on a walking while talking task was associated with greater functional connectivity between the prefrontal and supplementary motor areas (SMA) of the FPN and sensorimotor network, respectively (Yuan et al., 2015).

Much like the FPN, the SMA is a key structure for the execution and control of voluntary movement, motor planning (Roland et al., 1980; Eccles, 1982), and for maintaining attention on a motor task (Johansen-Berg and Matthews, 2000). Unlike some of the deeper brain structures, functional activation of the SMA can be assessed while walking using imaging techniques such as near infrared spectroscopy. Research using this technique, has found that an increase in activation of the SMA was associated with declines in gait performance under dual task conditions (Lu et al., 2015). Thus, the SMA may play an important role in the maintenance of gait performance during dual task walking.

Research focusing on the relationship between the DMN and the task-related networks responsible for motor functioning is lacking. However, one study did find that greater functional connectivity between the DMN and FPN was associated with reduced performance on a finger-tapping task in older fallers, indicative of an inability to focus attention on the task (Hsu et al., 2014). It is theorized that in older adults, this greater connectivity between the DMN and taskrelated networks at rest may suggest the DMN is remaining active on task, competing for resources with the task-related networks, and subsequently causing a decline in performance. This competition for resources is likely exacerbated under dual task conditions (Tombu and Jolicœur, 2003) and is hypothesized to underlie the relationship between cognitive and motor functioning decline, leading to an increased risk

of falls and mobility disability in older adults and those with MCI. However, to our knowledge, no research has yet investigated the effect of the functional between-network connectivity of the DMN in relation to dual task walking paradigms.

In addition to reduced dual task performance, slower gait speed and increased postural sway are considered major factors contributing to an increased risk of falls. Older adults with MCI were found to have significantly slower gait speed (Eggermont et al., 2010) and greater postural sway (Liu-Ambrose et al., 2008) compared to healthy older adults. Both the SMA and components of the FPN have been associated with having a functional role in controlling gait speed (Harada et al., 2009; Yuan et al., 2015) and maintaining postural stability (Mihara et al., 2008). Therefore, maladaptive connectivity between these regions and the DMN may also negatively impact gait speed and postural stability, further contributing to an increased risk of falling. No research has yet investigated this potential relationship.

Consequently, the primary objective of this study was to investigate the relationship between the functional connectivity within the DMN and the between-network connectivity of the DMN with both the FPN and SMA (i.e., DMN–FPN and DMN– SMA) during a dual task walking paradigm in older adults with MCI. In addition, the secondary objective was to explore the association of within-network and between-network connectivity with gait speed and postural sway.

We hypothesized that greater functional connectivity within the DMN, between DMN–FPN, as well as between the DMN–SMA will be associated with poorer dual task walking performance, slower gait speed and greater postural sway. The rationale and hypotheses for this study are highlighted in **Figure 1**.

### MATERIALS AND METHODS

### Participants

Forty community dwelling older adults with MCI are included in this cross-sectional study. MCI was defined as: (1) a Montreal Cognitive Assessment (MoCA) score <26/30; (2) have subjective memory complaints (SMC); (3) no significant functional impairment, as determined by a score >6/8 on the Lawton and Brody Instrumental Activities of Daily Living Scale; and (4) no dementia.

Participants were recruited from metropolitan Vancouver and interested individuals were telephone screened to confirm general eligibility according to the inclusion and exclusion criteria. We included those who: (1) were aged ≥60 years; (2) scored ≤26/30 on the MoCA (Nasreddine et al., 2005); (3) had SMC, defined as the self-reported feeling of memory worsening with an onset within the last 5 years, as determined by interview (Gifford et al., 2014); (4) preserved general cognition as indicated by a Mini-Mental State Examination (Folstein et al., 1975) score ≥24/30; (5) score ≥6/8 on the Lawton and Brody (Lawton and Brody, 1969) Instrumental Activities of Daily Living Scale; (6) were right hand dominant as measured by the Edinburgh Handedness Inventory (Oldfield, 1971); (7) were living independently in their own homes; (8) had visual acuity of at least 20/40, with or without corrective lenses; and (9) provided informed consent. We excluded those who: (1) had a formal diagnosis

of neurodegenerative disease, stroke, dementia (of any type), or psychiatric condition; (2) had clinically significant peripheral neuropathy or severe musculoskeletal or joint disease; (3) were taking psychotropic medication; (4) had a history indicative of carotid sinus sensitivity; (5) were living in a nursing home, extended care facility, or assisted-care facility; or (6) were ineligible for magnetic resonance imaging (MRI) scanning. All participants provided written consent and ethics approval was acquired from the Vancouver Coastal Research Health Institute and University of British Columbia's Clinical Research Ethics Board.

### Descriptive Variables

fnagi-09-00423 December 19, 2017 Time: 16:18 # 4

Age was quantified in years and education level was assessed by self-report. Standing height was measured as stretch stature to the 0.1 cm per standard protocol. Weight was measured twice to the 0.1 kg on a calibrated digital scale. As previously stated, the MoCA was used as a classification tool for MCI. The MoCA is a 30-point test that covers multiple cognitive domains (Nasreddine et al., 2005). The MoCA has been found to have good internal consistency and test–retest reliability and was able to correctly identify 90% of a large sample of MCI individuals from two different clinics (Nasreddine et al., 2005).

### Dual Task

Participants were also asked to perform a dual task involving a cognitive task while walking. Better performance on dual task measures is associated with better task switching, working memory, and divided attention (Schaefer and Schumacher, 2010). The cognitive task was the serial subtracting sevens task whereby the participant is required to start subtracting sevens aloud from a randomly given number. The walking task was performed on a GAITRite mat (McDonough et al., 2001), which was used to record the time between the first step onto the mat and the last step off of the mat for each task. Participants were instructed to begin walking 1 m before they reached the mat and finish once they made it to 1 m past the end of the mat to control for acceleration and deceleration effects. The participants were first asked to walk three times at a self-selected pace across the mat to give a mean walking time. They then performed the subtracting sevens task standing still (off of the mat) for 30 s. For the dual task component they were asked to walk at a self-selected pace and begin the serial subtraction once on the mat. This was repeated three times to get a mean dual task time. Dual task cost was then calculated by subtracting the mean walking only time from the dual task time, divided by the walking only time [(dual task − walking time)/walking time]. A lower dual task cost score indicated better dual task performance.

### Usual Gait Speed

Participants walked at their usual pace along a 4-m path. To avoid acceleration and deceleration effects, participants started walking 1 m before reaching the 4-m path and completed their walk 1 m beyond it.

Usual gait speed (m/s) was calculated from the mean of two trials. The test–retest reliability of usual gait speed in our laboratory is 0.95 (ICC; Liu-Ambrose et al., 2006).

### Postural Sway

Balance was assessed using the eyes open floor and foam sway component of the Physiological Profile Assessment© 18 (Prince of Wales Medical Research Institute, Randwick, Sydney, NSW, Australia; Lord et al., 2003). Participants were asked to stand with their feet hip width apart and look straight ahead for 30 s, first on the hard floor and secondly on a 3-inch highdensity foam cushion. The task was stopped if the participant grabbed for support. A pen attached to a rod and connected to a band around the participants' waist was set to lie parallel to the ground and rest on a large paper grid (sway-meter). Sway was calculated as the largest distance covered across the grid.

### Functional Magnetic Resonance Imaging Acquisition

The MRI scans were conducted at the University of British Columbia (UBC) Hospital in Vancouver at the UBC MRI research center. A 3.0-Tesla Intera Achieva MRI Scanner with an 8-channel SENSE neurovascular coil was used. During the scanning procedure, the participants were told to rest with their eyes open, remaining completely still and thinking of nothing in particular for the duration of the session.

The scanning session consisted of an initial resting-state scan with 360 dynamic images of 36 slices (3 mm thick) with a repetition time (TR) of 2000 ms, an echo time (TE) of 30 ms, a flip angle (FA) of 90◦ , a field of view (FoV) of 240 mm and an acquisition matrix of 80 × 80. High-resolution anatomical T1 images were also acquired with 170 slices (1 mm thick), TR of 7.7 ms, TE of 3.6 ms, FA of 8◦ , FoV of 256 mm, and an acquisition matrix of 256 × 200.

#### fMRI Pre-processing

FSL (FMRIB's Software Library), MATLAB (Matrix Laboratory), and toolboxes from Statistical Parametric Mapping (SPM) were used to carry out image processing. The Optimized Brain Extraction Tool (optiBET) (Lutkenhoff et al., 2014) was used to remove any excess unwanted structures in high resolution T1 images, e.g., skull, eyes, etc. The rigid body motion correction was done using MCFLIRT with the absolute and relative mean displacement extracted and included as covariates in the statistical analysis. The Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC) further removed any additional artifacts. FSL Motion Outliers was used to determine any data points that were corrupted with a large amount of motion. A confound matrix was used to remove the effects of these time points on any subsequent analyses. The Gaussian kernel of full-width-half-maximum (FWHM) 6 mm was used for spatial smoothing and temporal filtering was applied to create a signal between 0.008 < f < 0.08 Hz, which is the optimal range for analyzing resting-state functional connectivity data.

The functional MRI data was registered to the corresponding high resolution T1 anatomical image for each participant, which was then registered to standardized 152 T1 Montreal

Neurological Institute (MNI) space. Regression of the cerebral spinal fluid, white matter and global brain signal was used to remove noise from any physiological or non-physiological sources. Finally, to account for the delay of hemodynamic response, the first four volumes of data were discarded.

#### Functional Connectivity Analysis

The main focus of the functional connectivity analysis was to investigate the connectivity firstly within the DMN and secondly the inter-network connectivity between the DMN and the FPN and SMA independently. Previous studies guided the region of interest (ROI) selection for the analysis of the DMN, FPN, and SMA (Voss et al., 2010; Hsu et al., 2014). The ROIs within each network and their respective MNI space coordinates can be seen in **Table 1**. In order to analyze the overall interconnectivity between brain networks, the average of all the pairwise ROI–ROI correlations for each network was calculated. Preprocessed time-series data were extracted for each ROI with 14 mm diameter spherical ROIs drawn around their respective MNI coordinates in standard space. The time-series data for each ROI were then crosscorrelated with every brain voxel to create functional connectivity maps of each neural network. Ordinary least-squares regression using FSL's flameo (Beckman et al., 2003) was used to calculate group-level between subject results. The statistical map thresholding was set at Z = 2.33, with a cluster correction of p < 0.05.

### Data Analysis

Statistical analysis was conducted using the IBM SPSS Statistic 19 for Windows (SPSS Inc., Chicago, IL, United States). Descriptive data are reported for variables of interest. Alpha was set at p ≤ 0.05 for all analyses.

Partial correlations, adjusted for age and MoCA, were performed to investigate the association between: (1) dual task performance and DMN functional connectivity; and (2) measures of gait speed and postural sway and DMN functional connectivity

Additional partial correlational analyses were also conducted to determine whether functional between-network connectivity between the DMN–FPN and DMN–SMA were correlated with dual task performance, gait speed, and postural sway.

### RESULTS

#### Participants

A total of 40 participants were included in this study (see **Table 2**). However, five participants did not complete the dual task measure leaving 35 participants for analysis.

Based on visual inspection of the data there appeared to be a few outliers in the dual task cost results. Upon further analysis of all of the behavioral measures, two extreme outliers (mean ± 3 SD) were found in the dual task cost results. However, the removal of these outliers did not affect the results and as such, all data were included in the final analyses.

### Partial Correlations: Dual Task Performance, Gait Speed, Postural Sway, and DMN Functional Connectivity

Partial correlation analyses showed a significant association between mean DMN functional connectivity and dual task cost (r = 0.427; p = 0.013), such that greater DMN connectivity was associated with poorer dual task performance. No other significant correlations were found (p ≥ 0.05). The resting state connectivity map for the DMN can be seen in **Figure 2**.


DMN, default mode network; PCC, posterior cingulate cortex; FMC, frontal medial cortex; BMTG, bilateral middle temporal gyrus; BPHG, bilateral parahippocampal gyrus; LMFG, left middle frontal gyrus; BLOC, bilateral occipital cortex; FPN, frontoparietal network; RIPS, right inferior parietal sulcus; BVV, bilateral ventral visual; RSMG, right supramarginal gyrus; BLSOC, bilateral superior occipital cortex; BFEF, bilateral frontal eye field; SMN, sensorimotor network; SMA, supplementary motor area.

TABLE 2 | Participant characteristics.


MMSE, Mini-Mental Status Examination; MoCA, Montreal Cognitive Assessment.

#### Partial Correlations: Dual Task Performance, Gait Speed, Postural Sway, and Functional Between-network Connectivity between the DMN–FPN and DMN–SMA

Mean resting state between-network DMN–SMA functional connectivity was significantly correlated with gait speed (r = −0.440; p = 0.01) as well as postural sway under the eyes open floor condition (r = 0.365; p = 0.037). Thus, greater between-network DMN–SMA functional connectivity was associated with slower gait speed and greater postural sway under the eyes open floor condition. Results of the partial correlation analyses are reported in **Table 3**. The resting state between network connectivity map for the DMN and SMA can be seen in **Figure 3**.

#### DISCUSSION

We found that greater resting state DMN functional connectivity was significantly associated with greater dual task cost among community-dwelling older adults with MCI. It was previously discussed that greater DMN connectivity at rest is indicative of a lack of deactivation on task in older adults with MCI (Lustig et al., 2003; Mevel et al., 2011), which would subsequently result in a decrease in task performance. The association between greater


Adjusted for age and MoCA; <sup>∗</sup>p < 0.05. EO, eyes open.

DMN connectivity and reduced dual task performance found in this study is therefore likely evidence for a lack of deactivation of the DMN on task.

One possible explanation for greater dual task cost being associated with greater DMN connectivity is provided by the capacity sharing theory. This theory states that because attentional resources are limited, when a person is required to perform two attention-demanding tasks simultaneously, the performance of at least one of these tasks will deteriorate (Tombu and Jolicœur, 2003). Therefore, if the greater DMN connectivity at rest were evidence of a lack of deactivation on task, this would suggest that the DMN is providing additional competition for resources, reducing those available for task-related functioning and subsequently reducing performance on the task. In addition to competing for resources generally, the DMN would be using these resources to fuel a functional process responsible for increasing distraction from the task, further amplifying its negative impact on task performance.

Walking in everyday life is not performed as a single task. Hence, the dual task walking measure has been associated with replicating the demands of the environment whilst walking (Tideiksaar, 1996). Findings from the current study indicate that patients with MCI are unable to meet the cognitive demands required to maintain walking performance in everyday life due to a lack of resources available and reduced ability to maintain attention on the task. It is plausible that an increase in distraction whilst attempting dual task walking may contribute to the increase in falls risk seen in older adults with MCI.

In addition, we also found that the functional connectivity between the DMN and SMA was significantly associated with slower gait speed and increased postural sway. Given that the DMN would normally deactivate on task (Raichle et al., 2001), intact resting state DMN–SMA connectivity would usually be an indication that during task performance, when the SMA is active, the DMN is deactivated. In this case, greater connectivity between these brain regions at rest is negatively correlated with gait speed and thus suggests there is interference between these networks

on task; most plausibly that the DMN is influencing the SMA by remaining active on task.

The SMA has been shown to play an active role in motor planning (Roland et al., 1980), and in maintaining attention on a motor task (Johansen-Berg and Matthews, 2000). Through the DMN remaining active, increasing competition for resources and reducing the ability to maintain attention on the task, the SMA is unable to create motor plans with as much accuracy or efficiency. This would likely result in reduced quality of gait control and subsequently, gait speed. In addition, this decline in gait speed will further contribute to an increased risk of falls (Verghese et al., 2009; Espy et al., 2010). It would be useful to establish whether interventions targeted at increasing gait speed are also found to alter this maladaptive connectivity in order to provide support for this theory.

The SMA has also been implicated in postural control. Findings from Viallet et al. (1992) lead to the suggestion that the SMA was responsible for selecting the relevant circuits of phasic postural adjustments in order to maintain posture. Consequently, if an increase in DMN activity on task were to add competition for resources to the SMA, it is likely that there would be a decrease in the ability to select the correct postural adjustments to maintain balance, leading to the increased sway evident in this study.

It is important to note that greater connectivity between the DMN and SMA only correlated with performance on the floor sway task and not performance on the foam sway task. It is thought this is because the foam sway task may require more overt attention. Whereas, the participants may be less inclined to intentionally focus on the floor sway task because they are not so obviously unstable. This may be better explained by the capacity model of attention, which states that when a task is less cognitively demanding, more resources remain available to be allocated to task-irrelevant networks (Kahneman, 1973). In this case, the additional resources are allocated to the DMN. It may therefore be that when performing the foam sway task, participants were able to overcome the DMN interference by actively allocating more resources to the task-related regions than when performing the floor sway task. However, this can only be considered a speculative explanation as participants were not asked about how much attention they paid to the task and taskrelated neuronal activity could not be measured. This theory is also contradictory to the capacity sharing theory that has been used to explain the findings in relation to dual task performance, known to be a cognitively demanding task. Investigating the effect of the DMN–SMA connectivity over progressively more challenging postural tasks would be beneficial to determine which theory is most likely to explain these results.

There was no relationship found between connectivity of the DMN and FPN at rest and any of the behavioral measures. Due to the lack of previous literature in this area it is not completely clear why this may be. One of the studies that did find increased connectivity between the DMN and FPN during the performing of a motor task, found the increase in connectivity was specific to older adults classified as fallers rather than non-fallers (Hsu et al., 2014). Consequently, it may be that the FPN is more resistant to the influence of the DMN until a later stage in mobility decline. Further research is required to investigate the effect of the DMN on the FPN across several stages of falls risk, ideally comparing healthy older adults to those with MCI and those classified as fallers in order to establish at what stage connectivity between the DMN and FPN may become detrimental, if at all.

This study only compared resting state brain network connectivity with behavioral measures and did not investigate brain activity on task. Due to the limitations of functional MRI (fMRI) scanning (i.e., sensitivity to movement artifacts), and the lack of MRI-safe apparatus it was not possible for us to assess activity of the DMN and associated networks while performing motor tasks such as walking. Thus, caution must be taken when interpreting these results, as they are only correlational. To support our findings, future studies should aim to investigate the functional connectivity both within and between these networks whilst simultaneously performing a motor task. It is also important to acknowledge that there may be other underlying factors that could account for these results. For example, this study did not investigate the impact of potential sex differences or the underlying structural integrity of these networks, which may have provided an alternative explanation of the findings. Consequently, future studies should aim to identify any sex differences and compare both the structural and functional associations between these networks and mobility factors in order to better understand the underlying neurobiological factors contributing to increased falls risk. In addition, our study consisted of community-dwelling older adults with MCI exclusively and we did not have ethical clearance to collect any data pertaining to the race or ethnicity of our participants, as this was not related to our hypotheses. Thus, we acknowledge that this may make the generalizability of our findings ambiguous. Furthermore, some evidence highlights the potential influence of the cerebellum on the networks discussed in this study (Habas et al., 2009). Although this was not one of the main focuses of the current study, it may be beneficial to additionally investigate the role of the cerebellum in relation to these networks in future research. Due to the exploratory nature of this study and the small sample size, no adjustments were made to the significance value in order to control for multiple comparisons. It is acknowledged that this increases the likelihood of a type I error and future studies with larger sample sizes are needed to confirm our current findings. Finally, walking was performed indoors on a GAITRite mat and a verbal subtraction task was used for the dual task, consistent with other studies using a dual task walking paradigm (Beurskens and Bock, 2012). However, it would be useful to determine the specific effects of other cognitive and motor tasks whilst walking over varying terrains as well, as this is likely to be more representative of real-world scenarios.

Our results show that increased resting state DMN connectivity is associated with a decrease in dual task performance, slower gait speed and increased postural sway in older adults with MCI; either via greater within network connectivity or between-network connectivity with the SMA. This supports the theory proposed initially in **Figure 1**. To our knowledge, this is the first study to investigate the role of the DMN on dual task performance and motor functioning in people with MCI. These findings can be used to determine if interventions targeted to improve gait, cognition, and/or dual tasking specifically, can reduce the maladaptive effects of greater DMN connectivity in older adults vulnerable to an increased risk of mobility disability.

#### AUTHOR CONTRIBUTIONS

fnagi-09-00423 December 19, 2017 Time: 16:18 # 8

RC, TL-A, and CLH were involved in designing and performing the study. All authors contributed to the data analysis. RC,

#### REFERENCES


TL-A, and CLH were involved in the interpretation of results. RC wrote the first draft of the manuscript. TL-A, JB, and CLH wrote portions of the manuscript and critically reviewed the manuscript. All authors have read and approved the manuscript.

#### FUNDING

Funding was provided by an Alzheimer Society Research Program Grant No. F13-05246 to TL-A.


Kahneman, D. (1973). Attention and Effort. Englewood Cliffs, NJ: Prentice Hall Inc.


Alzheimer type. Proc. Natl. Acad. Sci. U.S.A. 100, 14504–14509. doi: 10.1073/ pnas.2235925100


functional, and biochemical effects. Neurosci. Biobehav. Rev. 34, 721–733. doi: 10.1016/j.neubiorev.2009.10.005


**Conflict of Interest Statement:** TL-A is a Canada Research Chair in Physical Activity, Mobility and Cognitive Neuroscience. JB is a Canadian Institutes of Health Research and Michael Smith Foundation for Health Research Postdoctoral Fellow. CLH is an Alzheimer Society Research Program Doctoral Trainee.

The other author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Crockett, Hsu, Best and Liu-Ambrose. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Influences of 12-Week Physical Activity Interventions on TMS Measures of Cortical Network Inhibition and Upper Extremity Motor Performance in Older Adults—A Feasibility Study

Keith M. McGregor 1, 2 \*, Bruce Crosson1, 2, Kevin Mammino<sup>1</sup> , Javier Omar <sup>1</sup> , Paul S. García1, 3 and Joe R. Nocera1, 2

<sup>1</sup> VA Rehabilitation R&D Center for Visual and Neurocognitive Rehabilitation, Atlanta VA Medical Center, Decatur, GA, United States, <sup>2</sup> Department of Neurology, Emory University School of Medicine, Atlanta, GA, United States, <sup>3</sup> Department of Anesthesiology, Emory University School of Medicine, Atlanta, GA, United States

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Jeffrey Thomas Cole, Uniformed Services University of the Health Sciences, United States Joyce Gomes-Osman, Beth Israel Deaconess Medical Center, Harvard Medical School, United States

> \*Correspondence: Keith M. McGregor keith.mcgregor@emory.edu

Received: 14 June 2017 Accepted: 08 December 2017 Published: 04 January 2018

#### Citation:

McGregor KM, Crosson B, Mammino K, Omar J, García PS and Nocera JR (2018) Influences of 12-Week Physical Activity Interventions on TMS Measures of Cortical Network Inhibition and Upper Extremity Motor Performance in Older Adults—A Feasibility Study. Front. Aging Neurosci. 9:422. doi: 10.3389/fnagi.2017.00422 Objective: Data from previous cross-sectional studies have shown that an increased level of physical fitness is associated with improved motor dexterity across the lifespan. In addition, physical fitness is positively associated with increased laterality of cortical function during unimanual tasks; indicating that sedentary aging is associated with a loss of interhemispheric inhibition affecting motor performance. The present study employed exercise interventions in previously sedentary older adults to compare motor dexterity and measure of interhemispheric inhibition using transcranial magnetic stimulation (TMS) after the interventions.

Methods: Twenty-one community-dwelling, reportedly sedentary older adults were recruited, randomized and enrolled to a 12-week aerobic exercise group or a 12-week non-aerobic exercise balance condition. The aerobic condition was comprised of an interval-based cycling "spin" activity, while the non-aerobic "balance" exercise condition involved balance and stretching activities. Participants completed upper extremity dexterity batteries and estimates of VO2max in addition to undergoing single (ipsilateral silent period—iSP) and paired-pulse interhemispheric inhibition (ppIHI) in separate assessment sessions before and after study interventions. After each intervention during which heart rate was continuously recorded to measure exertion level (load), participants crossed over into the alternate arm of the study for an additional 12-week intervention period in an AB/BA design with no washout period.

Results: After the interventions, regardless of intervention order, participants in the aerobic spin condition showed higher estimated VO2max levels after the 12-week intervention as compared to estimated VO2max in the non-aerobic balance intervention. After controlling for carryover effects due to the study design, participants in the spin condition showed longer iSP duration than the balance condition. Heart rate load was more strongly correlated with silent period duration after the Spin condition than estimated VO2.

Conclusions: Aging-related changes in cortical inhibition may be influenced by 12-week physical activity interventions when assessed with the iSP. Although inhibitory signaling is mediates both ppIHI and iSP measures each TMS modality likely employs distinct inhibitory networks, potentially differentially affected by aging. Changes in inhibitory function after physical activity interventions may be associated with improved dexterity and motor control at least as evidence from this feasibility study show.

Keywords: aging, motor control, physical fitness, TMS, interhemispheric inhibition, neuroimaging

### INTRODUCTION

Aging has been shown to be associated with a loss of interhemispheric inhibition that may negatively affect unimanual motor performance of the dominant hand (McGregor et al., 2012; Fujiyama et al., 2013; Heise et al., 2013, 2014; Levin et al., 2014; see also Spirduso, 1975; Salthouse, 1996; Talelli et al., 2008). Though the motor system is relatively spared as compared to other cognitive domains such as executive function, aging is associated with decreased upper extremity function (Salthouse, 1996). While impaired inhibitory function may not reach clinical significance for diagnostic purpose of motor dysfunction, it may reveal evidence of aging-related alteration of cortical function. This loss of interhemispheric inhibition can be assessed with transcranial magnetic stimulation (TMS). One TMS measure that has shown variability due to aerobic fitness and aging is the ipsilateral silent period (iSP) (McGregor et al., 2011; Davidson and Tremblay, 2013; Coppi et al., 2014). Briefly, the iSP is a stimulation-induced diminution or cessation of oscillation in electromyography (EMG) of a contracted muscle when stimulation is given to the motor cortex ipsilateral to the muscle target. This effect is believed to be mediated by alterations in inhibitory network function (Irlbacher et al., 2007; Lenzi et al., 2007), which may be sensitive to changes in aerobic capacity (Maddock et al., 2016). Previous cross-sectional work has shown that regular aerobic exercise may be associated with changes in interhemispheric inhibition and motor dexterity in older adults (Voelcker-Rehage et al., 2010; McGregor et al., 2011, 2013). This relationship may indicate that aging related motor declines might be mitigated or even reversed by the engagement in aerobic exercise. While the effect of acute exercise has been probed with respect to sensitivity to measures from TMS (Roig et al., 2012; Singh et al., 2014; Lulic et al., 2017), we know of few studies that have assessed the longitudinal effects of a longer-term aerobic exercise program on TMS measures of inhibitory function potentially sensitive to aging and motor control (see Gomes-Osman et al., 2017).

The iSP is a complicated measurement that involves a number of cortical and descending spinal inhibitory connections. The silent period onset is typically ∼38 ms after stimulation and can last anywhere from 10 to 70 ms depending on stimulation and level of muscle contraction (Giovannelli et al., 2009; Petitjean and Ko, 2013; Fleming and Newham, 2017; Kuo et al., 2017). The ipsilateral inhibition seen in the most commonly measured muscle, the first dorsal interosseous (FDI), certainly involves primary motor cortex (M1) callosal transfer, as degradation of the corpus callosum diminishes the measure (Meyer et al., 1995; Li et al., 2013). However, it is likely that inhibitory influences of the reticulospinal and propriospinal tracts provide an additive effect to muscle quiescence (Nathan et al., 1996; Ziemann et al., 1999). In addition to cortically mediated mechanisms the cause of the silent period is also influenced by the dynamics of the alpha-motoneurons themselves (Doherty et al., 1993). Refractory periods of the muscle spindle and the reactive involvement of spinal inhibitory interneurons due to descending ipsilateral corticospinal/corticobulbar/oligospinal input likely contribute to the duration of the iSP if not in its early phase, but later in its delayed resolution to baseline (Jung and Ziemann, 2006). The complexity of the iSP may be of relevance to its possible sensitivity to aging related change. Given that it is a volitional response, in contrast to another measure of interhemispheric using paired pulse parameters, aging related alteration of the iSP may reflect a functional change in motor capacity (Coppi et al., 2014).

A neurotransmitter system with strong influence on cortical inhibition and likely even motor control is the gamma aminobutyric acid system (GABA). The aging process may be responsible for a decrease in GABA tone in the neocortex (Gao et al., 2013; but see also Mooney et al. (2017). However, this decrease may or may not be responsible for changes in motor performance in aging, as GABA receptors can change endogenous sensitivity levels over time (Rozycka and Liguz-Lecznar, 2017). The TMS literature has approached GABA receptor function for many years, particularly in light of aging (Sale and Semmler, 2005; Stagg et al., 2011; Davidson and Tremblay, 2013; Opie et al., 2015). A significant question has arisen as to the role of a particular subtype of GABA receptor (GABAb) in the mediation of interhemispheric inhibition with respect to how it is assessed using TMS. GABAb receptors have been implicated in two distinct measures of interhemispheric inhibition: the iSP (described above) and paired-pulse interhemispheric inhibition (ppIHI). The employ of ppIHI requires the use of two stimulators and reflects the diminution of a motor evoked potential (MEP) in a muscle target when a conditioning pulse to the motor cortex ipsilateral to the target hand precedes a test stimulation pulse (Ferbert et al., 1992). It is yet unknown if these measures involve the same cortical circuitry (inhibitory networks) or reflect complementary findings in the estimation of the effects of pharmacological agents (see Ziemann et al., 2015). That is, the relationship between the two measures may be isolated to the administration of GABAb agents, and may not have a direct relationship with motor behavior. The current work seeks to investigate if physical activity, which has shown been associated with differences in silent period duration in previous cross-sectional work, shows proportional effects on measures of ppIHI.

The present work describes data collected from 21 older participants (60+ years) who engaged a 12-week exercise intervention comprised of either an interval-based aerobic spin program or a non-aerobic, balance and stretching condition. We employed a crossover design to compare the effects of the activity interventions on the same participants in alternate conditions. As such, participants crossed over into the alternate exercise condition for another 12-week intervention. We sought to test motor dexterity and assessments of cortical inhibition using both ppIHI and iSP paradigms that may be associated with improved motor function after increased aerobic activity. Based on our previous cross-sectional data, we hypothesized that participants completing the aerobic spin protocol would show improved upper extremity motor dexterity and increased levels of interhemispheric inhibition. We further hypothesized that changes in iSP after the intervention would indicate greater levels of interhemispheric inhibition as compared to ppIHI potentially due to the volitional and physiologically complex origin of the iSP.

### METHODS

#### Participants

In this 24-week randomized controlled crossover trial (RCT: NCT01787292), participants were randomized and divided into an aerobic, spin cycling exercise group (Spin) or a nonaerobic balance training group (Balance) to equalize contact and monitoring. Each intervention lasted 12-weeks (Arm 1), after which, the participant crossed over into the alternate arm (Arm 2) of the study for an additional, 12-week intervention (e.g., Arm 1, Balance–Arm 2, Spin). The crossover was an AB/BA uniformwithin-sequences design with a limited washout period (∼1 week). Data from both arms of the intervention are presented in this report.

Study personnel explained the purpose, potential risks of the experiment and completed the informed consent process with each participant following protocols approved by the Emory University's Institutional Review Board (IRB00059193) in compliance with the Helsinki Declaration. All participants gave written informed consent filed with both the Atlanta VA Research and Development Office and Emory University's IRB.

This report includes 21 participants that were recruited from a volunteer database, which included elderly individuals (60 years and over). An additional four older participants enrolled in the study, but chose to withdraw prior to completing the first arm of the intervention. To meet inclusion criteria participants had to (1) be between of 60 and 85 years of age, (2) report being sedentary, defined as not engaging in structured physical activity and/or not accumulating 30 min or more of moderate to strenuous weekly physical activity, assessed with a modified Godin Leisure Time Exercise Questionnaire— LTEQ (Godin and Shephard, 1997), (3) have no history of depression, neurological disease, including Parkinson's disease, Alzheimer's disease, multiple sclerosis or stroke, (4) report being right handed (using the Edinburgh handedness inventory Oldfield, 1971), (5) report being a native English speaker, and (6) obtain primary care physician's approval for study participation. Exclusion criteria included (1) conditions that would contraindicate TMS (e.g., seizure, stroke, tremor, etc.), (2) failure to provide informed consent, (3) hospitalization within the past 6 months, (4) uncontrolled hypertension or diabetes (reported non-compliance with prescribed management program), (5) inability to walk 400 m, and (6) significant cognitive executive impairment, defined as a score on the Montreal Cognitive Assessment (MoCA) of <24, (7) having a TMS measurement of lowest motor threshold (LMT) >66% of maximum stimulator output (MSO) (as stimulation for the paradigm was set to 150% of LMT). Due to the high incidence of prescription of hypertension medications in sedentary older adults (n = 12, six per group), we did not exclude individuals on these medications.

During intervention sessions, all participants wore a Polar FT7 chest strap heart rate monitor with paired monitor/wristwatch. Heart rate was taken from each participant every 2–3 min during the sessions and logged on a data sheet. On infrequent occasions (<2% of HR acquisitions), the chest strap monitor would fail to synchronize with the watch during the intervention session. In such instances, we interpolated the heart rate data from adjacent recordings within each session provided they were within reasonable ranges to each other (±∼5–10 bpm). If a heart rate monitor failed to synchronize at study outset (a problem with older adults with lower resting galvanic skin responses) we would use a battery-powered pulse oximeter or an Apple Watch (Cupertino, CA) to measure heart rate at the above described intervals. For both interventions, we recorded attendance, attrition, and heart rate. All participants completed the 36 assigned sessions for each intervention though we had to accommodate more absences for individuals in the balance condition.

#### Aerobic "Spin" Intervention Protocol

Consistent with our previous study (Nocera et al., 2015), the group exercise intervention began with 20 min of Spin aerobic exercise three times a week for 12 weeks on stationary exercise cycles and was led by a qualified instructor. Importantly, the time of each session progressed based on the recommendation of the instructor by 1–2 min as needed to a maximum time of 45 min per session. Heart rate reserve was assessed using the Karvonen method (220 bpm – age = maximum heart rate; heart rate reserve [HRR] = maximum heart rate – resting heart rate). Exercise intensity began at low levels (50% of HRR) and increased by 5% every week (as deemed appropriate by the instructor) to a target maximum of 75% HRR. Participants wishing to exceed this capacity could do so for limited exercise intervals if they so choose. Target exercise intensities were adapted for participants on diuretics, ACE-inhibitors, beta-blockers based on recent recommendations in the literature (Diaz-Buschmann et al., 2014; Taubert et al., 2015) to produce equivalent aerobic capacity improvement as non-medicated individuals. These included the "talk-test" and relative physical exertion estimation using the Borg 6–20 difficulty scale (6 = lowest effort; 20 = maximum effort).

The Spin intervention took place in a climate controlled fitness facility. The instructor guided the participants through a light effort 5-min warm up (not included in data analysis), then a workout phase that included steady up-tempo cadences, sprints (increased rpm), and climbs (increased resistance). As such, the exercise routine employed an interval-based training approach. During the workout phase the target HRR reserve was maintained by averaging increases and decreases in intensity/HR. The goal was to maintain within a 10% offset from the HRR goal during the workout phase. Thus, participants were within target HRR on average across the session despite the intervals of increased and decreased workload. All participants wore HR monitors throughout each session and were instructed to attain their respective HR target range at 5-min intervals. Staff members also monitored and tracked the HR to ensure adequate intensity throughout each session. Brief weekly meetings in which each participant's HR was reviewed served as a way to encourage those with lower attendance or HR measurements to improve their performance for the next week.

### Balance/Light Strength Training Intervention Protocol

The main purpose of the Balance and strength training group was to have participants engage in non-aerobic physical activities that may help reduce fall risk. Participants in the balance group were equalized to the Spin group with regards to contact and monitoring frequency. As such they reported to the same facility with the same interventionists; however, instead of progressive aerobic exercise they participated in group balance, stretching and light muscle toning exercises. Beginning at the outset of the intervention, a baseline balance assessment was taken for each individual to titrate task difficulty depending on intake stability risk. This was formally measured using the short physical performance battery (SPPB), which is a measure consisting of a top score of 12 (scores lower than 10 indicate moderate fall risk). All participants in this study had a score of 11 or greater, indicating low fall risk from the SPPB. Participants began the intervention by practicing balance exercises on foam pads using a chair for support (if necessary). Balance exercises included single leg stand, dual-task (counting backwards) and eyes closed conditions lasting about a total of 10 min. Participants increased difficulty when able to perform the balance session without use of the support chair. In place of foam pads, participants stood on less-stable air-filled pads as they advanced through the 12 week intervention. Participants were also challenged to learn to step on moveable friction pads (six-inch diameter "dots") with variable positions on the floor. Instructors changed the positions of these pads as the session progressed to challenge participants to safely deviate center of mass location during foot placement in order to improve proprioception during gait. In addition, light strength training exercises included instructor-led bodyweight and resistance training using Theraband (Akron, OH) stretch bands. These exercises focused on improving postural support with an emphasis on abdominal engagement and lateral hip abduction. As above, we held brief weekly meetings to discuss progress within the program and workload.

Similar to the aerobic intervention time from the initial 20– 45 min over the course of the 12-week intervention with a light 5-min warm up at the onset of each session. Additionally, heart rate was consistently monitored (also using the Polar FT7 chest strap monitors) to assess general intensity during each session and to advise participant to keep HR below aerobic levels (50% of HRR).

#### Crossover and Attrition

After completing the assessments within a 10-day period following the 12-week intervention, participants crossed over into the opposite arm of the study (e.g., exercise to spin). The participants then completed the second arm of the study for 12-weeks (36 sessions). We did not incorporate a full 12-week "washout" period (to potentially mitigate carryover effects) due to potential attrition of participants. We included a covariate model for carryover effects in our statistical analysis to attempt to account for the lack of washout.

Of note, we enrolled an additional four participants who completed baseline testing, but did not complete the first intervention arm choosing to withdraw from the study. Three of these participants were in the Balance arm, while one was in the Spin. Reasons for attrition were schedule conflicts or exigent family circumstances.

#### Assessments

All assessments were done no more than 10 days before the start of or 10 days after the conclusion of each 12-week intervention period. In total there was: one baseline measurement and two post-intervention measurements. Assessment sessions did not exceed 2 h to alleviate participant fatigue, so testing was spread across two nearly consecutive days (1–3 days). We assessed behavioral performance and cardiovascular fitness on the first day and TMS measures on the second assessment day. In all cases but two, participants began behavioral assessments during mid-morning hours. All TMS sessions were completed during morning hours.

#### Cardiovascular Fitness Assessment

To assess aerobic capacity, participants performed a YMCA submaximal fitness test on a Monark 828e (upright) or RC4 (recumbent) cycle ergometer (Vansbro, Sweden). This submaximal test was used to estimate the participant's maximal oxygen uptake (VO2max) prior to and after interventions. The selected submaximal test is much better tolerated than a maximum exertion treadmill test in the study's population (sedentary older adults). The YMCA-test uses an extrapolation method in which heart rate workload values are obtained at 2–4 points during stages of increasing resistance and extrapolated to predict workload at the estimated maximum heart rate (e.g., 220 age). VO2max is then calculated from the predicted maximum workload. Prior to beginning the test, the procedures were briefly explained and participants completed a 2-min warm-up consisting of pedaling without load so that they could adapt to the ergometer for the first minute and then pedaling with a 0.5 kg.m load during the second minute. The YMCA submax test has an R = 0.86 with VO2max and a SEE = 10% of the predicted VO2max (Beekley et al., 2004).

#### Heart Rate Workload Assessment

As a submaximal exercise estimate may be limited in determining the effectiveness of a physical activity intervention in a smaller sample size, we additionally calculated the average intrasession heart rate as compared to the target goal of 75% HRR. This was done during training intervals physical exertion starting in the sixth week of the spin program where participant HR target zone meet this criteria. Intervals in this zone (75% HRR) increased in frequency as training progressed up to six intervals per session in the final 3 days of the program. To analyze these data, we scaled the HR-values by 75% of HRR (6– 10 assessments per session × 16–19 sessions) per participant (with the previously denoted adjustments for BP medications). As such, we divided each HR assessment by 75% of adjusted participant HRR and averaged each assessment across sessions within each participant. For example, if a given participants achieved a HR average of 114 bpm for work intervals their 75% HRR target value was 130 bpm, the score would be 0.87. We completed this HRload analysis for all participant sessions and interventions. For the Balance group HRload assessment, we chose the 75% HRR time blocks in mirror of their Spin intervention (n = 6–10 per session × 16–19 sessions). We acknowledge that this estimate may ignore gradual improvement in the Spin intervention as it relies on a single fixed resting HR for baseline.

#### TMS EMG

Electromyography (EMG) was taken from the FDI muscle on both hands using Ag-Ag Cl electrodes using BrainSight (BrainSight 2, Rogue Research) EMG pods. EMG is continuously acquired and stimulator driven TTL triggers a 150 ms acquisition window post TTL with 50 ms of pre-trigger baseline. A LabJack U3-LV analog to digital converter acquired amplified EMG traces with a 12-bit dual-channel analog input sampled at 3 kHz. These data were bandpass filtered from 10 to 10,000 Hz. Muscle activation was monitored with oscilloscope software package integrated into a BrainSight 2 neuronavigated positioning system. Motor evoked potential and other EMG data was exported for statistical analysis using ADInstruments LabChart. A MagVenture X100 magnetic stimulator (MagVenture, Alpharetta, GA) and a MagVenture B-60 60 cm butterfly coil were used to stimulate the left primary motor cortex during the initial mapping procedure. Maximum stimulator output (MSO) for this model is 2.2 tesla. All stimulations were biphasic and stimulation and recording devices were synchronized using TTL pulses. The coil was placed tangential to the scalp with the handle pointing backwards and 45◦ away from the midline for stimulation. The scalp site corresponding to the lowest stimulator output sufficient to generate a magnetic evoked potential of at least 50 mV in six out of 10 trials was defined as the area of resting motor threshold (RMT), also known as the "hotspot." This was the site that was stimulated for the TMS assessments. It is worthy of note that this threshold determination is different from the currently accepted standard employ of a stimulus response curve analysis for measuring cortical excitability (Chang et al., 2016). We do not report on cortical excitability in the current manuscript as estimation of this according to the previous citation optimally uses more than 10 pulses.

#### Ipsilateral Silent Period

For iSP, the left FDI muscle was contracted via pinch grip at 25% maximal voluntary contraction (MVC) measured by pinch grip dynamometer and a stimulator output equivalent to 150% RMT was delivered to the left FDI hotspot. Recent work by Fleming and Newham (2017) has shown that these stimulation parameters are reliable in older adults. The highest acceptable RMT for participation in the current study was 66% of MSO. All participants had a RMT of 66% or less in the current study. Twenty silent period assessments were taken with brief rest breaks after every five trials to alleviate potential muscle fatigue. Participants were also instructed to request rest breaks as needed at any time during the stimulation. The iSP was determined using a longstanding visual inspection method (Garvey et al., 2001). Similar to our previous work (McGregor et al., 2011, 2013), we rectified EMG data and we determined silent period onset at background EMG activity during active pinch squeeze dropped below 20% of pinch baseline (assessed with pre-stimulus acquisition of 50 ms).

#### Paired-Pulse Measures

The long interhemispheric inhibition (LIHI) paired pulse procedure involved interhemispheric inhibition assessment (Ferbert et al., 1992) using a second MagVenture magnetic stimulator (R30) and a matching B-60 (60 cm butterfly coil) for stimulation of the right motor cortex. For this procedure, a conditioning TMS pulse set at 150% of RMT was applied to the right motor cortex FDI hotspot at 40 ms prior to a "test" pulse's administration of 130% of RMT to the left motor cortex. As a result of the conditioning stimulation, the test MEP's response amplitude (in the right FDI muscle) is lowered due to interhemispheric inhibitory processes (denoted as LIHI or long interval interhemispheric inhibition). The inter-trial interval was varied randomly between 4 and 6 s to reduce anticipation of the next trial and mitigate repetitive stimulation effects. Averages of MEP latencies and peak-topeak amplitudes were calculated for each stimulation condition (baseline, IHI). Twenty baseline stimulations (test pulses without conditioning pulse) were compared with 20 conditioned LIHI stimulations for this procedure. Baseline and conditioned stimulations were interleaved to mitigate systematic cortical modulation.

During behavioral assessment sessions, participants performed a battery of cognitive and upper extremity motor tests. Results from the cognitive battery will be addressed in later report. Participants completed motor assessments of the dominant hand including: grip strength, the Halstead-Reitan Finger Tapping task (Reitan and Wolfson, 2013) simple reaction time, the Purdue Pegboard (peg and assembly) (Tiffin and Asher, 1948), and the Nine-Hole Pegboard task (Mathiowetz et al., 1985). Additionally, to test distal motor dexterity, participants engaged in a coin rotation task with two conditions. In the first condition (unimanual), the participant rotated a coin (U.S. quarter) 20 times as quickly as possible using the index finger, middle finger, and thumb with duration as the outcome measure. This test is used for assessment in routine neurological screening and has been shown to be diagnostic of distal motor function both in cases of suspected pathology and aging in the absence of pathology (Hanna-Pladdy et al., 2002; Hill et al., 2010). In the second condition (bimanual), the participant maintained an isometric pinch force of 20–30% of maximum voluntary force with a Jamar brand pinch grip dynamometer using a lateral grip during the rotations. Coin rotation tasks were performed with both the left and right hands. Both the hand used for coin rotation and trial condition (unimanual or bimanual task) were pseudo-randomized and counterbalanced across participants to account for potential order effects across eight runs (two left unimanual, two left bimanual, two right unimanual, two right bimanual). Accidental coin drops were noted, but excluded from consideration and the trial repeated should a drop occur. Participants were allowed 5 min of practice to acclimate to the rotation task in each task condition. Data acquisition began if the participant reported that they believed that additional practice time would not improve task performance. No participants requested additional time beyond the 5-min practice period. The difference score between the bimanual and unimanual task conditions was calculated to assess the effect of bimanual activity on rotation performance.

#### Data Analysis

The current study was a uniform-within-sequences mixedeffects 2 × 2 crossover design with intervention type held as between subjects and intervention sequence (AB/BA) and period (A1B1/A1B2/A2B1/A2B2) as within subjects. A Shapiro-Wilks test was completed across measures to test data for normality. In the event of violation of normality of data, we employed non-parametric Wilcoxon rank sum tests (between subjects) or Mann-Whitney rank sum test (within subjects).

To analyze data from this design, we employed a mixed model approach (PROC MIXED in SAS) using a simple carryover (AB/BA) design with carryover adjustment for session sequence. To account for sequence carryover, we employed analysis of covariance (ANCOVA) in SAS 9.4 (Cary, NC) inclusive of sequence by period covariates against treatment effects. Least square means were adjusted for carryover from the crossover design and Tukey-Kramer mean comparisons for between subjects effects were analyzed with a Kenward-Roger degrees of freedom approximation (Kenward and Roger, 2010). Mauchly's test for sphericity was computed for session as a within subjects variable, and we applied a Greenhouse-Geisser correction to accommodate any violation. In addition, we completed a significance test for the carryover effect between sequences using a delta G∧2 likelihood ratio and Chi-square parameter estimation at alpha of 0.05.

We also performed a mixed-model split-plot ANOVA in JMP 12 (Cary, NC) using a restricted maximum likelihood design holding subjects as random and nested in sequence (i.e., AB/BA) to examine interaction effects of dependent variables based on sequence of presentation. This reduced model did not account for carryover covariates, but was employed to show main effects and interactions of treatments respective of change from each measurement (i.e., intervention at time A vs. intervention at time B; baseline assessment vs. intervention at time A; baseline assessment vs. intervention time at B). Comparisons of intervention effects on dependent variables are shown graphically in Bland-Altman repeated measures plots with t-test for intervention (Altman and Bland, 1983). In addition, we completed correlation analyses on dependent variables across sessions with output statistics reported with the non-parametric Spearman's rho due to the low sample size.

### RESULTS

Our screening measure of physical activity (Godin LTEQ) showed a moderate relationship with estimated VO<sup>2</sup> max, p = 0.42, p = 0.06. Baseline demographic data and neurophysiological correlations at the pre-session across all participants are shown in **Table 1**. Of note, VO<sup>2</sup> was positively correlated with education and inversely correlated with BMI and RMT. Resting motor threshold was also inversely correlated with level of education across all participants in the selected sample. Interestingly, we found no significant correlation between the TMS measures at baseline. There was an effect on gender at baseline with women having slightly longer silent periods t(20) = 1.99, p = 0.05 as compared to men. Baseline data for TMS and motor performance along with their correlations are shown in **Tables 2**–**4**, respectively.

#### Intervention Effects—Spin vs. Balance VO<sup>2</sup> Measures

Depicted in the repeated measures Bland-Altmann plot in **Figure 1**, change in estimated VO2max was significant both

TABLE 1 | Baseline demographic and exercise metrics: age, education, body mass index (BMI), Handedness (as assessed by Edinburgh Handedness Inventory: Right = 1.0, Left = −1.0, assessed level of oxygen consumption during exercise (VO2), Modified Godin Leisure Time Exercise Questionnaire (Self-report of physical activity) and Montreal Cognitive Assessment (MoCA).


accounting for carryover covariates t(20) = 4.90, p < 0.001, and in the reduced model [t(20) = 5.29, p < 0.001]. Interestingly, there was a significant carryover effects in the Spin First Intervention (AB/BA), χ 2 (0.05, 1) <sup>=</sup> 6.89 (<sup>p</sup> <sup>&</sup>lt; 0.03) as compared to the Balance First Intervention (BA/AB), which had no carryover effects for VO<sup>2</sup> change, χ 2 (0.05, 1) <sup>=</sup> 3.28, ns. We found no gender differences in change measures.

#### Heart Rate Workload

Heart rate workload (HRload) was computed as a function of participants 75% heart rate reserve during intervention sessions. Heart rates in the target interval blocks were expressed as a percentage of the goal of 75% HRR. As expected, HRload was higher for the Spin intervention as compared to the Balance intervention, Z(20) = 2.27, p < 0.04. A significant carryover effect

TABLE 2 | Baseline transcranial magnetic stimulation measures between groups—std. dev.


No differences were evident in comparisons of resting motor threshold (RMT), ipsilateral silent period (iSP), and paired pulse interhemispheric inhibition (ppIHI).

TABLE 3 | Baseline motor comparisons—std. dev.


Tests were with dominant (right) hand unless otherwise specified. Higher score on Purdue, Halstead are better. Lower scores on 9-Hole peg and coin rotation are better. Bimanual difference score is the difference between unimanual dominant hand coin rotation and dominant coin rotation when non-dominant hand is engaged in 25% maximum voluntary contraction squeeze task.

was evident for HRload was shown χ 2 (0.05, 2) <sup>=</sup> 8.03 (<sup>p</sup> <sup>&</sup>lt; 0.01). Interestingly, using a median split within interventions at time A, we determined that individuals who performed with highest HRload when performing Spin first continued to with higher HRload at crossover (n = 5), while those performing with the lowest HRload in Balance first had lower HRload at crossover (n = 5). No gender effects were evident for heart rate data either (See **Figure 2**).

#### TMS Measures Ipsilateral Silent Period

Depicted in repeated measures Bland-Altman plot in **Figure 3** are change scores respective of intervention shown in the **Table 5**

FIGURE 1 | Repeated measures Bland-Altman plot of VO2max estimate comparisons between interventions sessions as plotted in JMP12. Ordinate axis denotes difference score between treatments, while abscissa denotes. The central axis (in red) is offset to depict the mean value between interventions A+B/2. Thus, vertical gain (from red axis) indicates greater improvement in VO2 in Intervention A, while rightward gain indicates greater improvement after crossover. Circles represent Spin participant in Spin first condition while boxes represent participants in balance first (Between groups comparison—means represented by dotted lines: t = 5.29, p < 0.01).



Significant correlations were evident between estimated volume of oxygen consumption (VO2), education, resting motor threshold (RMT). Ipsilateral silent period (iSP) and paired pulse interhemispheric inhibition (ppIHI) were not correlated with baseline demographics. Correlations use Spearman rho statistic. BOLD denotes statistical significance below p = 0.05.

below, there were significant effects of intervention type on change in iSP duration t(20) = 2.11, p < 0.05 in the full model (inclusive of carryover), and in the reduced model, t(20) = 4.93, p < 0.01. Individuals in the Spin intervention had longer iSPs than those in the balance intervention. Significant sequence carryover was present in silent period assessment for both interventions, χ 2 (0.05, 2) <sup>=</sup> 8.89 (<sup>p</sup> <sup>&</sup>lt; 0.01). We found no gender effects for change score in silent period duration, though women had a slightly higher baseline duration than men, t(20) = 1.99, p = 0.05.

Paired Pulse Interhemispheric Inhibition: No significant differences were denoted for ppIHI changes in the full model t(20) = 2.13, ns, though a trend was shown in the reduced model with greater interhemispheric inhibition in the spin intervention t(20) = 1.94, p = 0.07.

#### Behavioral Changes

Body mass index did not change respective of either intervention. Across a battery of motor indices, individuals completing the Spin Intervention improved on measures of dominant upper extremity, as compared to no change in the Balance condition. These data are shown in **Table 6** and were derived from the reduced model comparison as computed by JMP12. Notably, significant differences were shown in the bimanual coin rotation task, during which the participant actively squeezes a dynamometer while rotating a coin. Participants completing the balance intervention performed the task significantly faster during the bimanual task condition as compared to little change in individuals completing the Spin intervention.

participant in Spin first condition while boxes represent participants in balance first (Between groups comparison—means represented by dotted lines: t = 2.11, p < 0.05).

TABLE 5 | TMS change measures after interventions.


RMT, Resting motor threshold; iSP, ipsilateral silent period; IHI, paired pulse interhemispheric inhibition. iSP is measured in ms, while IHI is percentage change from baseline pulse to preconditioned pulse. BOLD denotes statistical significance below p = 0.05.

### Correlations

As there do not exist ideal methods to index aging-related changes in upper extremity motor control, we performed a battery of tests. We were interested in how our TMS measures of interhemispheric inhibition related to these assessments. The data in **Table 7** show significant correlations between the silent period duration and measures of distal dexterity (9-hole pegboard, Purdue, and coin rotation tasks) in aggregate after both interventions. Again, carryover considerations somewhat lessen the extensibility of these results.

In addition, we performed correlations on VO2, HRload, TMS measures and motor performance to investigate relationship of the dependent measures. We performed correlations on VO2, HRload, TMS measures, and motor performance to investigate relationship of the dependent measures. Interestingly, whereas in the Balance intervention, estimates of VO<sup>2</sup> were strongly inversely related to iSP (more so than HRload) the strongest predictor of change in iSP was HRload. These data are shown in **Tables 8**, **9** for Spin intervention and Balance intervention, respectively. We did not show any relationship between ppIHI and estimates of physical fitness/activity.

It is important to note that these data are underpowered with respect to sample size per intake group. Respective of the change in iSP, the effect size at alpha of 0.05 is 0.6. Ideally, this would require 11 participants per group. As such, we consider these data preliminary.

#### DISCUSSION

The present study demonstrates that an aerobic spin exercise intervention appears to increase the duration of the iSP in older adults and improve measures of distal upper extremity dexterity. Increased iSP duration was correlated with improved performance across multiple distal dexterity measures. Additionally, we found that the aerobic spin condition had no effect on a paired pulse measure of long interval interhemispheric inhibition (ppIHI).

Previous research has shown that engagement in regular physical activity considered aerobic in nature is associated with increased activity of inhibitory networks within the brain (McGregor et al., 2011, 2013; Nocera et al., 2015). The current

TABLE 6 | Change metrics in behavioral performance comparing intervention groups—std. dev.; Purdue Peg—Higher score is better; 9-Hole pegboard and Unimanual coin rotation—lower is better.


Bimanual difference is the difference between unimanual and bimanual coin rotation tasks. Data is from reduced model as implemented in JMP12. BOLD denotes statistical significance below p = 0.05.

work presents the first evidence that previously sedentary individuals who engage in a relatively short-term (12-week, 36 sessions) aerobic exercise program show changes in a measure of interhemispheric inhibition, the iSP previously shown to be sensitive to aging-related change (Sale and Semmler, 2005; McGregor et al., 2011; Davidson and Tremblay, 2013; Coppi et al., 2014). Moreover, a longer silent period duration was associated with improved unimanual performance on distal dexterity in our study, potentially indicating an association of cortical inhibition with motor dexterity.

#### Aging-Related Motor Performance and the Ipsilateral Silent Period

One of the most interesting findings in the current study relates to the relationship between iSP duration and motor performance.

TABLE 8 | Post Intervention Correlations accounting for carryover effects after Spin Intervention.


Values are Spearman Rho with alpha value in parentheses. BOLD denotes statistical significance below p = 0.05.

TABLE 9 | Post intervention correlations accounting for carryover effects after balance intervention.


Values are Spearman Rho with alpha significance. BOLD denotes statistical significance below p = 0.05.

TABLE 7 | Relationship between TMS measures (iSP change, ppIHI % change) and motor dexterity change across all participants after both interventions regardless of order.


Purdue Peg—Higher score is better; 9-Hole pegboard and Unimanual coin rotation—lower is better. Bimanual coin difference is calculated as unimanual coin rotation – bimanual coin rotation. Values are Spearman Rho calculation with df = 20. BOLD denotes statistical significance below p = 0.05.

This work has demonstrated that an aerobic spin exercise program can increase the iSP duration in concert with improving motor performance on dexterity tasks. These findings may relate to previous cross-sectional reports in our lab that found that physically active older adults had longer silent period durations than sedentary individuals in the same age cohort (McGregor et al., 2011, 2013). The question arises as to the functional relationship between the iSP and distal upper extremity dexterity. What does a silent period increase after an intervention actually infer respective of motor capability, particularly for the purposes of rehabilitation? As the iSP is complicated involving alpha motor neuronal dynamics, callosal transfer with multiple inhibitory internetworks (including spinal), and muscular control, determination of the exact mechanism driving the change is not possible in the current work. However, it is likely that cortical changes account for more of the change than changes in the periphery (muscle capacity/tone), as previous work has repeatedly shown that increased levels of exogenous stimulation alters silent period duration more so than increased motor load (Giovannelli et al., 2009; Kuo et al., 2017). Therefore, the silent period may be a reflection of an intrinsic cortical inhibitory framework that serves to regulate interhemispheric transfer.

Toward this, the relationship between the iSP and the coin rotation task is worthy of additional discussion, as it is one the few bimanual motor tasks probed in the current report (a noted limitation). We have previously reported aging-related differences in the coin rotation task when subtracting unimanual performance from bimanual performance. Younger adults complete the coin rotation task faster than sedentary older adults in either the unimanual or bimanual conditions. However, when sedentary older adults engage in the bimanual task condition (i.e., ∼25% MVC pinch grip), their coin rotation speed improves. We have previously postulated that aberrant interhemispheric transfer during a unimanual task may interfere with dexterous task performance (McGregor et al., 2012). However, during a bimanual task, the engagement of the ipsilateral motor areas (to the hand performing the rotation) might act to either improve the signal-to-noise dynamics between hemispheres resulting in improved motor dexterity. Additionally, this bimanual performance effect is sensitive to differences in physical activity levels in middle-aged and older adults (McGregor et al., 2011). In the present study, our participants who completed the aerobic spin training showed improved performance on the unimanual coin rotation task. Moreover, the difference score between unimanual and bimanual conditions was lower, particularly for the non-dominant hand. Interestingly, these data were correlated with improved silent period duration. This may indicate that improved interhemispheric inhibition after the aerobic exercise in older adults could restore motor dexterity by improving signal-to-noise characteristics in the task active cortex.

#### Aerobic Spin Intervention

The main contrast condition in the present crossover study was the type of intervention either aerobic "Spin" or Balance. As a result of our interval-based Spin program our participants, regardless of sequence of intervention, improved estimated VO<sup>2</sup> as a result of increased workload. The physical performance metrics were highly correlated with both silent period duration and tests of motor dexterity. That individuals improved on motor dexterity is notable in the present study because we did not employ a manual task training component to the study, and indeed, our intervention was largely driven by the activity of the lower extremity. With respect to the mechanism of change, increased levels of brain derived neurotrophic factor (BDNF) has received dominant attention in the literature (Kleim et al., 2006; Tang et al., 2008; Schmolesky et al., 2013; Szuhany et al., 2015). Our lab only recently began to assay serum BDNF levels, but our preliminary data indicates that our aerobic Spin interval training program increases serum BDNF levels by 17% with peaks achieved 15 min post the 45-min spin session (unpublished data). BDNF is believed to promote synaptic plasticity possibly through facilitating signaling cascades after its dimers bind to its preferential receptor, TrkB (Phillips et al., 2014). As a result, multiple proteins associated with cell survival and proliferation are produced if the TrkB receptor has the beneficial Val66Val polymorphism (Kleim et al., 2006). It is unknown what benefit BDNF or other potential modulatory neurotrophins might have on either the TMS or behavioral measures employed in the current study. It is likely that increased HR workload is associated with a higher release of BDNF (Schmolesky et al., 2013) and this would predict greater motor performance and longer silent periods. However, this postulation requires additional study to vet. As such, the specific mechanism by which aerobic exercise alters cortical function remains largely unidentified with respect to both systems physiology and molecular neuroscience. Clearly, much more work is required address this critical issue.

It is important to note that the contrast condition of Balance training was not of detriment to our participants in terms of functional outcome despite a negative correlation between silent period duration and physical activity measures. Indeed, while the participants' aerobic capacity and heart rates were similar at post-assessment as compared to immediately beginning the Balance program, a crossover effect was evident in this condition. That is, participants continued to maintain gains in the currently reported metrics if they crossed over from Spin into this condition. Moreover, beyond the contrast to Spin, the Balance and light strength training condition improved core strength in participants and improved proprioception. As such, the negative correlations to motor dexterity shown in the current study may rather reflect the dominant improvement in the Spin comparison, rather than functional declines in the Balance condition. Respective of this, the Balance condition employed in the current report served as a contact control, but perhaps not an ideal control. We previously attempted to employ a wait-list control and an education-only program and washout periods to this project, but due to the study environment and recruitment dropout, we instead chose to directly enroll participants into the Spin or Balance interventions with immediate crossover. Future work will certainly employ a cleaner study design, though the results from the current, albeit non-ideal design are extremely encouraging.

#### Differences between iSP and ppIHI

The change in iSP after aerobic exercise is notable insofar that it differed from an alternative measure of interhemispheric inhibition assessed with the paired-pulse LIHI stimulation, which showed no differences after the intervention and was not correlated with motor performance. This is curious and worth some exploration since both the iSP and ppIHI protocols have been reported to involve similar neurotransmitter receptor systems and are considered complementary measures of interhemispheric inhibition (Chen, 2004; Di Lazzaro et al., 2007; Wischnewski et al., 2016). In a recent study, Li et al. (2013) identified patients with callosotomy or callosal agenesis and tested iSP duration and magnitude of ppIHI. The authors reported that both ppIHI and iSP are impaired after lesion of the corpus callosum. The inhibitory effects should not be considered identical, however, as ppIHI and iSP paradigms show different changes during pharmacological manipulations (Siebner et al., 1998; McDonnell et al., 2007; Ziemann et al., 2015) and when employed immediately after or in concert with other TMS paradigms (e.g., LICI, SICI) (Udupa et al., 2010). Based on pharmacologic investigations, it has been suggested that the interhemispheric inhibition underlying both the ppIHI and iSP paradigms involve GABA Type B (metabotropic) receptor (Ziemann et al., 2015). Our results show a difference between iSP and ppIHI and therefore suggest that it is unlikely that GABAb receptor activity is the sole mechanism of this interhemispheric inhibition. Were this true, the iSP and ppIHI measures should have been directly related in the current study. Much more study is required to elucidate the neurophysiological metabolism of inhibition using TMS methodology.

There are some noted limitations with the current work. Carryover effects in the crossover design from one intervention to the other limits the extensibility of these. Future work should employ a more appropriate control condition such as an education-only arm with equivalent frequency of participant contact. As the participants in the current study do not have motor pathology, the extensibility of these findings to clinical populations may be limited. With regard

#### REFERENCES


to the TMS procedures, additional metrics such as cortical excitability would have been useful to report. We additionally acknowledge that the motor assessment battery was somewhat limited. Largely due to time considerations for testing, we could only administer a relatively small number of upper extremity tests in our sessions. In addition, most tests involved engagement of the dominant hand. Given the coin rotation findings, intermanual differences should be assessed with better granularity in future work. Finally, we did not track extramural activity and lifestyle habits in the current study. As such, we cannot account for variance from various unmeasured factors (i.e., overall daily activity, inflammatory biomarkers) in our statistical models, which should be tracked closely in future work.

In conclusion, we believe the current work is the first to show that a 12-week aerobic exercise intervention may affect the duration of the iSP duration in older, sedentary adults. In addition, change in silent period duration is correlated with improvements in motor dexterity. These findings are in concert with previous data collected from cross-sectional work involving middle-age and older adults of varying physical fitness levels (McGregor et al., 2011, 2013).

#### AUTHOR CONTRIBUTIONS

KMM, JN, and BC: conceptualized the experiment; KMM, KM, and JN: completed data collection and handled recruitment of participants. KMM, KM, JO, and PG: analyzed the data for the work. KMM, PG, JN, and BC wrote the manuscript.

#### ACKNOWLEDGMENTS

The views expressed in this work do not necessarily reflect those of the United States Government or Department of Veterans Affairs. All authors contributed significantly to the production of this work. This work was supported by VA grants: E0956-W, 5IK2RX000744, and C9246C. The authors would like to thank Paul Weiss, MS for statistical consultation.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 McGregor, Crosson, Mammino, Omar, García and Nocera. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Brain Network Modularity Predicts Exercise-Related Executive Function Gains in Older Adults

Pauline L. Baniqued1, 2 \*, Courtney L. Gallen<sup>1</sup> , Michelle W. Voss <sup>3</sup> , Agnieszka Z. Burzynska<sup>4</sup> , Chelsea N. Wong<sup>2</sup> , Gillian E. Cooke2, 5, Kristin Duffy <sup>2</sup> , Jason Fanning6, 7, Diane K. Ehlers <sup>6</sup> , Elizabeth A. Salerno<sup>6</sup> , Susan Aguiñaga<sup>6</sup> , Edward McAuley 2, 6, Arthur F. Kramer 2, 8 and Mark D'Esposito<sup>1</sup>

<sup>1</sup> Helen Wills Neuroscience Institute, University of California, Berkeley, Berkeley, CA, United States, <sup>2</sup> Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL, United States, <sup>3</sup> Department of Psychological and Brain Sciences, University of Iowa, Iowa City, IA, United States, <sup>4</sup> Department of Human Development and Family Studies, Colorado State University, Fort Collins, CO, United States, <sup>5</sup> Interdisciplinary Health Sciences Institute, University of Illinois at Urbana-Champaign, Urbana, IL, United States, <sup>6</sup> Department of Kinesiology and Community Health, University of Illinois at Urbana-Champaign, Urbana, IL, United States, <sup>7</sup> Department of Internal Medicine-Gerontology, Wake Forest School of Medicine, Winston-Salem, NC, United States, <sup>8</sup> Psychology Department and Mechanical and Industrial Engineering Department, Northeastern University, Boston, MA, United States

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

#### Reviewed by:

Keiichi Onoda, Shimane University, Japan Richard Betzel, Department of Bioengineering, University of Pennsylvania, United States Veena A. Nair, University of Wisconsin-Madison, United States

> \*Correspondence: Pauline L. Baniqued paulineb@berkeley.edu

Received: 25 July 2017 Accepted: 11 December 2017 Published: 04 January 2018

#### Citation:

Baniqued PL, Gallen CL, Voss MW, Burzynska AZ, Wong CN, Cooke GE, Duffy K, Fanning J, Ehlers DK, Salerno EA, Aguiñaga S, McAuley E, Kramer AF and D'Esposito M (2018) Brain Network Modularity Predicts Exercise-Related Executive Function Gains in Older Adults. Front. Aging Neurosci. 9:426. doi: 10.3389/fnagi.2017.00426 Recent work suggests that the brain can be conceptualized as a network comprised of groups of sub-networks or modules. The extent of segregation between modules can be quantified with a modularity metric, where networks with high modularity have dense connections within modules and sparser connections between modules. Previous work has shown that higher modularity predicts greater improvements after cognitive training in patients with traumatic brain injury and in healthy older and young adults. It is not known, however, whether modularity can also predict cognitive gains after a physical exercise intervention. Here, we quantified modularity in older adults (N = 128, mean age = 64.74) who underwent one of the following interventions for 6 months (NCT01472744 on ClinicalTrials.gov): (1) aerobic exercise in the form of brisk walking (Walk), (2) aerobic exercise in the form of brisk walking plus nutritional supplement (Walk+), (3) stretching, strengthening and stability (SSS), or (4) dance instruction. After the intervention, the Walk, Walk+ and SSS groups showed gains in cardiorespiratory fitness (CRF), with larger effects in both walking groups compared to the SSS and Dance groups. The Walk, Walk+ and SSS groups also improved in executive function (EF) as measured by reasoning, working memory, and task-switching tests. In the Walk, Walk+, and SSS groups that improved in EF, higher baseline modularity was positively related to EF gains, even after controlling for age, in-scanner motion and baseline EF. No relationship between modularity and EF gains was observed in the Dance group, which did not show training-related gains in CRF or EF control. These results are consistent with previous studies demonstrating that individuals with a more modular brain network organization are more responsive to cognitive training. These findings suggest that the predictive power of modularity may be generalizable across interventions aimed to enhance aspects of cognition and that, especially in low-performing individuals, global network properties can capture individual differences in neuroplasticity.

Keywords: executive function, cognitive control, functional connectivity, exercise, brain network modularity

## INTRODUCTION

Aging is accompanied by changes in cognition and brain function, yet there is individual variability in the extent to which older adults experience such effects (Wilson et al., 2002; Raz et al., 2005; Fabiani, 2012; Burzynska et al., 2015; Salthouse, 2016). Individual differences in age-related cognitive decline, particularly in executive function processes, are related to changes in structural and functional connectivity between brain regions (Andrews-Hanna et al., 2007; Damoiseaux et al., 2008; Kennedy and Raz, 2009; Madden et al., 2009, 2012). One method to quantify these complex interactions is to conceptualize the brain as a network comprised of sub-networks, or modules (Newman and Girvan, 2004; Newman, 2006b; Chen et al., 2008; Bullmore and Sporns, 2009; Meunier et al., 2010; Betzel et al., 2014; Bertolero et al., 2015). The extent of a module's segregation from the rest of the network can be quantified with a modularity metric (Newman and Girvan, 2004), where networks with high modularity have many connections within modules and fewer connections between modules. Computational models suggest that a modular network organization allows for a system that is more adaptable to new environments (Kashtan and Alon, 2005; Clune et al., 2013; Tosh and McNally, 2015), suggesting a role for network modularity in supporting complex behaviors like executive function. Compared to young adults, older adults have less modular brain networks (Chen et al., 2011; Onoda and Yamaguchi, 2013; Betzel et al., 2014; Geerligs et al., 2015) with pronounced age-related differences in sub-networks that support "associative" processes, such as executive function (Chan et al., 2014). Taken together, these findings suggest that more modular brain networks enable complex cognitive processes and neuroplasticity and, further, may provide insight into the mechanisms underlying the effectiveness of interventions geared toward ameliorating age-related cognitive decline.

Recent work has demonstrated that individual differences in brain network modularity can predict the extent to which individuals improve after cognitive interventions aimed to improve executive function. Specifically, higher baseline modularity (i.e., measured prior to the intervention) quantified during a task-free "resting state" predicted greater improvements after cognitive training in patients with traumatic brain injury (Arnemann et al., 2015) and more recently, in healthy older (Gallen et al., 2016) and young adults (Baniqued et al., 2015). Importantly, modularity predicted training gains even after controlling for baseline cognitive performance. These findings suggest that the informative nature of such individual differences in brain network organization can be used to maximize intervention effectiveness, such as by modifying training intensity or duration, especially in populations where behavioral measures may be difficult to collect (Gabrieli et al., 2015). Previous studies have examined other neural metrics in relation to learning and training responses (Erickson et al., 2010; Basak et al., 2011; Vo et al., 2011; Mathewson et al., 2012), but have often focused on specific brain regions related to specific types of interventions. As modularity has been shown to be reliable in individuals across sessions (Stevens et al., 2012; Cao et al., 2014) and predictive of cognitive gains across a variety of populations and training protocols, modularity may be a unifying biomarker that indexes an individual's potential for adaptive reorganization with intervention.

In addition to cognitive training interventions, cost-effective and easily accessible physical activity interventions involving brisk walking have been shown to have rehabilitative and protective effects on brain function in older adults (Kramer et al., 2006; Voss et al., 2013c). Further, there are significant individual differences in responsiveness to exercise training, with factors such as initial levels of heart rate and blood pressure determining gains in cardiorespiratory fitness (Bouchard and Rankinen, 2001). Although we have previously found that individual differences in brain network modularity can predict training-related gains after cognitive training (Arnemann et al., 2015; Baniqued et al., 2015; Gallen et al., 2016), it is not yet known whether the relationship between modularity and training gains is generalizable to interventions aimed to enhance executive function in older adults. Although there are several graph theoretical metrics, we were specifically interested if this relationship between pre-intervention brain modularity and training gains can also be found in a different, non-cognitive training intervention, such as a physical exercise intervention.

Specifically, we hypothesize that modularity reflects an individual's readiness to engage in and benefit from training. A recent study demonstrated that individuals with higher general intelligence show smaller connectivity changes between a resting state and task states, suggesting the existence of a more "optimal" network organization that provides more efficient reconfiguration during performance of various tasks (Schultz and Cole, 2016). Similar to this idea, we hypothesize that a more optimal—more modular network configuration is better able to transition to task states demanded by the interventions; it is more adaptable. In the context of the current study, a more modular brain network may potentiate the rehabilitative and protective effects of physical exercise on the aging brain, leading to greater improvements in executive function.

Here, we examined brain network modularity in older adults who underwent a 6-month exercise training intervention. Specifically, we tested the hypothesis that higher baseline modularity predicts larger exercise-related gains in cognition. The current study employed a broad battery of cognitive tests to assess intervention-related gains in executive function, episodic memory, vocabulary and perceptual speed. Here, we focused on the relationship between baseline modularity and improvements in executive function, as these processes show pronounced agerelated decline and exercise-related changes (Hillman et al., 2008; Voss et al., 2013c; Kawagoe et al., 2017).

#### MATERIALS AND METHODS

#### Participants

Healthy, low active, older adults (N = 247) aged 60–80 from the Urbana-Champaign community participated in a randomized controlled exercise trial (https://clinicaltrials.gov/ ct2/show/NCT01472744; see Voss et al., 2016; Burzynska et al., 2017; Ehlers et al., 2017a,b; Fanning et al., 2017, for data published from this same cohort). All participants provided

informed consent and the University of Illinois Institutional Review Board approved all procedures used in the study. Selection criteria consisted of the following, (1) >75% righthanded on the Edinburgh Handedness Questionnaire; (2) normal or corrected-to-normal vision of at least 20/40; (3) no colorblindness; (4) no history of stroke, transient ischemic attack, or head trauma; (5) >23 score on Mini-Mental State Examination (MMSE); (6) >21 score on Telephone Interview of Cognitive Status (TICS); (7) <10 score on Geriatric Depression Scale (GDS); (8) reported that they engaged in moderate intensity exercise for 30+ min no more than twice a week in the last 6 months and 9) screened for safe participation in an MRI environment (e.g., no claustrophobia or metallic implants). In all analyses presented here, we further excluded participants with MMSE scores less than 27 (N = 26), as a more stringent criterion is recommended in highly educated samples such as in the current study (O'Bryant et al., 2008). Summary demographics for included participants are provided in **Table 1**. Additional data were excluded on a case-by-case basis during data quality procedures applied to each behavioral measure. Specifically, cognitive measures greater than 3 SD from the mean were excluded. After this step, to reduce the influence of remaining extreme values, scores greater than 3 SD from the recomputed mean were winsorized (Tukey, 1962; Wilcox, 2005) to the appropriate cut-off value (3 SD below or above the mean). Analyses involving only fitness or behavioral scores were performed on the larger sample (N = 188), prior to exclusion due to MRI data quality, but effects were similar in the MRI sample (N = 128).

For the MRI data, we excluded one participant with incomplete resting state data, one participant with structural abnormalities (see section MRI Acquisition and Processing for more details), and 39 participants who reported taking medications known to influence the central nervous system. Thirty-five participants whose resting state scans contained more than 10% of volumes with movement greater than 0.50 framewise displacement (FD) or any volume with a maximum absolute displacement of 4.0 mm were excluded. MRI data were not collected for five subjects. Demographics for this reduced sample are provided in **Table 1**.

### Protocol Summary

All participants underwent MRI, behavioral, and fitness testing sessions before and after a 6-month long physical exercise intervention. Participants were paid for the pre- and post-testing sessions at a rate of \$10/h. Participants were randomly assigned to one of four intervention groups, which met for an hour three times a week. All group sessions were led by trained exercise specialists. In the walking group (Walk), participants were instructed to walk within their target heart rate (50–60% of their maximal heart rate for first 6 weeks, 60–75% for last 18 weeks). A second group was also instructed to walk within the same target heart rate and was provided with a daily milk-based supplement formula provided by Abbott Nutrition that contained betaalanine (Walk+). A third group was instructed in exercises focusing on stretching, strengthening and stability (SSS). A fourth group (Dance) was instructed in social dance sequences (i.e., Contra and English country dancing) by experienced dance instructors. Since the focus of this study is on the utility of brain modularity in predicting intervention-related gains, we limit our discussion of the intervention approach and choice of training regimen (for detailed information, see Ehlers et al., 2016; Burzynska et al., 2017).

#### Cardiorespiratory Fitness Testing

Participants underwent cardiorespiratory fitness (CRF) testing before and after the intervention. CRF reflects the integrated ability of the cardiovascular and respiratory systems to deliver oxygen during sustained physical effort (Ross et al., 2016), and regular physical exercise increases the efficiency of these systems (Wenger and Bell, 1986). CRF testing involves gradually increasing exercise intensity to tax the aerobic system and measuring the corresponding increase in oxygen consumption. Physician's approval was solicited prior to testing. CRF, operationally defined as peak oxygen consumption (VO2peak in mL/kg/min, relative rate in milliliters of oxygen per kilogram of body mass per minute), was measured with indirect calorimetry during a modified Balke graded maximal exercise test on a motordriven treadmill (Balke and Ware, 1959; Froelicher et al., 1975). Participants walked on a treadmill at a constant pace while the incline was increased 2–3% every 2 min. Expired air was sampled at 30-s intervals until maximal VO<sup>2</sup> was reached or the test was terminated due to volitional exhaustion and/or symptom limitation. Maximal VO<sup>2</sup> was determined after two of three criteria were met: (1) a plateau in VO<sup>2</sup> after increase in workload; (2) a respiratory exchange ratio (ratio of CO<sup>2</sup> production and O<sup>2</sup> consumption, reflecting limits of cardiovascular system) >1.10, and (3) a maximal heart rate within 10 bpm of their age-predicted maximum. VO2peak was the highest VO<sup>2</sup> recorded during the test. For the correlation analyses, we calculated a standardized CRF gain score for each individual by taking the difference between post-and pre-scores and dividing this by the standard deviation of pre-test scores (SD collapsed across groups).

#### Behavioral Testing

Participants underwent cognitive testing before and after the interventions. With the exception of the Switching Task and the Spatial Working Memory Task, all tests were taken from the Virginia Cognitive Aging Project (VCAP) (Salthouse and Ferrer-Caja, 2003; Salthouse, 2004, 2005, 2010). The VCAP tests were categorized into four categories: vocabulary, perceptual speed, episodic memory, and fluid reasoning. In the analyses, we grouped the Switching Task and Spatial Working Memory Task together with the fluid reasoning tasks to create an "executive function" component score, given previously demonstrated relationships between cognitive control and fluid reasoning abilities (Kane et al., 2005; Salthouse, 2005). We also performed a principal components analysis (PCA) on all the pre-test measures to confirm the VCAP construct groupings and to confirm that the Switching and Spatial Working Memory Tasks were related to performance on the fluid reasoning tests (**Table 2**, Supplementary Table 1). For each pre-test and post-test measure, we calculated standardized scores (z-scores) and averaged these z-scores according to the task groupings specified above, resulting in

## TABLE 1 | Demographics.


Mean (SD) and range for age, education, MMSE and VO2peak. \*Full sample excludes participants with MMSE scores lower than 27. Two participants are missing VO2peak data.

four component scores representing baseline cognitive abilities in vocabulary, perceptual speed, episodic memory and executive function (fluid reasoning plus switching and working memory). For each test, we also calculated standardized gain scores by subtracting pre-test performance from post-test performance, and dividing this value by the standard deviation of raw pre-test scores (collapsed across groups). We averaged the standardized gain scores accordingly to create composite gain scores in vocabulary, perceptual speed, episodic memory, and executive function. The following sections have brief descriptions of each test and the specific measure used for analyses.

#### Task-Switching (Kramer et al., 1999; Voss et al.,

2010a,b, 2013b; Leckie et al., 2014)

On each trial, participants were shown a number between 1 and 9 (except 5) against a colored background: (1) on a pink background, participants were instructed to determine whether the number was odd or even, (2) on a blue background, they were to determine if the number was higher or lower than 5. Participants completed a high/low practice block (40 trials) an odd/even practice block (40 trials), a single high/low task block (40 trials), a single odd/even task block (40 trials), a mixed practice block (64 trials) and a mixed task block (160 trials). We analyzed performance on the mixed task block and extracted (1) local switch cost (mixed switch reaction time; RT—mixed non-switch RT) and (2) task switching bin score (combination of accuracy and RT measures) (Draheim et al., 2016). The task switching bin score was used in the principal components and correlation analyses to better TABLE 2 | PCA standardized loadings (pattern matrix) based upon correlation matrix of baseline scores.


Performed varimax rotation and extraction of 4 components, which accounted for 67% of total variance. Italics denote component groupings.

examine the relationship between task switching performance and performance on other tests (Draheim et al., 2016). Local RT switch cost was used in the analyses of intervention effects, consistent with previous studies (Voss et al., 2010a, 2013b). The two measures were correlated (Supplementary Table 1; baseline measures: r(211) = 0.322, p < 0.001, two-tailed; standardized gain scores: r(159) = 0.267, p < 0.001, two-tailed), and the intervention effects were similar when using bin score instead of local RT switch cost.

#### Spatial Working Memory (Erickson et al., 2011)

On each trial, an arrangement of two, three, or four black dots was briefly presented on the screen. After a delay, a red dot appeared and participants were instructed to determine if the red dot matched the position of one of the black dots presented earlier in that trial (match or non-match). Participants performed a practice block of 12 trials, and a task block of 120 trials (40 trials per condition). We analyzed mean accuracy during the task block for the more difficult three-dot and four-dot trial conditions.

#### Shipley Abstraction (Zachary, 1986)

Participants were given a list of word, letter, or number sequences on a piece of paper and were instructed to write the missing item/s (word, letter or number) in each sequence. Participants were given 5 min to answer 20 items. We analyzed the total number of correctly answered items.

#### Matrix Reasoning (Ravens, 1962)

On each trial, participants were shown a 3 × 3 grid, with each cell except for one containing an abstract pattern. Participants were instructed to select which among eight options best completes the matrix along both the rows and columns. Participants performed two practice trials and were then given 10 min to complete a maximum of 18 items. We analyzed the total number of correctly answered items.

#### Paper Folding (Ekstrom et al., 1976)

On each trial, participants were presented with images that show a sheet of paper folded in a certain sequence and a hole punched through the folded sheet. Participants were asked to select which among five options matched the pattern of holes that would result when the paper was unfolded. They were given 10 min to complete a maximum of 12 trials. We analyzed the total number of correctly answered items.

#### Spatial Relations (Bennett et al., 1997)

On each trial, participants were presented with a 2-dimensional object pattern and instructed to identify which among four threedimensional figures would match the 2-dimensional pattern when folded. Participants were given 10 min to complete a maximum of 20 trials. We analyzed the total number of correctly answered items.

#### Form Boards (Ekstrom et al., 1976)

On each trial, participants were presented with a specific shape and instructed to choose which pieces (five total options) will exactly fill the space inside the shape. They were given 8 min to complete a maximum of 24 trials. We analyzed the total number of correctly answered items.

#### Letter Sets (Ekstrom et al., 1976)

On each trial, participants were presented with five sets of fourletter strings and asked to determine which set was different from the other four. Participants were given 10 min to complete a maximum of 15 trials. We analyzed the total number of correctly answered items.

#### Digit-Symbol Coding (Wechsler, 1997a)

Participants were presented with a sheet of paper containing a series of numbers between 1 and 9, were asked to fill in the corresponding symbol based on a digit-symbol key provided. Participants completed 7 practice items and were given 2 min to complete a maximum of 133 items. We analyzed the number of correctly answered items.

#### Pattern Comparison (Salthouse and Babcock, 1991)

Participants were given a sheet of paper with a set of line patterns and were tasked to determine whether a pair of line patterns was the same or different. Participants completed three practice items, followed by two task sets, each set with a maximum number of 30 items to be completed within 30 s. We analyzed the number of correctly answered items, averaged across two sets of problems.

#### Letter Comparison (Salthouse and Babcock, 1991)

Participants were given a sheet of paper with a set of non-word letter strings and were tasked to determine whether a pair of letter strings was the same or different. Participants completed three practice items, followed by two task sets, each set with a maximum number of 30 items to be completed within 30 s. We analyzed the number of correctly answered items, averaged across two sets of problems.

#### Logical Memory (Wechsler, 1997b)

Participants listened to stories narrated by an experimenter and after each reading, were asked to recall each story in detail. We analyzed the number of correctly recalled story details, summed across three story-tellings (first story, second story, re-reading of second story).

#### Paired Associates (Salthouse et al., 1996)

Participants listened to a list of six word pairs read aloud by an experimenter. The experimenter then read the first word of each pair and asked participants to recall the paired second word. We analyzed the number of correctly recalled items, averaged across two sets of six pairs each.

#### Word Recall (Wechsler, 1997b)

Participants listened to a list of words and were given 90 s to recall the words in any order. Participants listed to the same list three more times and were asked to recall as many words as possible after each reading. Participants were then read a new list of words, asked to recall as many words as possible from the new list, and then asked to recall words from the old list. We analyzed the total number of correctly recalled items.

#### Word Vocabulary (Wechsler, 1997a)

Experimenters read aloud a list of 33 words and asked participants to verbally give the meaning of each word. Responses are scored 0–2 points according to the quality of the definition (based on provided word and phrase guidelines). The test is discontinued after six consecutive scores of 0. We analyzed the total number of points.

Picture Vocabulary (Woodcock and Johnson, 1989)

Experimenters present a maximum of 30 images and participants are tasked to name the objects presented. The test is discontinued after a participant fails to name six consecutive items. We analyzed the total number of correctly named items.

#### Synonym-Antonym (Salthouse, 1993)

On each trial, participants are presented a target word and are tasked to select which among five word options is most similar (synonym) or opposite (antonym) in meaning to the target word. Participants completed a synonym block followed by an antonym block, each with a maximum of 10 items to be completed within 5 min. We analyzed the total number of correctly identified words across the synonym and antonym blocks.

#### MRI Acquisition and Processing

Participants underwent MRI scanning on a 3 Tesla Siemens Trio Tim System with a 12-channel head coil before and after the intervention; however, only the pre-intervention scans were analyzed in this study given our hypotheses regarding correlations between baseline brain modularity and cognitive gains. The anatomical scan consisted of T1-weighted MPRAGE images acquired with the following parameters: GRAPPA acceleration factor 2, voxel size = 0.9 × 0.9 × 0.9 mm, TR = 1,900 ms, TI = 900 ms, TE = 2.32 ms, flip angle = 9 ◦ , FoV = 230 mm. To analyze network properties during a taskfree "resting state," a 6-min functional scan was obtained using a T2<sup>∗</sup> -weighted echoplanar imaging (EPI) pulse sequence with the following parameters: GRAPPA acceleration factor 2, 180 volumes, in-plane resolution = 3.4 mm<sup>2</sup> , TR = 2,000 ms, TE = 25 ms, flip angle = 80◦ , 35 4 mm ascending slices, no slice gap. Participants were instructed to lie still with their eyes closed.

Brain extraction from anatomical scans was performed with Advanced Normalization Tools (ANTs; Avants et al., 2010, 2011) using the Kirby/MMRR template (Landman et al., 2011). When this skull-stripping procedure failed, brain extraction was instead performed using the IXI template (Heckemann et al., 2003; Ericsson et al., 2008). The skull-stripped anatomical images and raw functional images were preprocessed through the Configurable Pipeline for Connectomes (CPAC; Giavasis et al., 2015). Anatomical images were registered to the MNI152 template (Fonov et al., 2009) using ANTs and segmented into gray matter (probability threshold = 0.7), white matter (probability threshold = 0.98) and cerebrospinal fluid (CSF; probability threshold = 0.98) using FSL/FAST (Zhang et al., 2001). Functional images were slice-time corrected, motioncorrected (Friston et al., 1996) and co-registered to the anatomical images. Nuisance signal removal was performed by regressing out the aforementioned motion parameters, signals from the first five components from white matter and CSF voxels (Compcor; Behzadi et al., 2007; Muschelli et al., 2014), and linear and quadratic trends. Signals were bandpass filtered at 0.009– 0.08 Hz. Participants whose resting state scan contained (1) more than 10% of volumes with framewise displacement (FD) greater than 0.5 mm (N = 23) or (2) maximum absolute displacement greater than 4.0 mm were excluded from subsequent analyses (additional N = 12). One participant was excluded because structural abnormalities caused anatomical-to-MNI registration to fail (spatial warping) during preprocessing, such that we could not reliably extract ROIs.

### Functional Connectivity and Modularity Analyses

Functional scans were warped to the MNI template and parcellated into 264 regions of interest (Power et al., 2011). Due to uneven partial coverage of the cerebellum across subjects in the functional data, we excluded the four cerebellum module ROIs prior to analysis. Eight additional ROIs were excluded due to lack of functional coverage in at least one participant, leaving a total of 252 ROIs. For each individual, time series from all voxels within each ROI were averaged together. Average ROI time series were correlated between each pair of ROIs (Pearson's coefficient), and the resulting ROI-to-ROI correlation matrices were Fisher z-transformed. Matrices were binarized over a range of connection density thresholds (costs): 2–10% of all possible connections, in 2% increments, following (Power et al., 2011; Power and Petersen, 2013). These thresholded matrices were used to create unweighted, undirected whole-brain graphs for each participant, from which network metrics were derived using the BrainX (https://github.com/nipy/brainx) and NetworkX Python package (Hagberg et al., 2008). Network modularity was quantified separately for each connection threshold to examine the consistency of results across thresholds. We use the middle 6% threshold for all our primary analyses, but verified effects at the other thresholds (Supplementary Material).

For our primary analysis, we quantified modularity, a network measure that compares the number of connections within modules to the number of connections across modules (Newman and Girvan, 2004). Modularity is defined as <sup>P</sup><sup>m</sup> i = 1 (eii − a 2 i ), where eii is the fraction of connections that connect two nodes within module i, a<sup>i</sup> is the fraction of connections connecting a node in module i to any other node, and m is the total number of modules in the network (Newman and Girvan, 2004). There are multiple methods for identifying network modules. Here, we used a spectral algorithm (Newman, 2006a) to identify the partition that maximizes modularity for each participant at each threshold.

Further, to confirm that our effects were not driven by a specific partitioning algorithm, we also computed modularity using partitions identified in Power et al. (2011) using the Infomap algorithm (Rosvall and Bergstrom, 2008; Fortunato, 2010). Here, every node was assigned to one of thirteen modules (as identified in Power et al., 2011): default mode (DMN), frontoparietal (FP), cingulo-opercular (CO), salience (Sal), dorsal attention (DAN), ventral attention (VAN), auditory (Aud), visual (Vis), memory (Mem), sensory/somatomotor hand (SM-hand), sensory/somatomotor mouth (SM-mouth), subcortical (Subcort)

and a module containing unassigned nodes. The modularity values derived from the Power partition were highly correlated with the modularity values obtained using the spectral clustering partition (all r > 0.761, all p > 0.001, two-tailed for all five cost thresholds).

#### Potential Confounds

Before examining the relationship between brain modularity and intervention-related gains, we examined relationships between potential confounding variables and our measures of interest (i.e., baseline modularity and intervention-related gains), including age, in-scanner motion (i.e., frame-wise displacement or FD; Power et al., 2012; Satterthwaite et al., 2012; Siegel et al., 2016), and baseline cognitive performance. All the analyses include only subjects with usable baseline MRI scans, baseline EF scores, and EF gain scores (N = 128). If a significant relationship between potential confounding variables and our dependent measures was found, we then used these variables as covariates in our primary analyses examining correlations between modularity and intervention-related gains. For all analyses, we also controlled for age and in-scanner motion (i.e., FD). For all correlation analyses, we computed bias-corrected and accelerated (BCa) confidence intervals (CI) using 5,000 bootstrapped samples.

There is considerable variability in brain volume in older adults (Salat et al., 2004; Raz et al., 2005; Raz and Rodrigue, 2006). Thus, for participants with structural volume data, we also tested whether the pattern of brain-behavior relationships from the network analyses could have been confounded by gross individual differences in brain structure. We extracted measures of brain volume using Freesurfer v5.3 (Dale et al., 1999); http:// surfer.nmr.mgh.harvard.edu), which performs segmentation of cortical and subcortical matter using automated and probabilistic algorithms (Fischl et al., 2002, 2004a,b; Desikan et al., 2006). AZB inspected the segmentation output and performed appropriate corrections. Using the anatomical scans obtained at baseline, we obtained measures of total intracranial volume, white matter, and total gray matter volume, described in more detail on the Freesurfer website (https://surfer.nmr.mgh.harvard.edu/fswiki/ MorphometryStats). We included estimated intracranial volume as a covariate in volumetric analyses to control for differences in overall brain volume (Jack et al., 1989; Buckner et al., 2004). Since not all participants had high-quality structural scans for volumetric analysis (N = 15), we conducted this analysis as a follow-up to the primary analyses of modularity vs. interventionrelated gains.

#### RESULTS

#### Exercise-Related Changes in Cardiorespiratory Fitness (CRF)

We first verified that the groups demonstrated the expected patterns of fitness improvements. At baseline, the groups did not differ in CRF F(3, 182) = 0.199, p = 0.897, η 2 <sup>p</sup> = 0.003. A mixed ANOVA with VO2peak scores over time (pre- and post-testing) as a within-subjects factor and group as a between-subjects factor revealed a main effect of time F(1, 182) = 21.737, p < 0.001,

extend to ±1.58 IQR/sqrt(n). The upper and lower hinges correspond to the first and third quartiles. The whiskers extend from the hinge to ±1.5\*IQR of the hinge. IQR, inter-quartile range.

η 2 <sup>p</sup> = 0.107, and an interaction of group and time F(3, 182) = 2.792, p = 0.042, η 2 <sup>p</sup> = 0.044. Follow-up analyses showed that the Walk and Walk+ groups showed greater improvements in CRF compared to the SSS and Dance groups (**Figure 1**). Separate comparisons of pre- and post-test scores within each group showed significant gains in the Walk and Walk+ groups (both p = 0.001, both d ≥ 0.539) and marginal gains in the SSS group (p = 0.059, d = 0.225). There were no significant gains in the Dance group (p = 0.345, d = 0.062).

#### Exercise-Related Changes in Cognitive Function

To determine the effects of the exercise intervention on cognitive function and to minimize measurement error and multiple comparison issues in analyzing each test separately, we analyzed cognitive effects at the construct level using composite scores. The creation of composite scores was guided by previous literature (Kane et al., 2005; Salthouse, 2005), correlations (Supplementary Table 1), and a PCA on the baseline test scores (**Table 2**), which confirmed the grouping of the cognitive tests into categories of vocabulary, episodic memory, perceptual speed, and executive function.

At baseline, the groups did not differ in EF, F(3, 187) = 1.191, p = 0.315, η 2 <sup>p</sup> = 0.019, perceptual speed F(3, 185) = 0.525, p = 0.665, η 2 <sup>p</sup> = 0.008, episodic memory, F(3, 187) = 0.098, p = 0.961, η 2 <sup>p</sup> = 0.002, and vocabulary, F(3, 184) = 0.619, p = 0.604, η 2 <sup>p</sup> = 0.010. With the exception of a correlation between vocabulary gain and perceptual speed gain

[r(186) = 0.152, 95% CI [−0.002, 0.304], p = 0.039, two-tailed], and between vocabulary gain and EF gain [r(188) = 0.147, 95% CI [0.011, 0.275], p = 0.043, two-tailed], the composite scores of cognitive change scores (i.e., gains) were not significantly correlated with each other (all others p > 0.05, two-tailed). Therefore, we conducted separate ANOVAs on the composite gain scores to examine the differential effects of the intervention on cognitive function (Huberty and Morris, 1989).

An ANOVA on the EF gain score with intervention group as a between-subjects factor yielded a significant group effect F(3,187) = 3.899, p = 0.010, η 2 <sup>p</sup> = 0.059. The Walk, Walk+ and SSS groups showed greater EF gains compared to the Dance group (all p < 0.05 for each group vs. Dance, Dance vs. Walk+ is p < 0.10 with Bonferroni correction; **Figure 2**). Tests for significance of gain scores (i.e., test against a comparison value of zero) separately in each group showed significant EF gains in the Walk, Walk+ and SSS groups (all p < 0.001), but not in the Dance group (p = 0.703). Since baseline EF was moderately correlated with EF gain, r(189) = −0.140, 95% CI [−0.255, −0.026] p = 0.053, two-tailed, we verified that the group effect in EF gain remained significant after controlling for baseline EF, F(3, 186) = 3.498, p = 0.017, η 2 <sup>p</sup> = 0.053. No group effects were observed for gains in perceptual speed, F(3, 185) = 0.129, p = 0.943, η 2 <sup>p</sup> = 0.002, episodic memory, F(3, 187) = 0.086, p = 0.968, η 2 <sup>p</sup> = 0.001, and vocabulary, F(3, 184) = 1.376, p = 0.251, η 2 <sup>p</sup> = 0.022 (**Figure 2**).

#### Relationship between Fitness and Cognitive Effects

Given that the groups that improved in EF were also those that showed larger CRF gains (i.e., Walk, Walk+, SSS groups), we tested whether the degree of CRF improvement was related to EF improvement. Across the whole sample with CRF data and behavioral data, there was no significant relationship between CRF gain and EF gain, r(184) = 0.104, 95% CI [−0.045, 0.251], p = 0.157, two-tailed, even when excluding the Dance group which did not show CRF and EF gains, r(138) = 0.080, 95% CI [−0.045, 0.251], p = 0.352, two-tailed. The correlations were also not significant within each group (all |r| <0.25, all p > 0.05, two-tailed).

#### Examination of Potential Confounds

Across the whole sample with quality MRI data, we first examined relationships between group assignment (i.e., to confirm that groups did not differ in baseline characteristics), potential confounding variables (i.e., age, years of education, mean FD) and our measures of interest (i.e., baseline modularity and EF gain). In the case of a non-significant relationship between variables when analyzing the whole MRI sample, we also verified that the relationship was not significant when analyzing each group separately, as the primary analyses of baseline modularity and EF gain were conducted within group.

Age did not differ across groups (**Table 1**), but was significantly correlated with baseline modularity, r(126) = 0.239, 95% CI [0.102, 0.370], p = 0.007, two-tailed, and was not correlated with EF gain, r(126) = −0.008, 95% CI [−0.211, 0.197], p = 0.932, two-tailed. We verified that there was no significant relationship between age and EF gain within each group (all |r| <0.314, all p > 0.097, two-tailed).

Years of education did not significantly differ across groups (**Table 1**), even after accounting for age F(3,123) = 2.117, p = 0.101, η 2 <sup>p</sup> = 0.049. Education was not significantly correlated with baseline modularity, r(126) = −0.031, 95% CI [−0.183, 0.124], p = 0.730, or EF gain, r(126) = −0.041, 95% CI [−0.216, 0.137], p = 0.649, even after accounting for age, and when examining within each group separately (all |r| < 0.295, all p > 0.101, two-tailed).

Mean FD did not differ across groups, F(3, 124) = 0.938, p = 0.425, η 2 <sup>p</sup> = 0.022, even after controlling for age, F(3, 123) = 0.935, p = 0.426, η 2 <sup>p</sup> = 0.022. Mean FD was not correlated with baseline modularity, r(126) = −0.087, 95% CI [−0.272, 0.092], p = 0.328, two-tailed or with EF gain, r(126) = 0.039, 95% CI [−0.138, 0.197], p = 0.666, two-tailed, even after controlling for age (all |r| <0.122, all p > 0.173, twotailed). When inspecting these relationships within each group however, we found a trending relationship between mean FD and modularity in the Walk group, r(27) = −0.359, 95% CI [−0.666, 0.005], p = 0.056, two-tailed.

Baseline modularity differed across groups, F(3, 124) = 4.628, p = 0.004, η 2 <sup>p</sup> = 0.101, even after accounting for age, F(3, 123) = 4.495, p = 0.005, η 2 <sup>p</sup> = 0.099, with the Walk+ group showing significantly lower baseline modularity compared to the SSS (p = 0.005) and Dance (p = 0.011) groups, but not compared to the Walk group (p = 0.116).

Lastly, given previously documented relationships between modularity and cognitive function (Kitzbichler et al., 2011; Stevens et al., 2012; Stanley et al., 2014; Sadaghiani et al., 2015), we examined whether baseline modularity was related to baseline EF. Across the whole MRI sample, there was no significant relationship between baseline modularity and baseline EF, r(126) = 0.023, 95% CI [−0.167, 0.207], p = 0.798, two-tailed, even after accounting for age and/or mean FD and examining each group separately (all |r| < 0.319, all p > 0.098, two-tailed). Thus, potential relationships between modularity and EF gains cannot be attributed to correlations between modularity and EF performance at baseline.

The above results were similar when using modularity values derived from other thresholds and when using modularity values derived from the Power partition (see Supplementary Material). Thus, given these findings that age and mean FD showed some relationship with modularity, and given that baseline EF was moderately related to EF gain, we used age, mean FD and baseline EF as covariates in the primary analyses of modularity and exercise-related gains.

#### Relationship between Baseline Modularity and Exercise-Related Gains

We next examined the relationship between baseline modularity and intervention-related effects on EF, having confirmed EF and CRF improvements in the Walk, Walk+ and SSS groups (**Figure 3**). For each group, we first performed linear regression analyses with EF gain as the dependent variable, age, mean FD and baseline EF as covariates, and independent variables of baseline EF, baseline modularity, and an interaction term of baseline EF and baseline modularity. Importantly, the interaction term was included to test whether the relationship between baseline modularity and EF gain was moderated by baseline EF (i.e., whether the modularity-gain relationship was stronger in high or low performing individuals at baseline).

The model (**Table 3**) with all three terms and covariates was significant in the Walk [R <sup>2</sup> = 0.450, Adjusted R <sup>2</sup> = 0.330, F(5, 23) = 3.756, p = 0.012] and Walk+ groups [R <sup>2</sup> = 0.375, Adjusted R <sup>2</sup> = 0.239, F(5, 23) = 2.762, p = 0.043].

In the Walk group, age, mean FD, modularity, and the interaction term of modularity and baseline EF were significant predictors of EF gain (**Table 3**). Critically, modularity positively predicted EF gain, while the interaction showed that individuals with lower baseline EF showed a stronger relationship between modularity and EF gain.

In the Walk+ group, age and the interaction term of modularity and baseline EF were significant predictors of EF gain, with baseline EF as a marginal predictor (**Table 3**). Similar to the Walk group, individuals with lower baseline EF showed a stronger relationship between modularity and EF gain.

In the SSS group, the full model was not significant [**Table 3**; R <sup>2</sup> = 0.094, Adjusted R <sup>2</sup> = −0.048, F(5, 32) = 0.661, p = 0.656]. Modularity was not a significant predictor, although it explained the most variance and was related to EF gain in a similar positive direction. Given that there were no significant predictors in the full model, we performed a reduced model with only baseline modularity. This model was marginally significant [R <sup>2</sup> = 0.083, Adjusted R <sup>2</sup> = 0.057, F(1, 36) = 3.246, p = 0.080], with modularity marginally related to EF gain (B = 1.218, p = 0.080, BCa 95% CI [−0.439, 2.276]).

As expected, in the Dance group, the full model [**Table 3**; R <sup>2</sup> = 0.217, Adjusted R <sup>2</sup> = 0.066, F(5, 26) = 1.437, p = 0.244], and a reduced model with only baseline modularity [R <sup>2</sup> = 0.002, Adjusted R <sup>2</sup> = −0.032, F(1, 30) = 0.045, p = 0.833] were not significant, with no factor emerging as a significant predictor.

In summary, we find that baseline modularity was related to EF gains in groups that showed training-related gains. For illustrative purposes, **Figure 3** shows the relationship between baseline modularity and EF gain with and without controlling for age, mean FD and baseline EF.

### Controlling for Individual Differences in Brain Volume

Age-related differences in white and gray matter volume loss may influence brain function (Persson et al., 2006; Chadick et al., 2014; Pudas et al., 2017), functional connectivity patterns (Meunier et al., 2014), and in turn, the pattern of brain-behavioral results we find here. On the sample of participants with highquality anatomical data, we ran partial correlation analyses of baseline modularity and EF gain within each of the four groups (one-tailed tests to confirm initial results), controlling for estimated intra-cranial volume, gray matter volume, and white matter volume in addition to age, mean FD and baseline EF. Critically, the pattern of relationships remained the same, Walk: rp(16) = 0.369, 95% CI [−0.339, 0.884], p = 0.066; Walk+: rp(20) = 0.098, 95% CI [−0.519, 0.628], p = 0.331; SSS: rp(25) = 0.408, 95% CI [−0.024, 0.709], p = 0.017, Dance: rp(20) = −0.017, 95% CI [−0.496, 0.635], p = 0.469, suggesting that individual differences in brain volume did not contribute to the relationship between baseline modularity and EF gain.

### Exploratory Analyses: Sub-network Contribution to Relationship between Baseline Modularity and Training-Related Gains

Brain modules show distinct age-related connectivity changes (Chan et al., 2014), and modularity in the association systems



\*\*\*p < 0.001, \*\*p < 0.01, \*p < 0.05, ∼p < 0.10.

(DMN, FP, CO, Sal, DAN, VAN) has been found to drive the correlation between global modularity and training-related gains (Gallen et al., 2016). Given this, we examined whether specific networks contribute to the modularity vs. gain relationship. Similar to previous findings, sensory-motor modularity was higher than association cortex modularity both when analyzing the whole sample, t(127) = 24.954, p < 0.001, and each group separately (all p < 0.001). We then examined the contribution of each sub-network to the modularity vs. EF gain relationship. For these analyses, we performed partial correlation analyses with age, mean FD and baseline EF as covariates. To reduce the number of analyses, we combined the three groups (Walk, Walk+ and SSS) given the similarity in their intervention-related effects.

Across the three groups, EF gain was marginally correlated with baseline association sub-network modularity r(91) = 0.159, 95% CI [−0.053, 0.346], p = 0.064, one-tailed, but not sensorymotor cortex modularity r(91) = 0.003, 95% CI [−0.188, 0.209], p = 0.488, one-tailed. Given the trending relationship between EF gain and association modularity and previous findings (Gallen et al., 2016), we examined the relationship between EF gain and each association sub-network. After Bonferroni correction however, none of the six modules showed a significant relationship with EF Gain (Supplementary Material).

We also quantified module segregation (Chan et al., 2014), defined as (Zw-Zb)/Zw, where Z<sup>w</sup> is the average Fishertransformed correlation between nodes in the same module (within-module connectivity) and Z<sup>b</sup> is the average Fishertransformed correlation between nodes in a module to nodes in any other module (between-module connectivity). Importantly, this metric retains the weights of all connections (lower than 2– 10% of connections). Given previous findings, we focused our analyses on the association cortex modules. When controlling for age, mean FD, and baseline EF, whole-brain segregation and association module segregation were not significantly related to EF gain, although the results were in the same direction as the modularity results (Supplementary Material).

### DISCUSSION

We examined whether baseline brain network modularity predicts cognitive improvements in older adults after an exercise intervention. We found that in the groups that showed gains in fitness and cognitive function (Walk, Walk+, and SSS), higher baseline brain modularity predicted greater gains in executive function, even after accounting for individual differences in baseline performance, age, in-scanner motion, and individual differences in brain volume. These results parallel findings in TBI patients (Arnemann et al., 2015), older adults (Gallen et al., 2016), and young adults who underwent cognitive training (Baniqued et al., 2015). Given that we find a similar relationship between modularity and cognitive gains after an exercise intervention in older adults suggests that the predictive power of brain modularity may be generalizable across populations and interventions aimed to enhance executive function. Moreover, these findings point to the potential of global network properties to capture individual differences in neuroplasticity.

#### Modularity and Exercise-Related Gains in Executive Function

Our findings demonstrating a relationship between baseline brain network modularity and EF improvements with exercise training add to a series of studies that find a similar relationship with cognitive gains from cognitive training interventions (Arnemann et al., 2015; Baniqued et al., 2015; Gallen et al., 2016). Importantly, the current study shows that the pattern of results holds after controlling for factors such as baseline cognitive performance, age, and individual differences in brain volume—the latter of which can present a confound, especially when analyzing measures of brain function in older adults, who show considerable variability in age-related atrophy and lesions (Hedden et al., 2012; Grady, 2013). In the current study, the modularity-gain correlations were found in two (Walk, SSS) out of the three groups that showed some improvement in CRF and EF. In the Walk and Walk+ groups, the modularity-gain relationship was moderated by baseline EF, which together with previous findings in older adults (Arnemann et al., 2015; Gallen et al., 2016) underscores the utility of the network modularity measure in lower-performing individuals. These results suggest that the two measures of baseline performance and modularity together may be a better predictor of training-related gains than either alone.

The relevance of the modularity metric in neuroplasticity, specifically, in predicting response to an intervention, can be linked back to computational models showing that modular networks more rapidly reconfigure in response to new environments (Kashtan and Alon, 2005; Clune et al., 2013; Tosh and McNally, 2015), such that reorganization is more efficiently achieved by slight modifications within and between relatively specialized modules than by a large-scale overhaul of a highly interdependent network. Moreover, individuals with disrupted modular brain organization (Fornito et al., 2015), such as those with focal lesions to brain regions important for between-module connectivity (Nomura et al., 2010; Gratton et al., 2012; Warren et al., 2014) show widespread cognitive dysfunction and thus underscore the role of a modular structure in enabling brain processes that support a wide range of behaviors. In a recent study, individuals who scored higher on general intelligence tests tended to show smaller functional connectivity changes between a "resting state" and task performance states (Schultz and Cole, 2016), suggesting that they adapt more efficiently to task demands. In this sense, the architecture of brain networks at rest guides the connectivity patterns that emerge during the performance of various tasks. Indeed, modularity measured during "resting states" has been found to predict working memory performance (Stevens et al., 2012), and stimulus detection in a perceptual task (Sadaghiani et al., 2015). Taken together, these findings suggest that an "optimally" organized network requires less reorganization to be receptive to new input encountered during learning or training, or to capitalize from intervention-related changes in brain function. In the context of the current study, a more modular brain network may potentiate the rehabilitative and protective effects of physical exercise on the aging brain. In fitness interventions, for example, exercise-associated up-regulation in neurotrophic factors has been related to greater exercise-related changes in brain connectivity (Voss et al., 2013a). Given previous findings and the results of the current study, an optimal network for intervention-related cognitive gains is modularly organized at rest, with a balance of within-module connections that support local processing and across-module connections that support global processing (Meunier et al., 2009, 2010). Indeed, recent studies have shown that increased brain modularity post-therapy correlated with greater speech improvement in aphasic patients (Duncan and Small, 2016), and that greater structural modularity prior to carotid artery intervention predicted reduced risk of cognitive decline after carotid intervention (Soman et al., 2016). Additionally, connectivity measures obtained during preclinical stages, when combined with biomarkers such as amyloid-beta, have been shown to predict later cognitive decline (Buckley et al., 2017), suggesting that these metrics have the potential to provide actionable information when clinical symptoms have yet to manifest.

We found that modularity predicted training gains, beyond the baseline behavioral EF measure. This is a promising finding given that behavioral or cognitive measures may be confounded in certain populations (Gabrieli et al., 2015), such as in older adults, where factors such as mobility or visual acuity interact with task performance. While typical behavioral measures may not reliably distinguish between individual differences in cognitive ability, brain network measures provide a way to gauge training responsiveness. Although this study involved a fairly large sample, functional connectivity was assessed during a relatively short resting-state scan. More reliable measures and more information regarding network structure, particularly in higher performing individuals, may be gleaned from a longer scan period (Birn et al., 2013; Laumann et al., 2015; Gordon et al., 2017). Nonetheless, the pattern of higher baseline modularity predicting intervention-related cognitive gains is now consistent across four studies (Arnemann et al., 2015; Baniqued et al., 2015; Gallen et al., 2016). Using brain network measures in combination with behavioral, demographic, lifestyle, and other brain measures could also help customize intervention protocols to maximize effectiveness, especially in the context of doseresponse relationships, for example by increasing the intensity, frequency, or duration of exercises, or including pre-intervention lifestyle or behavioral protocols geared to promote or maintain optimal levels of brain modularity. Nonetheless, future work may identify behavioral measures that are sensitive to the information captured by network measures; the relationship between baseline modularity and future behavior (i.e., training gains) suggests that modularity may be reflected in baseline behavior to some extent, a brain characteristic that the current study's behavioral measures are not designed to capture. In addition, more work is needed to examine the mechanisms in which a modular architecture interacts with changes in neural and vascular function to enable benefits from cognitive and fitness interventions, and whether such interventions lead to changes in brain modularity. In the current study, we found a marginal correlation between EF gain and baseline association cortex modularity, which suggests that association sub-networks drive the relationship between baseline modularity and EF gain, similar to our previous study (Gallen et al., 2016). Relatedly, association sub-networks have also been shown to increase in functional connectivity after a physical exercise intervention (Voss et al., 2010b), concomitant with improvements in EF.

In our dataset, we found a positive correlation between age and baseline modularity, unlike previous studies that found lower modularity in older adults compared to young adults (Meunier et al., 2009; Betzel et al., 2014; Song et al., 2014; Geerligs et al., 2015). Importantly, our study only included older adults, whereas reductions in modularity are typically found when comparing older and young populations. In addition, some studies show no correlation between modularity and age within older adults (Geerligs et al., 2015; Gallen et al., 2016), no difference in modularity per se when comparing young and old adults (Meunier et al., 2009) and observations that modularity variability was higher in older adults (Song et al., 2014). Moreover, our older adult sample may not be representative of the general population, as participants were relatively healthy and free of major health incidents despite being generally inactive or sedentary prior to participating in the study. Notably however, the relationship between baseline modularity and training gain in the current study remained even after accounting for age.

Neurovascular coupling is an important issue to consider when conducting fMRI studies in older adults, where age-related vascular changes may lead to age-related BOLD differences in the absence of "true" neural differences (D'Esposito et al., 2003; Samanez-Larkin and D'Esposito, 2008). The current study however, does not compare heterogeneous groups (e.g., young vs. old, low-fit vs. high-fit)—all participants were low-fit but relatively healthy older adults, and all analyses controlled for age. Moreover, across the whole sample, baseline VO<sup>2</sup> and baseline modularity were not significantly correlated (all |r| <0.067, all p > 0.457, two-tailed), even after controlling for mean FD. In addition, controlling for baseline VO<sup>2</sup> in the modularity vs. training gain analyses does not change the results. Future studies can include taking into account indicators of cerebrovascular health such as cerebral blood flow (Brown et al., 2010; Zimmerman et al., 2014) to determine whether and/or to what extent it relates to connectivity measures. In the current study, we controlled for measures such as age, medication, and structural brain measures to examine the potential effects of confounds common to studying an older population. Nonetheless, methodological considerations such as the use of population-specific brain templates may help increase the reliability of brain measures (Buckner et al., 2004).

#### Fitness and Cognitive Gains after Exercise Intervention

The cognitive improvements in the current study are similar to previous studies that find the largest gains in executive function after aerobic exercise training (Colcombe and Kramer, 2003; Guiney and Machado, 2013; Voss et al., 2013c; Kelly et al., 2014). Here, we used a composite score to analyze training effects instead of assessing group by time interactions in each cognitive task, which can be problematic given the multitude of tasks which requires multiple statistical comparisons. Nonetheless, it is possible that the cognitive effects of the current intervention are driven by specific tasks. For example, the task-switching and spatial working memory tasks in the current study are similar to previous tasks that are sensitive to fitness-related improvements (Hawkins et al., 1992; Kramer et al., 2001; Colcombe and Kramer, 2003; Erickson et al., 2011). On the other hand, improvements in reasoning tasks have been less studied in fitness interventions, although aerobic-related gains in visuospatial processes have been documented in younger populations (Stroth et al., 2009; Monti et al., 2012), and improvement in reasoning skills have been found after cognitive training interventions in older adults (Ball et al., 2002; Willis et al., 2006; Lustig et al., 2009). Moreover, compared to previous studies (Colcombe and Kramer, 2003; Voss et al., 2010b; Erickson et al., 2011), the current intervention lasted only 6 months, and it is likely that larger cognitive effects would result from a longer intervention (Colcombe and Kramer, 2003; Kelly et al., 2014). In addition, aerobic exercise has been shown to improve hippocampal function in animal and human studies (Berchtold et al., 2010; Voss et al., 2013c), increase hippocampal volume in humans (Erickson et al., 2011) and to relate to hippocampaldependent functions such as spatial memory (Erickson et al., 2011) and relational memory (Chaddock et al., 2010). Given these previous findings, we would have expected exerciserelated effects not only in the spatial working memory task, but also in the episodic memory tasks. The null findings of the current study may in part reflect a lack of sensitivity in these relatively brief memory tasks in measuring interventionrelated change, but may also stem from comparable effects across the four groups, with similar improvements from the different interventions.

In the current study, the SSS group performed exercises that involved some form of resistance training, which has also shown to be beneficial for executive functioning in older adults when performed at a higher intensity (Liu-Ambrose et al., 2010, 2012). Although the strength portion of the SSS exercises in the current study is not comparable to the intensive strength training regimens of other studies, the similarity in exercise style may present an issue for analyzing the effects of interventions such as these, since strength training exercise and aerobic-walking exercise may benefit cognitive function in both differential and overlapping ways. Thus, "null effects" in terms of a lack of differential improvement (i.e., group by time interaction) in other cognitive domains may instead partly reflect comparable gains from the different types of interventions (in addition to gains attributable to test-retest effects) and contamination effects from self-initiated exercise (Ehlers et al., 2016). The Dance group, despite the cognitive demands thought to be involved in the learning and execution of dance steps, showed the smallest effects post-intervention; the group as a whole did not improve in CRF and showed the smallest changes in cognitive function. These findings may in part stem from the heterogeneity and lack of intensity in the Dance sessions, which varied in form (i.e., type of dance) across sessions, and may have thus failed to consistently and intensively train specific physical and cognitive skills. Indeed, Dance participants perceived their inclass sessions as less intensive (Ehlers et al., 2016). Nonetheless, the Dance intervention in the current study has been shown to improve white matter microstructure in the fornix, with baseline fornix fractional anisotropy correlating with baseline processing speed (Burzynska et al., 2017). This paper focuses on EF and connectivity in gray matter, and it is likely that different brain measures reflect distinct aspects of cognitive function. Moreover, the sixth month duration before pre- and post-testing may not adequately reflect longer-term neural and behavioral effects of each intervention.

EF improvements were not directly related to CRF improvements. Combining the test scores into a composite score may have diluted any relationship between CRF gain and gains in a specific test, but no robust correlations were found when examining the relationship between CRF gain and gain on each test measure. In addition, it is possible that intervention-related gains in CRF per se does not lead to cognitive improvements, and that indirect effects of exercise on stress, sleep and overall health lead to positive cognitive outcomes (King et al., 1997; Etnier et al., 2006; Cotman et al., 2007; Bherer et al., 2013; Awick et al., 2015). Furthermore, CRF as measured using VO2peak in the current study, indexes an array of bodily functions (Dustman et al., 1984; Etnier et al., 2006; Jain et al., 2010) and may not adequately capture cerebrovascular changes.

### ETHICS STATEMENT

This study was carried out in accordance with the recommendations of the University of Illinois Institutional Review Board, with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the University of Illinois Institutional Review Board.

### AUTHOR CONTRIBUTIONS

Conceptualization and study design: PB, CG, MV, EM, AK, and MD. Data collection: PB, MV, AB, CW, GC, KD, JF, DE, ES, and SA. Data analysis: PB and CG. Writing—original draft: PB and CG. Writing—review and editing: PB, CG, MV, AB, JF, DE, ES, EM, AK, and MD.

### ACKNOWLEDGMENTS

This work was supported by the National Science Foundation (IGERT Grant 0903622 to PB), Beckman Institute for Advanced Science and Technology (Graduate Fellowship to PB), Department of Defense (NDSEG to CG), National Institutes of Health (Grant R37 AG025667 to AK and EM, Grant NS079698 to MD), Center for Nutrition Learning and Memory, UIUC (Grant 2012-04673 to AK and EM), and the American Cancer Society (Postdoctoral Fellowship Grant PF-16-021-01-CPPB to DE). We thank Anya Knecht and Susan Houseworth for coordinating the intervention, Nancy Dodge and Holly Tracy for assistance in MRI data collection, Kathleen Kramer and Kishan Patel for assistance with Freesurfer processing, and members of the Lifelong Brain and Cognition Laboratory and Exercise Psychology Laboratory for assistance in data collection.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi. 2017.00426/full#supplementary-material

### REFERENCES


of human brain functional networks. J. Neurosci. 31, 8259–8270. doi: 10.1523/JNEUROSCI.0440-11.2011


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Baniqued, Gallen, Voss, Burzynska, Wong, Cooke, Duffy, Fanning, Ehlers, Salerno, Aguiñaga, McAuley, Kramer and D'Esposito. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Aging and Network Properties: Stability Over Time and Links with Learning during Working Memory Training

Alexandru D. Iordan<sup>1</sup> \*, Katherine A. Cooke<sup>1</sup> , Kyle D. Moored<sup>2</sup> , Benjamin Katz <sup>3</sup> , Martin Buschkuehl <sup>4</sup> , Susanne M. Jaeggi <sup>5</sup> , John Jonides <sup>1</sup> , Scott J. Peltier <sup>6</sup> , Thad A. Polk <sup>1</sup> and Patricia A. Reuter-Lorenz <sup>1</sup>

*<sup>1</sup> Department of Psychology, University of Michigan, Ann Arbor, MI, United States, <sup>2</sup> Department of Mental Health, Bloomberg School of Public Health, Johns Hopkins University, Baltimore, MD, United States, <sup>3</sup> Department of Human Development and Family Science, Virginia Tech, Blacksburg, VA, United States, <sup>4</sup> MIND Research Institute, Irvine, CA, United States, <sup>5</sup> School of Education, University of California, Irvine, Irvine, CA, United States, <sup>6</sup> Functional MRI Laboratory, Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States*

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Huali Wang, Peking University Sixth Hospital, China Bin Jing, Capital Medical University, China Betty M. Tijms, VU University Medical Center, Netherlands*

> \*Correspondence: *Alexandru D. Iordan adiordan@umich.edu*

Received: *23 August 2017* Accepted: *07 December 2017* Published: *04 January 2018*

#### Citation:

*Iordan AD, Cooke KA, Moored KD, Katz B, Buschkuehl M, Jaeggi SM, Jonides J, Peltier SJ, Polk TA and Reuter-Lorenz PA (2018) Aging and Network Properties: Stability Over Time and Links with Learning during Working Memory Training. Front. Aging Neurosci. 9:419. doi: 10.3389/fnagi.2017.00419* Growing evidence suggests that healthy aging affects the configuration of large-scale functional brain networks. This includes reducing network modularity and local efficiency. However, the stability of these effects over time and their potential role in learning remain poorly understood. The goal of the present study was to further clarify previously reported age effects on "resting-state" networks, to test their reliability over time, and to assess their relation to subsequent learning during training. Resting-state fMRI data from 23 young (YA) and 20 older adults (OA) were acquired in 2 sessions 2 weeks apart. Graph-theoretic analyses identified both consistencies in network structure and differences in module composition between YA and OA, suggesting topological changes and less stability of functional network configuration with aging. Brain-wide, OA showed lower modularity and local efficiency compared to YA, consistent with the idea of age-related functional dedifferentiation, and these effects were replicable over time. At the level of individual networks, OA consistently showed greater participation and lower local efficiency and within-network connectivity in the cingulo-opercular network, as well as lower intra-network connectivity in the default-mode network and greater participation of the somato-sensorimotor network, suggesting age-related differential effects at the level of specialized brain modules. Finally, brain-wide network properties showed associations, albeit limited, with learning rates, as assessed with 10 days of computerized working memory training administered after the resting-state sessions, suggesting that baseline network configuration may influence subsequent learning outcomes. Identification of neural mechanisms associated with learning-induced plasticity is important for further clarifying whether and how such changes predict the magnitude and maintenance of training gains, as well as the extent and limits of cognitive transfer in both younger and older adults.

Keywords: intrinsic activity, functional connectivity, graph theory, reliability analysis, intraclass correlation

#### INTRODUCTION

Aging is associated with cognitive decline that may be linked in part to altered communication among various brain regions (Reuter-Lorenz and Park, 2014). Indeed, aging has been shown to affect the integration of information both within and between functional brain networks (Ferreira and Busatto, 2013; Dennis and Thompson, 2014; Damoiseaux, 2017), which may have implications for cognitive performance. Despite accumulating evidence suggesting age effects on the configuration of large-scale functional brain networks (Achard and Bullmore, 2007; Meunier et al., 2009a; Onoda and Yamaguchi, 2013; Betzel et al., 2014; Cao M. et al., 2014; Chan et al., 2014; Song et al., 2014; Geerligs et al., 2015; Ng et al., 2016), the stability of these effects over time remains poorly understood. One goal of the present study was to clarify this issue by assessing age differences in functional network properties at two different time points.

A substantial body of evidence suggests that aging influences the functional organization of the brain, both globally and at the level of individual brain networks (reviewed in Ferreira and Busatto, 2013; Dennis and Thompson, 2014; Sala-Llonch et al., 2015; Damoiseaux, 2017). The functional organization of the brain has traditionally been studied using fMRI-based "resting-state" functional connectivity (Greicius et al., 2003; Power et al., 2011) and more recently, with graph-theoretic analyses (Bullmore and Sporns, 2009; Rubinov and Sporns, 2010). The graph-theoretic approach enables characterization of the brain's connectivity structure and derives measures that assess global and local features that may be important for network function (Bullmore and Sporns, 2009; Rubinov and Sporns, 2010). One such measure is modularity (Newman and Girvan, 2004; Newman, 2006), which indexes the extent to which a graph is organized into separate modules with dense within- and sparse between-modules connections, a fundamental principle thought to support the brain's functional segregation and integration (Dehaene et al., 1998; Sporns and Betzel, 2015). A number of prior investigations have identified lower modularity in aging (Onoda and Yamaguchi, 2013; Betzel et al., 2014; Cao M. et al., 2014; Song et al., 2014; Geerligs et al., 2015; but see Meunier et al., 2009a), with networks becoming less distinct due to increased between- and decreased withinmodule integration. This evidence is consistent with the idea of functional dedifferentiation (Park et al., 2004, 2010; Grady, 2012). Another set of measures characterizes the efficiency of information flow across the graph. Global efficiency indexes graph-wide integration and has been linked with the capacity for rapid information exchange among distributed regions, whereas local efficiency indexes integration at a regional level and has been linked with fault tolerance within specialized regions (Latora and Marchiori, 2003; Achard and Bullmore, 2007). Previous investigations have associated aging with lower local efficiency (Achard and Bullmore, 2007; Cao M. et al., 2014; Song et al., 2014; Geerligs et al., 2015), while global efficiency was reported to be similar irrespective of age (Cao M. et al., 2014; Song et al., 2014; Geerligs et al., 2015; but see Achard and Bullmore, 2007).

Importantly, differences in connectivity structure observed at a brain-wide level may be related to specific patterns at the level of individual networks, and current evidence suggests differential effects of aging on particular brain networks (Ferreira and Busatto, 2013; Dennis and Thompson, 2014; Sala-Llonch et al., 2015; Damoiseaux, 2017). Although the majority of investigations have targeted the default-mode network (DMN), showing lower functional connectivity between its different sub-components with aging (Andrews-Hanna et al., 2007; Damoiseaux et al., 2008), recent evidence also points to age effects in other brain networks, such as the cingulo-opercular/salience and sensorimotor networks (Meier et al., 2012; Onoda et al., 2012; He et al., 2014; Geerligs et al., 2015; La Corte et al., 2016) 1 . Thus, to complement information provided by brain-wide network assessments, metrics applied at the level of individual networks can also be employed. This includes the participation coefficient, which indexes the relation between intra- and inter-network connectivity for each node (Guimerà and Amaral, 2005).

In sum, although there are some inconsistencies across studies, available evidence points to lower within- and higher between-network connectivity with aging. This is expressed topologically as lower modularity, and is associated with lower local efficiency and preserved global efficiency, compared to younger age (see Damoiseaux, 2017 for a recent discussion). The first main goal of the present study was to assess the replicability of these previously reported age effects on functional network configuration.

Inconsistencies across investigations of age differences in network properties may stem from methodological differences but also from variability of network measures over time (van Wijk et al., 2010; Zalesky et al., 2016; Ciric et al., 2017; Geerligs et al., 2017). One way to assess reliability is by measuring the same subjects at two or more time-points, while using the same methodology, and quantifying the level of agreement between measurements by calculating the intraclass correlation coefficient (ICC) (Shrout and Fleiss, 1979; McGraw and Wong, 1996). A meta-analysis of test-retest reliability of graph-theoretic brainnetwork metrics identified overall good reliability (Welton et al., 2015). However, the available evidence related to aging is very limited. Investigations of age differences in network properties have typically used singular assessments, and hence the reliability of such effects over time is not clear (but see Geerligs et al., 2017). Thus, the second main goal of the present investigation was to extend the assessment of age differences in network properties to multiple time points within the same individuals and to evaluate reliability.

Clarification of age differences in network properties and their stability over time is important for further assessment of changes associated with cognitive training in older adults. Specifically, if aging influences relations between functional network properties and training outcomes, then these effects

<sup>1</sup>While graph theory has been typically employed to assess global and local measures of connectivity, much evidence regarding aging effects on specific brain networks has been derived using complementary approaches, such as seed-based functional connectivity and independent component analysis. Although these approaches differ in important ways (see Ferreira and Busatto, 2013; Dennis and Thompson, 2014 for recent discussions), results have been overall convergent (see Geerligs et al., 2015 for a recent graph theory investigation at the level of individual networks).

need to be disentangled from variability of network measures in the absence of intervention. Recent evidence suggests potential links between baseline properties of functional brain organization and benefits accrued over the course of cognitive training in older adults (Gallen et al., 2016a), although at this point such evidence is only preliminary. Although a growing body of studies suggests that some working memory (WM) interventions may alter functional network organization and have beneficial, albeit limited, effects on cognitive functioning (Buschkuehl et al., 2008; Lustig et al., 2009; Brehmer et al., 2014; Karbach and Verhaeghen, 2014; Stepankova et al., 2014; Ballesteros et al., 2015; Bherer, 2015; Mewborn et al., 2017; Román et al., 2017), evidence linking baseline functional network characteristics with training is limited (Arnemann et al., 2015; Gallen et al., 2016a). In one investigation of this topic, Gallen et al. (2016a) showed that older adults displaying greater network modularity at baseline also showed greater improvements in gist reasoning, following a strategic memory and reasoning training intervention (Vas et al., 2011). However, the potential role of other network properties in learning remains largely unknown. Thus, the third main goal of this investigation was to assess relations between baseline network properties and subsequent learning during training in older adults.

These questions were investigated in a sample comprising both healthy younger and older adults, using resting-state fMRI data acquired in 2 different sessions, both preceding a WM training intervention. A complete treatment of training outcomes and other behavioral data will be reported separately. Based on the extant evidence, we expected to find lower modularity and local efficiency in older compared to younger adults, and similar global efficiency across groups. We also expected these differences to be stable over time. Finally, the limited evidence linking network properties with training effects suggests that modularity is beneficial (Gallen et al., 2016a); therefore, we expected that network properties, in particular modularity (i.e., as reflected in the modularity index), would be linked to learning rates.

### METHODS

### Participants

A sample of 23 younger (YA) and 23 healthy, cognitively normal older adults (OA) were recruited from the University of Michigan campus and community surrounding Ann Arbor, Michigan to participate in an adaptive verbal WM training study. All participants were right-handed, native English speakers with normal or corrected-to-normal hearing and vision and were screened for history of head injury, psychiatric illness, or alcohol/drug abuse. Data from 3 OA were excluded due to technical issues related to brain-imaging data acquisition. Thus, the sample for fMRI analyses consisted of 23 YA (age range: 18–28; 9 females) with a mean age (±S.D.) of 21.3 (±2.5) years and 20 OA (age range: 64–76; 9 females) with a mean age of 68.3 (±3.6) years. For analyses linking fMRI with behavioral results, 2 additional participants (1 OA) were excluded, due to technical issues related to behavioral task assessments, and thus these analyses were reported on 22 YA and 19 OA. Older adult participants completed the Short Blessed Test (Katzman et al., 1983) over the phone prior to inclusion in the study to screen for potential mild cognitive impairment, and additional neuropsychological assessments using the Montreal Cognitive Assessment (Nasreddine et al., 2005) confirmed normal cognitive function for all participants (scores ≥ 26). Additionally, participants were screened for depressive symptoms that could affect cognitive functioning using the depression module of the Patient Health Questionnaire (Kroenke et al., 2001). The University of Michigan Institutional Review Board approved all procedures, and all participants provided informed consent prior to participating.

#### Imaging Protocol

Functional MRI data were acquired during 8 min of resting state, following completion of a verbal WM task, in 2 sessions 2 weeks apart (t1, t2) (see Supplementary Figure 1 for an illustration of the study timeline). Participants were instructed to view a fixation cross in the center of the screen while keeping their mind calm and relaxed. Imaging data were collected using a 3 T General Electric MR750 scanner with an eight-channel head coil. Functional images were acquired in ascending order using a spiral-in sequence, with MR parameters: TR = 2,000 ms; TE = 30 ms; flip angle = 90◦ ; field of view = 220 × 220 mm<sup>2</sup> ; matrix size = 64 × 64; slice thickness = 3 mm, no gap; 43 slices; voxel size = 3.44 × 3.44 × 3 mm<sup>3</sup> . After an initial 10 s of signal stabilization, 235 volumes were acquired. A high-resolution T1-weighted anatomical image was also collected following the WM task and preceding resting-state acquisition, using spoiled-gradient-recalled acquisition (SPGR) in steadystate imaging (TR = 12.24 ms, TE = 5.18 ms; flip angle = 15◦ , field of view = 256 × 256 mm<sup>2</sup> , matrix size = 256 × 256; slice thickness = 1 mm; 156 slices; voxel size = 1 × 1 × 1 mm<sup>3</sup> ). Images were de-spiked in k-space and reconstructed using an in-house iterative reconstruction algorithm with fieldmap correction (Sutton et al., 2003), which has superior reconstruction quality compared to non-iterative conjugate phase reconstruction.

### Preprocessing

Preprocessing was performed using SPM12 (Wellcome Department of Cognitive Neurology, London). Functional images were slice-time corrected, realigned, and co-registered to the anatomical image using a mean functional image. A study-specific anatomical template was created (younger and older adults together; Geerligs et al., 2015), using Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) (Ashburner, 2007), based on segmented gray matter and white matter tissue classes, to optimize inter-participant alignment (Klein et al., 2009). The DARTEL flowfields and MNI transformation were then applied to the functional images and to the segments, and the functional images were resampled to 3 × 3 × 3 mm<sup>3</sup> voxel size. To minimize artificial local spatial correlations, no additional spatial smoothing was applied (Salvador et al., 2005; Achard et al., 2006; Achard and Bullmore, 2007; Wang et al., 2010, 2011; Liao et al., 2011; Zalesky et al., 2012; Alakorkko et al., 2017).

Identification of outlier scans was performed using Artifact Detection Tools (ART; www.nitrc.org/projects/artifact\_detect/), as follows. Scans were classified as outliers if frame-to-frame difference exceeded 0.5 mm in composite motion (combination of translational and rotational displacements) or 3 standard deviations in the global mean signal. On average, the proportion of outliers was below 5% in both YA (t1: 4.42%; t2: 2.72%) and OA (t1: 3.68%; t2: 3.74%). There were no significant differences between the two groups in the number of outlier scans (p's > 0.4), or in the average (p's > 0.1) or maximum (p's > 0.5) motion, either before or after correcting for outlier scans (see "scrubbing" below).

### Graph Construction

#### Functional Connectivity Analysis

Brain-wide functional connectivity analyses were performed using the Connectivity Toolbox (CONN; Whitfield-Gabrieli and Nieto-Castanon, 2012). To construct a brain-wide graph, we employed a commonly used functional atlas (Power et al., 2011), which comprises 264 meta-analytically defined coordinates, including cortical and subcortical areas; a 5 mm-radius sphere was centered at each of these coordinates. To ensure that the graph comprised regions that were not susceptible to fMRI signal drop-out, each sphere was filtered through a sample-level signal intensity mask, calculated as follows: First, binary masks were calculated for each subject, at each time point, thresholded at >70% mean signal intensity (Geerligs et al., 2015), computed over all voxels, using ART. Then, a sample-level mask was calculated, across all subjects and time points, using logical conjunction (see Supplementary Figure 2 for an illustration of the mask). Regions with fewer than 8 voxels (∼50% volume) overlap with the samplelevel mask were excluded, leaving 234 regions of interest (ROIs).

To remove physiological and other sources of noise from the fMRI time series we used linear regression and the anatomical CompCor method (Behzadi et al., 2007; Chai et al., 2012; Muschelli et al., 2014), as implemented in CONN. Each participant's white matter and cerebrospinal fluid segments, eroded by 1 voxel to minimize partial volume effects, were used as noise ROIs. The following temporal covariates were added to the model: signal extracted from each participant's noise ROIs (5 principal component analysis parameters for each<sup>2</sup> ), motion parameters (3 rotation and 3 translation parameters, plus their first-order temporal derivatives), regressors for each outlier scan (i.e., "scrubbing"; one covariate was added for each outlier scan, consisting of 0's everywhere but the outlier scan, coded as "1"), and a session-onset regressor (a delta function convolved with the hemodynamic response function plus its first-order temporal derivative). The residual fMRI time series were band-pass filtered (0.01 Hz < f < 0.1 Hz). Pearson correlation coefficients were computed between the time courses of all pairs of functional ROIs, followed by Fisher-z transformation, and the diagonal of the connectivity matrix was set to zero. Graph construction and analyses were performed separately for each group and time point, using tools from the Brain Connectivity Toolbox (Rubinov and Sporns, 2010).

#### Group-Level Consensus Partitions

To achieve a community structure representative of each group, we used the Louvain community detection algorithm (Blondel et al., 2008), in conjunction with consensus clustering (Lancichinetti and Fortunato, 2012). This approach capitalizes on the consistency of each node's module affiliation across a set of partitions, to circumvent the known degeneracy of the Louvain algorithm (i.e., multiple partitioning solutions) (Good et al., 2010). To obtain a unique (i.e., threshold-independent) solution for each group, the Louvain algorithm was applied on weighted graph edges (positive only); see Cohen and D'Esposito (2016) for a similar approach. The group-level consensus partitions were employed to derive node–module assignments used for analyses at the level of individual modules/networks (see Network Measures sub-section below) and for display purposes (**Figure 1**).

Consensus clustering was applied first at the individual level, to generate a robust partition for each participant, and then at the group level, to generate a representative partition for each group and at each time point; see Dwyer et al. (2014) for a similar approach. First, to generate a robust partition for each participant, the Louvain algorithm was run 500 times. Because the algorithm is susceptible to local maxima, each initial partition was optimized using iterative community fine-tuning (Sun et al., 2009), which maximizes modularity by reassigning the nodes to modules and iterating the Louvain algorithm. For each participant, we constructed an agreement matrix representing the fraction of runs in which each pair of nodes was assigned to the same module. The Louvain algorithm was then iteratively run on the agreement matrix (500 Louvain runs at each step), to generate a consensus partition for each participant. For each iteration, the agreement matrix was recalculated and thresholded, until a single representative partition was obtained for each participant. Second, to generate a group-level representative partition, an agreement matrix was calculated based on the consensus partitions of all participants in one group. The Louvain algorithm was then run on the agreement matrix to obtain a consensus partition for each group, as described above. The resolution parameter of the Louvain community detection algorithm (γ ) and the thresholding parameter for the agreement matrix (τ ) were determined using a procedure that maximized modularity over all group-level partitions. Specifically, we ran the procedure described above for typical ranges of values for both parameters and chose those values that, on average, maximized modularity across all 4 group-level partitions (see below for a formal description of modularity). The value ranges were γ between 1 and 1.5 and τ between 0.2 and 0.5, with increments of 0.05 for each parameter. The maximum average modularity was Q = 0.71, achieved for γ = 1.25

<sup>2</sup>Regressing out multiple principal components from noise ROIs typically leads to better noise correction than regressing out the average noise signal because physiological noise (including motion) is not spatially homogeneous across the brain (Chai et al., 2012; Muschelli et al., 2014).

and τ = 0.5, and these parameters were used for subsequent analyses.

#### Connection Density Thresholding

We used density-based thresholding, which equates the number of edges across graphs and allows proper between-groups comparisons (van Wijk et al., 2010; Garrison et al., 2015). To ensure that results were not due to any specific threshold, calculations were performed for a range comprising 2–10% of the strongest connections, in 1% increments. This threshold range is similar to that used in generating the Power et al. (2011) functional atlas and matches the range previously employed by Geerligs et al. (2015), thus enabling comparison of results. In general, stringent threshold ranges are preferable because inclusion of false-positive connections is more detrimental to network measures computation than exclusion of false-negative connections (Zalesky et al., 2016). The average number of disconnected nodes at each threshold in the 2–10% range was as follows: 47, 27, 17, 10, 7, 4, 3, 2, and 1. Because average connectivity was similar across groups (as assessed by permutation testing on positive edges; t1, 2: p's > 0.2), density-based thresholding was likely unbiased across groups (Zalesky et al., 2016; van den Heuvel et al., 2017). To calculate network measures, connectivity values were binarized for each threshold (i.e., 1 if above, 0 if below threshold). Betweengroups comparisons of graph-theoretic measures used binarized graphs and reported graph metrics are values averaged across all thresholds, unless specified otherwise.

#### Network Measures

To assess the strength of module segregation, we calculated the modularity index (Q) (Newman and Girvan, 2004; Newman, 2006), which compares the observed intra-module connectivity with that which is expected by chance. Higher modularity values indicate stronger separation of the graph's modules. The modularity index is formally defined as follows:

$$Q = \frac{1}{2E} \sum\_{\vec{\eta}} \left[ A\_{\vec{\eta}} - \chi e\_{\vec{\eta}} \right] \delta(m\_i, m\_{\vec{\eta}})$$

where E is the number of graph edges, A is the adjacency matrix, γ is the resolution parameter, e is the null model [e = kikj/2E, where k<sup>i</sup> and k<sup>j</sup> are the degrees (i.e., number of connections) of the nodes i and j], and δ is an indicator that equals 1 if nodes i and j belong to the same module and 0 otherwise. The modularity score for each participant was calculated as the average over 500 runs of the Louvain algorithm with iterative community finetuning. For consistency with the consensus clustering procedure described above, the same resolution parameter (γ = 1.25) was used.

To assess the integration of information, we calculated global and local efficiency (Latora and Marchiori, 2003). Global efficiency indexes integration at the level of the entire graph and it is defined as follows:

$$E\_{\text{glob}} = \frac{1}{N(N-1)} \sum\_{i \neq j} \frac{1}{L\_{ij}}$$

where N is the number of nodes in the graph and Lij is the shortest path length between nodes i and j. By contrast, local efficiency is a node-specific measure, and is defined relative to the sub-graph comprising the immediate neighbors of a node. Local efficiencies for all nodes were averaged to provide an estimate of the mean local efficiency of the entire graph or of a module. Local efficiency of a node i is defined as follows:

$$E\_{loc}(i) = \frac{1}{N\_{G\_i}(N\_{G\_i} - 1)} \sum\_{j,h \in G\_i} \frac{1}{L\_{jh}}$$

where G<sup>i</sup> is the sub-graph comprising all the immediate neighbors of the node i.

Another node-specific measure is the participation coefficient (Guimerà and Amaral, 2005), which indexes inter-network connectivity by quantifying the distribution of each node's connections across different modules. Participation coefficients of all nodes within a module were averaged to provide an estimate of mean participation for a module. Participation coefficient of a node i is defined as follows:

$$P\left(i\right) = 1 - \sum\_{m=1}^{M} \left[\frac{k\_i(m)}{k\_i}\right]^2$$

where M is the number of modules in the graph, and ki(m) is the degree of node i within its own module m, and k<sup>i</sup> is the degree of node i regardless of module membership.

Finally, to assess the convergence of results based on the graph-theoretic measures described above with simpler connectivity analyses, we calculated within- and between-module connectivity using an approach similar to Geerligs et al. (2015). For completeness, this procedure was performed separately for positive and negative connectivity values. First, the initial connectivity matrices were thresholded by retaining values that survived a false discovery rate (FDR) correction (q < 0.05) (Benjamini and Hochberg, 1995) and setting all the other values to zero. Then, for each module and pair of modules, we computed the sum of all connectivity values and divided by the number of possible connections to estimate within- and between-modules connectivity. Of note, this procedure was used only for the analysis of within- and between-networks connectivity, and it did not influence the previously introduced graph-theoretic measures, which were all calculated on unweighted (i.e., binary) graphs.

#### Statistical Methods

As a general strategy, assessments were performed on metrics averaged across all thresholds, and significant results were followed-up with tests for each threshold, to assess consistency across the threshold range.

#### Age Differences in Community Structure

To assess age differences in community structure, we compared module composition between groups using normalized mutual information (NMI) (Kuncheva and Hadjitodorov, 2004) and permutation testing. NMI measures how much information about the structure of one partition reduces uncertainty about the structure of another partition, and is a relative measure that varies from 0 (completely independent) to 1 (identical partitions). Because individual similarity measures are not independent, we used an unbiased procedure that compared the average betweengroups similarity in the actual data with a null distribution based on randomizing group memberships; see Alexander-Bloch et al. (2012) for a similar approach. Between-groups similarity in the actual data was calculated for each density threshold, by averaging the pair-wise partition similarity for all subjects across the two groups, separately at each time point. For each subject, we used the partition with the highest modularity for each threshold, calculated over 500 Louvain repetitions with community fine-tuning and resolution γ = 1.25. The null distribution was calculated in a similar way, using the randomly divided groups over 5,000 permutations, while retaining original group sizes. If the actual between-groups NMI was smaller than the 5th percentile of the null distribution, the difference was considered significant. Furthermore, to determine whether one group showed more similar partitions than the other, we examined within-group partition similarity. This analysis was performed in a similar way, by averaging pair-wise partition similarity separately for subjects in each group. Finally, to examine differences in the stability of partitions over time, we calculated within-subject partition similarity over the two sessions. A between-group difference in partition similarity over time was tested directly, using permutation testing (Groppe et al., 2011).

#### Age Differences in Network Measures

Age differences in network measures were first assessed brainwide (modularity, global efficiency, and local efficiency) and then significant results were followed-up by analyses at the level of each module or network (participation coefficient and local efficiency). To ensure comparability at the level of individual networks, each module was represented only by those nodes that were consistently assigned to the same module, both across groups and time points, based on the group-level consensus partitions; see Geerligs et al. (2015) for a similar approach. Between-groups differences in network properties were assessed using permutation testing, and a family-wise error (FWE) correction for multiple comparisons based on the "max statistic" method (Blair and Karniski, 1993; Groppe et al., 2011) was applied to account for simultaneous testing of the five main modules identified (see Results section). As mentioned above, an ancillary analysis of within- and between-modules connectivity was also performed and the same FWE correction for multiple comparisons was applied for this analysis as well.

#### Reliability Analysis

The intraclass correlation coefficient (ICC) was employed to measure the absolute agreement for each graph metric between the two sessions (McGraw and Wong, 1996; Welton et al., 2015). We used a mixed model<sup>3</sup> ICC(A, <sup>k</sup>) to estimate the degree of absolute agreement of measurements that are averages of k = 2 independent measurements on randomly selected subjects.

<sup>3</sup>We made no assumption of interchangeability of t<sup>1</sup> and t<sup>2</sup> assessments because the resting-state data were acquired following completion of a verbal WM task inside the scanner, and thus potential differences in task performance at the two time points might have differentially influenced resting-state recordings. We also expected, however, that these effects would be mitigated by a ∼6 min break (recording the T1-weighted anatomical image) following the WM task and preceding the resting-state acquisition (Breckel et al., 2013).

ICC was calculated as follows: ICC = (MSR–MSE)/[MS<sup>R</sup> + (MS<sup>C</sup> – MSE)/n], where MS<sup>R</sup> is mean square for rows/subjects, MS<sup>E</sup> is mean square error, and MS<sup>C</sup> is mean square for columns/assessments (Shrout and Fleiss, 1979; McGraw and Wong, 1996). We used the following guidelines for ICC interpretation: <0.20, poor; 0.21–0.40, fair; 0.41–0.60 moderate; 0.61–0.80 strong; >0.8, almost perfect (Montgomery et al., 2002; Telesford et al., 2010).

#### Links with Learning during WM Training

The second scanning session was followed by 10 days of computerized verbal WM training (Supplementary Figure 1). The adaptive training task consisted of a modified WM itemrecognition task that required participants to encode and retain consonant letters of variable set size for several seconds (Sternberg, 1966; see also Stepankova et al., 2014); set size changed adaptively depending on participants' performance. Participants completed 6 blocks of 14 trials during each training session. Here, we focus on training-related improvements in WM performance specifically, as measured by mean set size achieved during each training session for each participant, to evaluate their relationship with network properties. Furthermore, we focused on early and late learning rates, defined as the performance change between training sessions 1 and 2 (early learning rate), and as the performance change across training sessions 2–10 (late learning rate), respectively, modeled for each individual using a linear spline term with a knot at the second training session (see Appendix for details). YA had a higher mean early slope than OA [t(39) = 3.59, p = 0.001], but late slope did not differ by age group [t(39) = 1.64, p = 0.109].

To assess links between network measures and learning rates, we calculated correlations between brain-wide network measures and early learning slopes, separately for each group and at each time point. We focused on early learning rates because age differences were identified in early but not in late learning slopes. Due to relatively small sample sizes, we employed Spearman's rank correlation coefficient (ρ) to minimize influence from extreme values. Significant brain-wide results were followed by assessments at the level of each module/network, corrected for multiple comparisons using the permutation-based "max statistic" method (Groppe et al., 2011). We took multiple steps to assess the robustness of our findings, using a procedure similar to Gallen et al. (2016a). First, to assess whether the relations between network measures and learning rates were constantly present over the threshold range, we tested these relations separately for each threshold. Second, given the absence of differences in motion across groups and time points (see Preprocessing subsection above), we performed partial Spearman correlations (ρp) to examine whether controlling for motion altered the relations between brain-wide network measures and learning rates.

#### RESULTS

#### Age Differences in Community Structure

Functional networks were identified separately for each group and time point, using consensus partitioning (see Methods section for details). Similar modules were identified in both YA and OA, consistent with the main functional networks described in the literature (Power et al., 2011; Yeo et al., 2011): fronto-parietal (FPN), cingulo-opercular/salience (CON), default-mode (DMN), visual (Vis) and somato-sensorimotor (SMN) (**Figure 1**). The community structure of the partitions for each age group was examined using normalized mutual information (NMI). Results showed differences in node-module assignment between YA and OA, at both time points (**Figure 2**). First, analysis of between-groups partition similarity showed that similarity of community structure between YA and OA was significantly lower than expected based on the permuted data (t1, 2: p's < 0.001; **Figure 2A**). Second, analysis of within-group partition similarity showed less similarity for OA as a group. Specifically, partition similarity for YA was higher (t1: p = 0.003; t2: p < 0.001) whereas for OA was lower (t1: p = 0.007; t2: p = 0.003) than expected based on the permuted data (Supplementary Figure 3). This indicates that there is greater heterogeneity in OA's partitions, i.e., their partitions are less similar to one another than YA's partitions. Finally, analysis of within-subject similarity across time showed less within-subject consistency for OA (p = 0.001; **Figure 2B**), indicating more variability in node-module assignment in OA across time. In summary, although similar functional networks were identified in both YA and OA, their composition differed between groups, and OA showed less similarity, both as a group and across time, compared to YA.

#### Age Differences in Network Measures and Reliability Analysis

To complement the comparisons of community structure presented above, we assessed age differences in several network measures. Network measures were first calculated brain-wide, followed by an assessment of their reliability over time. Then, significant brain-wide differences were followed-up by assessments at the level of each of the five modules/networks.

#### Brain-Wide Network Measures

At a brain-wide level, OA showed lower modularity indices at both time points (t1: p = 0.046, t2: p < 0.001), indicating lower intra-module connectivity compared to YA. Furthermore, OA consistently showed lower local efficiency (t1, 2: p's < 0.001), while global efficiency was not significantly different across groups (t1, 2: p's > 0.1), suggesting age differences in local but not global integration of information (**Figure 3**). Ancillary correlation analyses between age and brain-wide network measures within the OA group revealed no significant results (p's > 0.05).

#### Reliability Analysis of Brain-Wide Network Measures

Reliability of brain-wide network measures was assessed using intraclass correlation (ICC), by calculating the absolute agreement of each graph metric across sessions. Brain-wide measures showed overall moderate to strong ICC over time (range 0.51–0.74), with the highest agreement for local efficiency (**Figure 4**). For each group, the agreement ranged from fair (>0.2) to strong (>0.6), with YA showing lowest agreement for global efficiency. Examination of the profiles of ICC values

across the range of thresholds indicated that the reproducibility of network measures was generally stable across thresholds, with the exception of global efficiency for YA.

older adults (red color). Magenta asterisks indicate *p* < 0.05 for each threshold. \*\**p* < 0.01, across all thresholds.

#### Individual Network Measures

Network properties were also assessed at the level of each individual network (**Figure 5**). To ensure comparability across groups and time points, each network was represented by only those nodes that were consistently assigned to the same network, both across groups and time points (see Methods section for details). OA showed greater participation coefficient for CON (t1: pFWE < 0.001; t2: pFWE = 0.002) and SMN (t1, <sup>2</sup>: pFWE's < 0.001), indicating that, compared to YA, a larger proportion of the nodes in these networks had connections outside the networks they belonged to. OA also showed lower local efficiency for CON (t1: pFWE = 0.014; t2: pFWE = 0.008) at both time points, and for DMN (pFWE = 0.029) and SMN (pFWE = 0.01) at t2. We also examined within- and between-network connectivity, using a procedure similar to Geerligs et al. (2015). Regarding within-networks connectivity, OA showed lower connectivity within DMN (t1: pFWE = 0.04; t2: pFWE = 0.018) and within CON (pFWE = 0.035) at t1, compared to YA. Regarding between-networks connectivity, OA showed greater positive connectivity between FPN and SMN (pFWE = 0.005) and between CON and SMN (pFWE = 0.048), as well as lower negative connectivity (anticorrelation) between CON and SMN (pFWE = 0.043), at t2. No other age differences in between-networks connectivity survived FWE correction for multiple comparisons. Ancillary correlation analyses between age and individual network measures within the OA group identified a significant negative correlation between age and within-DMN connectivity at t<sup>1</sup> (ρ = −0.65, pFWE = 0.015).

In summary, OA showed lower brain-wide modularity and local efficiency compared to YA, with the difference in local efficiency showing most consistency across time. At the level of individual networks, CON showed substantial differences between groups, reflected in all examined properties. Additionally, DMN and SMN were characterized by lower intranetwork connectivity and greater participation, respectively, in OA.

#### Links with Learning during WM Training

To assess links between network measures and performance during cognitive training, we calculated Spearman correlation coefficients, separately for each group and at each time point. Similar to the assessment of age differences in network measures, significant brain-wide results were followed by analyses of robustness and assessments at the level of individual networks.

#### Brain-Wide Network Measures

Interestingly, significant relations between network measures and learning rates were detected only for OA and only at t1. Specifically, modularity (ρ = 0.51, p = 0.028) and local efficiency (ρ = 0.59, p = 0.01) were positively correlated with early learning rates, whereas global efficiency (ρ = −0.61, p = 0.007) was negatively correlated with early learning rates (**Figure 6**, top panel). Ancillary analyses were performed to test for influences of educational level and sex on these results. There were no significant correlations between the number of years of education and networks measures (p's > 0.5), and controlling for the

FIGURE 3 | Age differences in brain-wide network measures. At a brain-wide level, OA showed lower modularity and local efficiency compared to YA, whereas global efficiency was not significantly different across groups. Boxplots in the left panel depict values averaged across all thresholds. YA, younger adults (blue color); OA, older adults (red color); t1, time point 1; t2, time point 2. Magenta asterisks indicate *p* < 0.05 for each threshold. \**p* < 0.05, \*\*\**p* < 0.001, across all thresholds.

number of years of education did not substantially influence the relations between network measures and learning rates. Also, Spearman correlations performed separately by sex showed similar trends in both males and females, and there were no sex differences in correlation strengths (p's > 0.6).

#### Robustness Analysis

We took multiple steps to assess the robustness of our findings, using a procedure similar to Gallen et al. (2016a). First, we assessed whether the relations between network measures and learning rates were constantly present over the threshold range, and the results confirmed that all these relations were fairly consistent across thresholds (**Figure 6**, bottom panel). Second, given the absence of differences in motion across groups and time points (see Methods section), controlling for motion (i.e., partial correlations) did not substantially alter the relations between any of the brain-wide network measures and learning rates (modularity: ρ<sup>p</sup> = 0.49, p = 0.04; local efficiency: ρ<sup>p</sup> = 0.55, p = 0.019; global efficiency: ρ<sup>p</sup> = −0.59, p = 0.01).

#### Individual Network Measures

To further elucidate the relations between network characteristics and early learning rates, significant results at the brain-wide level were followed-up by analyses at the level of individual networks. The results showed that participation of CON at t<sup>1</sup> was negatively correlated with learning rates in OA (ρ = −0.81, pFWE < 0.001), consistent with the brain-wide results. No other correlations survived FWE correction for multiple comparisons.

In summary, brain-wide network measures at t<sup>1</sup> were linked to learning rates during training in OA but not in YA. At the

level of individual networks, participation of CON showed links with training effects consistent with the patterns identified by the brain-wide analyses.

coefficient; ALL, all subjects (magenta color); YA, younger adults (blue color); OA, older adults (red color).

#### DISCUSSION

The goals of the present investigation were to assess the replicability of previously reported age effects on resting-state networks, to examine their reliability over time, and to assess their relation to behavioral outcomes (namely learning rates during a cognitive training intervention). Similar to previous investigations, we identified both consistencies in network structure and differences in module composition between groups. Notably, OA showed less similarity of their network partitions compared to YA, both as a group and across time. Regarding brain-wide network measures, OA showed lower modularity and local efficiency compared to YA, with the difference in local efficiency showing most consistency across time. At the level of individual networks, OA showed substantial differences in CON, reflected in all examined metrics, as well as lower intra-network connectivity in DMN and greater participation of SMN. Finally, baseline brain-wide network measures were linked to early learning rates in OA but not in YA, and the participation of CON showed links with early learning rates consistent with the patterns identified by the brain-wide analyses. The main findings are discussed, in turn, below.

The present results replicate previously reported age differences in functional network properties (Achard and Bullmore, 2007; Meunier et al., 2009a; Onoda and Yamaguchi, 2013; Betzel et al., 2014; Cao M. et al., 2014; Song et al., 2014; Geerligs et al., 2015) and extend these findings to multiple time points (Welton et al., 2015). Regarding community structure, the present results showing age differences in module composition, but overall similar modules are consistent with previous evidence (Geerligs et al., 2015) and suggest age-related topological changes in the context of overall similar functional configuration, irrespective of age. Furthermore, the results showing less similarity of network partitions in OA, both as a group and across time, are in line with recent evidence suggesting reduced baseline stability of network activity with aging (Tsvetanov et al., 2016).

Regarding age differences in network measures, we identified reliable age differences in brain-wide modularity and local efficiency, consistent with previous investigations (Achard and

network; DMN, default-mode network; Vis, visual network, SMN, somato-sensorimotor network; YA, younger adults (blue color); OA, older adults (red color).

Bullmore, 2007; Onoda and Yamaguchi, 2013; Betzel et al., 2014; Cao M. et al., 2014; Song et al., 2014; Geerligs et al., 2015). Modularity indexes the degree to which a graph can be partitioned into multiple communities, and is considered a central principle of brain organization, supporting functional segregation and integration through communication within- and between-modules, respectively (Dehaene et al., 1998; Sporns et al., 2000; Meunier et al., 2009b; Sporns and Betzel, 2015). Thus, results showing lower modularity in OA compared to YA suggest loss of functional specificity of the brain networks with aging (Ferreira and Busatto, 2013; Damoiseaux, 2017; Naik et al., 2017). Global efficiency indexes graph-wide integration and has been linked with information exchange among distributed regions, whereas local efficiency indexes regional-level integration and has been linked with fault tolerance within specialized areas (Latora and Marchiori, 2003; Achard and Bullmore, 2007). In general, the argument is that brains maximize cost-efficiency by favoring dense short-range connections and sparse longrange connections, because the latter are more costly (Achard

\**p* < 0.05, \*\**p* < 0.01, \*\*\**p* < 0.001, across all thresholds and corrected for multiple comparisons. *†*

and Bullmore, 2007; Bullmore and Sporns, 2012). Thus, results showing lower local efficiency in OA compared to YA suggest a reduction of cost-efficiency in aging; under conditions of similar connection density, which is considered a proxy for wiring cost, efficiency is lower in OA compared to YA (Achard and Bullmore, 2007; Geerligs et al., 2015). It should be noted, however, that wiring costs can only be approximated in functional networks, because two functionally connected regions do not necessarily share a direct structural link (Rubinov and Sporns, 2010; Zalesky et al., 2012). In fact, modularity and local efficiency are related measures, such that a system with denser local connections also tends to be more modular (Bullmore and Sporns, 2012). On the other hand, similar global efficiency irrespective of age has been explained by a greater number of inter-module connections in OA; specifically, more inter-module connections may counterbalance less intra-module connections, resulting in similar amounts of shortest path lengths between distant nodes (Song et al., 2014; Geerligs et al., 2015). In sum, these findings are consistent with overall patterns of decreased within- and

*p* = 0.016, across all thresholds, uncorrected.

increased between-system connectivity, suggesting decreased "system segregation" in aging (Betzel et al., 2014; Chan et al., 2014; Ferreira et al., 2016).

The present findings may also be relevant for better understanding task-related neural over-activation in OA relative to YA, which has been linked with both compensation and dedifferentiation (Cabeza, 2002; Park et al., 2004, 2010; Davis et al., 2008; Grady, 2008; Reuter-Lorenz and Cappell, 2008; Reuter-Lorenz and Park, 2014). Task-related over-activation in OA may be related to altered intrinsic network dynamics, reflected in differences in modularity and local efficiency "at rest." Whereas the loss of functional specificity in aging (reflected by the decline in modularity) is consistent with the idea of dedifferentiation, reduced cost-efficiency (reflected by the decline in local efficiency) may be linked to compensatory processes that are overall less efficient than the primary computations (Reuter-Lorenz and Park, 2014). Thus, dedifferentiation and compensation may both be expressions of the same process of functional recalibration due to declining structure with aging (Naik et al., 2017). This also highlights the critical need for better integrating resting-state and task-related approaches, in order to develop a practical understanding of neurocognitive function and age-related change (Iordan and Reuter-Lorenz, 2016; see also Gallen et al., 2016b).

To assess the reliability of age differences in network properties, in the present study we measured the same participants over 2 sessions 2 weeks apart and calculated ICC of network properties between the 2 sessions (McGraw and Wong, 1996). Results showed consistent age differences in network properties over time, with overall strong to moderate ICCs, comparable to previous investigations (Telesford et al., 2010; Wang et al., 2011; Braun et al., 2012; Park et al., 2012; Cao H. et al., 2014; Welton et al., 2015), thus suggesting that the observed age differences are reliable. Interestingly, results showed relatively higher reliability for local compared to global efficiency (see also Park et al., 2012). This effect was driven by YA, who showed more global efficiency variability between the 2 sessions, and it might have been linked to residual effects from the WM tasks performed prior to the resting-state recordings (Barnes et al., 2009; Breckel et al., 2013; Gordon et al., 2014). In line with our findings, a study by Park et al. (2012) also identified low reliability of global efficiency in a test-retest investigation of resting-state data in YA, assessed over a 24-h period. The authors concluded that this was likely due to high variability of long-range connections (given the dependence of global efficiency on this topological feature), and may reflect greater influence of cognitive control on this measure, compared to local efficiency (Honey et al., 2009).

The results showing age differences in network properties at the level of individual modules complement and further elucidate the patterns of brain-wide results. Although DMN has traditionally been the most investigated resting-state network (Ferreira and Busatto, 2013; Damoiseaux, 2017), recent investigations also point to CON changes as prominent features of healthy aging (Meier et al., 2012; Onoda et al., 2012; He et al., 2014; La Corte et al., 2016). The cingulo-opercular network (or salience network, in alternative taxonomies) is anchored in the anterior cingulate and frontal operculum/anterior insula regions, and has been implicated both in stable set-maintenance (Dosenbach et al., 2006, 2007, 2008; Power and Petersen, 2013) and multimodal sensory integration (Seeley et al., 2007; Bressler and Menon, 2010; Menon, 2011). The present results, showing both higher participation coefficients and lower local efficiency and intra-module connectivity for CON in OA, suggest agerelated dedifferentiation of this network and support the idea of changes in CON functionality as a hallmark of healthy aging (Meier et al., 2012; Onoda et al., 2012; He et al., 2014; La Corte et al., 2016). Greater participation coefficients for CON and SMN in OA indicate greater propensity of the nodes within these two networks to form links outside their own modules, and suggest that CON and SMN may drive the observed age differences in brain-wide modularity. Furthermore, local efficiency in CON was also consistently lower in OA, suggesting an age-related decline in local integration of information at the level of this network. In addition to CON, DMN showed consistent lower intra-module connectivity in OA relative to YA, in line with previous evidence (Andrews-Hanna et al., 2007; Damoiseaux et al., 2008; Ferreira and Busatto, 2013; Geerligs et al., 2015; Grady et al., 2016; Damoiseaux, 2017). Interestingly, our results did not show greater FPN-DMN inter-network connectivity in OA relative to YA (Geerligs et al., 2015; Turner and Spreng, 2015), which might have been related to the inclusion of relatively younger, high-functioning OA in our sample. Supporting this interpretation, a recent longitudinal study in OA (Ng et al., 2016) identified a u-shaped trajectory in which FPN-DMN inter-network connectivity initially decreased and then increased with age, with a turning point around 65–70 years of age. An alternative interpretation is that the functional interactions between FPN and DMN in OA might have been influenced by residual task-effects, as outlined above.

Regarding links between network measures and learning rates during training, the present results showed that higher restingstate modularity and local efficiency, as well as lower global efficiency prior to training, were associated with better early learning in OA. Early learning rates are thought to reflect the initial attainment of peak performance within an individual's baseline performance range, rather than plasticity per se (Lövden et al., 2010). Notably, associations between network properties and early learning rates were observed only for OA and only at t1. While the presence of these associations only in OA could be interpreted in line with evidence pointing to agerelated dissociations in the relations between network efficiency and cognitive performance (Stanley et al., 2015), the lack of consistency of these relations across time might be attributable to differences in residual task-effects related to the phenomenon of task exposure which may have, in turn, influenced the reliability of network measures across time. Specifically, if task exposure altered strategies for WM task performance across the two sessions for older but not for younger adults, the resting-state activity, which was always recorded subsequent to task performance, may have been differentially affected. Future analyses comparing the effects of task exposure on differences between task-related and subsequently recorded resting-state network configurations are needed to further clarify this aspect of the results.

Although evidence linking network properties with benefits accrued over the course of cognitive training is scarce, the present results are in line with previous findings showing positive relations for modularity in OA (Gallen et al., 2016a) and in patients with traumatic brain injury (Arnemann et al., 2015). Consistent with the idea that modularity supports both functional segregation and integration, previous evidence has positively linked modularity with cognitive performance (Stevens et al., 2012; Sadaghiani et al., 2015), and thus greater modularity during resting-state may reflect a more "optimal" functional organization that promotes cognitive improvements with training (Gallen et al., 2016a). Results at the level of individual networks add specificity to this interpretation, by associating lower CON participation coefficients with higher learning rates in OA. Combined with evidence showing greater participation coefficients for this network in OA as a group, the present findings provide preliminary evidence for a link between preserved CON segregation and better learning in OA.

The present results linking network properties at rest with learning rates can be interpreted in the light of evidence from investigations of task-related performance (Stanley et al., 2015; Cohen and D'Esposito, 2016; Bolt et al., 2017). Specifically, investigations comparing network properties across resting and task contexts have shown that cognitive task states are characterized by overall lower modularity and local efficiency, as well as greater global efficiency, and that such levels are positively associated with cognitive performance at the individual level in YA (Cohen and D'Esposito, 2016; Bolt et al., 2017). Consistent with this evidence, age-related investigations have linked lower local efficiency during task performance with better WM performance irrespective of age, whereas greater global efficiency was associated with better WM performance in YA but relatively worse WM performance in OA (Stanley et al., 2015). By contrast, prior investigations (Gallen et al., 2016a), as well as the present results, point to a seemingly inverse pattern characterizing the relationship between modularity "at rest" and learning rates in OA, whereby greater modularity and local efficiency, as well as lower global efficiency "at rest", are associated with better learning. Although relations between resting and task-related network configurations are still not well understood, the present evidence suggests that the potential for dynamic network reconfiguration across different states might play an important role for understanding cognition and its plasticity in aging (Cole et al., 2014, 2016; Krienen et al., 2014; Iordan and Reuter-Lorenz, 2016). However, the exploratory nature of these findings advises their interpretation with caution.

#### Limitations and Future Directions

Reliance on extreme groups to understand effects of aging has clear limitations, and thus future work assessing a broader age range (e.g., Chan et al., 2014), as well as longitudinal assessments of the same individuals over periods of years (e.g., Ng et al., 2016), are necessary to provide more comprehensive insights. Regarding the timing of resting-state acquisition, whereas a 6-min break from a preceding task can be a sufficient "wash-out" period for certain individuals under certain task conditions (Breckel et al., 2013), it is not as efficient as longer breaks, and thus restingstate recording before any task should be preferred. Finally, our investigations at the level of individual modules have been partly exploratory. Future studies with strong a priori hypotheses are needed to further elucidate effects of aging on specific within- and between-networks interactions.

#### CONCLUSIONS

In conclusion, we successfully replicated previously reported age effects on resting-state networks, demonstrated their reliability over time, and identified links with initial learning during WM training. We identified both consistencies in network structure and differences in module composition between YA and OA, suggesting topological changes and less stability of functional network structure with aging. Lower modularity and local efficiency in OA suggests age effects on both functional segregation and integration of brain networks, consistent with the idea of age-related functional dedifferentiation. Importantly, these differences were replicable over time, with the difference in local efficiency showing most consistency. On the other hand, global efficiency did not differ between the two age groups and showed low reliability in YA. At the level of individual networks, specific differences were identified for CON, DMN, and SMN, suggesting age-related differential effects at the level of specialized brain modules. Finally, associations between network properties and early learning rates were identified for OA only at t1, suggesting that baseline network configuration

#### REFERENCES


may be informative in predicting aspects of learning in OA, albeit with some limitations. The present findings advance our understanding of the effects of aging on the brain's large-scale functional organization and provide preliminary evidence for network characteristics associated with learning during training. Continued identification of neural mechanisms associated with training-induced plasticity is important for further clarifying whether and how such changes predict the magnitude and maintenance of training gains, as well as the extent and limits of cognitive transfer in both younger and older adults.

#### AUTHOR CONTRIBUTIONS

PR-L, JJ, TP, MB, SJ, BK, KC, KM, and SP designed the study. KC and KM collected the behavioral and brain imaging data, and analyzed the behavioral data. AI analyzed the resting-state brain imaging data and wrote the original draft. All authors reviewed and edited the final manuscript.

#### ACKNOWLEDGMENTS

This research was supported by National Institute on Aging [R21- AG-045460] grant to PR-L. The authors thank Sneha Rajen and KyungJun Kim for assistance with data analysis.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi. 2017.00419/full#supplementary-material


Frontiers in Aging Neuroscience | www.frontiersin.org

Campbell and Schacter (2016). Lang. Cogn. Neurosci. 32, 674–677. doi: 10.1080/23273798.2016.1242759


using aCompCor. Neuroimage 96, 22–35. doi: 10.1016/j.neuroimage.2014. 03.028


**Conflict of Interest Statement:** MB is employed at the MIND Research Institute, whose interest is related to this work. SJ has an indirect financial interest in the MIND Research Institute.

The other authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Iordan, Cooke, Moored, Katz, Buschkuehl, Jaeggi, Jonides, Peltier, Polk and Reuter-Lorenz. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

### APPENDIX

#### Learning Rates Calculation

Participants completed 6 blocks of 14 trials during each of the 10 training sessions (s1-s10). The number of letters in each memory set (i.e., set size) remained constant for each block and was determined by the participant's performance in the previous block. The set size increased by one letter if participants' accuracy was >93% on the preceding block and decreased by one letter if their accuracy was <70%. Thus, for each participant, we calculated the average set size across the 6 blocks for each session.

Learning rates were calculated using those average set-sizes achieved during each training session for each participant. According to Lövden et al. (2010), early learning is thought to reflect the initial attainment of peak performance within an individual's baseline performance range, and is therefore associated with a steep slope. Here, we defined early learning rate as the performance change (i.e., slope) between training sessions 1 and 2. On the other hand, later learning requires prolonged training, and is thought to reflect plasticity. Thus, we used the average performance change (i.e., slope) across training sessions 2–10 to reflect later learning rate.

Early and late learning rates were modeled for each individual using a linear spline term with a knot at training session 2 (s2). The knot location was selected by first fitting a negative exponential growth curve to the observed data using the NLME package in R (Pinheiro and Bates, 2000; Rhodes and Katz, 2017). This procedure was used as an alternative to the assignment of knot location based on only visual inspection of the overall training performance curves. The equation for the negative exponential growth curve is: y(t) = ϕ3– (ϕ<sup>3</sup> − ϕ1) exp[-ϕ<sup>2</sup> ∗ t], where t represents the training session, ϕ<sup>1</sup> is the performance at asymptote, ϕ<sup>2</sup> is performance at session 1 (t = 0), and ϕ<sup>3</sup> is the growth rate. The growth rate parameter can be used to determine the inflection point of the curve, which is the session by which half of the total growth has occurred, according to the formula: t(0.5) = log 2/exp(ϕ3). The parameter estimates for the observed data were ϕ<sup>1</sup> = 9.0 (SE = 0.11), ϕ<sup>2</sup> = 5.0 (SE = 0.23), ϕ<sup>3</sup> = -0.17 (SE = 0.14). Using the equation above, the inflection point occurred slightly before s<sup>2</sup> (t = 0.8), and was similar across both age groups. Individual linear spline models with a knot at s<sup>2</sup> were then computed for each participant.

# Age-Dependent Modulations of Resting State Connectivity Following Motor Practice

Elena Solesio-Jofre1,2 \*, Iseult A. M. Beets<sup>1</sup> , Daniel G. Woolley<sup>1</sup> , Lisa Pauwels<sup>1</sup> , Sima Chalavi<sup>1</sup> , Dante Mantini1,3,4 and Stephan P. Swinnen1,5

<sup>1</sup> Movement Control and Neuroplasticity Research Group, Department of Movement Sciences, KU Leuven, Leuven, Belgium, <sup>2</sup> Department of Biological and Health Psychology, Autonomous University of Madrid, Madrid, Spain, <sup>3</sup> Department of Health Sciences and Technology, ETH Zurich, Zurich, Switzerland, <sup>4</sup> Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom, <sup>5</sup> Leuven Research Institute for Neuroscience and Disease, KU Leuven, Leuven, Belgium

Recent work in young adults has demonstrated that motor learning can modulate resting state functional connectivity. However, evidence for older adults is scarce. Here, we investigated whether learning a bimanual tracking task modulates resting state functional connectivity of both inter- and intra-hemispheric regions differentially in young and older individuals, and whether this has behavioral relevance. Both age groups learned a set of complex bimanual tracking task variants over a 2-week training period. Resting-state and task-related functional magnetic resonance imaging scans were collected before and after training. Our analyses revealed that both young and older adults reached considerable performance gains. Older adults even obtained larger training-induced improvements relative to baseline, but their overall performance levels were lower than in young adults. Short-term practice resulted in a modulation of resting state functional connectivity, leading to connectivity increases in young adults, but connectivity decreases in older adults. This pattern of age differences occurred for both inter- and intra-hemispheric connections related to the motor network. Additionally, long-term training-induced increases were observed in intra-hemispheric connectivity in the right hemisphere across both age groups. Overall, at the individual level, the longterm changes in inter-hemispheric connectivity correlated with training-induced motor improvement. Our findings confirm that short-term task practice shapes spontaneous brain activity differentially in young and older individuals. Importantly, the association between changes in resting state functional connectivity and improvements in motor performance at the individual level may be indicative of how training shapes the shortterm functional reorganization of the resting state motor network for improvement of behavioral performance.

#### Keywords: aging, resting state functional connectivity, motor learning, motor network, bimanual coordination

### INTRODUCTION

A large body of research has shown reduced abilities in motor skill performance and learning with age (Swinnen, 1998; Serrien et al., 2000; Bangert et al., 2010; Seidler et al., 2010; Solesio-Jofre et al., 2014; Pauwels et al., 2015; Serbruyns et al., 2015). As almost every motor skill used in daily life requires practice before being efficiently implemented, it is crucial to understand

#### Edited by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece

Reviewed by: Rui Li, Institute of Psychology (CAS), China Arun Bokde, Trinity College, Dublin, Ireland

> \*Correspondence: Elena Solesio-Jofre elena.solesio@uam.es

Received: 19 September 2017 Accepted: 22 January 2018 Published: 06 February 2018

#### Citation:

Solesio-Jofre E, Beets IAM, Woolley DG, Pauwels L, Chalavi S, Mantini D and Swinnen SP (2018) Age-Dependent Modulations of Resting State Connectivity Following Motor Practice. Front. Aging Neurosci. 10:25. doi: 10.3389/fnagi.2018.00025

the neural mechanisms by which motor skills are learned and how they are modified with age. Behavioral research has shown that the motor learning process follows different stages. First, during an early acquisition stage considerable improvement in performance is achieved within a relatively short time, i.e., within a session. Second, in a later phase, performance becomes stable, with subtler training-induced improvement that involves consolidation processes. This is achieved over a longer period of time, i.e., between sessions spread over a period of several weeks (Ungerleider et al., 2002; Floyer-Lea and Matthews, 2005).

Converging evidence suggests that, although older adults are still able to acquire new motor skills, they may experience difficulties with the consolidation of acquired representations that occur in the later phase of learning (Voelcker-Rehage and Alberts, 2007; Brown et al., 2009; Wilson et al., 2012; Fogel et al., 2014; Pauwels et al., 2015; Mary et al., 2017). Neuroimaging research has shown a functional reorganization of different neural networks subtending motor performance in young individuals, involving neural plasticity mechanisms (Ma et al., 2010; Albouy et al., 2015). Among other regions, these networks include both inter- and intra-hemispheric connections between the supplementary motor area (SMA), the premotor cortex (PM), and the primary motor cortex (M1) (Donchin et al., 1998; Byblow et al., 2007; O'Shea et al., 2007; Talelli et al., 2008; Hinder et al., 2011). However, a reduced motor plasticity with progressing age may be responsible for the observed deficits in consolidation processes (Todd et al., 2010; Freitas et al., 2011; May and Zwaan, 2017).

Besides task training-induced functional activation changes, recent research has devoted increasing attention to resting state functional connectivity (rs-FC) as a reliable indicator of functional reorganization of brain networks supporting different mental processes (Biswal et al., 1995; Fox and Raichle, 2007; Greicius, 2008; van den Heuvel and Hulshoff Pol, 2010; Ma et al., 2011; van Dijk et al., 2017). Resting state functional magnetic resonance imaging (rs-fMRI) measures the large-scale covariance of low frequency spontaneous fluctuations in the blood oxygen level-dependent (BOLD) signal during rest. The strength of the correlation reflects the degree of functional connectivity between two or more brain regions.

Resting state functional magnetic resonance imaging studies investigating changes in functional connectivity following motor learning, have demonstrated that rs-FC can be modulated within and between training sessions in young individuals (Woolley et al., 2015). In one of the earliest studies, Albert et al. (2009) found that initial learning modulated both a fronto-parietal and a cerebellar resting state network. In a more confined motor network, Tung et al. (2013) showed that initial motor learning modulated functional connectivity between the right and left motor cortices, exhibiting increases in post-task compared to pretask periods. Looking into a later phase of learning, increases in rs-FC in the superior parietal cortex (Daselaar et al., 2010) and in the postcentral and supramarginal gyrus (Ma et al., 2011) were observed. In short, these and other studies have shown that both short-term (Waites et al., 2005; Barnes et al., 2009; Stevens et al., 2010; Gregory et al., 2014; Sami et al., 2014; Mary et al., 2017) and long-term (Buchel et al., 1999; Voss et al., 2008; Tambini et al., 2010; Taubert et al., 2011; Yoo et al., 2013; Hardwick et al., 2015; Sampaio-Baptista et al., 2015; Woolley et al., 2015; Mehrkanoon et al., 2016; Amad et al., 2017; May and Zwaan, 2017) learning effects can modulate rs-FC in young individuals.

Importantly, rs-FC has also been shown to correlate with motor improvement in young adults (Ma et al., 2011; Vahdat et al., 2011; Wu et al., 2014; Zhang et al., 2014), indicating that functional network reorganization can, to some extent, predict behavioral changes (Tambini et al., 2010; Deco and Corbetta, 2011; Wu et al., 2014). However, results about motor traininginduced modulation of resting state networks in older adults are very scarce, with only one study to date showing age-related rs-FC changes following motor sequence learning (Mary et al., 2017).

Here, we investigated whether and how motor skill acquisition and consolidation of a bimanual tracking task (BTT) (Solesio-Jofre et al., 2014; Serbruyns et al., 2015; Santos Monteiro et al., 2017) modulates rs-FC within a task-related motor network in young and older adults. Resting state activity was obtained across four scans (**Figure 1A**): two scans before a motor training protocol conducted over the course of 2 (one scan before a task-related fMRI scan and the other scan after the task-related fMRI scan), and two scans following completion of the motor training protocol (again, one scan before a task-related fMRI scan and the other scan after the task-related fMRI scan). The motor network was selected based on the results of a task-related fMRI study in which the same tracking task was used (Santos Monteiro et al., 2017). Due to the bimanual nature of the task, both inter- and intra-hemispheric functional connectivity were examined. It is well established that bimanual coordination relies on coupling between motor areas of both cerebral hemispheres (Serrien, 2008) through the corpus callosum (Gooijers et al., 2013; Gooijers and Swinnen, 2014). Moreover, learning a new bimanual coordination pattern results in changes in both intra- and interhemispheric coherence between pairs of motor regions, as shown by EEG studies (Andres et al., 1999; Gerloff and Andres, 2002; Serrien, 2009). In the current study, we specifically considered both homotopic (i.e., geometrically corresponding regions in each hemisphere) and non-homotopic inter-hemispheric, as well as right and left intra-hemispheric connectivity patterns. To the best of our knowledge, this is the first study examining motor training-induced modulations in both inter-and intrahemispheric rs-FC as a function of aging during the early and late learning phase. Based on the study from Mary et al. (2017), we predicted an age-dependent reorganization of the motor network, not only immediately but also weeks after initial practice, with more prominent changes anticipated after the former.

Finally, to investigate the behavioral relevance of traininginduced changes in functional connectivity, we correlated changes in rs-FC with bimanual task improvement over the course of learning. In accordance with previous studies (Ma et al., 2011; Taubert et al., 2011; Vahdat et al., 2011; Zhang et al., 2011, 2014; Wu et al., 2014), we expected both inter- and

intra-hemispheric connectivity increases to be associated with improvements on the bimanual task.

## MATERIALS AND METHODS

#### Participants

Twenty-six healthy young adults (YA) and 25 older adults (OA) participated in the study. All participants had normal or corrected-to-normal vision, and were right-handed according to the Edinburgh Handedness Inventory (Oldfield, 1971). They were naive with respect to the experimental paradigm. None of the participants reported a history of neurological, psychiatric, or vascular disease. Older participants above 60 years old (N = 25) were screened for cognitive impairments with the Montreal Cognitive Assessment test (MoCA) using the standard cutoff score of 26 (Nasreddine et al., 2005). All participants obtained a score within normal limits (≥26, mean = 28.02, SD = 1.18, range = 26–30). Three YA were excluded from the analysis due to technical problems with the scanner. Four OA were excluded due to either brain atrophy/lesions, or inability to comply with task instructions. As a result, our final sample included 23 YA (age range = 17–26 years, mean age = 21.19, SD = 1.99, 12 females) and 21 OA (age range = 61–81 years, mean age = 68.85, SD = 5.89, 12 females). Informed consent was obtained before testing and participants were financially compensated for participation. The experiment was approved by the local ethics committee for biomedical research of KU Leuven (Belgium), and was performed in accordance with the Declaration of Helsinki (1964).

### Experimental Setup

Magnetic resonance imaging (MRI) occurred twice: before (pretest session) and after (post-test session) five training sessions (training period), distributed across 2 weeks (see **Figure 1A**). Each training session had a total duration of 1 h. The resttask-rest fMRI design of both scan sessions was identical with a total duration of 1.5 h. Therefore, the overall experimental procedure was as follows: the pre-test session included a rest scan (rs1), followed by a task-related scan (tr1), after which another rest scan (rs2) was obtained. Hence, rs1 and rs2 referred to the within pre-test session rest scans corresponding to the early phase of learning. The pre-test session was followed by a bimanual task training period of 2 weeks (T1, T2, T3, T4, T5). Following completion of 2 weeks of task training, the post-test session included a rest scan (rs3), followed by a taskrelated scan (tr2), and subsequently another rest scan (rs4). Hence, rs3 and rs4 referred to the within post-test session rest scans corresponding to the late phase of learning. The present study mainly focused on rest scans (i.e., two runs within each scan session, four runs in total: rs1, rs2, rs3, and rs4). All rest scans had the same protocol and lasted 8 min, in which participants were instructed to keep their eyes open and to fixate a target point. Results regarding the task-related

fMRI study are published elsewhere (Santos Monteiro et al., 2017).

#### Bimanual Tracking Task (BTT)

The BTT task was performed during the task-related scans. It enables the evaluation of bimanual coordination accuracy, relying on the execution of complex bimanual patterns (Sisti et al., 2011). This task requires intensive practice to successfully integrate the two separate limb movements into one common spatiotemporal pattern. Learning such a task involves breaking away from the natural tendency to move both limbs in phase with the same velocity, i.e., a 1:1 frequency ratio (Swinnen et al., 1997; Swinnen, 2002).

The goal of the BTT was to track a white target dot over a blue target line, presented on a screen, by rotating two dials with both hands simultaneously in one of four directional patterns: both hands rotated inward (IN) or outward (OUT) together, or in a clockwise (CW) or counter-clockwise manner (CCW) (Sisti et al., 2011, 2012; Gooijers et al., 2013). The left (L) and right (R) hands controlled movements on the ordinate and abscissa, respectively. To increase complexity of the task, each directional combination was performed at five different relative frequency ratios: 1:1, 1:2, 1:3, 2:1, and 3:1 (L:R). For example, a 1:2 ratio indicated that the left hand was required to rotate twice as slow as the right hand. This resulted in 20 different bimanual patterns and target line directions (**Figure 1B**).

Each trial started with the presentation of the single blue target line with a distinct orientation. At the origin of this line, in the center of the PC display, the white target dot was presented, after which it began to move along the blue target line, toward the peripheral endpoint. The target dot moved at a constant rate and for a total duration of 9 s. The beginning and end of the trajectory were marked with an auditory cue (126 ms, begin: 525 Hz, end: 442 Hz). The inter-trial interval was 3 s. The goal was to match the target trajectory as closely as possible.

Each BTT fMRI session consisted of 144 trials, divided equally across six runs, each of which lasted 6 min, with an interrun interval of approximately 3 min. A run consisted of 24 target lines, presented in a pseudorandom order. The required frequency ratio was randomly distributed such that one third of trials required a 1:1 ratio, one-third required a 1:2 or 2:1 ratio and one-third required a 1:3 or 3:1 ratio. There were 96 "move" trials in which bimanual tracking was actively performed. The remaining trials were "no move" trials, containing the same information as the "move" trials but required no movement. They provided the baseline measure for the BOLD contrasts conducted in the task-related fMRI analysis (Santos Monteiro et al., 2017). Prior to the first MRI session, participants practiced the task briefly in a dummy scanner until the task was fully understood (∼10 min).

#### Training Sessions

In the training sessions, participants were seated in front of a PC-screen (distance approximately 0.5 m) and performed the BTT. For each of the five training days, 10 blocks of 20 randomized trials corresponding to 20 bimanual patterns, i.e., five different frequency ratios in four directions, were performed.

#### Kinematic Analyses

Data were recorded and analyzed with the Labview software (version 8.5, National Instruments, Austin, TX, United States). Offline analysis was carried out using Matlab R2011b. The x and y positions of the target dot and the cursor were sampled at 100 Hz. For each trial, we calculated the target deviation as a measure of accuracy, using the following multistep procedure: (a) Every 10 ms, the difference between the target position and the cursor position, d, was calculated, using the Euclidean distance:

$$d = \sqrt{(\varkappa\_2 - \varkappa\_1)^2 + (\wp\_2 - \wp\_1)^2}$$

Where x<sup>2</sup> and y<sup>2</sup> refer to the position of the participant's cursor on the x- and y-axis, respectively, and x<sup>1</sup> and y<sup>1</sup> correspond to the position of the target dot on the x- and y-axis, respectively. (b) At the end of each trial, the average of these distances was computed and defined as the trial's target deviation, expressed in units (U). A target deviation equal to 0 U would indicate that during the whole trial, the cursor was precisely on top of the white target dot, representing perfect performance. Accordingly, greater target deviation scores reflect greater error and, hence, poorer performance.

To determine whether participants generally met the task requirements, all data were transformed into z-scores [(X-MEAN/SD)]. Trials were discarded from the analysis when z > 3 (outlier) and/or when only one hand moved (2.7 and 1.1% of all trials during BTT fMRI 1 and BTT fMRI 2, respectively). For each participant, the average error scores were computed for both scan sessions with and without augmented visual feedback, and these error scores were used as an indicator of bimanual performance accuracy.

#### Statistical Analyses

Statistical analyses were performed using SPSS Version 22.0 (Armonk, NY, United States).

In accordance with previous results from our own group using the BTT task (Gooijers et al., 2013, 2016; Solesio-Jofre et al., 2014; Beets et al., 2015), movement directions (IN, OUT, CW, and CCW) were fully counterbalanced in the design and of no interest for the present analyses. Additionally, we collapsed trials into two levels: trials with the same (ISO, 1:1) and trials with different (N-ISO, 1:2, 1:3, 2:1, 3:1 collapsed) cycling frequency ratios.

Behavioral data acquired during both the pre- and posttest session were subjected to a 2 × 2 × 2 (age × scan session × frequency ratio) repeated measures ANOVA. Here, age (young, older) was the between-subject factor, and scan session (pre-test session and post-test session) and frequency ratio (ISO, N-ISO) were the within-subject factors.

The level of significance was set at p < 0.05. Significant effects were further explored using post hoc paired t-tests using Bonferroni correction for multiple comparisons. The partial eta squared statistic (η 2 p ) was calculated as the effect size measure for main and interaction effects in the repeated measures ANOVA. According to Cohen (1992), η 2 p values of 0.01, 0.06 and 0.13 represent small, medium and large effects, respectively.

#### MRI Data Acquisition

fnagi-10-00025 February 5, 2018 Time: 17:22 # 5

Data acquisition, pre-processing, and analyses followed the same steps for the four resting state runs (rs1, rs2, rs3, and rs4). A Siemens 3-T Magnetom Trio MRI scanner (Siemens, Erlangen, Germany) with a 12 channel head coil was used. For anatomical details, a 3D high-resolution T1-weighted image was obtained first (magnetization prepared rapid gradient echo, time repetition/time echo = 2300/2.98 ms, 1 mm × 1 mm × 1.1 mm voxels, field of view (FOV) = 240 × 256, 160 sagittal slices), lasting 8 min. Then a field map was acquired to address local distortions.

Functional resting state data were acquired with a descending gradient echo planar imaging (EPI) pulse sequence for T2 <sup>∗</sup> − weighted images (repetition time = 3,000 ms; echo time = 30 ms; flip angle = 90◦ ; 50 oblique axial slices each 2.8 mm thick; inter-slice gap = 0.028 mm; in-plane resolution 2.5 mm × 2.5 mm; 80 × 80 matrix, 160 volumes).

#### MRI Data Pre-processing

Standard preprocessing procedures were performed using SPM8 (Statistical Parametric Mapping software, SPM: Wellcome Department of Imaging Neuroscience, London, United Kingdom<sup>1</sup> ), which is implemented in Matlab 7.7 (The Mathworks, Natick, MA, United States).

Functional images were slice-time corrected to the middle slice (reference slice = 25), spatially realigned to the first image in the time series, normalized to the standard EPI template in Montreal Neurological Institute (MNI) space, and resampled into 3 mm isotropic voxels (Friston et al., 1995). Spatial smoothing was not applied in order to avoid introducing artificial local spatial correlations (Salvador et al., 2005; Achard et al., 2006; Achard and Bullmore, 2007).

We took additional preprocessing steps to remove spurious sources of variance. We defined a small, bilateral region of interest in the ventricles, a region of interest in the deep white matter, and one covering the whole brain; we then calculated the average signals in these three regions, which are typically referred to as cerebrospinal fluid, white matter and global signals (Fox et al., 2005, 2009). Next, we performed a regression analysis on the fMRI time-courses, modeling the three aforementioned signals and the parameters obtained by rigid body head motion realignment (Fox et al., 2005, 2009), as well as their temporal derivatives, as regressors.

Recently, there has been considerable discussion over the impact of head motion on rs-FC connectivity analyses. In addition to regressing out the three-dimensional motion parameters and their first derivatives, we also included regressors to deweight scans with a framewise displacement greater than 0.5 mm. A separate regressor was included for each outlier scan, with a 1 at the outlier time point and a zero at all other time points. Framewise displacement was calculated as the sum of the absolute scan to scan difference of the six translational and rotational realignment parameters (Power et al., 2014). Only 0.9% of all scans exceeded this threshold, and there was no significant difference in mean framewise displacement between the four resting state scans [one-way ANOVA: F(3,172) = 1.90, p = 0.14)].

<sup>1</sup>http://www.fil.ion.ucl.ac.uk/spm/

The BOLD time course in each voxel was then temporally band-pass filtered (0.01–0.08 Hz) to reduce low-frequency drift and high-frequency noise.

#### Region Definition

Candidate ROIs were generated from task-related fMRI scans (Santos Monteiro et al., 2017), in which the main aim was to explore the effects of aging on brain plasticity associated with motor learning while subjects performed the BTT. The main BOLD contrast of interest was bimanual visuomotor task performance (movement) vs. baseline condition (no movement) in young and older adults. Young and old z-score maps from the task-based fMRI study were combined to find overlapping ROIs by means of a conjunction analysis [young (bimanual visuomotor task > baseline) ∩ old (bimanual visuomotor task > baseline)] (Nichols et al., 2005). The statistical threshold was set to p < 0.05, FWE corrected for multiple comparisons and a minimal cluster size of 20 voxels.

To define regions for our resting state connectivity analysis, we chose the peak voxel with the highest z-score (z ≥ 4.10) in the positive group analysis. Our ROIs were composed of 6-mm radius spheres centered on these peak voxels and were created using the MarsBAR toolbox<sup>2</sup> . The size of the spheres was selected to ensure that they contained voxels that were significantly activated in all cases. We defined the following a priori ROIs: SMA (R, L); dorsal premotor area (PMd: R, L); ventral premotor area (PMv: R, L); primary motor cortex (M1: R, L); and primary somatosensory area (S1: R, L). ROI coordinates are listed in **Table 1**, and are illustrated in **Figure 2**.

#### Functional Connectivity Analysis

For each subject and within each of the four resting state scans, regional mean time series were extracted by averaging the functional MRI time series across all voxels within each ROI. Then, the correlation strength between every pair of ROIs was calculated using Pearson correlation coefficients creating a functional network captured by a 10 × 10 correlation matrix. These Pearson correlation values were converted to Z-scores by Fischer'sr-to-z transformation (Zar, 1998), correcting the degrees of freedom for the autocorrelation in the time series (Shumway and Stoffer, 2006). Group-level correlation matrices were created by using a random-effects analysis across subjects (Ebisch et al., 2011; Pravata et al., 2011).

Next, we calculated the average connectivity score for a group of ROI pairs, which is the average of the component Fisher Z-scores for the corresponding ROI pairs. Specifically, we report four kinds of average functional connectivity (FC) scores, including homotopic inter-hemispheric FC, heterotopic inter-hemispheric FC, right intra-hemispheric FC and left intrahemispheric FC. Average connectivity scores were subjected to repeated measures ANOVAs. We conducted a 2 × 2 × 2 × 2 (age × inter-hemispheric FC × scan session × scan location) repeated measures ANOVA, with age (young, older) as the between-subject factor and inter-hemispheric FC (homotopic, heterotopic), scan session (pre-test session, post-test session)

<sup>2</sup>http://marsbar.sourceforge.net

TABLE 1 | Regions defined for the resting state motor network.


Regions obtained after a conjunction analysis (young ∩ old) from a task-based fMRI study, pFWE < 0.05, z ≥ 4.10. Six-mm radius spheres centered on z-peak voxels. R, right; L, left; SMA, supplementary motor area; PMd, dorsal premotor area; M1, primary motor cortex; S1, primary somatosensory area; PMv, ventral premotor area.

and scan location (before task-related scan, after task-related scan) as the within-subject factors. Additionally, we conducted a 2 × 2 × 2 × 2 (age × intra-hemispheric FC × scan session × scan location) repeated measures ANOVA, with age (young, older) as the between-subject factor and intra-hemispheric FC (right, left), scan session (pre-test session, post-test session) and scan location (before task-related scan, after task-related scan) as the withinsubject factors. All statistical tests were completed with alpha set at 0.05, and significant interaction effects were further explored by post hoc paired t-tests using Bonferroni correction for each repeated measures ANOVA conducted. The partial eta squared statistic (η 2 p ) was calculated as the effect size measure for main and interaction effects in the repeated measures ANOVA and the size of the effects was interpreted according to Cohen (1992).

Finally, we calculated brain-behavior correlations to determine the extent to which training-induced changes in inter- and intra-hemispheric connectivity corresponded with the subsequent gain in behavioral performance, considering the entire sample of participants. We used two different calculations to quantify training-induced changes in inter- and intrahemispheric connectivity during the early phase of learning and the overall learning process: (1) the difference in the average rs-FC between the second (rs2) and the first (rs1) resting state scans to study the effect of task learning during the early phase of learning; and (2) the difference in the average rs-FC between the last (rs4) and the first (rs1) resting state scans to investigate the effect of long-term practice.

Similarly, we used two different calculations to quantify the behavioral gain during the early phase of learning and the overall learning process: (a) the difference in the average target deviation between the last block of trials (15 trials in total) and the first block of trials (15 trials in total) in the first task-related fMRI scan (tr1) to study the effect of task practice during the early phase of learning; (b) the difference between the average target deviation in the last block of trials (15 in total) of the second task-related fMRI scan (tr2) and the first block of trials (15 in total) in tr1 to investigate the effect of long-term practice.

Hence, we performed brain-behavior correlations between the following functional connectivity measures: (A) Homotopic connections extracted from rs2 minus rs1 (Hm FC short-term learning) difference 1; (B) Heterotopic connections extracted from rs2 minus rs1 (Ht FC short-term learning); (C) Homotopic connections extracted from rs4 minus rs1 (Hm FC longterm learning); (D) Heterotopic connections extracted from rs4 minus rs1 (Ht FC long-term learning); (E) Right hemispheric connections extracted from rs2 minus rs1 (R FC short-term learning); (F) Left hemispheric connections extracted from rs2 minus rs1 (L FC short-term learning); (G) Right hemispheric connections extracted from rs4 minus rs1 (R FC long-term learning); (H) Left hemispheric connections extracted from rs4 minus rs1 (L FC long-term learning),

BTT gain measures were defined as follows:

(A) The last 15 trials of tr1 minus first 15 trials of tr1 for N-ISO condition (BTT Gain 1); (B) The last 15 trials of tr2 minus the first 15 trials of tr1 for N-ISO condition (BTT Gain 2). We focused on the N-ISO conditions as these represented new unfamiliar patterns requiring practice to improve proficiency whereas the

ISO conditions reflected familiar patterns that constitute the default coordination modes (not requiring elaborate practice) (Sisti et al., 2011). Greater BTT gains reflect larger improvements in performance. **Figure 3** illustrates the correlations computed. Correlations surviving Bonferroni correction (p < 0.025) were considered significant.

#### RESULTS

#### Kinematic Data

#### Scan Sessions

Motor performance during the scan sessions was assessed with a 2 × 2 × 2 (age × scan session × frequency ratio) repeated measures ANOVA for average target deviation. There was a main effect of age [F(1,42) = 49.59, p < 0.0001, η 2 <sup>p</sup> = 0.54], with YA performing better than OA.

There was also a strong learning effect from pre- to posttraining period, reflected by the main effect of scan session [F(1,42) = 109.8, p < 0.0001, η 2 <sup>p</sup> = 0.72], indicating that performance improved as a result of training.

We also observed a main effect of frequency ratio [F(1,42) = 197.50, p < 0.0001, η 2 <sup>p</sup> = 0.83], suggesting that subjects had more difficulty in performing the most difficult (N-ISO) as compared with the easiest (ISO) frequency ratios.

A significant age × scan session interaction effect [F(1,42) = 20.71, p < 0.001, η 2 <sup>p</sup> = 0.33] indicated that, although both age groups were able to significantly improve their performance as a result of training [YA: t(22) = 8.95, p < 0.0001: OA: t(20) = 7.65, p < 0.0001], OA improved their performance to a higher degree as compared to YA from pre-test to posttest session. Furthermore, a significant age × frequency ratio interaction effect was observed [F(1,42) = 10.25, p = 0.003, η 2 <sup>p</sup> = 0.20], reflecting that OA, but not YA, had more difficulty in performing the most difficult (N-ISO) relative to the easiest (ISO) condition [t(42) = −9.03, p < 0.0001].

#### Training Sessions

A 2 × 5 × 2 (age × training session × frequency ratio) repeated measures ANOVA was conducted for the average target deviation scores obtained across training days.

There was a main effect of age [F(1,42) = 33.94, p < 0.0001, η 2 <sup>p</sup> = 0.48], indicating that the overall performance level of YA was better than the one of OA.

The main effect of training session was also significant [F(4,168) = 99.90, p < 0.0001, η 2 <sup>p</sup> = 0.70], suggesting a strong practice effect. Post hoc t-tests revealed that the five sessions differed from each other (all p < 0.001). However, greater differences were observed for the first two sessions as compared to sessions 3, 4, and 5, suggesting that the practice effect was strongest at the first training sessions and a plateau effect was reached toward the final two sessions.

A significant main effect of frequency ratio [F(1,42) = 120.57, p < 0.0001, η 2 <sup>p</sup> = 0.74] reflected greater error rates for N-ISO as compared to ISO ratio.

We observed a significant age × training session interaction effect [F(4,168) = 7.85, p < 0.0001, η 2 <sup>p</sup> = 0.16]. Post hoc t-tests revealed that, although YA had a better performance than OA in all the five training sessions, these age differences were statistically greater during training session 1 compared to sessions 4 [t(42) = −3.29, p < 0.004] and 5 [t(42) = −3.76, p < 0.001], suggesting that as training progressed, the differences in performance between YA and OA decreased.

As not much learning was required for the ISO condition, **Figure 4** focuses on the behavioral performance during both the scan and training sessions for the N-ISO condition.

15 trials of tr1 minus first 15 trials of tr1 (BTT Gain 1), and also last 15 trials of tr2 minus first 15 trials of tr1 (BTT Gain 2) for N-ISO conditions.

FIGURE 4 | Behavioral performance during the scan and training sessions for the N-ISO condition. There was an initial reduction in target deviation error during the pre-test session, indicative of initial learning. During the training period, BTT performance became more stable, particularly during the last two training days. YA showed a more stable performance during the post-test session than OA, especially in the most difficult task condition (N-ISO). Error bars represent the standard error of the mean (SEM). N-ISO, non-isofrequency.

#### Imaging Data

We studied low-frequency functional correlations associated with a task-specific motor network composed of 10 ROIs: left and right SMA, PMd, PMv, M1, and S1. We calculated Pearson correlation coefficients between each pair of ROIs across subjects, and within each of the four resting state scans. **Figure 5** shows the resulting 10 × 10 correlation matrix for each resting state scan, which reflects the strength of functional connectivity between each pair of regions. Next, we calculated average functional connectivity scores regarding four different groups of ROI pairs, including homotopic inter-hemispheric FC, heterotopic interhemispheric FC, right intra-hemispheric FC and left intrahemispheric FC.

#### Modulations of Resting State Inter-Hemispheric Connectivity in Young and Older Adults throughout the Learning Process

The 2 × 2 × 2 × 2 (age × inter-hemispheric FC × scan session × scan location) repeated measures ANOVAs revealed a main effect of inter-hemispheric functional connectivity [F(1,42) = 247.26, p < 0.0001, η 2 <sup>p</sup> = 0.86], with greater homotopic than heterotopic connectivity values (Hm: 6.68 ± 0.29, Ht: 2.52 ± 0.18) (**Figure 6**). Main effects of age group, scan session and phase were not significant.

A significant age × scan location interaction indicated that young and older adults showed a different pattern of rs-FC change as a function of performance of the BTT both during the pre- and post-test sessions [F(1,42) = 9.01, p = 0.005, η 2 <sup>p</sup> = 0.18]. Subsequent post hoc unpaired t-tests demonstrated that rs-FC increased after practicing the BTT, that is from rs1 to rs2 and from rs3 to rs4, in YA, whereas OA showed the opposite pattern [t(42) = −2.21, p = 0.002. YA, pre-task rs: 4.48 ± 0.41, posttask rs: 4.71 ± 0.28; OA, pre-task rs: 5.03 ± 0.33, post-task rs: 4.16 ± 0.30)]. This effect was true for both homotopic and heterotopic connections. **Figure 6** depicts the age × scan location interaction. None of the remaining interactions reached significance.

#### Modulations of Resting State Intra-Hemispheric Connectivity in Young and Older Adults throughout the Learning Process

The 2 × 2 × 2 × 2 (age × intra-hemispheric FC × scan session × scan location) repeated measures ANOVAs revealed no significant main effects of age, intra-hemispheric FC, scan session and learning phase.

There was a significant age × scan location interaction, in which young and older adults showed different patterns of rs-FC change as a function of task practice [F(1,42) = 6.21, p = 0.017, η 2 <sup>p</sup> = 0.13]. Subsequent post hoc unpaired t-tests demonstrated that rs-FC increased after BTT performance in YA, that is from rs1 to rs2 and from rs3 to rs4, whereas OA exhibited the opposite pattern [t(42) = −2.56, p = 0.003.

FIGURE 6 | Bar plots showing the age × scan location interaction effect for inter-hemispheric functional connectivity. (A) Changes in connectivity in homotopic pairs of regions. (B) Changes in connectivity in heterotopic pairs of regions. In both cases, functional connectivity increased after task performance in YA, whereas it decreased in OA and we observed this pattern of results within both the pre- and post-test sessions. Moreover, homotopic functional connectivity was greater than heterotopic functional connectivity. Error bars represent SEM. Hm rs1+rs3, homotopic rest scans before task-related scans; Hm rs2+rs4, homotopic rest scans after task-related scans; Ht rs1+rs3, heterotopic rest scans before task-related scans; Ht rs2+rs4, heterotopic rest scans after task-related scans.

YA, pre-task rs: 2.75 ± 0.22, post-task rs: 3.15 ± 0.23; OA, pre-task rs: 3.23 ± 0.39, post-task rs: 2.70 ± 0.32)]. Of note, this is the same pattern as previously observed for inter-hemispheric functional connectivity. Moreover, this age-related difference in the pattern of functional connectivity occurred for both left and right hemisphere connections. **Figure 7** depicts the age × scan location interaction.

We also observed a significant intra-hemispheric functional connectivity × scan session interaction effect, indicating that functional connectivity within the right and left hemisphere showed a differential change from pre- to post-test session. [F(1,42) = 4.55, p = 0.04, η 2 <sup>p</sup> = 0.10]. Subsequent post hoc paired t-tests revealed training-related increases in functional connectivity in the right hemisphere, but not in the left hemisphere, after the training period [t(43) = 2.62, p = 0.004; Right post-test session: 3.41 ± 0.32; Right pre-test session: 2.89 ± 0.24]. None of the remaining main and interaction effects reached significance.

#### Correlation between Resting State Functional Connectivity and Behavior

We tested whether changes in rs-FC corresponded with gains in behavioral performance as a general tendency across both age groups.

Increases in inter-hemispheric connectivity for both homotopic (Hm FC long-term learning) and heterotopic (Ht FC long-term learning) connections correlated with greater gains in BTT performance (BTT Gain 2 N-ISO) (Hm: r = 0.40, p = 0.010; Ht: r = 0.30, p = 0.04). Of note, the first result survived Bonferroni correction (p < 0.013), whereas the second did not. None of the remaining correlations



Increases in inter-hemispheric connectivity from the first to the last resting state scan for both homotopic and heterotopic connections (Hm FC long-term learning, Ht FC long-term learning) correlated significantly with greater gains in BTT performance from the first 15 trials of the first task-related scan to the last 15 trials of the second task-related scan in the N-ISO condition (Gain 2). Only the bolded values corresponding to the correlation between homotopic rs-FC change and Gain 2 survived Bonferroni correction (p < 0.025). r, Pearson coefficient; p, probability value.

reached significance. See **Table 2** and **Figure 8** for further details.

#### DISCUSSION

coefficient.

We investigated whether learning a bimanual tracking task modulates rs-FC in the early and late phase of practice and whether this has behavioral relevance. Compared with YA, OA showed a lower motor performance in general, but a larger improvement relative to baseline after 2 weeks of training. Shortterm practice effects achieved within pre- and post-test sessions modulated rs-FC, leading to connectivity increases in YA, but connectivity decreases in OA. This pattern of age differences occurred for both inter- and intra-hemispheric connections related to the motor network. We did not observe age-related modulations of long-term practice (i.e., from pre- to post-test session) on interhemispheric rs-FC at the group level. However, long-term training-induced increases were observed in intrahemispheric connectivity in the right hemisphere (across both age groups). Finally, changes in inter-hemispheric functional connectivity from the start to end of practice correlated with training-induced motor improvement, underscoring the behavioral significance of rs-fMRI for prediction of motor skill learning. We discuss these findings in detail below.

#### Aging Effects on Bimanual Coordination Learning

Kinematic analyses revealed that older individuals encountered particular difficulties with performing the bimanual tasks with different frequency ratios. This is a consequence of the higher complexity of these tasks. Strikingly, we observed greater performance improvement in older compared to young adults, which is in accordance with previous studies (Wishart et al., 2002; Voelcker-Rehage and Willimczik, 2006). This may be a consequence of the lower performance levels of older adults at baseline, giving rise to a larger window for improvement. However, both age groups showed a similar overall pattern of improvement: a stronger practice effect during the initial phase of learning (pre-test scan session), followed by a plateau toward the final phase (post-test scan session), reflecting relatively stable performance.

### Functional Connectivity Changes in Young and Older Adults as a Result of Short-Term Practice

We consistently observed an age-related change in rs-FC after short-term practice both during the early and later stages of learning (pre- and post-test sessions, respectively). Specifically, we observed increases in functional connectivity in young adults but decreases in older adults after task performance. Hence, task practice induced a differential short-term functional reorganization within the resting brain in young and older individuals. However, we did not observe an age-related functional reorganization of our motor network over the longer course of training across 2 weeks (from pre- to post-test session). This implies that the involvement in the training protocol did not induce long-term changes in rs-FC, at least not significantly different between both age groups. Hence, our pattern of results does not support our initial hypothesis of an age-dependent reorganization of the motor network that is more pronounced during the early (pre-test session) compared to the later stage of learning (post-test session), with the most dramatic changes in

motor performance occurring in the former stage (Mary et al., 2017).

The findings observed in young adults are generally in accordance with previous research showing short-term effects on rs-FC after task practice (Peltier et al., 2005; Waites et al., 2005; Boonstra et al., 2007; Houweling et al., 2008; Barnes et al., 2009; Stevens et al., 2010; Klingner et al., 2012; Tung et al., 2013). In this regard, it seems reasonable to suggest that the short-term rs-FC increases following the execution of the motor task observed in young adults may reflect tighter communication between motor network areas. Although speculative, these processes may entail sensorimotor integration (Loayza et al., 2011; Ma et al., 2011; Wu et al., 2014) and short-term storage of visuomotor skills (Johnson-Frey et al., 2005; Ma et al., 2011).

The rs-FC decreases due to short-term practice, as observed in older adults, are more difficult to understand. In an attempt to come to grips with this finding, it is important to bear in mind that network organization differs between young and older adults and this pertains to within-network as well as between-network rs-FC. Previous work has shown that rs-FC within the motor network is higher in older as compared to younger adults and this is negatively associated with motor performance (Solesio-Jofre et al., 2014). Furthermore, FC among the different resting state networks is also increased in older adults, pointing to reduced overall network segregation (Geerligs et al., 2014; Chan et al., 2017; King et al., 2017), associated with poorer cognitive performance or no performance benefits at all (Nashiro et al., 2017). This is supportive of a dedifferentiation process, referring to an age-related diminished specificity in the cortical response to some stimulus categories and a reduction in the fidelity of neural representations (Grady et al., 1992; Park et al., 2004; Voss et al., 2008). Against the backdrop of these age-related changes, it is conceivable that training led to a reorganization of the motor network by strengthening interactions with other networks that have functional relevance for the task (such as attention or executive networks) and reducing interactions with other networks that are less or not relevant for task performance (such as the default mode network). This process of increasing efficiency of brain activity may be associated with a temporary reduction in FC between areas of the motor network, as observed in our findings. However, this is a hypothetical account that requires confirmation in future research.

### Age-Related Modulations of Inter- and Intra-Hemispheric Functional Connectivity

Interestingly, we observed the same pattern of age group differences for both inter- and intra-hemispheric connections. Regarding the former, it is important to remark that we observed greater homotopic relative to heterotopic functional connectivity values. Increased functional and structural homotopic connections have been commonly reported during the acquisition of bimanual coordination skills. On the basis of other studies employing rs-fMRI, this tendency for homotopic regions to exhibit stronger functional connectivity relative to heterotopic regions is more prevalent for regions of the adult human brain, such as motor, somatosensory, visual and subcortical regions (Zuo et al., 2010; Ruddy et al., 2017). In the same vein, previous research in non-human primates (Rouiller et al., 1994; Dancause et al., 2007) and humans (Ruddy et al., 2017) has reported that the largest proportions of interhemispheric fibers connecting to M1, SMA and PM originate in their homologous region in the contralateral hemisphere.

Finally, we observed that the strength of rs-FC within the right hemisphere increased after 2 weeks of training in both age groups, highlighting the importance of increased interregional interactions in the right hemisphere during motor skill learning in general and bimanual skill learning in particular (Halsband and Lange, 2006; Ma et al., 2011). This finding may reflect increased functional interactions among the right hemisphere motor areas to control the less skilled non-dominant hand as part of the unified bimanual kinematic chain. Successful coordination is contingent on the cooperation between both hemispheres/hands and this requires elevation of non-dominant hand function to better interact with the dominant hand.

### Behavioral Relevance for Learning-Related Modulations in Functional Connectivity

Even though neither of the groups exhibited a change in interhemispheric rs-FC across long-term practice, we observed an association between FC change and behavioral change across long-term practice at the individual level. Increases in the strength of inter-hemispheric connections across the 2 week period were associated with higher motor improvement in both young and older individuals, also across the 2-week period, suggesting that the resting state motor network may support the functional reorganization of the motor system in order to improve behavioral performance and ultimately, motor consolidation. This is in accordance with previous studies suggesting a role for resting state networks in terms of keeping active relevant functional systems to improve behavioral performance (Raichle et al., 2001; Miall and Robertson, 2006; van den Heuvel and Hulshoff Pol, 2010). Specifically, resting state networks become relevant functional units that may be recruited whenever a task needs extra support to be successfully performed.

In the light of these results, we suggest that resting state networks do not simply reflect physiological markers of anatomical pathways, but represent highly efficient modules of brain organization, somehow capable of predicting behavioral performance and improvement following motor learning.

#### CONCLUSION

In this study, we provided behavioral evidence that motor learning capacity is preserved in aging. Furthermore, we demonstrated that shorter-term practice modulates the resting state differentially in young and older individuals. On the one hand, increased within-network connectivity after task

practice observed in young adults may be indicative of enhanced interactions between motor network areas that are engaged in motor learning. On the other hand, an agerelated reduction in within-network connectivity after task practice may be an indirect consequence of how the motor network interacts with other networks to optimize overall brain activity for task performance. Older adults may necessitate extra resources to learn the motor task, becoming more dependent on cognitive processes embodied in motor learning. Hence, further research is warranted to shed light on the behavioral relevance of functional interactions across the motor, memory and attentional networks during the resting state.

#### ETHICS STATEMENT

This protocol was approved by the local ethics committee for biomedical research and subjects gave written informed consent prior to participation.

#### REFERENCES


#### AUTHOR CONTRIBUTIONS

ES-J performed the data analyses and wrote the manuscript. IB conducted the data collection. DW helped with data analyses and language check. LP helped with behavioral analysis. SC helped with behavioral analysis. DM developed the pipeline for data analysis and helped with the writing of the manuscript. SS is the principal investigator of this project and helped with the writing of the manuscript.

#### FUNDING

This work was supported by the Research Program of the Research Foundation – Flanders (FWO) [G.0708.14, G.0898.18N], the Wellcome Trust [101253/Z/13/Z], the Interuniversity Attraction Poles Program of the Belgian Federal Government [P7/11], the Special Research Fund KU Leuven [C16/15/070], and Excellence of Science grant (EOS, 30446199, MEMODYN).




**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Solesio-Jofre, Beets, Woolley, Pauwels, Chalavi, Mantini and Swinnen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Aerobic and Cognitive Exercise Study (ACES) for Community-Dwelling Older Adults With or At-Risk for Mild Cognitive Impairment (MCI): Neuropsychological, Neurobiological and Neuroimaging Outcomes of a Randomized Clinical Trial

Cay Anderson-Hanley <sup>1</sup> \*, Nicole M. Barcelos <sup>1</sup> , Earl A. Zimmerman<sup>2</sup> , Robert W. Gillen<sup>3</sup> , Mina Dunnam<sup>4</sup> , Brian D. Cohen<sup>5</sup> , Vadim Yerokhin<sup>6</sup> , Kenneth E. Miller <sup>7</sup> , David J. Hayes <sup>1</sup> , Paul J. Arciero<sup>8</sup> , Molly Maloney <sup>1</sup> and Arthur F. Kramer <sup>9</sup>

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Despina Moraitou, Aristotle University of Thessaloniki, Greece Aristea Kyriaki Ladas, University of Sheffield, United Kingdom*

\*Correspondence:

*Cay Anderson-Hanley andersoc@union.edu*

Received: *08 September 2017* Accepted: *07 March 2018* Published: *04 May 2018*

#### Citation:

*Anderson-Hanley C, Barcelos NM, Zimmerman EA, Gillen RW, Dunnam M, Cohen BD, Yerokhin V, Miller KE, Hayes DJ, Arciero PJ, Maloney M and Kramer AF (2018) The Aerobic and Cognitive Exercise Study (ACES) for Community-Dwelling Older Adults With or At-Risk for Mild Cognitive Impairment (MCI): Neuropsychological, Neurobiological and Neuroimaging Outcomes of a Randomized Clinical Trial. Front. Aging Neurosci. 10:76. doi: 10.3389/fnagi.2018.00076* *<sup>1</sup> The Healthy Aging and Neuropsychology Lab, Union College, Schenectady, NY, United States, <sup>2</sup> Alzheimer's Disease Center, Albany Medical Center, Albany, NY, United States, <sup>3</sup> Sunnyview Rehabilitation Hospital, Schenectady, NY, United States, <sup>4</sup> Stratton VA Medical Center, Albany, NY, United States, <sup>5</sup> Department of Biology, Union College, Schenectady, NY, United States, <sup>6</sup> Biomedical Sciences Department, Oklahoma State University, Tulsa, OK, United States, <sup>7</sup> Department of Anatomy and Cell Biology, Oklahoma State University, Tulsa, OK, United States, <sup>8</sup> Department of Health & Human Physiological Sciences, Skidmore College, Saratoga Springs, NY, United States, <sup>9</sup> Beckman Institute, University of Illinois, Urbana-Champaign, Champaign, IL, United States*

Prior research has found that cognitive benefits of physical exercise and brain health in older adults may be enhanced when mental exercise is interactive simultaneously, as in exergaming. It is unclear whether the cognitive benefit can be maximized by increasing the degree of mental challenge during exercise. This randomized clinical trial (RCT), the Aerobic and Cognitive Exercise Study (ACES) sought to replicate and extend prior findings of added cognitive benefit from exergaming to those with or at risk for mild cognitive impairment (MCI). ACES compares the effects of 6 months of an *exer-tour* (virtual reality bike rides) with the effects of a more effortful *exer-score* (pedaling through a videogame to score points). Fourteen community-dwelling older adults meeting screening criteria for MCI (sMCI) were adherent to their assigned exercise for 6 months. The primary outcome was executive function, while secondary outcomes included memory and everyday cognitive function. Exer-tour and exer-score yielded significant moderate effects on executive function (Stroop A/C; *d*'s = 0.51 and 0.47); there was no significant interaction effect. However, after 3 months the exer-tour revealed a significant and moderate effect, while exer-score showed little impact, as did a game-only condition. Both exer-tour and exer-score conditions also resulted in significant improvements in verbal memory. Effects appear to generalize to self-reported everyday cognitive function. Pilot data, including salivary biomarkers and structural MRI, were gathered at baseline and 6 months; exercise dose was associated with increased BDNF as well as increased gray matter volume in the PFC and ACC. Improvement in memory was associated with an increase in the DLPFC. Improved executive function was associated with increased expression of exosomal miRNA-9. Interactive physical and cognitive exercise (both high and low mental challenge) yielded similarly significant cognitive benefit for adherent sMCI exercisers over 6 months. A larger RCT is needed to confirm these findings. Further innovation and clinical trial data are needed to develop accessible, yet engaging and effective interventions to combat cognitive decline for the growing MCI population.

ClinicalTrials.gov ID: NCT02237560

Keywords: cognitive, exercise, aging, MCI, dementia, neuropsychological, exergame, Alzheimer's disease

#### INTRODUCTION

Annual diagnoses of dementia, such as due to Alzheimer's disease (AD), are expected to approach 1 million in 2050, more than double the new cases diagnosed in 2000 (Alzheimer's Association, 2014). Lacking a cure for any of the many causes of this devastating decline in cognitive function and loss of independence, a rising public health outcry has helped spark initiatives like the National Alzheimer's Project Act (2011) and the Healthy Brain Initiative (Alzheimer's Association and Centers for Disease Control Prevention, 2013). Research continues to press on, seeking innovative and empirically-validated ways to prevent or ameliorate cognitive decline and impairment.

While behavioral interventions are unlikely to completely prevent or halt dementia, there is the potential for physical exercise to reduce the risk of dementia onset (Heyn et al., 2004; Hillman et al., 2008) or slow progression (Larson, 2010). It has been estimated that if the onset of dementia could be delayed a few years, the impact on the population over time could be dramatic, decreasing prevalence by 1 million cases after 10 years in the United States alone. By also delaying progression for 2 years this could reduce dementia incidence by an astonishing 18 million cases globally per year (Sano et al., 1997; Brookmeyer et al., 1998). Interventions that reach patients before they have declined into diagnosable dementia can have the greatest impact. Thus, it is especially desirable to develop interventions that target those who may be at an intermediate stage, perhaps diagnosed with or at risk for mild cognitive impairment (MCI; per DSM-IV in 1994; Jak et al., 2009) also known more recently as mild neurocognitive disorder (mNCD per DSM-V in 2013; henceforth MCI will be used for simplicity; Roberts and Knopman, 2013; Jak et al., 2016).

#### Neuropsychological Effects of Physical and Cognitive Exercise

Researchers have been investigating a variety of behavioral interventions, including exercise, to improve cognitive function and decrease risk of progression to dementia among individuals with MCI (e.g., Kramer et al., 1999; Colcombe and Kramer, 2003; Studenski et al., 2006; Angevaren et al., 2008; Baker et al., 2010; Geda et al., 2010; Martin et al., 2011). Physical exercise has garnered significant support in the literature for its potency in effecting cognitive benefits (e.g., Colcombe and Kramer, 2003). While the literature examining the relationship between physical exercise and cognitive benefit is not without issues (e.g., recent Cochrane Reviews; Forbes et al., 2015; Young et al., 2015, which focused on dementia and normative older adults, respectively), it is widely accepted that exercise benefits cognition, particularly executive function among symptomatic individuals, that are yet early in their cognitive decline (e.g., MCI), the very population on which this study focused (e.g., Colcombe and Kramer, 2003; Etnier and Chang, 2009; Geda et al., 2010; Gates, 2013; Hess et al., 2014; Zheng et al., 2016). Indeed, pre-clinical persons, ranging from asymptomatic older adults to those with early MCI, may experience the greatest benefit from exercise, due to the relatively preserved brain structures and functions (in contrast with more progressed dementia) that can support and supply components needed to realize neuroplasticity as triggered by exercise (e.g., Colcombe et al., 2004; Kramer and Erickson, 2007; Ahlskog et al., 2011; Teixeira et al., 2016). However, knowing that physical exercise may be good for one's body and brain may not be enough to induce regular practice of exercise. Most older adults do not meet the guideline from the American College of Sports Medicine (ACSM) which now recommends 45 min per day, 5–7 days per week, including vigorous exercise (Chodzko-Zajko et al., 2009).

In a previous randomized clinical trial (Anderson-Hanley et al., 2012), we aimed to induce a thorough dose of exercise (Vidoni et al., 2015) by assigning older adults to an exergame<sup>1</sup> . We expected that the engagement of a stimulating exergame would motivate regular exercise, therefore allowing older adults to achieve maximum cognitive benefit. Three months of pedaling a virtual-reality-enhanced stationary ergometer, displaying scenic bike tours which we referred to as a "cybercycle" (see **Figure 1**), was compared with a traditional stationary bike. In the end, the cybercycle condition yielded significantly greater cognitive benefit, as hypothesized. However, surprisingly, it was not due to a difference in dose; we found instead that both the cybercycle and traditional cyclists has been similarly adherent (e.g., no significant difference between groups in miles, minutes, power, heart rate, etc.).

This unexpected finding raised the possibility that the screen display was providing benefit, beyond that of basic entertainment

<sup>1</sup>Exergaming is a subset of health games, sometimes referred to as active games, that combine physical exercise with interactive gaming features, often through a virtual reality-type platform (Van Schaik et al., 2008; Read and Shortell, 2011; Stanmore et al., 2017).

"cybercycle" so named in our previous RCT; Anderson-Hanley et al., 2012). Former study participant demonstrating use of a cybercycle exergame; used with permission.

or motivation. We theorized the enhanced form of exercise was yielding findings much like those of the environmental enrichment literature (e.g., animals provided with greater physical and mental stimulation experience enhanced brain function; Pusic et al., 2016); thus it seems more (even two types at once) is generally better, although to a point. In humans the complexity of experience and the diversity of type, timing and compounding of "enrichment" makes this considerably challenging to study as a unique phenomenon. It seemed that for our participants, by interacting with the tour stimuli on screen (e.g., changing pedaling speed yielded a change in speed of scenery moving by), they were certainlyengaged in some additional amount of cognitive processing beyond the control condition. This layered cognitive stimulation may have combined with the physical exercise required of the exergame to create a "two-for-one" impact that resulted in the observed additional cognitive benefit for participants. Hence, we hypothesized that participants experienced better neuropsychological outcomes because of the potentially synergistic phenomena created by combining a type of "mental exercise" that was interactive with physical exercise (Anderson-Hanley et al., 2012; Frantzidis et al., 2014).

Several areas of the research literature speak to the possible validity of this hypothesized compounding or synergistic effect. As previously mentioned, there is a robust literature documenting the cognitive benefits of physical exercise. There is also a growing literature on cognitive training that purports significant effects (Willis et al., 2006; Karr et al., 2014; Toril et al., 2014; Wang et al., 2014; Huntley et al., 2015; Train the Brain Consortium, 2017). However, in our view, the claims are less convincing due to problems with reporting, limited generalizability, and other reasons that have been cited in various critiques of the literature and related "brain training" industry (e.g., Owen et al., 2010; Muijden et al., 2012; Redick et al., 2013; Boot and Kramer, 2014; Simons et al., 2016; Zokaei et al., 2016). Indeed, a Cochrane Review of 11 RCTs concluded there was "no indication of any significant benefits from cognitive training" (Bahar-Fuchs et al., 2013, p. 1). Nevertheless, considering some "mixed" results on the potential benefits of cognitive training, and as well as the animal literature on neurobiological impact of cognitive enrichments (reviewed below), there is potential for mental challenge to have its own benefit on cognition. Although perhaps it is most effectively implemented as part of a more naturalistic intervention, as when integrated with physical activity that taps additional neural mechanisms.

One way to examine this possibility further is to evaluate the body of literature that encompasses the cognitive effects of interventions with multiple components (e.g., cognitive training and physical exercise). Often these are referred to as:


Indeed, there is a vast literature on "dual-task" experiments and interventions (Kramer et al., 1985; Hiyamizu et al., 2012; Kramer and Wong et al., 2015; Hosseini et al., 2017), which sometimes focus on the limitations of human abilities to manage two separate (non-interactive tasks) at once (e.g., walking and counting numbers backwards). Traditionally, the aim of dual-task training might focus on interference effects (Al-Yahya et al., 2011) and improving physical outcomes (e.g., decreasing falls) for impaired individuals (e.g., post-stroke), especially while managing dual tasks. More recently, studies have examined cognitive outcomes in dual tasks wherein there is more naturalistic interactivity (as in the cybercycle scenario above; Anderson-Hanley et al., 2012; as well as other interactive modalities such as dance: Foster, 2013; Kattenstroth et al., 2013; Dhami et al., 2015; Schoene et al., 2015; Burzynska et al., 2017; Marquez et al., 2017; Müller et al., 2017; Rehfeld et al., 2017). These recent studies seem to be converging on a similar cognitive-enhancing phenomenon that lends empirical support to the synergistic hypothesis outlined above and in part, elsewhere (Fissler et al., 2013). Recent reviews of the literature on the cognitive benefits of exergaming for older adults conclude the preliminary evidence is promising (Smith et al., 2010; Schoene et al., 2013; Bamidis et al., 2014; Barry et al., 2014; Chao et al., 2014; Ogawa et al., 2016; Zhu et al., 2016; Zilidou et al., 2016), but further research of the effects of these interactive interventions and the impact of specific components (e.g., passive stimulation vs. active cognitive training, aerobic vs. non-aerobic activity, etc.) is necessary.

We have previously hypothesized that the benefits of interactive physical and cognitive exercise may increase with increased mental effort and our lab published a pilot trial as a precursor to this present RCT, the Aerobic and Cognitive Exercise Study pilot (ACES-pilot; Barcelos et al., 2015). The ACES-pilot sought to investigate whether greater cognitive challenge while exergaming would yield differential outcomes in executive function and generalize to everyday functioning. Sixtyfour community-dwelling older adults (mean age = 82) were randomly assigned to pedal a stationary bike, while interactively engaging on-screen with: (1) a low cognitive demand task (pedaling a steering along a virtual bike tour, referred to as exer-tour), or (2) a high cognitive demand task (pedaling and steering to chase dragons and score points in a 3D video game, referred to as exer-score). Executive function (indices from Trails, Stroop and Digit Span) was assessed before and after a single-bout and 3-month exercise intervention. A significant group × time interaction after 3 months of exergaming (Stroop; among 20 adherents). Those in the high cognitive demand group performed better than those in the low cognitive dose condition. Self-reported everyday function improved across both exercise conditions. These pilot data indicated that for older adults, cognitive benefit while exergaming increased concomitantly with higher doses of interactive mental challenge (Barcelos et al., 2015). The present study aims to replicate and extend that finding to MCI.

#### Neurobiological and Neuroimaging Effects of Physical and Cognitive Exercise

It is our hypothesis that interactivity can produce greater cognitive effects due to synergistic processes that compound or magnify the benefit in a non-linear way, perhaps by activation and utilization of neurobiological substrate of the mind-body interface that is evolutionarily adapted for success in naturalistic tasks (in this case, goal-directed motion through 3D space). While theory and research connecting cognitive processes to neurobiological pathways triggered by exercise in humans is expanding rapidly, there are indications of important links with various biomarkers (e.g., brain-derived neurotrophic factor [BDNF], insulin-like growth factor [IFG-1]; Cotman et al., 2007; Knaepen et al., 2010), including expression of microRNA (miRNA) in circulating exosomes (Pusic et al., 2016; Bertoldi et al., 2017), and brain structure and function (e.g., gray and white matter regions: Colcombe et al., 2006; hippocampus; Thomas et al., 2012; Erickson et al., 2015; Sexton et al., 2016). Animal models have more plainly linked certain differential, compounding or synergistic cognitive benefits with neurobiological phenomenon when both physical and mental exercise are provided (e.g., in laboratory mice exposed to enriched environments with physical and/or cognitive challenges; Greenough et al., 1999; Churchill et al., 2002; Olson et al., 2006; Galvan and Bredesen, 2007; Fabel et al., 2009; Voss et al., 2013; Jessberger and Gage, 2014). These studies and others, detail a variety of mechanisms and impacts linking mental and physical exercise to improved neuronal and brain health (e.g., via increased cerebral perfusion, neurogenesis, angiogensis, synaptogenesis; Greenough et al., 1999; Gage, 2002; Trachtenberg et al., 2002; Tsai et al., 2016; Kleemeyer et al., 2017), even detailing specific differential benefits (Black et al., 1990; Olson et al., 2006; Suo et al., 2016) such as cell proliferation with physical exercise, and cell survival with mental exercise (Van Praag et al., 2005; van Praag, 2008).

### The Present Research

Since only a tiny fraction of older adults exercise at levels recommended by the ACSM and American Heart Association (AHA; Chodzko-Zajko et al., 2009), and physician recommendation appears to do little to change sedentary behaviors (Grandes et al., 2009), the novel exergame utilized in this study holds the promise of increasing compliance by virtue of the distracting components of the virtual reality environment and the interactive and challenging features of the videogame. Besides being a potentially effective way of engaging older adults in exercise, the present RCT augments prior research by exploring the neuropsychological benefits of interactive cognitive and physical training combined, rather than focusing on either alone (Colcombe and Kramer, 2003; Anguera et al., 2013), or in tandem, but lacking interactivity (Fabre et al., 2002; Oswald et al., 2006; Kraft, 2012; Suzuki et al., 2012; Hotting and Roder, 2013; Shatil, 2013; González-Palau et al., 2014; Satoh et al., 2014; Shah et al., 2014; Suo et al., 2016). By examining the neuropsychological effects some of the components separately (e.g., physical exercise and mental exercise alone), this study attempts to tease apart the unique contributions of each component to any benefit to cognition, as well as the potentially differential effect of low- vs. high-cognitive demands. The present research specifically examines the benefits of these behavioral interventions for community-dwelling older adults with or at risk for MCI and thus, attempts to extend previous findings to this more vulnerable population.

Finally, this RCT examines potential underlying mechanisms of effects of combined physical and mental exercise on cognition through the collection of neurobiological and neuroimaging data. Prior research has begun to clarify the role a number of biomarkers (primarily from serum, CNS, or saliva) that change in accord with exercise interventions and appear to be linked improvements in brain health that and/or promote improvements in cognition. For example, exercise has been shown to yield: increases in Brain-Derived Neurotrophic Factor (BDNF; Cotman et al., 2007; Coelho et al., 2013a; Leckie et al., 2014; Vaughan et al., 2014; Dinoff et al., 2016; Maass et al., 2016) and variable responses of insulin-like growth factor 1 (IGF-1; Lorens-Martín et al., 2009). Research linking traditional physical exercise and cognition in older adults has identified BDNF and IGF-1 as biomarkers linking cognitive function and healthy aging (Voss et al., 2010; Bellar et al., 2011; Erickson et al., 2012; Hotting and Roder, 2013; Phillips et al., 2014; Szuhany et al., 2014; Geerlings et al., 2015; Lara et al., 2015; Maass et al., 2016). BDNF plays an important role in neuronal growth and survival, it modulates neurotransmitters, and is involved in the cascade to promote neuronal plasticity, while IGF-1 has also been implicated as a mechanism in exercise-induced neuronal plasticity (Voss et al., 2010; Brown et al., 2013; Phillips et al., 2015; Maass et al., 2016). Cotman et al. (2007) has presented a model in which exercise induces activity of BDNF and IGF-1 in multiple pathways that facilitate: (a) learning, (b) neurogenesis, and (c) angiogenesis. Furthermore, some research has found increases in vascular endothelial growth factor with exercise (VEGF; Vital et al., 2014). While growth factors are typically said to increase with exercise, there are number of reports of decreases in inflammatory markers: C-reactive protein and Interleukin-6 (CRP and IL6; Thielen et al., 2016; Monteiro-Junior et al., 2017). Additionally, recent progressive research has also been examining the potential for circulating exosomes to serve as early markers of neuropathology, such as AD, to act as delivery vehicles for potentially reparative miRNA expressions, and even to play a positive role the neurbiological effects of exercise (e.g., Zhang et al., 2014; Kumar and Reddy, 2016; Pusic et al., 2016; Batistela et al., 2017; Bertoldi et al., 2017). Some promising research has pointed to miRNA-9 and miRNA-193 as linked to benefit in neurobiological processes such that they play a therapeutic role in dementia (e.g., Russell et al., 2013; Párrizas et al., 2015; Li et al., 2016; Riancho et al., 2017).

In sum, exercise appears to be able to trigger or enhance the function of complex cascades of neurobiological processes that have sometimes been linked to cognitive benefits or other expressions of brain or CNS health. It is likely that a number of those processes act on neurophysiological substrate to foster improved brain health that leads to improved cognition; some of those mediating phenonmena may even be visible through structural imaging.

Structural changes in certain regions of the brain have been found to follow exercise and are often linked to concomitant changes in cognition; for example, increases in the hippocampus (Lorens-Martín et al., 2009; Erickson et al., 2011), as well as the anterior cingulate cortex, dorsal lateral prefrontal cortex, and more broadly the prefrontal cortex (ACC, DLPC, and PFC; Gordon et al., 2008; Erickson et al., 2011; Weinstein et al., 2012; Curlik and Shors, 2013; Hayes et al., 2013; Nishiguchi et al., 2015; ten Brinke et al., 2015; Jonasson et al., 2017; Li et al., 2017). While the body of research has been growing with respect to linking physical exercise and cognition via biomarkers, scant literature has yet explored how these indicators and mechanisms react in the case of combined or interactive mental and physical exercise interventions in humans, wherein there might be somewhat differential or compounding beneficial effects given a two-forone intervention (for a promising exception, see Eggenberger et al., 2016). The present study aims to address a number of these gaps and replicate or extend other findings in the literature.

### Hypotheses

This study is a partial replication and extension of the ACESpilot (Barcelos et al., 2015) and it was hypothesized that, similar to that which was found in that prior pilot study: physical exercise interactive with effortful cognitive challenge (exer-score; **Figure 2**), would produce greater cognitive benefit than physical exercise that was interactive with relatively passive cognitive processing (exer-tour). We expected this previously observed phenomenon to extend to MCI specifically. The primary aim of the trial was for community-dwelling older adults with or at risk for MCI to engage their assigned intervention regularly for 6 months (longer than the ACES-pilot, which extended only to 3 months). Baseline to 6-month intervention (exer-tour vs. exer-score) were hypothesized to yield these effects:

	- a. additionally, a comparison would be made with archival/normative exercise-only data (available only for a 3-month window, this trial's midpoint)
	- a. increased growth factors: BDNF, IGF-1, and VEGF
	- b. decreased inflammatory markers: CRP and IL-6
	- c. increased exosomal expression of miRNA-193 and miRNA-9
	- a. increased gray matter volume in the ACC, DLPFC, hippocampus, and PFC.

## METHODS

#### Participants

All study procedures were approved by the appropriate Institutional Review Boards (i.e., including the PI's institution and the three medical centers where the study was conducted; see section Author Notes). Older adults with a diagnosis of MCI were initially sought via referrals from neurology and neuropsychology at three medical centers in a specific region of upstate New York where participants were to exercise (using a cybercyle placed at each center's physical therapy clinic). However, given difficulties enrolling patients (especially due to travel requirements, but also due to challenges targeting patients with MCI before progression to dementia), further IRB approval was obtained to expand recruitment by locating additional cybercycles at several more locations around the region (e.g., retirement communities, YMCAs, etc.) for a total of 14 sites. Enrollment was thus expanded to include self-referred community-dwelling older adults, aiming to include especially those that might be undiagnosed (because symptoms had not affected function), but yet who might meet criteria for, or could be said to be at risk for MCI (e.g., thus those that would meet screening criteria for MCI, herein referred to as "screened as MCI" or sMCI<sup>2</sup> ). Participants were recruited using fliers,

<sup>2</sup> Screened as MCI (sMCI) based on a MoCA cut-off score for MCI of <26 (Nasreddine et al., 2005). This screening approach was utilized after initial planned attempts to enroll clinically diagnosed patients with MCI (e.g., recruited from neurologists, neuropsychologists, and geriatricians, yielded patients that had often

FIGURE 2 | Exer-tour (relatively cognitively passive) vs. Exer-score (cognitively effortful). Exer-tour (pedaling controls speed on screen and progress along scenic bike paths; involves steering, but relatively passive compared to exer-score; for example, can't leave road or crash into anything or any rider which one can steer through; could cease steering without consequence other than tilted view, bike will follow curb). Exer-score (requires navigating in 360◦ radius to locate colored coins and matching colored dragons of varying speed/difficulty to steer through; the goal is to score points and strategy may be employed to avoid losing points by avoiding hazards, while also seeking out bonus points available via tagging specialized objects one can choose to explore). Former study participant demonstrating exer-tour condition; used with permission.

newspaper ads, and information sessions. The timeline for the conduct of the study was 2014–2016. Volunteers (n = 220; see CONSORT **Figure 3** for details) were screened by phone and were excluded if they had known neurologic disorders (e.g., Alzheimer's, Parkinson's, or seizures) or functional limitations that would restrict participation in cognitive testing or exercise. After reviewing the informed consent form (ICF) and answering all questions, a signed document was obtained from all participants and their surrogate, if appropriate, based on the Impaired Decision Making Capacity screen (IDMC; Karlawish, 2008). Physician approval (PCP and if applicable, cardiologist) was obtained for all participants to engage in exercise as assigned. Enrollees (n = 111) were communitydwelling older adults with a mean age of 78.1 (SD = 9.9), who were predominantly female (66%), well educated (average years of education = 16.2, SD = 2.4) Causasians (three individuals selfreported minority ethnic/racial status). Participants scored on average 23.7 (SD = 3.1) on the Montreal Cognitive Assessment (MoCA; 30-point multidimensional scale, with higher scores indicative of better global cognitive function; Nasreddine et al., 2005; Freitas et al., 2012; Julayanont et al., 2014), such that 75% of enrollees were categorized at meeting screening criteria for sMCI ("screened as MCI" based on MoCA < 26; Nasreddine et al., 2005).

To provide an additional point of comparison of the relative effect of physical exercise alone vs. interactively combined with mental exercise, archival data was culled from our prior Cybercycle Study. In that study, 33 participants were randomly assigned to a traditional exercise or "pedal-only" condition in which participants pedaled a stationary bike (in many cases the same model as used in the present study), but without the virtual reality display (see below for between group comparisons at baseline; for additional details see Anderson-Hanley et al., 2012, 2014; Dimitrova et al., 2016).

#### Baseline Comparisons Between Groups

Between the three randomly assigned groups in the present study (exer-tour, exer-score, and game-only; n = 46, 45, 20, respectively), cognitive measures revealed that the groups were comparable (no significant differences) on: overall cognitive function (MoCA), rate of sMCI, and performance on individual tests of cognition including: executive function, visual spatial skill, and memory (verbal, nonverbal, list, story, immediate or delayed).

On behavioral measures, those who were in the game-only condition were somewhat less motivated to be in the study, were not as ready for exercise, and reported they were more sedentary (recall that some participants were randomly assigned to game-only, but others were recruited specifically to the condition, including with remuneration, due to a high dropout rate once randomly assigned participants learned they would not be engaging in a physical exercise component). However, selfreported cognitive engagement (both recent and lifetime), was similar across all three groups.

Baseline comparisons between the four groups, including the above three, but also a pedal-only condition (archival control group data from our prior study noted above; n = 33), revealed that all four groups were similar (no significant differences) on: the three primary outcome measures of executive function to be the focus of our study (i.e., ratio scores for Digits, Trails,

already progressed in their decline toward dementia). Indeed, by the time some patients had been referred for clinical evaluations, received a formal diagnosis and subsequently were referred to our study, their disease had progressed such that they no longer fit our target population of MCI. As a result, attempts to enroll such referrals led to exclusion as the patient sometimes would be unable to understand or participate in the research and/or complex intervention procedures.

and Stroop; see below for details). The four groups were also comparable in proportion of men and women and BMI, but the pedal-only condition was significantly older and had fewer years of education<sup>3</sup> (confirming use of age and education covariates in statistical analyses).

#### Procedures

Participants were initially randomly assigned to one of three conditions for 6 months: (1) exer-tour: physical exercise interactive with relatively passive, low cognitive load, virtual scenic bike tour; or (2) exer-score: physical exercise interactive with a relatively effortful, high cognitive demand, videogame (see **Figure 2** for sample screen shots); or (3) game-only: the same videogame operated by a joystick or keyboard (no physical exercise was required). Exercise participants pedaled a virtual reality enhanced, recumbent stationary bike (Expresso S3R, from Interactive Fitness Holdings, LLC). Individuals in the exer-tour condition pedaled through various virtual scenic bike tours while

<sup>3</sup>This difference is surmised to be likely due to differences in recruitment as the prior, Cybercycle Study, only included participants exercising at their residential facility, whereas the present study included those that could travel to exercise equipment, often at locations across town.

steering along the path (but without significant consequence for not steering; that is the bike would coast along the curb and one could pass through other riders fluidly). Those in the exerscore condition pedaled through a scenic landscape where the goal was to score points by collecting different colored coins and corresponding colored dragons. Throughout the videogame (either exer-score or game-only), participants could navigate through bonus items to increase their speed and score, while avoiding penalties (e.g., steering into water). The exer-tour condition was presumed to require less mental engagement than the videogame conditions, as the game called for scoring and thus, potentially effortful planning, tracking, multi-tasking, and strategizing (a validity check of this assumption was confirmed by those in the exer-score condition who reported expending significantly greater "mental effort" than those in the exerscore condition; see **Table 1**). Participants were provided "thank you gifts" (e.g., water- bottle, mug, etc.) at each of evaluation (baseline, 3-month, and 6-month).

Given difficulty retaining participants in the game-only condition (due to preference for active/physical exercise, boredom once started, etc.), the protocol was again adjusted, with IRB approval, to allow randomly assigned game-only participants to switch to the exer-score condition at the 3-month TABLE 1 | Exer-tour vs. exer-score baseline demographics for adherent completers (0–6M).


*<sup>a</sup>Comparison between groups.*

*<sup>b</sup>Comparison with baseline.*

*<sup>c</sup>Repeated meas ANCOVA controlling for age and education.*

*Bold value indicate statistically significant (p* ≤ *0.05).*

mid-point evaluation. Additionally, game-only participants were also recruited separately with modest periodic remuneration provided. Upon enrollment, participants were phone screened for eligibility and administered the Impaired Decision-Making Capacity structured interview (IDMC; Veterans Administration Medical Center (VAMC), 2007) and provided informed consent (co-signed by a surrogate or legally-authorized representative as applicable and/or per the IDMC). Once scheduled for an initial evaluation, participants were mailed the ICF, demographic, health and fitness history questionnaires. If they were willing, an MRI (3T) of the brain was scheduled. At the baseline evaluation, a saliva sample was collected (passive drool), and they were administered a neuropsychological test battery (specified below). Individuals were trained in the use of the equipment for their assigned condition and underwent a single bout of exercise for 20 min. Participants were asked to maintain a target heart rate level, which was calculated using the Karvonen equation (McAuley et al., 2011) and measured via steering hand-grips built into the cybercycle (displayed on screen for participant monitoring and captured by onboard computer for later data extraction and analysis). Post-testing was conducted following the single bout<sup>4</sup> (SB) of exercise.

Following the baseline (BL) evaluation, individuals were instructed to exercise for at least 20 min at least twice a week, and to gradually increase exercise duration to 45 min and frequency to at least three to five times per week until the 3-month (3M) mid-point evaluation, and then maintain that pattern though a 6-month (6M) final evaluation. Participants were instructed to aim toward their individualized specific target heart rate (per above) throughout the intervention period. Participants were also asked to document their exercise sessions on log pages provided in a binder and used to calculate adherence statistics (e.g., frequency, heart rate average, etc.); additionally, the computer onboard the cybercycle captured similar data which was used for spot verification. A final evaluation, including repeat

<sup>4</sup> single-bout results are not reported herein in part due to space constraints, but also given the focus of the RCT on long-term intervention results which

necessarily analyzes a different/smaller sample of those adherent for 6-months; furthermore the single-bout was not a part of a priori hypotheses in the study design, but rather included to accompany the training session and provide a repeat testing experience/familiarization with tests before the follow-up evaluations (3 M and 6 M) which would partly help "wash out"/reach sooner the plateau of practice/learning effects and thus more plainly isolate intervention effects at the follow-up evaluations; finally, the single-bout assessment provides a consistent pattern with our prior RCTs (the Cybercycle Study and the ACES-pilot) in which similarly a second round of testing before the long-term outcome for similar reasons, as such this increases comparability of our data across trials and timepoints (both in terms of amount of familiarization with testing and matching of alternate forms at each time-point).

neuropsychological, biomarker, and neuroimaging assessments, was conducted at the end of the 6-month intervention period.

### MEASURES

#### Primary Measures Cognitive Function

Paper and pencil tests comprised the battery and alternate forms of cognitive tests were utilized at each evaluation to minimize practice/learning effects from serial testing. Cognitive testing was done at baseline (BL) before starting the intervention, after the 20-min training single-bout session, at a 3-month (3M) mid-point, and after concluding the 6-month (6M) intervention period.

#### Executive Function (Primary Outcome)

Three tests were selected to assess aspects of executive function, which has been shown to be a complex, yet overarching domain of cognitive function. The following three tests were chosen as they are thought to tap overlapping, yet distinct components of executive function, particularly mental flexibility (i.e., Stroop for inhibition, Trails for set shifting, and Digit Span for working memory; Wecker et al., 2000; Strauss et al., 2006). Executive function is critical cognitive function for maintaining independence in later life and thus an important focus of this study's intervention efforts (Pereira et al., 2008; Marshall et al., 2011).

#### **Stroop (Van der Elst et al., 2006)**

A short, 40-item version of Stroop was administered. Colored squares (red, green, blue) were presented in rows first (Stroop A), followed by those color words typed in black ink (Stroop B), followed by incongruent color words (Stroop C; in which participants were asked to name the color of the ink while ignoring the written word). A ratio was computed to isolate the executive function component of the task (Stroop A/C; Lansbergen et al., 2007). Among older adults, the ratio score has reasonable test-retest reliability (0.68) over 1–2 months (Ettenhofer et al., 2006) and strong test-retest reliability (0.80) over 2 weeks using the shortened version used herein (Houx et al., 2002). Higher ratios indicate better executive function.

#### **Color trails (D'Elia et al., 1996)**

Color Trails 1, requires participants to connect numbered circles in ascending order. Color Trails 2, requires individuals to connect numbered circles in consecutive order while also alternating the color of the circle (pink or yellow). Following the pattern set by Lansbergen et al. (2007) above of isolating the executive function component by dividing basic and faster processing speed by the slower interference trial, a ratio score (Trails1/2) was computed. Reliability and validity are adequate (D'Elia et al., 1996). Higher ratios represent better executive function.

#### **Digit span (Strauss et al., 2006)**

Digit Span Forward requires participants to first listen to a list of numbers and repeat them, with the string length increasing to the maximum of their ability. Digit Span Backward, requires repeating a string of numbers in reverse order. Continuing the pattern above to isolate the executive function component, the ratio of the typically smaller sum of correct interference trials on Digit Span Backward, divided by the typically greater sum of correct basic attention trials on Digit Span Forward (DigitsB/F). Good reliability and validity have been reported for Digit Span (Strauss et al., 2006). An increase in the ratio was the desired outcome.

#### Secondary Measures (Characterization of Sample and Possible Additional Outcomes) Montreal Cognitive Assessment (MoCA; Nasreddine et al., 2005)

The MoCA was administered at baseline to characterize the sample as either normative aging or "screened as MCI" (sMCI; see details above). The MoCA consists of eight different subtests to assess overall cognitive impairment. Scores below 26 out of 30 were used to categorize sMCI (Nasreddine et al., 2005).

#### Ecological Validity (EV; Self-Reported Cognitive Function; Klusmann et al., 2010)

In an attempt to measure any generalized, clinically relevant, effects of the interventions (beyond any training to the tests that might occur when using standardized, serial testing), a participant self-report of perceived cognitive function was utilized. Participants were asked about their self-perceptions and beliefs related to their everyday functioning in regards to their cognitive functioning (e.g., memory and concentration). The Ecological Validity Questionnaire (Klusmann et al., 2010) was administered at BL, 3M and 6M. This measure consists of eleven statements that ask about participants perceptions of memory, concentration, and everyday function. Using a 5-point Likert scale (1 = absolutely wrong / bad to 5 = absolutely true / good), participants rated how much each of the statements described their behavior over the past 2 weeks. The total score can range from 11 to 55, with higher scores representing perceptions of better cognitive functioning.

#### Verbal Memory (Alzheimer's Disease Assessment Scale; ADAS Word List—Immediate and Delayed

Recall; Harrison et al., 2007; Podhorna et al., 2016). Participants are shown a list of 11 words on cards and they recall as many as they are able immediately and also after a delay interval. The number of errors/omissions comprises the score, so lower scores are better.

## Get-up-and-go Test (Podsiadlo and Richardson,

#### 1991)

Participants rise from sitting, walk 10 feet, turn around and return to sitting position. The time it takes to complete the task is the score. Lower scores are better.

#### Other Measures Include

The Physical Activity Readiness Questionnaire (PAR-Q; 0–7, higher scores indicates barriers to readiness; NASM) and Selfreported Physical Activity (SRPA); mental and physical effort thermometers; Exercise Induced Feeling Inventory (EIFI); the Confusion and Vigor subscales of the Brunel Mood Scale (Terry et al., 2003); experience of Flow (Payne et al., 2011), and other neuropsychological measures (e.g., Rey-O, NAART, etc.) not expected to be affected by the intervention, but included for separate use in full characterization of enrollees for later crosssectional analyses. Due to the focus of this report on a priori hypotheses, and that the small adherent sample available for final analysis and circumscribes statistical power, some of these measures will be reported separately in cross-sectional or other limited analysis.

### Exploratory Measures

The following biomarker and neuroimaging data were collected from some willing and able participants, who comprised a small subset of the enrolled participants. Due to the small and incomplete sample, we report these as pilot data to examine comparisons that were planned a priori and provide tentative findings, which may be useful in directing future research.

#### Biomarkers

Saliva samples (passive drool per Salimetrics protocol) were collected at: BL, SB, 3M, and 6M from willing and able participants. Samples were centrifuged, with a portion reserved for protein analyses and a portion further processed for exosome analyses. Exosomes were isolated from whole saliva via differential ultracentrifugation as described in previously published protocols (Thery et al., 2006; Gallo and Alevizos, 2013; Witwer et al., 2013). Briefly, whole saliva was subjected to differential centrifugation steps at 300 × g, 1,500 × g, 17,000 × g, and 160,000 × g. The exosome pellet was suspended in 1X PBS. All samples were stored at −80◦C until analyses were conducted.

#### Protein Assays

Frozen samples were shipped overnight on dry ice for analysis (to RayBiotech, Inc., Norcross, GA). Protein concentrations of BDNF, CRP, IGF-1, IL-6, and VGEF were obtained via ELISAs (e.g., Mandel et al., 2011). Each patient sample was run in duplicate using 100 ml of plasma diluted by a factor of 2. A commercial multiplexed sandwich ELSIA-based array was used (Quantibody custom array, RayBiotech Inc., Norcross, GA, USA). All of the samples were tested using a panel of cytokines per above. The antibody array is a glass-chip-based multiplexed sandwich ELISA system designed to determine the concentrations all cytokines simultaneously. One standard glass slide was spotted with 16 wells of identical biomarker antibody arrays. Each antibody, together with the positive and negative control, was arrayed in quadruplicate. The samples and standards were added to the wells of the chip array and incubated for 3 h at 4 uC. This was followed by three to four washing steps and the addition of primary antibody and HRP-conjugated streptavidin to the wells. The signals (Cy3 wavelengths: 555 nm excitation, 655 nm emission) were scanned and extracted with an Innopsys laser scanner (Innopsys, Carbonne France), and quantified using Quantibody Analyzer software (Ray Biotech Inc). Each signal was identified by its spot location. The scanner software calculated background signals automatically. Concentration levels, expressed in picograms per milliliter (pg/ml), were calculated against a standard curve set for each biomarker from the positive and negative controls.

#### Exosome Analyses

Frozen samples were shipped overnight on dry ice for analysis (to VY at Oklahoma State University). The exosome/PBS solution was lyophilized for 9 h and RNA extraction was performed using TrizolTM reagent (Life Technologies <sup>R</sup> , catalog # 15596026.PPS) following the manufacturer's protocol. A fixed volume of 1,000 ng RNA was reverse-transcribed using the High Capacity cDNA Reverse Transcription Kit (Appled Biosystems, catalog #: 4368814). PCR forward and reverse primers for the human miRNA-193 (forward: CTTTTGGAG GCTGTGGTCTCAGAATC; reverse: CCAGTTGGATAAAAC ATAAACTCATCTCGCC) and miRNA-9 (forward: AGGCGGG GTTGGTTGTTATCTTTG; reverse: CTAGCTTTATGAAGACT CCACACCACTCATAC) were used for real-time quantitative RT-PCR (qRT-PCR) amplification. miRNA-193 and miRNA-9 expression was normalized against the U6 housekeeping miRNA (forward: GTGCCTGCTTCGGCAGC; reverse: TATGGAACGC TTCACGAATTTGCGTG).

The real-time qRT-PCR was performed using SYBR <sup>R</sup> Select Master Mix (ThermoFisher, cat #: 4472919). 100 ng of cDNA was loaded per well and amplified for 45 cycles using Opticon <sup>R</sup> 2 real-time PCR instrument (Biorad, model #: CFB-3220) operated by Opticon Monitor 3 software. All samples and references were run in triplicate and each well contained 20 µL total volume. Raw florescence readings were directly imported into, and baseline corrected with, LinregPCR software package (Version 2016.1). Linear regression was performed on the baseline-corrected data to calculate efficiency using a common window-of-linearity for each primer pair (Ruijter et al., 2009). Between-session variations were estimated using a maximum likelihood approach (Ruijter et al., 2006) and inter-plate variability was minimized using factor correction calculated using Factor qPCR software (Version 2016.0). Taking into account the reaction efficiency of each primer set, we used REST 2009 software package (Version 2.0.13, QIAGEN, Valencia, CA, USA) to calculate the relative RNA expression ratios in samples by a Pair Wise Fixed Reallocation Randomization model with 8000 iterations and a combination of randomization and bootstrapping techniques (Pfaffl et al., 2002).

#### Neuroimaging

Brain imaging T1-weighted 3D MP-RAGE axial images were acquired with a Siemens MAGNETOM Verio 3T MRI scanner running syngo MR B19 software, using an 32-channel bodycoil using the following parameters: 0.5 × 0.5 × 1.3 mm<sup>3</sup> voxel size; 256 × 256 matrix; 25 cm field of view; flip angle = 8 ◦ ; echo time = 2.9 ms; repetition time = 1,640 ms; inversion time = 828 ms. Gray matter (GM) volume was analyzed using FreeSurfer software v. 5.3.0<sup>5</sup> (Fischl et al., 2002), in a Linux environment, as described in detail previously (Destrieux et al., 2010). Briefly, cortical reconstruction and volumetric segmentation require the initial motion correction and averaging of T1 images, signal intensity normalization, the removal of non-brain tissue, automated Talairach transformation and segmentation of gray and white matter<sup>6</sup> , and the parcellation

<sup>5</sup>http://surfer.nmr.mgh.harvard.edu/

<sup>6</sup> total gray matter volumes were the focus of this study (white matter not considered herein)

of the cortex into well-described units, using gyral and sulcal landmarks for automatic differentiation. Each segmented brain volume was then inspected visually for processing errors which can occur during automation, and these were corrected using manual edits and re-processed. The Longitudinal Stream pipeline, available in Freesurfer v.5.3.0 (see website<sup>7</sup> ), was used to create an unbiased within-subject template for the two timepoints in the present study (Reuter et al., 2012). Moreover, we expect good reliability and a minimal risk of bias-introduction with the automated pipeline, given that the within-subject scans took place in the same scanner using the same parameters and our focus was on cortical ROIs, which show somewhat better scan-rescan reliability than subcortical structures (e.g., Mills and Tamnes, 2014).

Given the literature cited above and a prioi hypotheses, the regions-of-interest (ROIs) included: hippocampus, dorsolateral prefrontal cortex (DLPFC), prefrontal cortex (PFC) and anterior cingulate cortex (ACC)<sup>8</sup> . Freesurfer does not directly extract these ROIs as they are not singular regions of the brain (see discussion above as to why they are structurally and functionally significant in the present context). As such, we defined our ROIs by combining the following Freesurfer labels: DLPFC = frontal middle gyrus and sulcus from the Destrieux et al. (2010) atlas; ACC = rostral and caudal anterior cingulate cortices from the Desikan et al. (2006) atlas; PFC = ACC + medial orbitofrontal and transverse frontopolar regions (Domenech and Koechlin, 2015). The DLPFC ROI that was chosen is one of the most conservative and non-controversial groupings, although some others have included more medial and/or caudal regions in their work (Vijayakumar et al., 2014; Jonasson et al., 2017). The volumes extracted are a measure of total gray matter only (white matter was parcellated separately and excluded). For the final analysis, GM volumes of each region were extracted and considered as a ratio to total intracranial volume to account for variations in brain size. Given the small sample size, volumetric ratios were used in parametric partial correlations with cognitive outcomes (controlling for age and sex).

#### Statistical Analyses

Data were analyzed using SPSS version 23 for Windows (IBM Corporation). The primary goal of the RCT was to examine, for those with sMCI, the possible differential effects (via neuropsychological, neurobiological, and neuroimaging measures) of long-term use of different combinations of physical and cognitive exercise (exer-tour vs. exer-score). Analyses focused on those sMCI participants that adhered to the protocol (approximating the minimum prescribed dose by averaging 3×/wk for 6 months, allowing for 2 weeks of vacation, illness, or equipment failure, in each of the 3-month evaluation windows). Given the small sample of completers, there was concern about insufficient power since a priori power analysis suggested a larger sample was indicated (Faul et al., 2007). Additionally, small samples often violate assumptions of standard parametric tests. However, the data for the primary outcome variables was found to be normally distributed (Shapiro-Wilk tests for StroopA/C, Trails1/2, and DigF/B were 0.52, 0.71, 0.16, respectively; Normal Q-Q plots were also reviewed which appeared to be normally distributed, with data following a linear pattern) and error variances were consistent with parametric assumptions (Levine's tests were not significant). Furthermore, given the importance of covariates (more readily implemented in parametric tests), along with the possibility that effects could be magnified in the sMCI sample, and following guidance in the literature on statistical strength of ANCOVA above nonparametric approach, especially in clinical trials (wherein a focus on change from baseline facilitates a trend toward normality; Vickers, 2005), it was decided to proceed with standard parametric analyses. Repeated measure ANOVAs, controlling for age and education were used to examine group (exer-tour vs. exer-score) × time (BL vs. 3M and 6M) interaction effects on the primary outcome measures of executive function. Covariates entered into the repeated measures analyses included: age and education, due to links to outcomes noted in the theoretical and empirical literature (Hannay and Lezak, 2004; Lam et al., 2013). Paired t-tests were used to evaluate within group change from baseline to the 3 month mid-point, and baseline to conclusion of the 6-month intervention.

Partial correlations, controlling for age and sex (based on prior literature demonstrating links to secondary outcomes; e.g., BNDF decreases with age: Lommatzsch et al., 2005; brain region variability with sex: Ruigrok et al., 2014), were conducted to explore the relationship between cognitive change and possible corresponding change in biomarkers (proteins and exosomes) and neuroimaging regions of interest, per above hypotheses. Further correlations were conducted to explore relationships between baseline memory and exercise dose as represented by ride frequency.

#### RESULTS

#### Effect on Primary Outcome: Executive Function

Full (6-Month) Intervention (Exer-Tour vs. Exer-Score) Fourteen participants who met criteria for sMCI (MoCA < 26) were adherent to the protocol approximating the minimum assigned weekly average of 3×/week over 6 months. Half had been randomized to exer-tour and half to exer-score. After 6 months of interactive physical and cognitive exercise, the repeated measures ANCOVA omnibus test of executive function measures revealed a non-significant group x time interaction effect, F(3, 7) = 0.85, p = 0.51, η<sup>p</sup> <sup>2</sup> = 0.27. Univariate tests for StroopA/C, Trails1/2, and DigB/F were also not significant (p = 0.82, 0.11, 0.99, respectively). Paired t-tests were conducted to examine the change within groups over 6 months and StroopA/C improved significantly in both the exer-tour, t(1, 5) = −2.6, p = 0.049, and exer-score conditions; t(1, 6) = −5.5; p = 0.001 (**Figure 4**); Trails1/2 and DigB/F did not change significantly in either group (**Table 2**). Paired t-tests of the

<sup>7</sup>https://surfer.nmr.mgh.harvard.edu/fswiki/LongitudinalProcessing

<sup>8</sup>While the PFC and ACC overlap, they are reported separately here, in an attempt to replicate prior reports in the literature which identified change after exercise in each ROI.

change from baseline to the mid-point (3M) evaluation revealed no significant differences between the exer-tour and exer-score conditions at that point among sMCI participants (n = 14) that were adherent through 6 months (**Table 2**). sMCI participants who had been assigned to either exer-tour or exer-score and were not adherent over 6 months (n = 54) did not differ significantly from adherent sMCI participants (n = 14) on any baseline variables (including demographics, cognitive, physiological and functional variables; **Table 1**).

#### Mid-point (3-Month) Intervention (Exer-Tour vs. Exer-Score vs. Game-Only vs. Pedal-Only)

Given the high level of attrition from the game-only condition, as noted above, a mid-course adjustment was made to enroll game-only participants into a 3M-only intervention window. Doing so provided some comparative data regarding the possible relative effect of the mental exercise component separate from physical exercise (videogame practice only). Additionally, we were able to extract comparative 3M control data from our prior Cybercycle Study regarding the effect of a pedal-only condition for older adults. These 3M analyses utilized all older adults (sMCI + normative<sup>9</sup> ) that were adherent through 3M in their assigned condition (thus the n's for exer-tour, exer-score, gameonly, and pedal-only were: 15, 19, 8, 31, respectively). Paired ttests were conducted and revealed significant gains on StroopA/C by 3M for exer-tour and pedal-only, with statistically significant moderate effects (d = 0.49 and 0.35, respectively; **Figure 5**), while exer-score and game-only yielded non-significant small effects (d = 0.14 and 0.13, respectively). TrailsA/B and DigF/B did not change significantly. The 6M effect sizes for the two conditions that progressed to 6M (exer-score and exer-tour) are also included in **Figure 4** and reveal that both the exer-tour and exer-score conditions yielded statistically significant moderate effects by 6M (d = 0.52 and 0.47, respectively); thus, the exerscore group seems to "catch up" to the exer-tour from 3 to 6M.

#### Effect on Secondary Outcomes

Full (6-Month) Intervention (Exer-Tour vs. Exer-Score) There were significant group x time interactions among the 14 sMCI adherents for: (1) immediate verbal memory (exertour improved significantly more than exer-score; p = 0.018), (2) self-report of everyday cognitive function (as assessed on the Ecological Validity Questionnaire), with exer-tour reporting significantly more improvement while exer-score reported decline; p = 0.046), and (3) physical ability (exertour increased significantly more than exer-score; p = 0.001; see **Table 3**).

#### Biomarkers<sup>10</sup>

Biomarker samples were sought from all participants, but not all were willing or able to provide saliva samples (data from completers was available from a maximum of n = 16 for analysis). Target analytes included: BDNF, CRP, IGF-1, IL-6, and VEGF. Due to low concentrations of IGF-1 in the samples collected, the assay results were below the limit of detection on too many samples and this assay was excluded from analyses. Partial correlations (controlling for age and sex; **Table 4**) were conducted to examine the change from baseline to 6M in biomarkers related to change in the primary outcome measure, executive function (StroopA/C) which, per above, changed significantly with the intervention. Similarly, change in biomarkers were correlated with a secondary outcome that was impacted significantly by both interventions, was verbal memory, and we chose to focus on delayed verbal memory specifically (ADAS errors), since one of the more salient cognitive functions to impact for MCI patients. Finally, the exercise dose, based on average number of rides, was examined for association with the biomarkers change to evaluated the possibility of a dose effect. Exercise dose (rides) was found to correlate significantly with an increase in BDNF (r = 0.50, p = 0.04). Delayed recall errors showed a trend toward a moderate inverse correlation with increased VEGF such that better memory performance tended to coincide with increasing VEGF (r = −0.42, p = 0.08). Exosome analyses revealed a significant positive correlation between miRNA-9 expression with improved StroopA/C performance (r = 0.98, p = 0.002).

### Neuroimaging<sup>7</sup>

Neuroimaging (3T structural MRI) was grant-funded and sought from all participants, but not all were willing or able to undergo imaging (data from completers was available from a maximum

<sup>9</sup>The 3M analyses were not limited to sMCI, because added comparison of the fourth group at 3M (the pedal-only condition from the archival dataset; Anderson-Hanley et al., 2012), could not be sorted by the same sMCI criteria used in the present study (MoCA scores were not available for pedal-only participants). Thus, the decision was made to utilize all adherents (normative + sMCI) at 3M from the four groups to retain comparability within the 3M comparisons.

<sup>10</sup>Results for bioassay and neuroimaging data analyses are to be interpreted tentatively as exploratory pilot results, given the small sample sizes even with collapsing across conditions.


*cRepeated meas ANCOVA controlling for age and education.*

*Bold values indicates statistically significant (p* ≤ *0.05).*

of n = 8 for analysis). Partial correlations, just as described above, are reported in **Table 5**. Greater exercise dose (number of rides over 6M) was positively correlated with increasing PFC (r = 0.80, p = 0.003) and ACC (right; r = 0.70, p = 0.05; **Figure 6**). Verbal memory (delayed recall) errors were inversely related to increasing DLPFC, such that improvement in memory corresponded to increasing volume (r = −0.80, p = 0.01;

reference, illustrating the magnitude of effect from physical exercise alone.

### DISCUSSION

**Figure 6**).

This RCT was an attempt to replicate and extend findings of prior research, which found greater cognitive benefit for older adults engaged in interactive mental and physical exercise (i.e., while exergaming using a cybercycle; Anderson-Hanley et al., 2012) and furthermore, given pilot data which suggested greater mental challenge may increase cognitive benefit of exergaming (Barcelos et al., 2015). This RCT examined whether similar effects would be found among community-dwelling older adults with or at risk for MCI (screened as MCI; sMCI). We hypothesized that a more mentally challenging exergame would yield greater benefit to executive function than a more passive exergame. Community dwelling older adults, with a focus on those with sMCI, were randomly assigned to: exer-tour (pedaling along a virtual bike path), exer-score (pedaling to score points by navigating an interactive videogame), or a game-only condition (the latter having to be adapted due to issues with participant adherence, resulting in a shorter and remunerated, quasi-experimental arm).

Fourteen older adults with sMCI were adherent to the minimum assigned dose (average 3×/wk of exercise) for 6 months. Both conditions were found to yield significant improvement on one of three measures of executive function (Stroop A/C), producing similar moderate effect sizes at 6 months (d = 0.52 and 0.47, respectively). There was no interaction effect; however, from informal examination of the graph of the data (**Figure 4**), it appears that the exer-score condition did not produce incremental gains at the midpoint (3M), while the exer-tour condition did appear to yield benefit sooner, although by 6 months the exer-score and exertour gains were similar. It is hypothesized that since the exer-score condition does require significant mental effort to play, it may have taken longer for exer-score participants (with sMCI) to master the interactive pedaling and gameplay. Thus, there may have been delay in triggering the synergistic effects of the interactive exercise that would lead to a neurobiological cascade of events ultimately linked to any cognitive benefit. Furthermore, it may be different brain networks are activated by various aspects of the interactive physical and mental tasks in each condition, thus resulting in differential enhancement of network functions; for example, the effect on Stroop herein is consistent with activation of


TABLE 4 | Neurobiological change (0–6M) correlates with exercise dose/effort and change in cognitive function for participants in exer-tour and exer-score: Exploratory pilot data.


*<sup>a</sup>Controlling for age and sex.*

*Bold values indicates statistically significant (p* ≤ *0.05).*

the fronto-parietal network used in inhibition and switching and the fact that the volume of the ACC as shown on MRI imaging was enlarged proportionally to exercise dose further lends support to the impact on that specific network (Grandjean et al., 2012).

The differential relative impact of the two forms of interactive physical and cognitive exercise, exer-tour and exer-score, was apparent when comparing mid-point (3M) moderate effect with small effect (d = 0.52 and 0.14, respectively), while the gameonly condition showed little impact (i.e., of cognitive training; d = 0.13). By contrast, the pedal-only condition by 3M was more consistent with the literature on the benefits of physical exercise alone for cognition as it yielded a significant withingroup improvement (Colcombe and Kramer, 2003), although it was a more modest effect (d = 0.35) than the exer-tour, which is also consistent with past research in which the cybercycle effect exceeded exercise alone (Anderson-Hanley et al., 2012). It may be that the measures of executive function that were chosen were not as sensitive to the particular impact of each of the conditions as they could be; while Stroop A/C revealed an effect, Trails 2/1 and Digit Span B/F did not, perhaps because the components of executive function were not particularly enhanced (e.g., task switching and working memory, respectively). Future research, could strategically select measures that might better detect change in selective attention or inhibitory control which arguably might be more affected by repeated use of the exer-score condition in which one is chasing dragons while scanning the horizon for additional opportunities and avoiding pitfalls, and so forth (e.g., Flanker or Go/no-Go tasks might have greater sensitivity to detect for specific and relevant changes in executive function; Diamond, 2013).

Analysis of secondary outcomes from the present study yielded promising indications that this type of exergaming

*cRepeated meas ANCOVA controlling for age and education.*

*Bold values indicates statistically significant (p* ≤ *0.05).*

Anderson-Hanley et al. Aerobic and Cognitive Exercise Study for MCI

TABLE 5 | Neuroimaging change (0–6M) correlations with exercise dose/effort and changes in cognitive function (0–6M) for participants in exer-tour and exer-score: Exploratory pilot data.


*<sup>a</sup>Controlling for age and sex.*

*Bold values indicates statistically significant (p* ≤ *0.05).*

activity can yield cognitive gains in other realms of cognition<sup>11</sup> (e.g., verbal memory) among individuals with sMCI (with exertour outperforming exer-score). Moreover, these effects seem to generalize beyond standardized tests of cognition, to selfreported perceptions of everyday cognitive function and also vigor<sup>12</sup> (again, with exer-tour outperforming exer-score). The latter finding may benefit from further probing, perhaps via focused exit interview questions regarding possible reasons for this curious finding. Perhaps the regular practice of a mentally challenging videogame led participants in that condition to subjectively feel less capable (even though by 6M the gains exerscore made in executive function matched that of exer-tour). It is also likely that there would be individual variability in response to the gaming scenarios (e.g., prior research has found that one's trait competitiveness can interact with benefit from presumably facilitative features of an exergame, such as the presence or absence of an avatar (Snyder et al., 2012). Indeed in conduct of this trial, anecdotally, it was noted that some individuals were keenly enthused by and motivated to keep track of their score, while others were lackadaisical in their approach to pedaling around for exercise but without much care for scoring. Indeed, research shows the salience to the individual of the storyboard of a videogame, as well as other features may affect enjoyment or persistence, and can facilitate or limit use (Wang and Goh, 2017); all relevant issues to consider when designing and planning a long-term clinical trial with a goal of ensuring a thorough dose, via consistent adherence.

These findings of significant effects with both types of exergaming: exer-tour (low mental challenge) and exer-score (high mental challenge) on executive function and memory, mirrors one of the few similar studies, by Eggenberger et al. (2015), who reported similar trends when comparing an interactive dance exergame, with dual-task walking (noninteractive), and physical exercise alone. It is interesting to note that effect sizes were larger in our study and thus reached statistical significance, while the trial above had a much larger sample, but the effects were smaller leading to marginal results, despite having exergaming conditions that were in many ways similar. Perhaps this was due to the fact that the sample herein was sMCI, specifically, while the Eggenberger study utilized normative older adults. It seems likely that general, MCI participants have already begun some cognitive decline and so they may have more room to grow, so to speak, or at least more distance from the ceiling on cognitive tests to reveal change when attempting to attenuate the anticipated downward slope of cognitive decline, and thus the effect may be more readily apparent. Future research could further evaluate the possibilities by enrolling and comparing effects of both normative and MCI participants, with care taken to avoid possible ceiling effects on baseline assessments.

Finally, possible mechanisms linking exercise to enhanced cognition via presumed improved brain health were explored with promising findings from pilot data sampled from willing participants. Biomarker changes (e.g., BDNF and exosomes) were found to correspond to improvements in cognition and dose of exercise, each consistent with prior research (e.g., BDNF has been found to promote the differentiation of new neurons and synapses, and therefore, has been proposed to be a mediator of adult neuroplasticity; Huang and Reichardt, 2001; Leßmann and Brigadski, 2009; Flöel et al., 2010; Park and Poo, 2013; Edelmann et al., 2014). Furthermore, neuroimaging revealed promising gains in key ROIs (e.g., DLPFC and ACC) that corresponded with improvements in cognition and dose of exercise, again consistent with reports in the literature (Burdette et al., 2010; Weinstein et al., 2012; Chapman et al., 2013, 2016; Hayes et al., 2013; Ehlers et al., 2017; Li et al., 2017). While we retain a cautious stance for interpreting these findings given the limitations already noted, we are heartened that changes found are largely consistent with changes in gray matter volume that have been noted in other studies of learning and physical training. Some of which have reported statistically significant changes as early as 1–2 weeks (e.g., Ceccarelli et al., 2009; Zatorre et al., 2012) and specifically regarding training-induced changes in older adults before the 6-month time point (e.g., Boyke et al., 2008).

Strengths of this RCT include a study design crafted to clarify whether cognitive benefit could be maximized with a more challenging videogame component integrated with the physical exercise. In contrast to the findings of the ACES-pilot study (Barcelos et al., 2015), which found that after just three months

<sup>11</sup>Additional cognitive domains that might be sensible to assess in a follow-up study, might include visuospatial skills, given the integration and regular use of visuospatial tracking in the conditions with virtual reality screens.

<sup>12</sup>The two mood subscales (vigor and confusion from the BMS) were probed retrospectively on suggestion of an anonymous reviewer and the ex-tour reported a significantly greater sense of vigor at 6M, than exer-score.

ACC (green) and DLPFC (blue) in 3 sagittal (A), coronal (B), and axial (C) images aligned by crosshairs. ROI 3D model reconstructions give a sense of their structure in 3D space, as seen from the anterior (D) and left (E) sides.

of exer-score resulted in greater cognitive benefit than the exertour. The present findings suggest that for a more impaired MCI population the benefit of exer-tour shows up by three months, but it may be necessary for participants to stick with the interactive exercise through a longer window, six months, in order to master the challenges and reap the benefits of the enhanced exercise experience. Additionally, this study design began to tease apart the impact the of component parts of exergaming (e.g., physical exercise only vs. mental exercise only), as well as explore possible neurobiological mechanisms and neuroimaging markers that tentatively shed light and nudge future research regarding the ways in which the body and mind might interact for neuroprotection or amelioration of cognitive decline. A particular strength of this study is the use of alternate forms and additional serial testing (via single bout and 3M evaluation) that preceded the final 6-month evaluation such that practice or learning effects should have largely "washed out" or likely plateaued by that point (Beglinger et al., 2005), thus greater confidence can be placed in the results as representing a genuine effect.

A significant limitation of this study is the difficulty participants experienced in engaging a new behavioral approach to preventing or curbing cognitive decline. Attrition far exceeded our expectations based on past pilots and standard rates noted in the literature. The fact that this study focused on persons with sMCI, who face more than the usual challenges in starting a new exercise regimen, was likely a major contributing factor. Additionally, the requirement set out initially by grant reviewers and medical center IRBs that persons with MCI be considered a vulnerable population and must exercise in a limited environment with professional oversight (i.e., physical therapy clinics) proved daunting as participants were unwilling or unable to arrange transport 3–7×/wk to engage in regular exercise. Obtaining approval to make the equipment more accessible within retirement communities and neighborhood

<sup>13</sup>As noted above, both ACC and PFC were analyzed in order to replicate prior research which reported each ROI correlated with outcomes (e.g., exercise dose or cognition), but because of the overlap of these ROIs only one is shown here for clarity.

wellness centers did help somewhat, but nevertheless, retention remained a significant issue. The resultant small sample does diminish statistical power, although despite this, effects were sizeable enough to detect. Additionally, a small sample limits our ability to ask nuanced questions of the dataset which the study design would have facilitated if a larger sample had been obtained. For example, it was an a priori goal to compare known subtypes of MCI (i.e., amnestic vs. not) to see what, if any, impact subtype of MCI might have had on participants' abilities to learn and engage the more challenging exer-score videogame. This will have to wait for a future trial with a larger sample, perhaps by recruiting, as was necessary herein, those yet undiagnosed, but going beyond a screening tool, a future study could prospectively build in the mechanisms for a complete clinical workup of MCI as well. Nevertheless, much was learned regarding the needs of MCI patients and families in the conduct of such a trial, and will enhance future trial designs and interventions (see below). It will remain challenging however, to balance ideals about statistical power with the reality and importance of continuing to conduct research with clinical samples.

Future research is needed to address limitations of this study and improve retention and adherence to these types of behavioral interventions which have promising findings despite the variety of hindrances. For example, to try to address the challenges of enrolling and retaining MCI participants, our lab has been developing a more strategic and tailored in-home approach to facilitate adoption of this type of long-term behavioral intervention. We have developed a portable, adaptive, tabletbased neuro-exergame<sup>14</sup> , Memory LaneTM, specifically designed to target executive function by way of integrating specific mental challenges presented via a relatively simple, in-home interactive Physical and Cognitive Exercise System (iPACESTM15). While still only available for research purposes, the design had limited bulkiness of equipment and used budget components such that a large clinical trial will remain feasible (for example, an under-table pedaler paired with a budget tablet). Preliminary results from single bout use of our prototype iPACESTM have been promising (Anderson-Hanley et al., 2017), and we hope these various studies will encourage others to identify, design, and implement novel behavioral interventions to combat and ameliorate cognitive decline. Indeed there are a number of other open protocols currently recruiting to innovative behavioral interventions (Legault et al., 2011; Lee et al., 2016); for example, the Aerobic Exercise and Cognitive Training Trial (ACT) out of the University of Minnesota is enrolling patients with MCI to better understand the benefit of combined physical and mental activities (NCT03313895), and the Multidomain Alzheimer Prevention Trial (MAPT) combines physical and cognitive exercise, while also layering a nutritional component that may be able to further maximize any cognitive benefit derived (Vellas et al., 2014). Some research teams have been pushing the envelope on the possibilities given advancing technologies and have even explored web-connected exergaming to facilitate participation (Bamidis et al., 2011; Konstantinidis et al., 2017). The availability of such innovative interventions through data-yielding clinical trials is heartening to many individuals and families already facing the challenges of cognitive decline.

Given the encroaching dementia epidemic, it behooves researchers and society at large to continue to press for innovative and effective solutions that aim to maximize cognitive benefits. In this way we can hope to make meaningful contributions to the important cause of slowing or preventing the onset of cognitive decline. It is likely that a multi-factorial approach will yield the greatest impact, but there is much work to be done to clarify which components and intensities (e.g., of physical and mental exercise, nutritional and/or social supports, etc.), to meld and leverage for the best and most meaningful impact on brain health so as to ensure sustained cognitive functions in later life. Furthermore, it seems likely that there is a great deal to learn about factors that predict response to any intervention, including physical and mental exercise. It may be that certain genetic factors, such as ApoE status may moderate the body's ability to spark a given neurobiological cascade and thus reap the benefits of a given intervention such as interactive physical and cognitive exercise (Raichlen and Alexander, 2014). The more we learn about these underlying mechanisms and the nuanced effects exposed by analyzing larger samples, the more effective we can be in designing and tailoring interventions (Morrison-Bogorad et al., 2007; Lieberman, 2009; Larson, 2010; Read and Shortell, 2011; Gerling and Mandryk, 2014; Barha et al., 2016) to do the most good for the greatest number of people as fast as we can, especially now that so many are faced with sliding into the long suffering.

#### ETHICS STATEMENT

This study was carried out in accordance with the recommendations of the Institutional Review Board (AMC, SV, VA, and Union College) with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the IRB.

#### AUTHOR NOTES

Thank you to:


<sup>14</sup>Neuro-exergaming (Anderson-Hanley et al., 2016, 2017), is an adaptation of existing terms and used herein to indicate the integration of neurogaming with exergaming for the express purpose of benefiting brain health and fostering improved cognitive function

<sup>15</sup>iPACESTM 2013-present, patent pending No. 62/140,991; BSK Ref. No.: 416P1700A

Altman, Zachariah Arnold, Amelia Denney, Casey Terzian, Jennifer Vu, and Emily West.


Earlier versions of these results were presented at the annual meetings of: the Society for Neuroscience Chapter, the International Neuropsychological Society, and the American Academy for Clinical Neuropsychology.

#### REFERENCES


#### AUTHOR CONTRIBUTIONS

CA-H, NB, MM, VY, DH wrote pieces of the manuscript; CA-H, NB, EZ, RG, MD contributed to study design, implementation and interpretation; BC contributed to biomarker data analysis and interpretation; KM contributed to exosome study design, data analysis and interpretation; DH contributed to MRI data analysis and interpretation; PA contributed to study design, exercise science factors, implementation and interpretation; AK contributed to study design, interpretation.

### FUNDING

This research was funded by a grant from the NIA (1R15AG042109-01A1) with additional support provided by Union College via Faculty and Student Research Grants.

mild cognitive impairment: a controlled trial. Arch. Neurol. 67, 71–79. doi: 10.1001/archneurol.2009.307


process: impact of aerobic exercise. Mol. Cell. Biochem. 440, 115–125. doi: 10.1007/s11010-017-3160-4


6-month randomised controlled trial. Br. J. Sports Med. 49, 248–254. doi: 10.1136/bjsports-2013-093184


growth factor (VEGF) in elderly: a systematic review. Arch. Gerontol. Geriatr. 59, 234–239. doi: 10.1016/j.archger.2014.04.011


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer DM and handling Editor declared their shared affiliation.

Copyright © 2018 Anderson-Hanley, Barcelos, Zimmerman, Gillen, Dunnam, Cohen, Yerokhin, Miller, Hayes, Arciero, Maloney and Kramer. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Acupuncture Modulates the Cerebello-Thalamo-Cortical Circuit and Cognitive Brain Regions in Patients of Parkinson's Disease With Tremor

Zhe Li 1,2†, Jun Chen3†, Jianbo Cheng<sup>4</sup> , Sicong Huang<sup>5</sup> , Yingyu Hu<sup>6</sup> , Yijuan Wu<sup>1</sup> , Guihua Li <sup>1</sup> , Bo Liu<sup>3</sup> , Xian Liu<sup>3</sup> , Wenyuan Guo<sup>1</sup> , Shuxuan Huang<sup>1</sup> , Miaomiao Zhou<sup>1</sup> , Xiang Chen<sup>7</sup> , Yousheng Xiao<sup>7</sup> , Chaojun Chen<sup>8</sup> \*, Junbin Chen<sup>9</sup> \*, Xiaodong Luo<sup>2</sup> \* and Pingyi Xu<sup>1</sup> \*

#### Edited by:

*Christos Frantzidis, Aristotle University of Thessaloniki, Greece*

#### Reviewed by:

*Bing Zhang, Nanjing Drum Tower Hospital, China Aristea Kyriaki Ladas, Faculty-CITY College, Thessaloniki, University of Sheffield, United Kingdom Eugene Golanov, Houston Methodist Hospital, United States*

#### \*Correspondence:

*Pingyi Xu pingyixujd@163.com Xiaodong Luo luoxiaod@126.com Junbin Chen cjbcl0397@163.com Chaojun Chen ccjbs@126.com*

*†These authors have contributed equally to this work.*

> Received: *21 July 2017* Accepted: *18 June 2018* Published: *05 July 2018*

#### Citation:

*Li Z, Chen J, Cheng J, Huang S, Hu Y, Wu Y, Li G, Liu B, Liu X, Guo W, Huang S, Zhou M, Chen X, Xiao Y, Chen C, Chen J, Luo X and Xu P (2018) Acupuncture Modulates the Cerebello-Thalamo-Cortical Circuit and Cognitive Brain Regions in Patients of Parkinson's Disease With Tremor. Front. Aging Neurosci. 10:206. doi: 10.3389/fnagi.2018.00206* *<sup>1</sup> Department of Neurology, The First Affiliated Hospital, Guangzhou Medical University, Guangzhou, China, <sup>2</sup> Department of Neurology, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine, Guangzhou, China, <sup>3</sup> Department of Radiology, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine, Guangzhou, China, <sup>4</sup> Department of Radiology, The People's Hospital of Gaozhou, Gaozhou, China, <sup>5</sup> Department of Laboratory, The Second Affiliated Hospital, Guangzhou Medical University, Guangzhou, China, <sup>6</sup> Department of Business Development, Zhujiang Hospital, Southern Medical University, Guangzhou, China, <sup>7</sup> Department of Neurology, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou, China, <sup>8</sup> Department of Neurology, Guangzhou Hospital of Integrated Traditional and West Medicine, Guangzhou, China, <sup>9</sup> Department of Neurology, Yuebei People's Hospital, Shaoguan, China*

Objective: To investigate the effect of acupuncture on Parkinson's disease (PD) patients with tremor and its potential neuromechanism by functional magnetic resonance imaging (fMRI).

Methods: Forty-one PD patients with tremor were randomly assigned to true acupuncture group (TAG, *n* = 14), sham acupuncture group (SAG, *n* = 14) and waiting group (WG, *n* = 13). All patients received levodopa for 12 weeks. Patients in TAG were acupunctured on DU20, GB20, and the Chorea-Tremor Controlled Zone, and patients in SAG accepted sham acupuncture, while patients in WG received no acupuncture treatment until 12 weeks after the course was ended. The UPDRS II and III subscales, and fMRI scans of the patients' brains were obtained before and after the treatment course. UPDRS II and III scores were analyzed by SPSS, while the degree centrality (DC), regional homogeneity (ReHo) and amplitude low-frequency fluctuation (ALFF) were determined by REST.

Results: Acupuncture improved the UPDRS II and III scores in PD patients with tremor without placebo effect, only in tremor score. Acupuncture had specific effects on the cerebrocerebellar pathways as shown by the decreased DC and ReHo and increased ALFF values, and nonspecific effects on the spinocerebellar pathways as shown by the increased ReHo and ALFF values (*P* < 0.05, AlphaSim corrected). Increased ReHo values were observed within the thalamus and motor cortex of the PD patients (*P* < 0.05, AlphaSim corrected). In addition, the default mode network (DMN), visual areas and insula were activated by the acupuncture with increased DC, ReHo and/or ALFF, while the prefrontal cortex (PFC) presented a significant decrease in ReHo and ALFF values after acupuncture (*P* < 0.05, AlphaSim corrected).

Conclusions: The cerebellum, thalamus and motor cortex, which are connected to the cerebello-thalamo-cortical (CTC) circuit, were modulated by the acupuncture stimulation to alleviate the PD tremor. The regulation of neural activity within the cognitive brain regions (the DMN, visual areas, insula and PFC) together with CTC circuit may contributes to enhancing movement and improving patients' daily life activities.

Keywords: acupuncture, Parkinson's disease, tremor, functional magnetic resonance imaging, neuromechanism

#### INTRODUCTION

Parkinson's disease (PD) is an age-related neurodegenerative disorder of unknown origin that is characterized by the selective loss of dopaminergic neurons in the substantia nigra pars compacta (Miller and O'Callaghan, 2015). Tremor is usually the first clinical sign of PD, and approximately 70% of PD patients manifest conspicuous tremor at rest and/or during the maintenance of posture (Wang, 2006). The management of PD tremor presents a number of challenges to clinicians (Jiménez and Vingerhoets, 2012). Medication, which is the first line of treatment, often has unpredictable side effects. Stereotactic surgery provides better clinical results than medication but is poorly accepted, due to its invasiveness and high cost (Jiménez and Vingerhoets, 2012). Thus, many physicians and patients desire a complementary alternative strategy for tremor management. Acupuncture is a promising traditional Chinese medicine therapy that can be used to treat PD, and ∼7–10% of Asians choose acupuncture for tremor improvement (Lam et al., 2008). Due to its better adaptability, fewer side effects and lower cost, acupuncture has been widely used. Clinical studies have showed a positive benefit of acupuncture in treating PD tremor (Jiang et al., 2006; Wang et al., 2015). However, the mechanism underlying the effects of acupuncture on tremor associated with PD remains unknown.

Due to advances in brain neuroimaging technologies, recent functional magnetic resonance imaging (fMRI) studies have demonstrated that the basal ganglia (i.e., the pallidum and putamen) are active at the onset of tremor and the cerebellar circuit displays activity that is correlated with the magnitude of the ongoing tremor (Hallett, 2012). Both the basal ganglia and the cerebellum are connected to the motor cortex because the motor cortex is a component of both circuits, indicating the presence of pathology in the cerebello-thalamo-cortical (CTC) circuit in PD patients with tremor (Hallett, 2012).

fMRI is also a versatile tool to investigate the mechanism of acupuncture. According to previous animal studies, acupuncture plays a potential neuroprotective and restorative role in neuron survival (Kim et al., 2011; Sun et al., 2012; Rui et al., 2013; Xiao, 2015). This disease-modifying effect was reported to be similar to the effects of certain neuroprotective agents with antioxidative stress, anti-inflammatory and anti-apoptosis effects that improve motor performance in PD patients (Kim et al., 2011; Sun et al., 2012; Rui et al., 2013; Xiao, 2015). However, due to the physiological differences between humans and animals, conclusions based on animal experiments might differ from those based on clinical investigations involving human patients. fMRI can be used to visually measure the specific impact of acupuncture on the human brain (Deng et al., 2008, 2016; Zhang et al., 2012; Zhou et al., 2014; Zhang Q. et al., 2015; Zhang S. Q. et al., 2015). Although several fMRI studies have investigated the medical effect of acupuncture on PD symptoms, studies investigating tremor are limited (Chae et al., 2009; Su, 2009; Shang, 2010; Ye, 2011; Yeo et al., 2012, 2014). For example, Yeo et al. (2012, 2014) found that acupuncture stimulation on GB34 (Yanglingquan) activated substantia nigra, basal ganglia, precentral gyrus and prefrontal cortex in PD, but the authors were unable to determine the mechanism by which acupuncture decreased tremor.

Based on the abovementioned knowledge, we speculated that acupuncture might alleviate tremor and improve motor function in PD patients by modulating the CTC circuit or other pathways. Thus, here, we investigated the effectiveness of acupuncture paratherapy on PD patients with tremor and explore its underlying neuromechanism by fMRI analyzing the degree centrality (DC), regional homogeneity (ReHo), and amplitudes of low-frequency fluctuation (ALFF). The analytical processes used in these three methods are very similar and, thus, are useful for identifying regions with consistent activity across fMRI studies.

#### MATERIALS AND METHODS

#### Subjects

This study was conducted at the 2nd Affiliated Hospital of Guangzhou University of Chinese Medicine between May 2014 and January 2016. The patients included in this study were diagnosed based on the UK PD Society Brain Bank clinical diagnostic criteria, and tremor at rest in at least one upper or lower extremity on either side was assessed by item 20 of the Unified Parkinson's Disease Rating Scale (UPDRS) (Gibb and Lees, 1988; UKNCCF, 2006; Prodoehl et al., 2013). The exclusion criteria included secondary Parkinsonism, atypical parkinsonian disease, advanced PD stage (H-Y ≥ 4), age less than 45 or greater than 80 years, history of other neurological disorders or head trauma, left-handedness, cognitive impairment (Mini Mental State Examination (MMSE) score <24), depression tendency (Beck Depression Inventory (BDI) score >4), and any contraindications for fMRI. The subjects were randomly assigned to a true acupuncture group (TAG), sham acupuncture group (SAG), or waiting group (WG) using a computer-generated list based on consecutive numbers that were distributed in sealed, opaque envelopes. All subjects provided written and verbal informed consent before participating in the study. They were informed what the study was about, including the possible risks and benefits to them, and were completely voluntary taking part in this study. They may also leave the study at any time. If they left the study before it was finished, there would be no penalty to them, and they would not lose any benefits to which they were otherwise entitled. This study was approved by the Ethics Committee of the 2nd Affiliated Hospital of Guangzhou University of Chinese Medicine.

### Acupuncture

All subjects in the three groups received conventional levodopa treatment for a course of 12 weeks. A single experienced acupuncturist, who was not blinded to the group assignment, performed acupuncture twice weekly. In the TAG, stainless steel needles were inserted to a depth of 2.0–3.0 cm into DU20 (Baihui), GB20 (Fengchi), and the Chorea-Tremor Controlled Zone to alleviate tremor according to traditional Chinese medicine documents. Chorea-Tremor Controlled Zone is located at the scalp above the front of precentral gurus, 1.5 cm before Motor Zone. The reinforcing-reducing method conducted by twirling was performed every 10 min within the 30-min needle retention time. In the SAG, needles were inserted to 0.2 cm deep and 0.5 Chinese cun next to DU20, GB20 and the Chorea-Tremor Controlled Zone, but no manipulation of the needle was performed during the needle retention time. In the WG, true acupuncture was performed for 12 weeks following the completion of the medication course and the acupuncture effect was not evaluated. To guarantee that the patients were blinded during the treatment period, patients in each group received acupuncture treatment in different independent single-rooms; and all patients received bilateral and equivalent number of acupoint each time.

### Clinical Evaluation

Clinical evaluators, who remained blinded throughout the study, assessed the UPDRS II and III of all subjects before and after the treatment course. In UPDRS III section, items 20 and 21 are for the tremor score, 22 for the rigidity score, 23, 24, 25, 26, and 31 for the hypokinesia score, and 27, 28, 29, and 30 for the postural instability/gait disorder (PIGD) score (Liu et al., 2011). These four sub-scores represent the four typical motor symptoms of PD. Adverse effects of acupuncture were recorded if they happened.

### Image Acquisition

The brains of all subjects were scanned using a 3.0 Tesla MRI (Siemens MAGNETOM Verio 3.0T, Erlangen, Germany) with an 8-channel phased-array head coil at the radiology department of the hospital before and after the treatment course. To eliminate the effect of levodopa on the brain, the fMRI scan was performed at least 4 h after the levodopa administration. During the data acquisition process, all subjects were asked to close their eyes and lie quietly for MR scanning. Resting-state functional images were acquired using a T2-weighted gradient-recalled echo-planar imaging (GRE-EPI) sequence with the following parameters: repetition time = 2,000 ms, echo time = 30 ms, flip angle = 90◦ , thickness = 3.5 mm, gap = 0.35 mm, field of view = 224 × 224 mm2, matrix = 64 × 64, 31 axial slices, and 240 time points. The structural images were analyzed using a three-dimensional T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence with the following parameters: repetition time = 1,900 ms, echo time = 2.27 ms, flip angle = 9 ◦ , thickness = 1.0 mm, field of view = 256 × 256 mm2, and matrix = 256 × 256.

### Data Preprocessing and Calculations

The fMRI data analyser was also blinded to the group assignment. The resting-state fMRI data preprocessing was performed using DPABI based on MATLAB (The Math Works, Natick, MA, USA). After removing the first four volumes of each participant, the functional images were corrected for the intra-volume acquisition time delay using slice-timing and realignment. None of the participants were excluded based on the criteria of displacement >2 mm or angular rotation >2 ◦ in any direction. All corrected functional data were then normalized to the Montreal Neurological Institute (MNI) space and resampled to a 3-mm isotropic resolution. The resulting images were further temporally band-pass filtered (0.01–0.08 Hz) to remove the effects of low-frequency drift and high-frequency physiological noise. Finally, 24 head-motion parameters, white matter signals, and cerebrospinal fluid signals were regressed using a general linear model, and linear trends were removed from the fMRI data. Spatial smoothing was also performed before the ALFF analysis using a Gaussian filter (6-mm full-width half-maximum, FWHM), but after the ReHo calculation.

REST (http://resting-fmri.sourceforge.net) (Song et al., 2011) V1.8 was used to calculate the values of the DC, ReHo, and ALFF. The DC represents the large-scale brain intrinsic connectivity related to the global information integration function at the voxel level (Buckner et al., 2009). We applied threshold to the correlation coefficients at r >0.25 to remove the weak correlations caused by noises. ReHo depicts the local synchronization of the time series of neighboring voxels, which is related to the local information integration function (Zang et al., 2004). ALFF measures the amplitude of time series fluctuations at each voxel and is thought to be associated with spontaneous neuronal activity (Zang et al., 2007). Thus, these three fMRI measures probe different aspects of brain activity (Wang et al., 2017).

#### Statistical Analysis

One-way ANOVA and the chi-squared test were performed to assess the baseline differences in the demographic and clinical data among the three groups, and paired-sample t-test was performed to evaluate UPDRS II and III scores before and after the treatment in each group using SPSS V22.0 (SPSS Inc., Chicago, IL, USA). The level of significance was set as P < 0.05.

The statistical analysis of the DC, ReHo and ALFF was conducted using REST V1.8. An ANCOVA was performed on the fMRI data to identify the DC, ReHo and ALFF maps among three groups with the pretreamtment images as covariates. Subsequently, the regions that showed significant differences were extracted as a mask, and the DC, ReHo, and ALFF values were subjected to post hoc analysis. Statistical comparisons of these values between each pair of groups were performed using a two-sample post-hoc t-test, and the LSD correction was applied for multiple tests to keep the overall type I error level of 0.05. Voxels with a P < 0.05 corrected by AlphaSim and a cluster size >2, 295 mm<sup>3</sup> (85 voxels) were considered significantly different.

### RESULTS

### Demographic and Clinical Characteristics

Of the 42 patients who were identified as potential participants, one patient refused the fMRI scanning. The remaining 41 patients were randomized in this study. No patients were withdrawn from the TAG. Two subjects, one in the SAG who was diagnosed with acute ischaemic stroke by fMRI and another in the WG who experienced an accidental fall, were excluded. Four subjects, 2 in the SAG and 2 in the WG, were withdrawn because of poor efficacy, failure to follow up or refusal to rescan fMRI after the week 4 visit call (**Figure 1**). The baseline demographic and clinical characteristics of the patients in three groups are presented in **Table 1**. All subjects were cognitively normal and free of depression according to the MMSE and the BDI. No statistically significant differences in gender, age, family history, onset age, PD duration, UPDRS (II and III) or levodopa usage were observed among three groups.

### Clinical Evaluation

TAG showed significant improvement in UPDRS II and III scores, while WG and SAG didn't. In UPDRS III, tremor score in TAG decreased obviously while in WG increased significantly, but no big change in SAG. However, rigidity, hypokinesia and PIGD scores before and after treatment in each group displayed no significant differences (**Table 2**). No obvious adverse effects of acupuncture in the patients were reported.

### DC

An ANCOVA revealed significant differences in the DC index between the TAG, SAG, and WG in the following regions: fusiform gyrus, cuneus, lingual gyrus, superior and middle occipital gyri, insula and cerebellum crus. Compared with the SAG, the TAG showed increased DC in the fusiform gyrus, cuneus, lingual gyrus, superior and middle occipital gyri, and decreases in the cerebellum crus. Compared with the WG, the TAG displayed increased DC in the cuneus and decreased DC in the cerebellum crus. In addition, compared with the WG, the SAG's DC values were significantly elevated in bilateral insula. The details of the peak coordinates and cluster sizes are listed in **Table 3**.

### ReHo

An ANCOVA exhibits significant differences in the ReHo index among three groups in the following regions: middle and inferior frontal gyri, rectus gyrus, precentral gyrus, supplement motor area (SMA), inferior parietal lobules, precuneus, cuneus, fusiform gyrus, superior and middle cccipital gyri, anterior cingulate gyrus, hippocampus, thalamus, insula and cerebellum. Compared with the SAG, a significantly increased ReHo values was detected in the SMA, inferior parietal lobules, precuneus, cuneus, fusiform gyrus, superior and middle cccipital gyri, and a significant decrease in the ReHo was observed in the middle and inferior frontal gyri, anterior cingulate gyrus and the cerebellum crus of patients in the TAG (**Figure 2**). Compared to the WG, patients of the TAG had an enhanced ReHo values in the

#### TABLE 1 | Baseline demographic and clinical data of all subjects.


TABLE 2 | UPDRS II and III score of all subjects before and after treatment.


*Data are shown as mean* ± *SD.* \**P* <*0.05.*

left precentral gyrus, SMA, precuneus, hippocampus, thalamus, insula and cerebellum 4\_5, and a reduced ReHo values in the middle and inferior frontal gyri, rectus gyrus, right precentral gyrus and cerebellum crus (**Figure 3**). What's more, the SAG patients displayed an increased ReHo in the left insula and a decreased regional activity in the right precentral gyrus compared to the WG. The details of the peak coordinates and cluster sizes are listed in **Table 4**.

### ALFF

The inferior and medial frontal gyri, fusiform gyrus and the cerebellum revealed significant differences in the ALFF values among the TAG, SAG and WG. Compared to the SAG, an enhanced ALFF in the cerebellum 4\_5 and 6, and a reduced ALFF in the orbital inferior frontal gyrus were observed in the patients of TAG. Furthermore, compared to the WG, significantly elevated spontaneous neural activities exhibited in the fusiform TABLE 3 | Brain regions exhibiting increased and decreased degree centrality among three groups.


*TAG, true acupuncture group; SAG, sham acupuncture group; WG, waiting group. A* > *B, Compared with B group, A group showed increased DC values; A* < *B, Compared with B group, A group showed decreased DC values (P* < *0.05, AlphaSim corrected).*

FIGURE 2 | Differences in ReHo values between the TAG and SAG. (*P* < 0.05, AlphaSim corrected). Warm colors represent positive ReHo values; blue (cold) colors represent negative ReHo values.

FIGURE 3 | Differences in ReHo values between the TAG and WG. (*P* < 0.05, AlphaSim corrected). Warm colors represent positive ReHo values; blue (cold) colors represent negative ReHo values.

gyrus, cerebellum crus and vermis of the TAG patients, and in the medial frontal gyrus and cerebellum crus of the SAG patients. The details of the peak coordinates and cluster sizes are listed in **Table 5**.

### DISCUSSION

In this neural imaging study, we investigated the effect and its neural substrates of acupuncture stimulation on PD patients with tremor by the DC, ReHo and ALFF methods. We proved that acupuncture could improve the daily life activities and motor symptoms in PD patients with tremor without placebo effect, and tremor was the only symptom that had been ameliorated in the four typical motor symptoms of PD. We then analyzed the fMRI data that the DC, ReHo and ALFF analyses were consistent in the spinocerebellum, thalamus, default mode network (DMN), insula, visual areas and prefrontal cortex (PFC), and inconsistent in the cerebrocerebellum and motor cortex.

Interestingly, the acupuncture administration exerted a specific activating effect on the cerebrocerebellum (cerebellum crus1 and 2, and cerebellum 6) with decreased DC and ReHo values and enhanced ALFF signals, and a nonspecific effect on the spinocerebellum (cerebellum 4\_5 and vermis) associated with the enhanced ReHo and ALFF values. A significant change of the ReHo value in the thalamus and motor cortex (precentral gyrus and SMA) was also observed. Anatomically, the cerebrocerebellum receives input neural signals from the cerebral cortex and sends output signals mainly to ventrolateral thalamus in turn connected to motor areas of the premotor cortex and the primary motor area of the cerebral cortex, which is thought to be involved in the initiation, planning and coordination of movements (Kandel and Schwartz, 1985; Schmahmann et al., 1999). The conflicting signal changes in the cerebrocerebellum may be due to a compensation by the activated left precentral gyrus and SMA for the function of the cerebellum in integrating information with the motor cortex. The inconsistent ReHo changes in the left and right precentral gyri could be attributed to the fact that all subjects in our study were right-handed; in these patients, the improved function of the left precentral gyrus might be expected to compensate for the decreased function of the right precentral gyrus. The connections of the cerebellum, thalamus and motor cortex to the CTC circuit have been shown to influence patients' functional motor activity (Wu and Hallett, 2013). Thus, our data demonstrate acupuncture improve the motor function and daily life activities of PD patients by a direct stimulatory effect on the CTC circuit. Moreover, several imaging

#### TABLE 4 | Brain regions exhibiting increased and decreased regional homogeneity among three groups.


*TAG, true acupuncture group; SAG, sham acupuncture group; WG, waiting group. A* > *B, Compared with B group, A group showed increased ReHo values; A*<*B, Compared with B group, A group showed decreased ReHo values (P*<*0.05, AlphaSim corrected).*

studies have suggested that PD tremor is strongly associated with the CTC circuit (Helmich et al., 2011, 2012; Zhang J. et al., 2015), which may be the reason that tremor was the only motor symptoms being improved. Although this study may be the first time to elucidate how acupuncture affects PD tremor, evidence for tremor control via the CTC network has been accumulated by deep brain stimulation (DBS) and repetitive transcranial magnetic stimulation (rTMS), and several reports indicate that the stimulating mechanism of DBS or rTMS is potentially similar to that of acupuncture (Fukuda et al., 2004; Mure et al., 2011; Popa et al., 2013; Coenen et al., 2014).

We also observed that acupuncture had a specific effect on brain regions relevant to cognitive activity, such as the DMN (the anterior cingulate gyrus, precuneus, cuneus, medial PFC, inferior parietal lobule and hippocampus), visual areas (the lingual gyrus, superior and middle occipital gyri and the fusiform gyrus), insula and PFC (gyrus rectus, middle and inferior frontal gyri). The effect of DMN and insula on cognitive processing has been confirmed in recent fMRI studies investigating aging individuals and individuals with neurodegenerative disorders (Tessitore et al., 2012; Zhang S. Q. et al., 2015; Jiang et al., 2016). The visual processing controlled by the visual areas and the executive function managed by the PFC (Yuan and Raz, 2014) are also parts of the cognition. Since the subjects enrolled in this study were cognitively normal, these cognitive brain regions is speculated to participate in the cognitive management of movement, as movement control includes motor and cognitive components (Prevosto and Sommer, 2013). Generally, the cognitive control of movement is achieved by motion perception, movement learning, movement memory, movement planning and motor inhibition (Lu et al., 2012). For instance, in humans, an environmental stimulus related to motion is first perceived by the visual system, and the produced visual information is conveyed from the visual areas to the motor cortex to generate motion perception. Subsequently, an empirical rule is formed from the learning and memory of movement, which results from neural activity in the primary motor cortex, cerebellum, DMN and insula. Then, movement planning is regulated by the TABLE 5 | Brain regions exhibiting increased and decreased amplitude of low frequency fluctuations among three groups.


*TAG, true acupuncture group; SAG, sham acupuncture group; WG, waiting group; A* > *B, Compared with B group, A group showed increased ALFF values; A* < *B, Compared with B group, A group showed decreased ALFF values (P* < *0.05, AlphaSim corrected).*

basal ganglia, SMA and PFC to enable the execution of precise action (Lu et al., 2012). Therefore, the modulation of the neural activity in cognitive regions may contribute to the movement improvement along with the CTC circuit.

Several limitations of this study should be addressed. First, this was a small sample study, and the authors did not continue to assess the clinical symptoms and fMRI after the treatment course was discontinued. Second, there was 6 drop-outs out of 41 randomized patients (15%), which might cause bias in the conclusion. Third, we used MMSE to screen for the cognitive impairments as it's quite specific and easy to complete. Nonetheless, MMSE may not be as sensitive as the Montreal Cognitive Assessment to detect mild cognitive impairments. A study with a larger patient population, long-term follow-up and reliable cognitive tests is needed to more conclusively determine the efficacy and its neuromechanism of acupuncture on PD tremor and find out whether acupuncture could work on the cognition for a longer period of time by modulating the cognitive brain regions.

In conclusion, our findings reveal that acupuncture has specific and nonspecific effects on different brain regions involved in PD tremor, and the motor and cognitive management of movement. The underlying mechanism of the effects of acupuncture on PD tremor may be related to a modification of the CTC circuit, and the modulation of the cognitive functional regions together with CTC circuit contributes to enhancing movement and improving the daily life activities of PD patients.

#### AUTHOR CONTRIBUTIONS

ZL, BL, and XDL conceived and designed the experiments. JBC and XL performed the fMRI scans. JC analyzed the fMRI data. GHL performed the acupuncture. WYG and YSX analyzed the clinical data. SCH, SXH, MMZ, and XC collected the clinical data. YJW, CJC, JBC, and XDL recruited potential participants. and ZL, YYH, and PYX wrote the manuscript. All authors read and approved the final manuscript.

#### ACKNOWLEDGMENTS

This study was supported by the Guangzhou Postdoctoral International Training Program Funding Project, the National Key Research and Development Projects of China (2016YFC1306600, 2017YFC1310300), the National Natural Science Foundation of China (81471292, U1603281, U1503222, 81430021, 81603681), the Science and Technology Project of Guangdong Province (2015A030311021, 2016A020215201), a Science and Technology Planning Project of Guangzhou (201504281820463, 2018-1202-SF-0019) and an international project of science and technology for Guangdong (2016A050502025).

#### REFERENCES

Buckner, R. L., Sepulcre, J., Talukdar, T., Krienen, F. M., Liu, H., Hedden, T., et al. (2009). Cortical hubs revealed by intrinsic functional connectivity: mapping, assessment of stability, and relation to Alzheimer's disease. J. Neurosci. 29, 1860–1873. doi: 10.1523/JNEUROSCI.5062- 08.2009

Chae, Y., Lee, H., Kim, H., Kim, C. H., Chang, D. I., Kim, K. M., et al. (2009). Parsing brain activity associated with acupuncture treatment in Parkinson's diseases. Mov. Disord. 24, 1794–1802. doi: 10.1002/mds.22673

Coenen, V. A., Allert, N., Paus, S., Kronenbürger, M., Urbach, H., and Mädler, B. (2014). Modulation of the cerebello-thalamo-cortical network in thalamic deep brain stimulation for tremor: a diffusion tensor imaging study. Neurosurgery 75, 657–669. doi: 10.1227/NEU.0000000000000540


frequency fluctuations of healthy people in resting state functional magnetic resonance imaging. Zhongguo Zhong Xi Yi Jie He Za Zhi 34, 1197–1201. doi: 10.7661/cjim.2014.10.1197

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Li, Chen, Cheng, Huang, Hu, Wu, Li, Liu, Liu, Guo, Huang, Zhou, Chen, Xiao, Chen, Chen, Luo and Xu. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

**533**

# Inhibitory Control, Task/Rule Switching, and Cognitive Planning in Vascular Dementia: Are There Any Differences From Vascular Aging?

Krystallia Pantsiou<sup>1</sup> , Ourania Sfakianaki<sup>1</sup> , Vasileios Papaliagkas<sup>2</sup> \*, Dimitra Savvoulidou<sup>1</sup> , Vassiliki Costa<sup>3</sup> , Georgia Papantoniou<sup>4</sup> and Despina Moraitou<sup>1</sup> \*

<sup>1</sup> Lab of Psychology, Department of Experimental and Cognitive Psychology, School of Psychology, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>2</sup> Laboratory of Clinical Neurophysiology, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>3</sup> 1st Neurology Department, AHEPA Hospital, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>4</sup> Department of Early Childhood Education, School of Education, University of Ioannina, Ioannina, Greece

Recent studies have shown that patients diagnosed with Vascular Dementia (VaD) exhibit deficits in executive functions. According to "vascular hypothesis of cognitive aging," community-dwelling older adults having risk factors for vascular disease development (RVD) may suffer from cognitive decline of the same type. The aim of the study was to assess the level of specific executive functions (EF) that have been revealed as most affected by vascular abnormalities, in older adults with incipient VaD and RVD. Subsequently specific ways of EF measuring could be suggested for more accurate diagnosis of early stage VaD. The study compared three adult groups (N = 60): (a) patients diagnosed with incipient VaD, according to DSM-5 criteria (n = 20); (b) community-dwelling older adults presenting cardiovascular risk factors (RVD; n = 20); (c) healthy young adult controls (n = 20). Three types of executive functions were examined: inhibitory control, cognitive flexibility as rule/task switching, and planning. The following D-KEFS subtests were administered for their evaluation: The 'Color-Word Interference Test,' the 'Verbal Fluency Test,' and the 'Tower Test.' Mixed-measures ANOVA, MANOVA, and one-way ANOVA as well as Scheffe post hoc test were applied to the data of the scores in each condition of each test. The results showed that VaD patients had significantly lower performance in test conditions requiring switching and planning, compared to RVD group and young controls. The specific deficits of VaD patients, compared to older adults presenting RVD according to multiple-group path analyses were: more uncorrected errors in inhibition, the use of semantic knowledge primarily instead of switching ability to switch between semantic categories, as well as a lower level of movement precision in planning.

Keywords: cold executive functions, D-KEFS subtests, early stage vascular dementia, periventricular white matter hyperintensities, vascular hypothesis of cognitive aging

### INTRODUCTION

Recent epidemiological studies showed that Vascular Dementia (VaD) is the second most common type of dementia in Western countries after Alzheimer's disease (AD). Specifically, it affects approximately 5.6% of people older than 60 years old (O'Brien and Thomas, 2015). A meta-analysis of the prevalence of VaD showed that 26% of dementia cases met the NINDS-AIREN criteria for VaD, with a rate for people over 65 years between 0.6 and 2.1% (Kalaria et al., 2008;

#### Edited by:

Ana B. Vivas, CITY College, International Faculty of the University of Sheffield, Greece

#### Reviewed by:

Anna Emmanouel, CITY College, International Faculty of the University of Sheffield, Greece Ioannis Mavroudis, University of Leeds, United Kingdom

\*Correspondence:

Vasileios Papaliagkas vpapal@auth.gr Despina Moraitou demorait@psy.auth.gr

Received: 04 February 2018 Accepted: 28 September 2018 Published: 17 October 2018

#### Citation:

Pantsiou K, Sfakianaki O, Papaliagkas V, Savvoulidou D, Costa V, Papantoniou G and Moraitou D (2018) Inhibitory Control, Task/Rule Switching, and Cognitive Planning in Vascular Dementia: Are There Any Differences From Vascular Aging? Front. Aging Neurosci. 10:330. doi: 10.3389/fnagi.2018.00330

**534**

Rodriguez et al., 2008; Rizzi et al., 2014; Elahi and Miller, 2017). Many studies suggest that the rates of VaD increase with advancing age, a finding which is related to both genders (O'Brien and Thomas, 2015).

The term 'VaD' is used to describe a group of dementias associated with vascular brain lesions, which can cause ischemic lesions including large vessel disease (e.g., CVAs and large-vessel atherosclerosis) or small vessel disease (e.g., lacunar -very smallinfarcts in the deeper white matter of the brain). The use of the term is also linked to the description of dementias associated with extensive intracerebral hemorrhages or to more focal ischemic strokes (cerebral infarctions) impairing the efficient blood supply in cortical and/or subcortical brain regions, as well as to abnormalities in hemodynamic status (Reilly et al., 2010; Korczyn et al., 2012). The presence of vascular lesions in the brain has not always been associated with cognitive decline. There are cases where large ischemic lesions are not associated with cognitive deficits whereas small vascular lesions are likely to lead to cognitive decline and loss of functionality (O'Brien and Thomas, 2015). Vascular lesions have also been observed during 'normal aging' but they do not always cause dementia phenomena.

Hence, many researchers are concerned about the suitability of the term 'VaD' for the description of the aforementioned group of dementias and have suggested to replace it with the term 'Vascular Cognitive Impairment (VCI),' which refers to cognitive deficits induced by various cerebrovascular lesions (Hachinski and Bowler, 1993; Román et al., 2004; Erkinjuntti, 2007; Dichgans and Leys, 2017). The main reason for the suggestion of this replacement is that the 'traditional' criteria for defining VaD were based on, and affected by the criteria established for AD. Hence, memory loss and impairment of everyday functioning were considered diagnostic criteria for VaD whilst none of them was necessarily the case in early VaD (Bowler, 2007). Therefore, VaD could only be identified in late stages. This underestimation of VaD incidence led to the development of the new concept of VCI, the criteria of which can be applied to diagnose VaD in early stage, and subsequently to lead to timely treatment (Bowler, 2007). Similarly, DSM-5 (American Psychiatric Association, 2013) uses the term 'Vascular Neurocognitive Disorder (VNCD).'

The VNCD is distinguished into probable or possible. Probable VNCD is diagnosed if one of the following is present: (a) clinical criteria supported by neuroimaging evidence which confirms the cerebrovascular disease; (b) cognitive deficits that are temporally related to one or more documented cerebrovascular events; and (c) an evidence of both clinical and genetic cerebrovascular disease. Diagnosis of possible VNCD occurs when although clinical criteria are met there is no available evidence from neuroimaging and no establishment of the temporal relationship of neurocognitive disorder with one or more cerebrovascular events.

### Executive Functions in Vascular Dementia

Cognitive profiles of dementias, and especially of VaD, have been best understood due to the findings of recent neuropsychological research (Lindeboom and Weinstein, 2004; Traykov et al., 2005; Jacova et al., 2007; Nelson, 2007; Epelbaum et al., 2011; Hoffman, 2013). In this context, dysexecutive impairments have been revealed as an early feature of VaD in its various types. 'Executive Functions (EF)' (or otherwise 'Cognitive control') is an umbrella term for higher-order cognitive processes, which are supported by frontal regions and their distributed networks. They are considered as higher-order abilities because they coordinate other, cognitive, affective, and motor abilities during the execution of complex tasks (Lippa and Davis, 2010; Moreira et al., 2017). At a theoretical level, they have been divided into cold or cool EF and hot EF (Lippa and Davis, 2010; Moreira et al., 2017). Cold or cool EF include 'pure cognitive' or logical, goal-directed abilities such as planning, inhibition, flexibility/switching, and working memory updating. Hot EF abilities are activated when affect (e.g., motivation, emotion-regulation, metacognitive feelings etc.) is involved. Nevertheless, this distinction might not be straightforward and it has been suggested that the two types of EF are rather interdependent (Lippa and Davis, 2010; Moreira et al., 2017). Cognitive psychology and clinical neuropsychology approach EF -mainly but not only in terms of measurement- in different ways. The latter has tried to develop all-inclusive tasks/tests/scales to measure EF, in terms of the various EF abilities that should be recruited to perform in a measure. On the other hand, cognitive psychology 'supports' the development of 'clear' measures of EF. In other words, it stipulates that each task should be designed to measure a primary EF as well as component functions that contribute to performance on this task (Lippa and Davis, 2010; Moreira et al., 2017).

With regards to VaD, cold EF abilities have been extensively studied, as compared to hot EF abilities, and they are revealed, along with processing speed, to be among the first cognitive abilities that decline. Regarding memory functioning, some studies showed that in VaD patients a decline is noticed in the ability to organize the information that is going to be recalled. According to them this decline is attributed to the relationship between this specific ability and EF. The aforementioned conclusion results from the findings of a study (Di Cesare et al., 2012) in which a group of VCI participants, who were outpatients with a positive history of chronic cerebral vascular disorder, were compared with healthy younger and older adults, by means of a serial learning task of concrete frequent words. Similar conclusions were obtained from another study (Levy and Chelune, 2007) in which, groups of participants diagnosed with different dementia types, including VaD, were compared with the use of a battery of tests which are consisted of learning tasks, a verbal memory test, tasks measuring attention and executive functions as well as instruments measuring affect and emotions. In a recent longitudinal study, Livner et al. (2009) investigated two aspects of episodic memory, retrospective (free and cued recall task) and prospective (task developed for the purpose of the study), in people with AD and VaD as multi-infarct or strategic dementia. Based on the findings, they concluded that deficits in prospective memory in VaD are directly related to frontal – subcortical regions. Lamar et al. (2007) studied the relationship between leukoaraiosis in patients diagnosed

with AD, VaD (probable/possible), and mixed AD/VaD, and alterations in working memory capacity measured via a modified Digit Span Backward paradigm. Their results suggest that high degrees of leukoaraiosis in people with working memory deficits are connected, to some extent, to their inability to inhibit over-learned and automatic procedural memories.

In the same vein, it seems that language skills remain intact during the various stages of VaD (even in the most progressive stages) in contrast with verbal fluency. In specific, studies which compared groups of participants diagnosed with different dementia types, including VaD, as well as groups of cognitively impaired, non-demented (CIND and vascular CIND) participants, showed that people with VaD had low scores in phonemic fluency, (Duff Canning et al., 2004) a finding which was mainly attributed to dysfunction of frontal lobes (but see Jones et al., 2006). In the same vein, during the first phase of a longitudinal study (Bagnoli et al., 2012) participants with preclinical AD and preclinical VaD as multi-infarct or strategic dementia, were examined in category and letter fluency with the use of a task of early and late word generation 3 years before the dementia diagnosis. It was found that the two groups had the same performance in letter fluency. However, preclinical VaD group outperformed their preclinical AD counterparts in category fluency which is considered to be more dependent on the medial-temporal lobe.

Visuospatial skills have been assessed by the Clock Drawing Test (CDT). Mild visuospatial deficits are observed in patients with VaD. Such deficits are depicted not only in construction abilities but also in executive control function (Graham et al., 2004; Lee et al., 2009). In particular, studies in which the CDT was used to assess the visuospatial skills of patients with subcortical VaD, confirmed spatial and planning deficits in these patients. A possible interpretation of this finding associates the observed deficits with frontal-subcortical circuit dysfunction (Graham et al., 2004). Lee et al. (2009) who studied the characteristics of CDT errors in different types of dementia, including subcortical VaD, concluded that errors in VaD are mainly associated with impairments of planning functions.

Hence, most studies show that there is a relationship between VaD in its various types and deficits in executive functions (Moorhouse et al., 2010; De Witte et al., 2011). Moreover, two review studies (Gunning-Dixon and Raz, 2000; Oosterman et al., 2004) on cognition in cerebral small vessel disease (CSVD; it refers to pathological processes that affect small arteries, capillaries, and small veins in the brain), which is considered the most common cause of subcortical vascular pathology and corresponding cognitive impairment, resulted in findings supporting executive dysfunction in CSVD. However, one of them (Oosterman et al., 2004) showed that there is indeed a significant relationship between white matter lesions and executive functions but only with regards to timed tests.

Moreover, a recent review study (Vasquez and Zakzanis, 2015) aiming at examining the cognitive profile of vascular CIND, showed that, when persons with vascular CIND are compared to persons with non-vascular CIND, they demonstrate impairments in executive functions and processing speed. However, their comparison with healthy controls suggests performance decrements across a broad range of cognitive domains., Phonemic fluency, Stroop interference, Wisconsin Card Sorting Test in various versions, Trail Making Test B and B-A (switching), and Clock Drawing are the tools mostly used in the aforementioned studies (Vasquez and Zakzanis, 2015) to measure executive functions in VaD and VCI.

Nevertheless, today, diagnosis of small vessel disease, at least at the clinical level, includes executive dysfunction manifested not only by impaired capacity to use complex information and formulate plans, but also by incapability to exercise self-control (Wallin et al., 2018). In this context, new batteries, such as the 'Executive Interview 25,' the 'Frontal Assessment Battery,' the 'INECO Frontal Screening,' and the 'Frontier Executive Screen' have been developed to measure executive functions (Moreira et al., 2017). However, more research is needed, mainly at the psychometrics level in order their usefulness to be established comparatively to traditional tasks measuring the same abilities, as well as to tests of general intelligence. Moreover, future studies would be appropriate to show the usefulness of the aforementioned tools for each neurodegenerative condition (Moreira et al., 2017; Wallin et al., 2018).

Hence, the latest diagnostic criteria for VaD (American Psychiatric Association, 2013) include the presence of cognitive impairments in frontal - executive functions (Rockwood, 2002; Desmond, 2004). Deficits in executive functions are mainly associated with subcortical regions (e.g., thalamus, caudate) and their connection with frontal–subcortical circuits (Desmond, 2004; Royall et al., 2004; Moorhouse et al., 2010; Meguro et al., 2013). Greater deficits in these functions are more frequently observed in people with severe white matter lesions (Gunning-Dixon and Raz, 2003; Buckner, 2004; Nakamizo et al., 2013).

In this vein, the present study aimed at developing a small and reliable tool to differentiate diagnosis of VaD from cognitive aging due to vascular risk factors. This tool will be comprised by a battery of tests measuring specific executive functions that have been revealed as the most vulnerable dimensions of cognition in VaD.

### Vascular Aging in Community Dwelling Older Adults

The theoretical approach of the 'vascular hypothesis of cognitive aging' (Spiro and Brady, 2008) posits that risk factors for the emergence of vascular disease, affect primarily cognitive functions that are supported by the frontal brain regions (van den Berg et al., 2009; Jellinger, 2013; Kling et al., 2013). A disruption to fronto-subcortical network seems to be one of the main factors that contribute to decline in executive functions. The network might be damaged by white matter lesions, microbleeds affecting connecting pathways of the network, as well as lacunes or microbleeds at subcortical structures of the network (van den Berg et al., 2009; Jellinger, 2013; Kling et al., 2013).

At the conceptual level, the 'vascular hypothesis' (Spiro and Brady, 2011) posits that age is only a descriptive variable of cognitive declines that occur as people age. This theoretical perspective suggests health rather than age as a candidate explanatory process. The term 'health' includes 'disease' as the opposite end of a continuum starting with 'health.'

Older adults usually have health issues, such as elevations in various risk factors for disease, higher prevalence of disease, including subclinical situations, pure control of disease, widespread or even wrong use of medication etc. Taking all these into account, the 'vascular hypothesis' posits that instead of focusing on diagnostic categories, such as 'dementia,' and introducing new ones, such as 'Mild Cognitive Impairment (MCI),' 'Subjective Cognitive Impairment (SCI)' etc., in order to move faster on diagnosis, it might be more useful to consider cognitive aging itself as a long pathophysiological process of 'brain at risk' (Hachinski, 2007; Spiro and Brady, 2011). Recent findings indicating that even 'healthy' older adults can demonstrate a degree of vascular abnormality, suggest that there is indeed a continuum of vascular degeneration in aging, which can result in different degrees of cognitive decline, ranging from trivial to clinically relevant changes (Hachinski, 2007; Spiro and Brady, 2011).

Hypertension is one of the most important risk factors for vascular dysfunction (Moss and Jonak, 2007; Knecht et al., 2009; Sharifi et al., 2011; Areosa Sastre et al., 2017). Several studies suggest that hypertension causes pathological changes in the brain before the presence of acute events (e.g., stroke). Cognitive functions that are more affected by hypertension seem to be attention and executive functions (switching mainly) (Knecht et al., 2009; Sharifi et al., 2011; Areosa Sastre et al., 2017). In the same context, the effect of diabetes mellitus on vascular function and cognitive abilities (Yen et al., 2010; Blom et al., 2013; Hugenschmidt et al., 2013; Spauwen et al., 2013; Zhang et al., 2014) has also been demonstrated by relevant research as an additional risk factor. High levels of cholesterol, homocysteine, and ethanol (Schreurs, 2010; Ehrlich and Humpel, 2012; Niu et al., 2013; Galper et al., 2015) are included to additional physiological risk factors for cerebral vascular disorders.

Given that community dwelling older adults, without a diagnosis of dementia or MCI, usually present risk factors for developing vascular disease, it was considered useful to compare this group with patients diagnosed with incipient VaD, since the existence of risk factors could be a prodromal stage to the development of VaD.

#### The Purpose and the Hypotheses of the Study

The aim of the study was to assess the level of some executive functions that have been revealed as the most affected by vascular abnormalities, (van den Berg et al., 2009; Moorhouse et al., 2010; De Witte et al., 2011; Jellinger, 2013; Kling et al., 2013) in patients with incipient VaD and community dwelling older adults presenting vascular risk factors, in order to suggest specific ways of measuring the executive functions for a more accurate diagnosis of VaD at the neuropsychological level. Since we could hardly find any healthy older adults, presenting no vascular risk factors to participate in the healthy controls' group of the study, and as our aim was to reveal and estimate the extent of the potential declines in executive functions due to vascular risk factors and VaD, we decided to include a group of healthy young adults in the sample of the study. We considered it useful to assess young adults' performance as an 'index' of the best possible performance in the tests, since it is well-established in the literature of lifespan development that fluid intelligence, including cognitive control, is at the highest possible level in this age-group (Salthouse, 2010; Spiro and Brady, 2011).

The hypotheses of the study were formulated as follows:

Hypothesis 1: we hypothesized that older adults diagnosed with incipient VaD would show a lower performance in inhibition tasks, compared to community dwelling older adults who are at risk of developing vascular disease. Both older adult groups were expected to have a lower performance in the same tasks compared to a healthy young adult group.

Hypothesis 2: Similarly, it was expected that patients with incipient VaD would show a lower performance in tasks requiring a combination of inhibition and switching abilities mainly, compared to older adults with risk factors for vascular disease. Both older adult groups were expected to have a lower performance in the same tasks compared to healthy young adult group.

Hypothesis 3: Planning is a complex process which integrates into the theoretical framework of cognitive control: in fact, it represents, among others, the combined use of inhibition, switching, and updating processes. Thus, it was expected that older adults with a diagnosis of incipient VaD would show a lower performance in planning tasks, compared to community dwelling older adults who are at risk of developing vascular disease. Both older adult groups were expected to have a lower performance in the same tasks compared to healthy young adult group.

### MATERIALS AND METHODS

#### Participants

The sample comprised a total of 60 adults (29 males, 31 females). Their age ranged from 20 to 85 years (M = 57.73, SD = 25.26). There were three groups: (a) healthy young adult controls (n = 20, 6 men and 14 women, age range: 20–25 years, M = 22.75, SD = 1.58); (b) community dwelling older adults who are at risk of developing vascular disease (RVD, n = 20, 13 men and 7 women, age range: 71–85 years, M = 75.85, SD = 4.47); (c) older adults with a diagnosis of incipient VaD. The diagnosis of VaD was made according to the criteria of probable VNCD (DSM-5, American Psychiatric Association, 2013; VaD, n = 20, 10 men and 10 women, age range: 68–83 years, M = 74.60, SD = 5.17).

The two groups of older adults did not differ significantly in age [t(38) = 0.81, p > 0.05] and gender [t(38) = 0.94, p > 0.05]. Female gender was overrepresented compared to male gender in the young adult group. The participants' educational level varied; 23 participants (38.3%) had a low educational level (LEL: 0–9 years of education), 13 (21.7%) had a medium one (MEL: 10–12 years education), and 24 (40%) were highly educated (HEL: 13 or more years of education). The groups of older adults did not differ in educational level [χ 2 (2,40) = 0.120, p > 0.05]. In

the VaD group, 11 participants had a LEL, 7 had a MEL, and 2 had a HEL. In the RVD group, 12 participants had a LEL, 6 had a MEL, and 2 had a HEL. All participants of the young adult group had a high educational level.

Young adults and older adults with risk factors came from the general population. History of psychosis and existence of addiction were exclusion criteria for the whole sample. Hypertension, hypercholesterolemia, and diabetes were also exclusion criteria for young adults. These criteria were selected because the authors wanted to ensure the absence of factors that could possibly provoke vascular disorders in this group. Moreover, none of the young participants was taking medication.

On the contrary, the inclusion criterion for older adults with risk factors (RVD) was a self-report of a diagnosis of hypertension, hypercholesterolemia and/or diabetes (one or more of the above), while a self-reported diagnosis of a growth-related cognitive disorder (any type of dementia or MCI) constituted an exclusion criterion for this group.

Older adults diagnosed with incipient VaD were recruited from the inpatient unit of Memory disorders and AD of the 1st Neurology Department of the Aristotle University of Thessaloniki which is located in the AHEPA Hospital. For the diagnosis of VaD (as probable VNCD), patients underwent clinical examination, laboratory tests (blood tests, biochemical tests, thyroid tests, B12, folic acid, etc.) and neuroimaging (CT/MRI). The inclusion criteria were the following: (a) clinical evidence that was consistent with a vascular etiology [(1) temporal relationship between the presence of cognitive deficits and the onset of one or more strokes, (2) evidence of early decline in cognitive abilities, such as complex attention and frontal-executive functions], (b) a cerebrovascular disease which could be indicated by medical history, physical examination, and/or neuroimaging examination, (c) cognitive deficits that could not be explained by another organic or brain disease, and (d) a degree of cognitive decline that did not prevent the daily functioning of the person.

In particular, the diagnosis of VaD was given after clinical and neuropsychological examination. Clinical examination consisted of neurological examination, biochemical and neuroimaging tests (MRI/CT/SPECT). Neuropsychological examination included the Mini Mental State Examination (MMSE), (Folstein et al., 1975; Fountoulakis et al., 2000) the Sort Cognitive Performance Test (SKT), (Lehfield and Erzigkeit, 1997) the Clinical Dementia Rating scale (CDR), (Berg, 1988) the Ishihara Tests for Color-Blindness, (Ishihara, 1992), and the Hamilton Depression Rating Scale (HDRS) (Hamilton, 1960). Based on these examinations, the neurologist was able to diagnose VaD.

The participants in this study were selected among the patients diagnosed with incipient VaD, by one of the authors (Prof. Vassiliki Costa). They were patients diagnosed with specific lesions in the brain. The majority of them had been diagnosed with multi-infarct dementia (bilateral thromboembolic events), some others with leukoaraiosis and decrease in brain volume. It is important to note that none of the participants had been diagnosed with severe hemorrhagic or ischemic stroke in the past, however, as it was found in the MRI, some of them suffered small ischemic attacks. The functional level of all participants was not affected in specific domains as clothing, dressing, feeding and use of toilet. 75% of the VaD patients reported that they had hypertension, 35% hypercholesterolemia, and 30% diabetes and for these reasons they were taking medication. The Geriatric Depression Scale - 15 (GDS-15) (Sheikh and Yesavage, 1986; Fountoulakis et al., 1999) was used to examine the incidence of depressive symptoms. No scores indicative of depressive symptomatology were observed in the overall rating of the GDS-15 (score ≤ 5). In MMSE (Folstein et al., 1975; Fountoulakis et al., 2000) a wider band of scores (25–30) as compared to RVD group was observed. However, these values do not indicate the presence of cognitive impairment related to dementia (see **Table 1**).

In the RVD group, 75% reported having hypertension, 30% high cholesterol, and 40% reported a diagnosis of diabetes, and therefore they were taking medication. None of the participants in this group showed scores indicating depressive symptomatology in the GDS-15 (score ≤ 5). Another exclusion criterion used in this group was the performance in the MMSE, as scores under 24 are indicative of the presence of dementia symptoms. The scores of the participants ranged from 28/30 to 30/30.


<sup>1</sup>VaD, older adults diagnosed with Vascular Dementia according to the criteria of probable Vascular Neurocognitive Disorder; RVD, community dwelling older adults having risk factors for vascular disease development; YA, young adults. <sup>∗</sup>χ 2 (2,40) = 0.120, p > 0.05. ∗∗p < 0.05.

The mean GDS-15 score of the RVD group (M = 1.95, SD = 1.43) did not differ [t(38) = 1.03, p > 0.05] from the mean GDS-15 score of the VaD group (M = 1.50, SD = 1.31). On the contrary, there was a significant difference [t(27.4) = 5.14, p < 0.05] in the mean scores of MMSE between RVD (M = 28.9, SD = 0.718) and VaD group (M = 27.0, SD = 1.48). The first group had a higher performance.

#### Tools

Based on the literature suggesting early impairments of cold EF abilities, such as inhibition and planning, in VaD and VCI, and adopting a cognitive perspective for measuring EF, we chose to assess cold EF abilities, namely inhibition, cognitive flexibility/switching, and planning with the use of specific tests of the Delis-Kaplan Executive Function System (D-KEFS) (Delis et al., 2001). The D-KEFS includes tests that are administered to determine whether poor performance is due to specific impairment in EF, or impairment in lower-order cognitive abilities. Contrast measure scores, completion time, correct answers, and error analyses provided by the D-KEFS tests are factors which enable the assessment of an individual's executive functioning (Delis et al., 2001; Homack et al., 2005)

#### Delis-Kaplan Executive Function System: Color – Word Interference Test, Standard Form – D-KEFS C-WIT, SF (Delis et al., 2001)

This test is based on Stroop's experiment (1935) and measures the ability to inhibit a dominant and automatic verbal response. The participants have to inhibit an overlearned verbal response (e.g., reading the printed words) to generate a contradictory response (e.g., naming the color of the word). The test comprises four conditions: (a) naming the three basic colors (green, red and blue), (b) reading words written in black that are naming colors, (c) naming the dissonant ink colors in which the words are printed (inhibition), and (d) switching between naming the color of the words and reading the words (inhibition/switching). There is a time limit for each condition, i.e., 90 s in the first two conditions and 180 s for the following two. The examiner interrupts the test when the participant cannot complete the task within the given time.

There are two training sessions in each condition to ensure that participants have understood the task. In the first condition, the participant has to name the colors of squares (green, red, and blue). In the second condition, he/she has to read black-colored words naming the above three colors. The third condition comprises of color words written in incompatible color from the one they name. The participant has to name the color of the word instead of reading the word. Finally, in the fourth condition the participant has to (1) name the color of the word and (2) read the framed word instead of naming its color. The instructions for these tasks are displayed at the top left side of the stimulus booklet that is given to the participant. The test follows some discontinue rules as well. That is, the examiner administers the third condition only if the participant succeeds in finishing the first two ones the examiner is entitled to interrupt the administration of the test in case that participants have difficulty in the training sessions or if participants perform three consecutive uncorrected errors. Given that the contrast scores in this test have been found to display poor reliability, (Homack et al., 2005) it was decided that in this study, the examiner would record the time until the completion of the task with the restriction that there should not be an overrun of the time limit. The examiner had also to record the total of wrong answers in the form of corrected and uncorrected errors in each condition.

Plenty of clinical studies examined the validity of D-KEFS – C-WIT and showed that people with frontal lobe epilepsy, cardiovascular disorder, chronic kidney failure, AD, and Huntington's disease have a low performance in the third and fourth condition (Kramer et al., 2002; Delis et al., 2004).

#### Delis-Kaplan Executive Function System: Verbal Fluency Test, Standard Form – D-KEFS VF, SF (Delis et al., 2001)

This test assesses crystallized intelligence and executive functions. It is comprised of three testing conditions: (a) phonemic fluency, (b) semantic (category) fluency, and (c) category switching. In the first condition, the participant has to say as many words starting with the letters F, A, and S as he/she can within 60 s. The rules of this condition are as follows: (1) Names of persons or locations are excluded of the words to be stated by the participants, (2) Numbers do not count as words, and (3) Derivatives of the words already stated by the participants are not credited with additional points. In the second condition, the participant is asked to generate as many words that belong to a designated semantic category (animals and men names) as he/she can within 60 s. Finally, in the third condition the participants are asked to generate as many fruit and furniture names as they can, switching between these two semantic categories within 60 s.

The score of phonemic fluency condition consists of the total sum of the correct answers for each letter. The ranking scale is from 0 to 1, each correct word is credited with 1 point and the wrong ones (words that do not fall within the rules or word repetitions) with 0. The score of the second condition consists of the total sum of the correct answers in each category. The ranking scale is the same as in the first condition, namely 1 point for each correct word and 0 points for irrelevant or repeated words. Lastly, the third condition score is derived from the total sum of correct words in each category (3a) and the total sum of correct switches between the semantic categories (3b). Correct answers are marked with 1 point while wrong ones with 0 points.

D-KEFS VF factor analysis revealed a three-factor solution. The first factor was labeled "conceptual flexibility," the second one "monitoring," and the third one "inhibition" (Latzman and Markon, 2010; Fine and Delis, 2011). The results showed that the first condition (phonemic fluency) of the D-KEFS VF had modest loadings on inhibition factor (0.35) in the 20–49 age range and a higher loading (0.61) in the 50–89 age range. The second condition (semantic fluency) had modest loadings on monitoring factor (0.39) in the 20–49 age range and higher loadings on inhibition factor (0.60) in the 50–89 age range. Lastly, the third condition (category switching) had an extremely high loading on monitoring factor (0.97) in both age ranges.

The test also assesses crystallized intelligence, the ability to use learned knowledge and experience. In other words, it is

associated with the accumulated knowledge that is derived from the individual's vocabulary depth and from the world knowledge.

#### Delis-Kaplan Executive Function System: Tower Test, Standard Form – D-KEFS – TT, SF (Delis et al., 2001)

In this test, participants have to make towers along three vertical pegs using five disks. Each participant has to solve the issue of the task using a specific minimum number of moves. The test consists of a wooden board with three vertical pegs, five wooden disks that vary in size and color (shades of blue), a timer, a recording and a stimulus booklet. The test comprises nine problems of increasing difficulty. Each problem has a specific time for completion, 30 s for the first three problems, 60 s for the fourth, 120 s for the fifth and sixth, 180 s for the seventh, and 240 s for the eighth and ninth problem. There are two rules in this test which are given both orally and in writing. The first rule is that only one disk can be moved at a time, and the second one is that no disk must be placed on top of a smaller disk. In each problem the examiner slides the disks onto the pegs, in a specific order, and shows the participant a picture of the tower he has to do (the final position of the disks). In the first problem, the participant has to use two disks the number of which is increased during the administration of the test. If the participant fails to solve the problem, the examiner shows him/her the solution. This aid only applies to the first and second problems. The examiner records the completion time of the first move (optionally), the total number of moves, the whole of rules violation (optionally), the completion time of the problem, and whether the problem was solved (yes or no). If the participant exceeds the time provided, the examiner interrupts the resolution of the problem and scores it with 0. The examiner interrupts the administration of the test in case that participants perform three consecutive failures (i.e., exceeded the given time, making an incorrect version of the tower, using fewer moves than the those origianally predefined).

The scores of D-KEFS TT used in this study are the following: (a) the total number of administered problems, (b) the total number of violations of the rules, (c) the total achievement score which is derived from the sum of the total of correct solutions (the scoring is based on the number of moves). The range of raw scores extends from 0 (zero problems solved) to 30 (all problems solved within the given time and with the minimum number of moves). In the present study, (d) the precision of movements was also measured, which results from the ratio of the total number of movements to the total number of the minimum predefined movements for each administered problem. This measurement assesses the state of planning awareness and the ability to learn heuristic strategies when solving problems.

D-KEFS TT is a tool which assesses complex executive functioning. Factor analysis (Latzman and Markon, 2010) showed positive loadings on spatial planning, learning rules, inhibition, and the ability to define and maintain cognitive sets. The validity of the tool has been studied by various studies in people with brain damage. A recent study showed that people with cardiovascular disease had achieved a lower overall score compared to control group (Jefferson et al., 2006; Fine and Delis, 2011).

#### Procedure

The above tools were merged to create a battery. Two different versions of this battery, as far as the order of administration is concerned, were designed to avoid order effects. The battery tests were individually administrated and the length of the test administration was almost an hour. The examination took place in a quiet and comfortable environment.

Young adults and older adults presenting risk factors for developing vascular disease were recruited from the community by the first author (Krystallia Pantsiou) via convenience sampling. They were examined at a place of their own choice. Patients diagnosed with incipient VaD, were recruited from the inpatient unit of Memory disorders and AD, of the AHEPA hospital, and they were examined by the same author in an office of the unit.

### Ethics Statement

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All participants participated voluntarily in the study. They were informed about the procedure and the aim of the study, and subsequently they provided their written consent for participation. The Ethics Committee of the School of Psychology of the Aristotle University of Thessaloniki, after reviewing the research protocol, confirmed that all ethical guidelines for research on human subjects were followed (011/16-06-2017).

### Statistical Analysis

The data analysis was conducted in SPSS version 21 (IBM Corp, 2012). The analyses carried out were (a) mixed-measures ANOVA, (b) repeated measures ANOVA, (c) multivariate analysis of variance (MANOVA), (d) one-way ANOVA. The aim of these analyses was to compare the performance between the three groups as well as the performance of each group in each condition of a test. Levene's test was used to assess the equality of variances, Box's Test for the assessment of the equivalence of covariance matrices, and Mauchly's test of sphericity for the assessment of within-subject factor. Authors used Greenhouse-Geisser for thecorrection of sphericity violations. Partial eta-squared (η 2 ) was used for the estimation of the effect size. Scheffe test (plus Bonferroni correction) was adopted for post hoc comparisons: even if there are issues related to test's conservatism, given the exploratory 'nature' of this study, we considered Scheffe procedure as the best choice due to the following reasons: it tests all possible comparisons, it is robust in relation to non-normality, and it provides maximum protection against type I error, which was our main concern in this study (Armstrong and Hilton, 2006).

In the next step, data analysis was conducted in EQS version 6.1 (Bentler, 2005). Specifically, structural equation modeling (SEM) on covariance matrices was used. The 'Robust maximum likelihood estimation' procedure was performed due to small sample size and data kurtosis. The specific SEM technique applied to the data was path analysis. This technique was used to examine

the directional relations between the constructs – observed variables which emerged from the measurements of each test (Kline, 2005). Lagrange Multiplier and Wald test (Brown, 2006) were conducted to investigate the elements of the model that didn't fit the data well. In specific, a series of multiple-group path analysis were conducted to examine and compare the relations between the variables in the three groups, in order to reveal potential differences in the way participants of each grouprecruit executive functions when needed. Regarding the confirmation of a path model, a significance level of Goodness of Fit Index χ 2 that is p > .05 is indicative of a good fit of the model to the data. In this study, authors calculated the Satorra-Bentler χ 2 index (Bentler, 2005) due to the statistical analysis they chose to perform (Robust procedure). In addition, when the value of Root Mean Square Error of Approximation (RMSEA) is <0.05, it is also an indication of the good fit of the model to the data. RMSEA values ranging from 0.06 to 0.08 indicate a reasonable and therefore acceptable approximation error. RMSEA value is relatively "expanded" in cases of small sample size (n < 100) and that is reflected in confidence interval range (90% CI). This means that RMSEA should be considered as a model fit index, however, with caution (Bollen, 1990). Comparative Fit Index (CFI) examines whether the data fit a hypothesized measurement model compared to the basic model. Values greater than 0.90 indicate adequate fit of the model to the data, whereas values close to 1.00 indicate a good fit.

#### RESULTS

#### Color – Word Interference Test, Standard Form: Inhibition and Switching

Initially, means and standard deviations (SD) of test scores in all C-WIT conditions were calculated for all groups. Subsequently, a series of analyses of variance (ANOVA) was applied in order to examine the 'quantitative' differences between the three groups in the variables under examination for the first three C-WIT conditions. Finally, a Multiple-Group Path Analysis was performed for the third C-WIT condition in order to examine the main differences between the three groups, in cases of the application of the executive function of inhibition.

#### Total Corrected and Non-corrected Errors

A 3 × 3 mixed design ANOVA (three groups of participants × three conditions of C-WIT) showed a significant main effect of the participant group, F(2,57) = 15.84, p < 0.0001, η <sup>2</sup> = 0.35, condition, F(1.07,61.04) = 53.21, p < 0.0001, η <sup>2</sup> = 0.48, as well as a significant group – condition interaction, F(2.14,61.04) = 21.21, p < 0.0001, η <sup>2</sup> = 0.42.

With regards to the significant main effect of the participant group, Scheffe's method of multiple comparisons showed that more statistically significant errors were made by the VaD group compared to the RVD group, p = 0.01, and to young adult group, p < 0.0001. The difference between the last two groups was non-significant, p = 0.057. A series of repeated measures ANOVA applied to the data in each group, showed that the group of young adults did not show statistically significant differences in the total number of errors in all three C-WIT conditions, F(1.23,23.49) = 2.13, p = 0.07. On the other hand, the group of adults with RVD showed differences in the errors in all three conditions, F(1.22,23.28) = 11.78, p = 0.001, with the highest number of errors (corrected and not corrected) being observed in the third condition, followed by the first and the second conditions. The VaD group showed significant differences in the total number of errors, F(1.00,19.17) = 39.55, p < 0.0001, with a greater number of errors being observed in the third C-WIT condition compared to the second condition (see **Table 2A**).

TABLE 2A | D-KEFS Color – Word Interference Test, Standard Form: comparisons of the three sample groups in total scores as well as of each group performance in the first three conditions.


<sup>1</sup>VaD, older adults diagnosed with Vascular Dementia according to the criteria of probable Vascular Neurocognitive Disorder; RVD, community dwelling older adults having risk factors for vascular disease development; YA, young adults. <sup>∗</sup>p < 0.017 (after Bonferroni correction).

#### Task Completion Time

fnagi-10-00330 October 15, 2018 Time: 19:27 # 9

A 3 × 3 mixed design ANOVA was applied with the participant group (young adults, RVD, VaD) as a between-group variable and the C-WIT condition (first three conditions) as a within-group variable. The results of the analysis showed that there is a main group effect, F(2,57) = 34.25, p < 0.0001, η <sup>2</sup> = 0.54, a main condition effect, F(1.01,62.94) = 212.84, p < 0.0001, η <sup>2</sup> = 0.78, as well as a significant group – condition interaction, F(2.20,62.94) = 32.04, p < 0.0001, η <sup>2</sup> = 0.52.

With regard to the main group effect, Scheffe's multiple comparison showed that the VaD group had a significantly longer task completion time compared to young adults, p < 0.0001. The VaD group also differed from the young adult group, p < 0.0001.

For further analysis of the main condition effect, repeated measures ANOVA was applied to the data of each group. For the group of young adults, F(1.28,24.47) = 145.63, p < 0.0001, significant differences were observed in the completion time between the three conditions. Similar results were observed for both the RVD group, F(1.22,23.17) = 55.06, p < 0.0001, as well as for the VaD patients, F(1.03,19.70) = 109.28, p < 0.0001. However, despite the fact that the prototype of differences observed is almost similar among all groups, the young adult group spent much less time to complete the 3rd condition, compared to older adults with RVD and VaD (see **Table 2A**).

#### Multiple-Group Path Analysis Models for the 3rd C-WIT Condition (Inhibition)

This analysis investigated the directed relations between the total number of errors (corrected and non-corrected), the total number of corrected errors, the total number of uncorrected errors and the completion time of the third condition in all three groups of participants. The results showed that different path models were confirmed for each group, Satorra-Bentler χ 2 (3, N = 60) = 2.20, p = 0.53, CFI = 1.00, RMSEA = 0.00 (90% CI: 0.00 −0.19). In specific, regarding the group of young adults, no relations were observed between the measurements. However, the model for RVD adults appears to be interesting. In this case, it was found that the number of uncorrected errors is positively associated with the completion time (i.e., associated with increased time), while the number of corrected errors has a compensatory function, as associated with reduced completion time. Furthermore, a different model was confirmed for VaD adults, since the total number of uncorrected errors is positively associated with both the total number of errors and the completion time of the condition (see **Figure 1**).

#### Findings Regarding the Score in the 4th C-WIT Condition (Inhibition and Switching)

In this condition, only 8 participants from the group of VaD adults completed the test. Following the instructions, the fourth condition was not administered, because participants experienced great difficulty in the practice lines of this condition. For this reason it was not included in the aforementioned analyses. Regarding the group of RVD adults, although only one participant could not complete the fourth condition, the differences from young adults in both the number of errors and the completion time of the condition were statistically

significant: Errors: F(1,38) = 21.71, p < 0.0001. Completion time: F(1,38) = 51.51, p < 0.0001 (see **Table 2B**).

### Verbal Fluency Test: Inhibition and Switching

Means and SD were calculated for VFT conditions for all groups. Then, a series of analyses of variance was applied in order to examine the differences between the three groups on the relevant variables. Finally, based on Pearson correlations between variables and taking into account the sample size, a multiplegroup path analysis was applied to the data of the second and third condition of VFT.

TABLE 2B | D-KEFS Color – Word Interference Test, Standard Form: comparisons of the sample groups in the fourth condition: Switching between naming the color of the words and reading the words.


<sup>1</sup>VaD, older adults diagnosed with Vascular Dementia according to the criteria of probable Vascular Neurocognitive Disorder; RVD, community dwelling older adults having risk factors for vascular disease development; YA, young adults. ∗∗∗p < 0.001.

#### Total Correct Answers-Words

fnagi-10-00330 October 15, 2018 Time: 19:27 # 10

Initially, a 3 × 3 mixed design ANOVA was performed with the participant group (young adults, RVD, VaD) as a between-groups factor and the condition of VFT (three conditions) as a within-subjects factor. A significant main effect of the participant group was found, F(2,57) = 78.20, p < 0.0001, η <sup>2</sup> = 0.73, as well as a main effect of the condition type, F(2,114) = 288.35, p < 0.0001, η <sup>2</sup> = 0.83, and a significant group – condition interaction, F(2,114) = 11.95, p < 0.0001, η <sup>2</sup> = 0.29.

Subsequently, analysis of variance was applied with the participant group as the independent variable and performance in each of the three conditions as the dependent variables: According to Pillai's trace, the group effect was found to be significant, V = 0.78, F(2,57) = 12.03, p < 0.0001, η <sup>2</sup> = 0.39. In particular, the group effect on the total number of correct responses in the phonemic fluency condition (1), F(2,57) = 42.95, p < 0.0001, η <sup>2</sup> = 0.60, in the semantic fluency condition (2), F(2,57) = 56.13, p < 0.0001, η <sup>2</sup> = 0.66, and in the semantic fluency condition under switching rules (3a), F(2,57) = 24.74, p < 0.0001, η <sup>2</sup> = 0.46, was statistically significant. Scheffe's post hoc comparison showed that young adults produced more words in the first condition, as compared to RVD and VaD groups, p < .0001. No significant differences were observed between the two older adult groups, p = 0.48. Similar results were observed in the second condition: young adults had a higher performance compared to RVD and VaD groups, p < 0.0001, whereas the difference between the two older adult groups was not significant, p = 0.13. It was found that in the condition '3a' VaD adults produced a reduced number of words, compared to RVD, p = 0.016, and young adult groups, p < 0.0001. A significant difference was also observed between the last two groups, p < 0.0001 (see **Table 3**).

For a further analysis of the main condition effect repeated measures ANOVA was used to compare the data of each group. Young adults showed a statistically significant difference in the responses produced in all three conditions, F(2,38) = 138.90, p < 0.0001. In particular, they produced a greater number of correct words in the first and second condition, compared to the third, p < 0.0001. Similar results were observed for the RVD group, F(1.48,28.17) = 50.57, p < 0.0001, and also for the VaD group, F(2,38) = 150.52, p < 0.0001 (see **Table 3**).

#### Total Repetitions and Total Errors in the Three Conditions of VFT

A univariate analysis of variance showed that the three groups did not differ significantly in the total number of repetitions they did, F(2,57) = 1.86, p = 0.16, as well as in the total number of their errors, F(2,57) = 1.76, p = 0.18.

#### Total Number of Switches in the Third Condition (Semantic Fluency Under Rules of Switching) of VFT (Condition '3b')

A univariate analysis of variance was performed, that showed a significant effect of the participant group on the total number of correct switches, F(2,57) = 13.09, p < 0.0001, η <sup>2</sup> = 0.31. Scheffe's multiple comparison showed that young adults achieved a significantly higher number of switches, compared to RVD and VaD groups, p = 0.006 and p < 0.0001, respectively (**Table 3**). However, there was no significant difference between the two groups of older adults, p = 0.27.

#### Multiple-Group Path Analysis Models for VFT

A multiple-group path analysis was used to examine the directed relations between semantic fluency (2nd condition), the words produced in the condition of semantic fluency with switching (condition '3a') and the total number of switches (condition '3b'). It must be noted that the error in measuring word production under switching condition was set to be equal in the two older adult groups. The data from young adults were not used in the analyses, because the aforementioned variable-scores were not significantly inter-correlated. The results for the two older adult groups, Satorra-Bentler χ 2 (1, N = 40) = 0.70, p = 0.39, CFI = 1.00, RMSEA = 0.00 (90% CI: 0.00 −0.39), showed that for the RVD group it was confirmed a model in which the number of switches (condition '3b') is positively associated with the production of words in the condition '3a'. On the other hand, a different model was confirmed for VaD adults. In this case, semantic fluency (2nd condition) appeared to be positively associated with the word production under switching rules (3a condition) (**Figure 2**).

#### **Tower test: planning**

Means and SDs were calculated for all measurements in all TT conditions for all groups. Subsequently, a series of analyses of variance was applied in order to examine the differences between the three groups of the sample on the variables under

TABLE 3 | D-KEFS: Verbal Fluency Test: Standard Form: comparisons of the three sample groups as well as of each group performance in the three conditions.


<sup>1</sup>VaD, older adults diagnosed with Vascular Dementia according to the criteria of probable Vascular Neurocognitive Disorder; RVD, community dwelling older adults having risk factors for vascular disease development; YA = young adults. <sup>∗</sup>p < 0.017; ∗∗∗p < 0.001.

examination, namely the total number of problems given, the total number of violations, the precision of movements and the total achievement score. Based on the calculated Pearson correlations between variables of TT and taking into account the small sample size, a multiple-group path analysis was applied to the data of TT.

In more detail, analysis of variance was applied with the participants' group as the independent variable and the four performance variables as dependent ones: the main effect of group was found to be significant, V = 0.85, F(2,57) = 10.33, p < 0.0001, η <sup>2</sup> = 0.42. In particular, the group effect was significant on (a) the total number of problems given, F(2,57) = 17.47, p < 0.0001, η <sup>2</sup> = 0.38, (b) the total number of rule violations, F(2,57) = 17.09, p < 0.0001, η <sup>2</sup> = 0.37, (c) the precision of movements, F(2,57) = 11.79, p < 0.0001, η <sup>2</sup> = 0.29, and (d) the total achievement score, F(2,57) = 33.45, p < 0.0001, η <sup>2</sup> = 0.54.

Scheffe's multiple comparison showed that VaD patients were given less problems (because of their inability to achieve the task or to resolve it correctly) compared to young adults and RVD group, p < 0.0001 and 0.001, respectively. Similar findings were found for the total number of violations, where significant differences were observed among all three groups, with the group of VaD patients showing the higher number of rule violations, p < 0.0001 and 0.004, respectively. A reduced precision of movements was observed in the VaD group compared to the other two groups, p < 0.0001 and 0.002, respectively. Finally, regarding the total achievement score, significant differences were observed among the three groups, with young adults to show higher score compared to RVD and VaD groups, p < 0.0001. The difference between the two older adult groups was also found to be significant, p = 0.008 (see **Table 4**).

#### Multiple-Group Path Analysis Models for TT

A multiple-group path analysis was applied to TT data to examine the directed relations between the total number of TABLE 4 | D-KEFS: Tower Test: comparisons of the three sample groups in four scores.


<sup>1</sup>VaD, older adults diagnosed with Vascular Dementia according to the criteria of probable Vascular Neurocognitive Disorder; RVD, community dwelling older adults having risk factors for vascular disease development; YA, young adults. <sup>∗</sup>p < 0.013 (after Bonferroni correction).

the problems given, the precision of movement and the total score achieved in all groups, for TT data. The results showed that different models were confirmed for each group, Satorra-Bentler χ 2 (1, N = 60) = 0.90, p = 0.34, CFI = 1.00 and RMSEA = 0.00 (90%CI: 0.00 −0.33). Regarding young adults, there was no relationship between measurements. For the RVD group, the confirmed model showed that only the total number of problems given was positively associated with the total score achieved. Regarding the VaD group, the confirmed model showed that both the total number of given problems and the precision of movements were positively linked to the total score

achieved. There was also a significant correlation between the precision of movements and the number of problems given (see **Figure 3**).

### DISCUSSION

### Cognitive Control in Vascular Dementia: Inhibition

Based on the extant literature supporting EF deficit in VaD, (Moorhouse et al., 2010; De Witte et al., 2011; Vasquez and Zakzanis, 2015; Moreira et al., 2017) the aim of this study was to go a step further in terms of examining whether older patients with incipient VaD would show a low performance in cold EF tasks mainly requiring the application of three types of executive functions, namely inhibition, switching (or set-shifting), and planning. The investigation of discrete EF abilities, in which older adults diagnosed with incipient VaD and community dwelling older adults with risk factors for vascular disease development (RVD) could show a pattern of 'clear' differences, was considered a subject with great interest, which would enable the development of a short and reliable diagnostic tool for early stage VaD.

'Inhibitory control' is the process of altering one's learned behavioral responses in a way that makes it easier to achieve a particular goal. In this vein, an inhibitory control task measures one's ability to overcome their habitual or dominant response to a stimulus in order to implement a different response. In this study, inhibition was examined with the D-KEFS Color - Word Interference Test (Delis et al., 2001) which is based on selective information processing in the presence of disruptive stimuli.

According to the findings, the number, and mainly the type of errors (mostly uncorrected) in D-KEFS – C-WIT inhibition condition (3rd) might be an indicator of the differentiation between incipient VaD patients and community dwelling older adults presenting risk factors for vascular disease development. Theoretically, it has been argued that the ability to correct errors requires control and adjustment of behavior and is related to metacognitive skills which are impaired in old age (Delis et al., 2001). Thus the inability of VaD patients to correct their wrong responses may reflect a deficit in their metacognitive control system. Based on the findings, the first signs of such a deficit may be evident in community dwelling older adults with RVD, as compared to healthy young controls. However, it seems that at this phase individuals maintain their inhibitory control at a relatively adequate level, since they can correct a part of their errors, displaying metacognitive control to some extent, and perform well in tasks requiring inhibition.

With regards to the completion time, the findings indicated that in the inhibition condition, VaD patients took more time to complete the task, compared to young adults. According to Delis et al. (2001) persons who demonstrate a relatively good performance in naming and reading (1st and 2nd conditions), but have an increased response time and mainly make uncorrected errors in the inhibition condition, tend to be characterized by severe deficits in inhibitory control. Given the aforementioned findings, this pattern of performance is fully confirmed for VaD patients. In addition, first signs of deficit in inhibitory control, in terms of response time, /become visible in older adults with RVD when they are compared to young adults.

Taking into account the significantly increased overall response time and the total error score of VaD patients, in the D-KEFS – C-WIT compared to the other two groups, it can be suggested that the Hypothesis 1 was partially confirmed at least in the case of examining inhibitory control via a Stroop-type task: older adults diagnosed with incipient VaD have generally shown lower performance in the D-KEFS – C-WIT compared to older adults with RVD and young adults, while there are some indications of inhibitory control deficit in RVD group compared to young adults. In verification of these findings, Kramer et al. (2002) suggest that subcortical ischemic vascular disease is associated with impaired inhibition even in individuals without a diagnosis of dementia. It is noteworthy that in previous studies a significant contribution of the frontal and subcortical regions has been found in the ability to inhibit, (Lenartowicz et al., 2010) and especially of the anterior cingulate cortex, (Alvarez and Emory, 2006) and frontal gyrus (Leung et al., 2000).

A recent systematic review and meta-analysis study (Sudo et al., 2017) suggested a specific inhibitory control deficit in small vessel cognitive impairment (SVCI), based on the findings of a series of studies that used the Stroop test. The results of this review showed significant differences between SVCI patients and healthy controls in the number of errors as well as in the completion time of the 3rd condition of the Stroop test. The suggested interpretation for this deficit is that it is potentially due to vulnerability to interlobar disconnection on account of periventricular white matter abnormalities (Banich et al., 2000). Neurophysiological studies have indicated that periventricular white matter receives blood supply from terminal vessels of long perforating arteries in areas where branches of large cerebral vessels are encountered. However, most such arteries are very tortuous and this makes their areas vulnerable to hypoperfusion due to arteriosclerosis (Banich et al., 2000). In the same vein, another study (Dunet et al., 2016) which examined the relationship of cognitive impairment with functional connectivity between the basal ganglia and cingulate cortex in vascular parkinsonism, showed that inhibitory control and errors in the Stroop test are related to increased caudate nucleus functional connectivity with the perigenual anterior cingulate cortex, and decreased caudate nucleus functional connectivity with the posterior cingulate cortex at resting-state, respectively. White matter lesions, indicative of small vessel disease, were found to partially contribute in this pattern.

The D-KEFS VF was also used in this study to examine inhibitory control (see Materials and Methods). The findings showed that it is not able to differentiate VaD patients and older adults with RVD. A possible reason for this may be that D-KEFS VF doesn't require the same level of inhibitory control as the D-KEFS C-WIT, while both older adult groups can function relatively well at the level required by D-KEFS VF. In any case, this finding is important, considering the widespread use of the verbal fluency task in neuropsychological examination as well as the findings of previous studies which showed a relationship of VaD and deficits in phonemic fluency (Duff Canning et al., 2004;

Jones et al., 2006). However, the aforementioned meta-analysis and a series of studies on VCI did not replicate this relationship (Banich et al., 2000). Methodological reasons (e.g., different stage of VCI) may be responsible for this inconsistency.

### Cognitive Control in Vascular Dementia: Inhibition plus Switching

'Task or rule switching' represents a kind of cognitive flexibility that involves the ability to shift attention between one task/rule and another. This ability allows a person to adapt efficiently to different environmental situations. In this study, switching was examined via the administration of the D-KEFS C-WIT (condition 4) and the D-KEFS VF (condition 3). Despite the fact that the two tests examine the same EF abilities, namely switching in conjunction with inhibitory control, it was considered that different results could be found due to the different lower-order cognitive abilities required to perform in each test.

In relation to the 4th condition of the D-KEFS C-WIT, the findings showed that community dwelling older adults with RVD presented a lower performance compared to young adults. The conclusion drawn is that the ability to control a situation declines when task requirements increase and this is particularly obvious in older adults. This finding seems to corroborate those of previous studies suggesting significant age-related effects on the combined application of inhibition and switching (Bryan and Luszcz, 2000; Plumet et al., 2005; Adornetti, 2016; Lam et al., 2017).

Nevertheless, in the present study, many VaD patients failed to complete the 4th condition. This is indicative of the increased difficulty of the patients with the specific neuropathology compared to community dwelling older adults, when they are asked to recruit two executive functions in combination, one of which had to be an expression of cognitive flexibility. Hence, it appears that failure to complete 'inhibition plus switching' condition of the D-KEFS C-WIT recommends a performance standard which can work as a 'clear indicator' of the differentiation between VaD patients and older adults with RVD.

However, with regard to the differentiation between the last group and young adults, confounding variables such as age-related factors and education should be taken into account, (Bryan and Luszcz, 2000; Plumet et al., 2005) in addition to potential early signs of neurodegeneration.

According to the findings, older adults with VaD and with RVD differentiated in the 3rd condition of the D-KEFS VF. Delis et al. (2001) claim that normal performance in phonological fluency (1st) and semantic fluency (2nd) conditions of the D-KEFS VF together with low performance in the semantic fluency under switching (3rd) condition indicate deficits in cognitive flexibility and not in verbal fluency. Based on the findings, it appears that such deficits are so severe in incipient VaD patients that they rely on their semantic knowledge to complete this condition without practically employing switching ability. Inversely, increased performance under switching recruitment (number of words generated) of older adults with RVD reflects that there is no such severe decline in set-shifting ability in this group, compared to VaD patients. On the other hand, the higher performance of young adults in the same condition compared to older adults with RVD, may indicate a progressive decline in cognitive flexibility along with age as well as its corresponding health deterioration (Bryan and Luszcz, 2000; Gunning-Dixon and Raz, 2003; Plumet et al., 2005; Spiro and Brady, 2008; Adornetti, 2016; Lam et al., 2017).

Several studies argue that there is an increase in interference and a decline in switching in people with severe white matter lesions (Lam et al., 2017). According to the findings of the same systematic review and meta-analysis study, (Sudo et al., 2017) which examined inhibitory control measured by the Stroop test, the decline of cognitive flexibility, as measured by the Trail Making Test – B in small vessel disease, also represents a specific EF deficit that might be attributed to the increased vulnerability to interlobar disconnection due to periventricular white matter hyperintensities.

Furthermore, Tsutsumimoto et al. (2015) used in their study the Trail Making Test to measure switching in MCI patients. According to their findings (gross) gray matter volume was both significantly and negatively associated with this EF ability. In the same light, another study in which gray matter correlates (regions of interest) of set-shifting in neurocognitive disorders was examined with the use of three D-KEFS tests [Design Fluency (non-verbal analog to Verbal Fluency), Trail Making, and Color-Word Interference] (Delis et al., 2001) and voxel-based morphometry, suggests that the switching performance correlated with focal regions in prefrontal and posterior parietal cortices. Interestingly, bilateral prefrontal cortex and the right posterior parietal lobe were identified as common sites for switching across all tasks. This finding is important as different types of tasks that measure set-shifting appear to have different underlying neuroanatomy to some extent (Kramer et al., 2007; Pa et al., 2010). Hence, VaD patients might experience lesions at least in the common areas that support switching, even from the initial stage of the disease.

In conclusion, the findings of this study showed that the level of switching ability can differentiate VaD even in the very first stage (Hypothesis 2). In fact, it appears that cognitive flexibility deficits are more severe and more obvious in incipient VaD compared to inhibitory control decline. Thus, the neuropsychological tests that are used to diagnose early stage VaD should include more than one conditions examining set-shifting.

### Cognitive Control in Vascular Dementia: Planning

'Cognitive planning' encompasses the neuropsychological processes involved in the formulation, evaluation and selection of a sequence of thoughts and actions in order a desired goal to be achieved. It is considered a complex EF ability in terms of involving functions such as the 'updating' component of working memory, inhibitory control, and task/rule switching (Jefferson et al., 2006).

In this study, young adults and the two older adult groups showed a significant difference between them in two of the four examined variables (total achievement score and total rule violations). Community dwelling older adults with RVD were

significantly worse than young adults in the aforementioned variables but they had almost the same level of precision in movements and solved the same number of Tower Test problems as young adults: this pattern of performance could be indicative of a progressive decline in planning along with age that starts with rule violations which indicate deficits in production and preservation (Miyake et al., 2000; Homack et al., 2005; Lee et al., 2009; Moreira et al., 2017). A step further, incipient VaD patients showed lower performance than the other two groups in all variables measured: this finding reflects a more generalized deficit of planning ability in VaD. Therefore, D-KEFS – TT appears to be a very promising tool for differentiating incipient VaD from vascular cognitive impairment in aging (Hypothesis 3).

Moreover, different patterns of variable associations were revealed in older adults with RVD and VaD patients: the more problems solved the higher the total score in the first older adult group. However, the performance of VaD patients was additionally affected by the precision in movements. According to Delis et al. (2001) a close-to-zero ratio in the precision in movements in combination with a low total achievement score indicate one's inability to find correct problem-solving strategies as well as impairments in careful planning ability. VaD patients systematically presented this behavioral pattern. Furthermore, they failed to solve more complex problems, while the administration had to be interrupted whenever high-level planning was required.

The findings of the present study corroborate the results of previous research aimed at examining cognitive changes associated with moderate to severe white matter hyperintensities and less than 5 lacunes (Sudo et al., 2013, 2015). These studies indicated that a series of 'impure' EF tasks that assess complex EF abilities, such as planning, appear to

TABLE 5 | Differences in inhibitory control, task/rule switching, and cognitive planning as indicators for differentiation between Vascular Dementia and vascular aging.

#### Vascular Dementia Vs. Vascular Aging

Color – Word Interference Test


Verbal Fluency Test: inhibition plus switching (3rd condition)


#### Tower test: planning


be able to distinguish patients from healthy adults in the earliest pathological neuroimaging of cerebrovascular disease. Functional neuroimaging studies suggested that complex EF employ areas of an extensive neural network including prefrontal, parietal, medial and superior temporal cortices as well as subcortical structures. Leukoaraiosis as well as the onset of excessive white matter hyperintensities may lead to disconnection of those areas (Sudo et al., 2013, 2015).

With regards to Tower Test, in specific, fMRI studies indicated the involvement of the dorsolateral PFC (dlPFC) and anterior cingulate cortex (ACC) in the TT performance (Moreira et al., 2017). According to Foster et al. (2014) selective pyramidal cell atrophy in the dlPFC might partially explain the lower performance in TT. Aberrant functional connectivity between medial prefrontal cortex and ACC might also underlie the impaired complex EF ability (Zhou et al., 2016).

In conclusion, the present study shows that the D-KEFS – TT is an important tool for differentiating between incipient VaD and vascular aging of community dwelling older adults (Hypothesis 3).

Nevertheless, given that all EF tools used in this study are timed tests, processing speed could also be considered a function that is seriously affected in VaD. This is well-established in the extant literature and the main reason for this impairment is periventricular white matter hyperintensities (Gunning-Dixon and Raz, 2000; Oosterman et al., 2004; Vasquez and Zakzanis, 2015; Moreira et al., 2017; Sudo et al., 2017). Finally, it should be taken into account that the D-KEFS tests which were used in the present study measure cold EF mainly supported by dlPFC. However, it has been claimed that damage, at least in some central locations of the dlPFC, lead to defects in general intelligence (g) and not exclusively in EF (Barbey et al., 2013; Keifer and Tranel, 2013).

#### CONCLUSION

At the theoretical level, the present study showed that deficits in complex or/and combined cold EF are more prominent in incipient VaD patients, compared to basic abilities of cognitive control. Specifically, the findings indicated that cognitive planning and cognitive flexibility are considerably affected by VaD progression even in the very first stages. On the contrary, inhibitory control does not seem to be a very useful indicator for the differential diagnosis of incipient VaD. At the applied level, this means that specific timed tests for the examination of planning and set-shifting should be 'in a prominent position' in the neuropsychological batteries administered for VaD evaluation. In this direction, the present study suggests some very specific indicators that can be used to differentiate incipient VaD from vascular aging experienced by community dwelling older adults (see **Table 5**), based on the findings which derived from a detailed examination of three EF abilities, that is planning, cognitive flexibility, and inhibitory control.

However, this study can be only considered preliminary and replication is deemed necessary to ensure the reliability and validity of the findings.

#### Limitations and Future Research

fnagi-10-00330 October 15, 2018 Time: 19:27 # 15

A number of limitations of the study have to be pointed out. First, the size of the sample was small due to the difficulty in recruiting participants diagnosed with incipient VaD as well as healthy controls of the same age, gender, and educational level. The administration of a broader battery of tools developed to measure specific EF abilities, not only cold but also hot EF, on a larger sample, may demonstrate more reliable and valid findings regarding executive functioning in vascular aging and incipient VaD. Moreover, it was also difficult to recruit a reliable number of patients with a diagnosis of a specific type of VaD. This is an important limitation, as different deficits have been reported on cognitive tests in specific types of VaD. As regards vascular aging, community dwelling older adult participants in the study reported that they have been diagnosed with specific risk factors. Hence, there was no objective examination and other risk factors for the development of vascular disease (e.g., diet, obesity) were not taken into account. The design of the study was cross-sectional and did not allow to 'capture' the trajectory of the change in vascular health during the lifespan. The cooperation of different scientific fields in a longitudinal study would enable a detailed observation of any changes occurring in the human

#### REFERENCES


brain, the way they are reflected on cognitive functions, and the time when the pathology starts. Other dementia groups should also be added to the sample, in order to differentiate VaD from other dementia types at the executive functioning level at an early stage of the disease. Moreover, more ecologically valid instruments or real life situations should be used to measure specific EF in the future.

#### AUTHOR CONTRIBUTIONS

KP designed the study under the supervision of DM, examined all participants, and participated in the statistical processing of the data and the writing of the manuscript. OS and VP participated in the writing of the manuscript. DS corrected the writing style of the manuscript. VC provided all patients diagnosed with VaD by her Clinic as well as all participants with vascular risk factors. GP participated in the statistical processing of the data and provided the main neuropsychological instrument of the study. DM brought the general supervision in all phases of the study which is a part of a broader research project of the same as principal investigator.



neuropsychological and behavioral implications suggested by involvement of the thalamic nucleus and the remote effect on cerebral cortex. the Osaki-Tajiri project. Psychiatry Res. Neuroimaging 213, 56–62. doi: 10.1016/j.pscychresns. 2012.12.004


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Pantsiou, Sfakianaki, Papaliagkas, Savvoulidou, Costa, Papantoniou and Moraitou. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Functional Re-organization of Cortical Networks of Senior Citizens After a 24-Week Traditional Dance Program

Vasiliki I. Zilidou1,2 , Christos A. Frantzidis <sup>1</sup> \*, Evangelia D. Romanopoulou<sup>1</sup> , Evangelos Paraskevopoulos <sup>1</sup> , Styliani Douka<sup>2</sup> and Panagiotis D. Bamidis <sup>1</sup>

<sup>1</sup>Laboratory of Medical Physics, Medical School, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>2</sup>Department of Physical Activity and Recreation, School of Physical Education and Sport Science, Aristotle University of Thessaloniki, Thessaloniki, Greece

Neuroscience is developing rapidly by providing a variety of modern tools for analyzing the functional interactions of the brain and detection of pathological deviations due to neurodegeneration. The present study argues that the induction of neuroplasticity of the mature human brain leads to the prevention of dementia. Promising solution seems to be the dance programs because they combine cognitive and physical activity in a pleasant way. So, we investigated whether the traditional Greek dances can improve the cognitive, physical and functional status of the elderly always aiming at promoting active and healthy aging. Forty-four participants were randomly assigned equally to the training group and an active control group. The duration of the program was 6 months. Also, the participants were evaluated for their physical status and through an electroencephalographic (EEG) examination at rest (eyes-closed condition). The EEG testing was performed 1–14 days before (pre) and after (post) the training. Cortical network analysis was applied by modeling the cortex through a generic anatomical model of 20,000 fixed dipoles. These were grouped into 512 cortical regions of interest (ROIs). High quality, artifact-free data resulting from an elaborate pre-processing pipeline were segmented into multiple, 30 s of continuous epochs. Then, functional connectivity among those ROIs was performed for each epoch through the relative wavelet entropy (RWE). Synchronization matrices were computed and then thresholded in order to provide binary, directed cortical networks of various density ranges. The results showed that the dance training improved optimal network performance as estimated by the small-world property. Further analysis demonstrated that there were also local network changes resulting in better information flow and functional re-organization of the network nodes. These results indicate the application of the dance training as a possible non-pharmacological intervention for promoting mental and physical well-being of senior citizens. Our results were also compared with a combination of computerized cognitive and physical training, which has already been demonstrated to induce neuroplasticity (LLM Care).

Keywords: dancing, electroencephalography, neuroplasticity, dementia, active aging, functional connectivity, neurodegeneration, brain networks

#### Edited by:

Panteleimon Giannakopoulos, Université de Genève, Switzerland

#### Reviewed by:

Soledad Ballesteros, Universidad Nacional de Educación a Distancia (UNED), Spain Maria Velana, Universitätsklinikum Ulm, Germany

> \*Correspondence: Christos A. Frantzidis christos.frantzidis@gmail.com

Received: 03 September 2018 Accepted: 04 December 2018 Published: 21 December 2018

#### Citation:

Zilidou VI, Frantzidis CA, Romanopoulou ED, Paraskevopoulos E, Douka S and Bamidis PD (2018) Functional Re-organization of Cortical Networks of Senior Citizens After a 24-Week Traditional Dance Program. Front. Aging Neurosci. 10:422. doi: 10.3389/fnagi.2018.00422

#### Zilidou et al. Dance Cortical Neuroplasticity

### INTRODUCTION

As the number of elderly people worldwide is constantly increasing, important changes regarding the mental, physical and emotional condition occurred. These changes refer to the brain structure, as well as to various functions related to aging. The alterations of the structure and function of the brain associated closely with changes in cognitive function. The most important cognitive functions affected by aging are considered to be memory, attention and movement that may lead to dementia or even Alzheimer's disease.

Different parts of the brain, frontal cortex and hippocampus, were mostly affected by aging leading to the reduction of blood flow in the vessels by causing changes in amyloid plaques. High level of education, healthy lifestyle and regular exercise seem to be essential factors contributing on keeping older adults healthy, preventing diseases and reducing associated complications (Erickson et al., 2011). In recent decades, studies show that regular physical activity and exercise could mediate the negative effects of aging on both body and mind (Billis et al., 2010). The growing proportion of elderly people demonstrates the need of maintaining the independent living, the social integration and the active life for a longer period.

Neuroscience research has provided insights into the brain functions and behaviors with the main focus on the Mild Cognitive Impairment (MCI) transition by contributing to the development of neurodegenerative diseases. Therefore, a plethora of non-pharmacological interventions have been developed. During the last decades there is a steady interest on the investigation of the aging negative effects that can be eliminated through the physical and/or cognitive training. These training procedures aim to induce the neuroplasticity of the mature human brain (Bamidis et al., 2014). The brain is adapted to the new environment (brain and cognitive alterations) by enhancing existing or creating new connections. It is reported that aerobic exercise induces neurogenesis of hippocampus (Erickson et al., 2011) and physical activity is associated with the reduction of risk factors that may accelerate neurodegeneration (Cotman et al., 2007). Research findings show that physical activity is a promising non-pharmaceutical intervention to prevent cognitive impairment associated with age and neurodegenerative diseases (Bherer et al., 2013). In another study, the computerized physical training presented positive effects in the physical and cognitive capacity as well as the participants' quality of life (Zilidou et al., 2016).

Long lasting memories (LLM) is developed as an integrated platform of existing technological achievements in the field of computer-aided physical/cognitive training and ambient assisted living (AAL) solutions for elderly people and vulnerable groups (Romanopoulou et al., 2015). LLM offers an opportunity to improve cognitive and physical condition, being able to continue feeling an active part of society. The core of the LLM service is an integrated ICT platform which combines state-of-the-art cognitive exercises with physical activity in the framework of an advanced AAL environment (Konstantinidis et al., 2010). LLM proved the capacities for improving the mental and physical condition of the elderly that adhered to it. The results from 322 individuals participated in the project shown an improvement in the working and episodic memory, as well as in the executive functions (Bamidis et al., 2015). Moreover, 70 elderlies with MCI were subjected to electroencephalographic (EEG) recording in a resting state to investigate whether or not the combination of cognitive and physical activity, induces changes in brain cortical activity. The results revealed a significant reduction in the delta, theta and beta brain rhythms, and a decrease in the activity of certain brain regions (PCU/PCC) associated with functional plasticity. They also discovered that the greater the decrease in the delta and the theta rate, the higher the neuropsychological assessment through Mini Mental State Examination (MMSE; Styliadis et al., 2015). Moreover, graph analysis in brain networks demonstrated the ability of the brain to be ''educated'' in order to reorganize the functional connectivity to prevent the effects of neurodegenerative phenomena and normal aging (Klados et al., 2016). An additional study was conducted to investigate whether the combination of physical and mental exercise may lead to the enhancement of the brain functional reorganization in the beta band which is considered to be particularly important for mental functioning. Also, an increase in the functional connection of two brain hemispheres of occipital, temporal, parietal and pre-frontal lobe has been occurred (Frantzidis et al., 2014a).

The aforementioned approach is based on a combination of computerized cognitive and physical training, which is performed intensively (more than four times) for a period of approximately 2 months. Both cognitive and physical training are widely used to induce neuroplasticity on the mature human brain and to delay or even prevent neurodegeneration (Heyn et al., 2004; Voss et al., 2010; Frantzidis et al., 2014b; Bamidis et al., 2015). However, there is an active debate on the optimal intervention design and its effect size and sustainability of the positive effects observed. It seems that improvement is greater when the intervention is performed in day care centers under supervision and not at home installations (Lampit et al., 2014). Similarly, multi-domain physical intervention demonstrated improvement or at least preservation of an adequate cognitive functioning in senior citizens (Ngandu et al., 2015). Combination of cognitive and physical training was proven to induce a more pronounced neuroplasticity effect (Frantzidis et al., 2014b; Bamidis et al., 2015). Although, the aforementioned studies demonstrated the applicability of cognitive and physical interventions as well as the robustness of their combination, the research community recently investigates new training approaches which would be more ecologically valid and applicable on non-clinical settings in order to further enhance the adherence rate and to improve drop-outs. Given also the fact that beneficial effects of the training are evident for a given time period after the intervention's cessation, novel approaches should be adopted as a life-style habit and be performed on a regular basis.

An ideal, non-pharmacological intervention that fulfils the aforementioned criteria is dance. A previous study showed 7% lower risk for developing dementia for elderly people who involved in dancing, playing a musical instrument, reading and playing board games once a week (Verghese et al., 2003). Previous research studies have investigated the potential benefits of dancing interventions to senior citizens. Dance is a less-threatening exercise for many elderlies, since they had many positive and enjoyable experiences in their youth. Also, dancing is an important factor for health and well-being aging. Dance offers the opportunity to maintain a connection with everyday life as it encourages fun and enjoyment, social interaction, team spirit, health, physical activity and mobility. Dance, is considered to be a form of physical exercise that may be adopted as part of many interventions for the elderly instead of other (Wikström, 2004; Lima and Vieira, 2007). Dance improves energy, increases mood and reduces stress or anxiety in ways similar to aerobic exercise (Kim and Kim, 2007). Also, studies have reported that the traditional dances as a physical activity prevent cognitive decline and improve the coordination and control of body movements (Balkus, 1989).

Dance training has been associated with positive behavioral effects in various populations. However, only few studies have evaluated the functional brain correlations in dance interventions as they have not investigated the structural correlations of brain changes due to the dance exercise. It is necessary to perform additional research on brain related networks and correlative behaviors of dance interventions, so as to better understand and validate the dance-related neuroplasticity mechanisms. Thus, there is a huge need to further study the neuroscience of dance in order to provide an increasing multidisciplinary area that may offer information on the interactions between the brain and arts (Karpati et al., 2015).

Thus, traditional dancing may induce many feelings as it is achieved in groups of 15–20 participants who simultaneously perform dancing movements which require balance, flexibility, rhythm, orientation and cognitive functions. More even it does not require any special equipment and can be applied easily by Day Centers with low cost.

Despite the promising results, the majority of studies that are employed dancing training have not been evaluated under medical and neuroscientific criteria. Therefore, we studied an experimental group who performed traditional dancing for 6 months. This group was compared with a control group which performed a cognitive stimulation program. Both groups were assessed in terms of a detailed neurophysiological assessments and neurophysiological (EEG) examination which aimed to assess the brain functioning of the participants before and after the training. This examination evaluates the brain functions in terms of each functional organization through a system level approach employed analysis of brain networks based on graph theory.

This piece of work aims to investigate whether traditional dancing could improve cognitive and brain functioning of elderly compared with a cognitive stimulation program (Active Control Group). We also attempt to answer the following research questions:

• Is traditional dancing intervention an effective approach for promoting active and healthy aging?


According to previous studies we hypothesize that traditional dancing may be more effective than active stimulation. It would also induce functional re-organization of the brain networks that would shift brain function to a more optimal mood. The latter is hypothesized to be achieved by more efficiency and faster communication among distant brain regions.

### MATERIALS AND METHODS

#### Participants

The study initially employed 54, non-demented, senior citizens, who were randomly assigned either to an active control (mean age 66 ± 5.51 years) or to a traditional dance training group (mean age 68.73 ± 4.73 years). However, five participants from the training-group refused to repeat the neurophysiological intervention at the post level and they excluded from the study. There were also five participants who did not complete the active control program and also excluded. So, the study resulted in 44 senior citizens. Both groups consisted of equal number of participants. Neuropsychological, physical status evaluation and neurophysiological by an electroencephalogram (EEG) were performed. The evaluation took place 1–14 days before/after the training initiation/finalization respectively. Prior to their enrolment in the study activities all participants were informed about the study hypotheses and aims as well as the underlying activities. Inclusion criteria were the following: (1) age to be greater than 60 years; (2) participants should be sedentary or mild sedentary with an upper threshold of 150 min of moderate to vigorous exercise per week; (3) lack of participation in any other physical training intervention within the last year; and (4) performance of medical examination assessing their safe participation in the study. People with heart failure, hypertension and respiratory insufficiency and patients suffering from dementia were excluded from the program as they who participated in other Greek Traditional Dances programs. Also, participants who did not complete at least 80% of the total of hours of program were excluded from further analysis. They had the opportunity to discuss any issues raised and then they signed a written informed consent form. The study was based on specific guidelines and regulations, which were approved by the ethics committee of the Aristotle University of Thessaloniki (AUTH) according to international principles. Group demographics and information about the participants' age, education, body mass index (BMI) and cognitive status, is depicted in **Table 1**. The groups did not significantly differ in baseline cognitive level, mean age, education and gender (all ps > 0.05).

#### Intervention

The active group participated in cognitive intervention with the use of Video GRade software that was developed by the Lab of Medical Physics at the AUTH. Video GRade is an inspired environment for cognitive training<sup>1</sup> , which consists of a set of

<sup>1</sup>www.llmcare.gr



50-min educational video from YouTube that includes content with history, art, culture, dance and music. At the end of each video, participants were asked to answer eight multiple choice questions. Participants performed 24 sessions during a period of 8–10 weeks.

The intervention group was conducted through traditional dances from all over Greece that were divided into three categories according to the number of steps and complexity, the intensity of the pace (slow-fast), as well as the upper limbs position and movement. A 24-week period program was performed for 60 min two times per week, taking place at Day Care Centers of Municipality of Thessaloniki located in Greece. The intervention duration was much longer than the active (Video Grade) control group, due to the annual organizational structure of the local Day Care Centers. So, the maximum number of available sessions was 48. However, the long duration of the intervention overlapped with Christmas and Easter holidays. During that period, many senior citizens preferred to stay at home and did not attend 1–4 (maximum two successive weeks) sessions per occasion. Considering the abovementioned pragmatic limitations and aiming to have a similar number of sessions between the control and the training group we kept the mean number of 28 sessions within the 24-week period.

### Neuropsychological and Physical Assessments

Cognitive functioning was estimated through the MMSE (Folstein et al., 1975), Trail Making B Test (Butler et al., 1991) and the Geriatric Depression Scale (GDS short version; Yesavage et al., 1982–1983).

Their physical status and functional ability was evaluated through the following tests: (i) Fullerton Senior Fitness Test which contains six domains: chair Stand (assessed lower body strength), 8 Foot Up and Go (assessed complex coordination, agility and dynamic balance), Back Scratch (assessed flexibility of upper body), Chair Sit and Reach (assessed flexibility of the lower back and hamstring muscles), Arm Curl (assessed upper body strength) and 2 Min Step (assessed aerobic capacity; Jones and Rikli, 2002); (ii) Berg Balance Scale which assessed balance and risk of falls (Berg et al., 2009); (iii) Tinetti Test which assessed walking and risk of falls (Tinetti, 1986); and (iv) Stork Balance Stand Test which assessed balance when standing on one leg (Johnson and Nelson, 1969). Body composition was also calculated.

### EEG Analysis Data Acquisition

The EEG data acquisition was performed through a Nihon Kohden JE-207A device and 57 active electrodes attached on a cap (EASYCAP) fitted directly to the participant's head. Two more electrodes were placed on positions behind the participant's ears and a third active electrode, placed on a left anterior position, served as ground electrode. The electrode impedances were kept lower than 5 KΩs. The sampling frequency was set at 500 Hz. The participants sat on a comfortable, armed chair, which was located in a quiet room with minimal, ambient light. They were instructed to close their eyes for 5 min and avoid any movement, especially of their upper part. The subsequent processing and analysis steps are visualized in **Figure 1** and are the following: (i) electroencephalographic (EEG) data acquisition; (ii) artifact removal and sensor data epoching; (iii) generic anatomy head and cortex modeling; (iv) estimation of the cortical activity for anatomically relevant regions of interest (ROIs); (v) quantification of the co-operative degree (functional connectivity analysis) among ROIs cortical activity through the notion of Relative Wavelet Entropy (RWE); and (vi) formation of functional brain networks regarding the ROIs as network nodes and their connectivity values as edges. The analysis is performed through contemporary mathematical tools derived from graph theory and involves both global and local network properties.

#### Pre-processing

The 57 active electrodes recording brain signals were re-referenced through the common average model. Then, digital filtering through 3rd order Butterworth filters was applied as follows: (a) high-pass filter with cut-off frequency at 1 Hz; (b) band-stop (notch) filter with cut-off frequency at the 47–53 range; (c) low-pass filter with cut-off frequency at 100 Hz; (d) band-stop (notch) filter with cut-off frequency at the 97–103 Hz range; and (e) band-stop (notch) filter with cut-off frequency at the 147–153 Hz range. Independent Component Analysis was then applied to identify and reject potentially contaminated source components due to eye blinks, muscle artifacts, bad electrode placement, high frequency noise, ECG modulation, linear trends. Finally, the data were epoched. The epoch duration was 1,024 sample points (2.048 s). Epochs still containing artifacts were visually rejected. The pre-processing was performed through built-in code based on the Matlab Signal Processing Toolbox and EEGLAB software (Delorme and Makeig, 2004).

#### Cortical Activity Estimation

The Brainstorm software package was used for the computation of the inverse problem, which results in the reconstruction of the cortical activity (Tadel et al., 2011). Default anatomy modeling was employed in terms of magnetic resonance imaging (MRI) volume (Colin 27 stereotaxic T1-weighted MRI volume). The head modeling was computed through the Open

MEEG Boundary Elements Method (BEM head model). The solution space was constrained to the cerebral cortex which was modeled as a 3-dimensional grid of 20,000 fixed dipoles oriented normally to the cortical surface. Then, the inverse solution was estimated by means of the sLORETA methodology (Pascual-Marqui, 2002). The study analysis, employed the entire cortex analysis through the definition of 512 cortical regions of interest (ROIs), which were defined according to the deterministic process implemented by the Brainstorm software. Since, the source analysis requires a symmetric electrode map of 64 electrodes, seven additional electrode activations were inserted through first-neighborhood activity interpolation.

#### Cortical Synchronization Analysis

The synchronization analysis aims to quantify the co-operative degree among the ROIs by comparing their cortical time series activity. Therefore, it was performed for each participant on the average scout cortical activity. The Orthogonal Discrete Wavelet Transform (ODWT) was applied and the analysis was based on the relative wavelet entropy (RWE) metric. Bi-orthogonal wavelets of 5th order were selected as the family wavelet class (Frantzidis et al., 2010, 2014b; Chriskos et al., 2018). Optimal time-frequency analysis is obtained since the mother wavelet is subjected to scaling and translation. The ODWT analysis framework is based on an iterative decomposition scheme, which employs recursive, low-pass filters for estimating the activity of the five frequency bands (delta, theta, alpha, beta and gamma) through the estimation of the wavelet coefficients. The decomposition scheme involves j = 1. . .5 layers. The amplitude of the wavelet coefficients quantifies the wavelet's correlation with the actual rhythmic activity, while the coefficient's sign is an index of either positive or negative correlation. The ODWT framework involves a periodization mode. So, the final decomposition layer (j = 5) contains one wavelet coefficient and corresponds to the lowest rhythms (delta and theta). Each other step contains the double number of the previous one. So, there are two coefficients for the alpha, four for beta and eight for the gamma rhythm. In this way, non-redundancy and optimal resolution is facilitated through a parameter-free methodology. The energy of each frequency band is estimated by squaring the amplitude of the corresponding wavelet coefficients. Then, the overall energy is estimated as following:

$$E\_{\text{tot}} = \sum\_{j<0} \sum\_{k} \left| \mathbf{C}\_{j}(k) \right|^{2} \tag{1}$$

In this way, the relative energy distribution of each frequency band is obtained and the synchronization degree among two ROIs was quantified as the similarity of their energy distribution (p<sup>j</sup> and q<sup>j</sup> , j = 1. . .5) given by the following formula of the RWE:

$$RWE = \sum\_{j=1}^{5} p\_j \times \ln\left(\frac{p\_j}{q\_j}\right) \tag{2}$$

The RWE metric may be then used as a dissimilarity matrix (the larger the value the greater the desynchronization), which may be used as an adjacency matrix for network analysis through graph theory.

#### Cortical Brain Network Analysis **Functional Connectome Estimation**

For each participant and for each time condition (pre/postintervention) all the available data epochs were used. For each epoch the resulting RWE matrix was used to create a functional brain network, which is defined as a graph G = (V, E). The graph contains a set V of nodes (here V = 512 cortical ROIs) and a number E of edges quantifying the co-operative degree as estimated through the RWE analysis. The RWE matrix is used as it is, but the auto-synchronization information of the main diagonal is discarded from further analysis. Therefore, the resulting adjacency matrix forms a weighted and directed network. Subsequent analysis was performed on binary networks. The thresholding was dynamically configured for each network instance in order to result in fixed density ranges of 10,000, 12,500 and 15,000 edges. The density range was similar as in previous studies (Frantzidis et al., 2014a).

#### **Global Network Characteristics**

Global network analysis was applied by computation of both the cluster coefficient for each node and the characteristic path length for each pair of ROIs. Then, their mean values were used for the estimation of the small-world property. While, the interested reader may find detailed information on these fundamental network properties in classical textbooks and studies, in brief the cluster coefficient is an index of local information processing, while the characteristic path length quantifies the integration of information and the network's flow. More specifically, the cluster coefficient (C) for a given node is the number of the existing connections of the node's direct neighborhood to the total number of the possible ones. The set of the node's direct neighborhood contains the ROIs with which the node under consideration is connected. So, C is actually a possibility of the local information capacity for each node. The characteristic path length (L) is applied on node pairs and quantifies the shortest distance among those nodes. The mean values of these two characteristics are computed for both the actual brain network and for a number (N = 100) of random networks with the same characteristics (nodes and density). Then, the small world property (sigma) is computed as follows:

$$
gamma = \frac{\frac{L}{L\_{\text{rand}}}}{\frac{C}{C\_{\text{rand}}}} \tag{3}
$$

In the case that the mean cluster coefficient of the actual brain network is much larger than the mean value of the random ones and simultaneously the characteristic path length of the actual and the random networks are quite similar, we have sigma >1. The greater the sigma the more prominent the smallworld property is. According to these, small-world is an index of non-random clustering and near-random path length. So, it is a property of the optimal network configuration since it implies strong local information processing which may be transferred from every individual node to every other, regardless of their physical distance, through a small set of intermediaries (Watts and Strogatz, 1998).

#### **Brain Hub Detection**

Functional hubs of the cortical connectome were identified based on the betweenness centrality (BC) metric (Bi). It is an index of the information amount transferred through a given node, since it is computed as the number of the network's shortest paths that pass through the specific node (Rubinov and Sporns, 2010). Hub detection was performed based on a criterion of Bi ≥1.5, which was firstly introduced by Seo et al. (2013) and previously applied by Frantzidis et al. (2014b).

#### **Functional Cartography and Node Roles**

Aiming to ensure that putative effects were not biased by the particular hub definition, we also followed an alternative approach of applying a previously proposed functional cartography technique (Guimerà and Nunes Amaral, 2005). This framework involves the computation of the community structure (modularity analysis) and then the estimation of the within-module z-score and the participation coefficient (PC). Based on these two metrics a specific role is assigned to each network node. The network's community structure was performed based on the Brain Connectivity Toolbox (BCT; Rubinov and Sporns, 2010). Then, the within-module z-score was used to estimate the connection strength of a given node with the other nodes of its own module. The greater the z-score is, the larger the connection strength. The PC (P) is a probability value, which quantifies the connection distribution of a given node across all the feasible modules of the specific network. The closer the coefficient's value to 1, the more uniform the distribution across modules is. The role assignment to each node is defined based on the node's position to the z-P parameter space. Heuristically, in case a node has within-module z-score greater or equal to 1.5 then it is regarded as a hub and otherwise as non-hub node. Both node categories

proper sample size was 48 (24 participants per group), whereas we have included 44 senior citizens (22 participant per group).

are then further characterized according to their PC value. More specifically non-hubs are further characterized as ultraperipheral (p ≤ 0.05), peripheral (0.05 < p ≤ 0.45), connector (0.45 < p ≤ 0.70) and kinless (p > 0.70). Hub nodes are assigned as provincial (p ≤ 0.25), connector (0.25 < p ≤ 0.50) and kinless (p > 0.50).

### RESULTS

#### Analysis of Global Network Characteristics

The statistical analysis was conducted via a 2 × 2 mixed model ANOVA. The groups of intervention (Dance and Active) served as between-subjects factor whereas the time (pre-post) as withinsubject factor and the three inter-related dependent variables (small-world value, cluster coefficient and characteristic path value).

Aiming to investigate whether the sample size was adequate for the statistical analysis we performed power analysis. It was conducted in G-Power 3.1 to determine a sufficient sample size using an alpha of 0.05, a power of 0.80, and a medium effect size (f = 0.21; Faul et al., 2013). Based on the aforementioned assumptions, a total sample size of 48 people would be sufficient to detect significant interaction effects. Specifically, the appropriate number for each group of intervention (Dance and Active) is 24, in order for group differences to reach statistical significance at the 0.05 level. Our sample size is estimated 44, 22 per group which is adequate for our experimental design and the analysis performed. The results obtained from the statistical analysis are visualized in **Figure 2**.

Results revealed an interaction between time and intervention for the small world property for 10,000 edges, F(1,42) = 20.56;

TABLE 2 | 2 × 2 mixed model ANOVA statistical analysis.


The groups of intervention (Dance and Active) served as between-subjects factor whereas the time (pre-post) as within-subject factor and the three inter-related dependent variables were the small-world value, cluster coefficient and characteristic path length. The p-values that reached statistical significance (p < 0.05) are displayed in bold.

p < 0.05, as well as the small world property for 12,500 edges F(1,42) = 19.53; p < 0.05. Furthermore, an interaction between time and intervention was also revealed for the characteristic path for 15,000, F(1,42) = 5.16; p < 0.05 as you can see in **Table 2**.

### Correlation With Cognitive and Physical Status

Pearson's correlations were computed, in order to test if there was a linear relationship between cognitive status as measured with the MMSE assessment and physical status as measured with the Fullerton Senior Fitness Test, as well as the Small-World property value. The analysis revealed a significant positive correlation between the Back-Scratch Test score and the Small World property at 12,500 edges, r = 0.301, n = 43, p = 0.05 and between the Chair Sit and Reach Test score and the Small World property at 15,000 edges, r = 0.385, n = 43, p = 0.01. Furthermore, a significant negative correlation was also revealed between the Back-Scratch Test score and the PC of the fronto-parietal network (FPN) score, r = −0.32, n = 43, p = 0.039.



Bold values denote statistical significance.

TABLE 4 | Statistically significant changes regarding local network (betweenness centrality, participation coefficient, within-module z-score) at pre-post level for both groups.


The p-values that reached statistical significance (p < 0.05) are displayed in bold.

#### Analysis of Physical Status Assessments

A paired-samples t-test was conducted to investigate the effectiveness of Greek traditional dancing on dependent measures of functional connectivity. **Table 3** presents the results (post–pre-intervention differences and p-value) concerning the physical assessment tests. More specifically, the results indicate that the intervention group evoked statistically significant improvement in the Chair Stand, Chair Sit and Reach and 8-Foot-Up-and-Go. Active Control group showed statistically significant improvement in Chair Stand test.

#### Local Network Analysis

From the 512 ROIs we further analyzed the 22 being part of the default mode network (DMN), FPN, and executive network (EN) according to Voss et al. (2010). For these three networks as well as for their combination we estimated the following local graph properties: PC, BC and within module z-score (ZM). We selected these metrics in order to define node hierarchy according to Guimerà and Nunes Amaral (2005) and Seo et al. (2013). The statistical analysis was conducted via a 2 × 2 mixed model ANOVA. The groups of intervention (Dance and Active) served as between-subjects factor whereas the time (pre-post) as within-subject factor and the three inter-related dependent variables (BC, PC and within Module z-score value). **Table 4** presents the results that revealed an interaction between time and intervention for the EN for the BC F(1,42) = 23.246 p < 0.05 and Within module z-score F(1,42) = 37.262 p < 0.05. Furthermore, an interaction between time and intervention was also revealed for FPN for BC F(1,42) = 9.291 p < 0.05, PC F(1,42) = 25.695 p < 0.05 and Within module z-score F(1,42) = 22.284 p < 0.05. Finally, PC was also statistically significant feature for DMN F(1,42) = 7.740 p < 0.05 and contribution of all F(1,42) = 11.069; p < 0.05.

#### Hub Identification

As described within the ''Methodology'' section (Brain Hub Detection), hubs were identified according to their relative BC value (Bi). The B<sup>i</sup> value for both groups and both conditions as well as their post-pre/intervention difference are visualized for each cortical node in the **Figure 3**.

#### Functional Cartography Analysis

The **Figure 4** describes the two-dimensional distribution in terms or PC and within-module z-score for each cortical node, for both groups and conditions.

Pearson analysis revealed statistically significant correlations among the node roles (both positive and negative) as visualized in the following **Figure 5**.

### DISCUSSION

This study investigated dancing induced neuroplasticity effects of the mature human brain. We employed an objective evaluation framework based on contemporary mathematical tools such as cortical functional connectivity and graph theory. Network neuroscience has been applied until now to quantify physiological aging mechanisms (Micheloyannis et al., 2009) and neurodegeneration phenomena (Sanz-Arigita et al., 2010; Tijms et al., 2013; Frantzidis et al., 2014a). However, there are only a few preliminary attempts on how cognitive and physical training

improved brain network functioning (Burdette et al., 2010; Voss et al., 2010; Klados et al., 2016). The pioneering work of Voss et al. (2010) employed functional MRI (fMRI) and showed that aerobic training increases the functional connectivity of the Default Mode and the Fronto-Executive resting state network. The authors found that this type of intervention may increase restoration mechanisms in fronto-temporal, brain networks affected by physiological aging. However, they did not investigate how aerobics improve the functional organization of these networks as an entire system. A similar type of training was also employed by Burdette et al. (2010) and validated that physical exercise increases hippocampal connectivity in comparison to a control group. The authors employed graph theory tools to evaluate these changes and showed that physical training enhanced the hippocampus degree and its connectivity with the anterior cingulate cortex. These two regions were part

of the same sub-network (module) for the training group. A more recent work by Klados et al. (2016) examined whether a combination of computerized cognitive and physical training would induce changes in a cortical functional network. They evaluated these changes in terms of alterations in network's density and node's strength. They focused on beta band and on cortical estimations of 305 regions defined by sLORETA software.

Despite the impact of these three studies, they introduced only a partial estimation of the network properties associated with neuroplasticity. Due to methodological limitations attributed either to the low temporal resolution of fMRI (Burdette et al., 2010; Voss et al., 2010) or to the cortical estimation employed by Klados et al. (2016), the previous studies failed to track a global perspective of how training changed the cortical functional organization of the mature human brain. Moreover, the local analysis they performed was minimal since it focused only on modularity analysis of central brain regions (Burdette et al., 2010) or node strength in beta band (Klados et al., 2016). Although these analyses were significant since they shed light at the prominent role of hippocampus and anterior cingulate cortex (Burdette et al., 2010) and the hub strength of cortical regions (Klados et al., 2016) they did not provide answers whether physical intervention increases local information processing or information flow resulting thus in better functional organization as estimated by small-world property (Watts and Strogatz, 1998; Frantzidis et al., 2014a). Moreover, the hub definition through node strength as defined in Klados et al. (2016) is mainly a connectivity-driven metric that may also incorporate compensatory mechanisms and did not regard the node's contribution to the global information flow. Therefore, we employed the BC metric as defined in Seo et al. (2013) and Frantzidis et al. (2014a). Finally, we proposed a more detailed network analysis through the notion of functional cartography as defined in the article of Guimerà and Nunes Amaral (2005). We used that type of analysis to assign to each cortical node a specific role, but we also investigated correlations (either positive or negative) among these nodes and how intervention changes the node distribution to these specific roles. The network analysis was applied to a detailed cortical parcellation (512 nodes) but also in known resting state networks (DMN, FPN and EN) as defined in Voss et al. (2010). So, we believe that the main novelty of our study is the proposal of a holistic intervention evaluation framework based on network neuroscience and covering most of the aspects of graph theory.

Our results regarding global network characteristics demonstrated significant time by intervention interactions for both small-world metric (in density ranges of 10,000 and 12,500 edges) and for characteristic path length regarding 15,000 edges. This was because the participants in dance training showed a significant increase of small-world property at two density ranges mainly attributed to a decrease of characteristic path length which was significant only for 15,000 edges. Although there are not previous global network results regarding intervention evaluation, there is a plethora of studies reporting disrupted functional organization of brain networks due to neurodegeneration. More specifically, the pioneering work of Stam et al. (2008) showed that both cluster coefficient and characteristic path length obtained through Phase Lag Index (PLI)-weighted connected networks were decreased for AD patients. Similarly, sensor-based EEG analysis showed decreased small-world property in the AD group within the beta band (de Haan et al., 2009). The first fMRI-based study performed by Supekar et al. (2008) showed loss of small-world property in AD patients due to decrease in local information processing (cluster coefficient). Another study through fMRI found that AD patients showed increased cluster coefficient and characteristic path length in comparison with a healthy control group (Zhao et al., 2012). Finally, a recent study employing healthy elderly and patients suffering either from amnestic MCI (aMCI) or from mild dementia (MD) reported functional disorganization of small-world brain networks for both pathological groups (Frantzidis et al., 2014a). This decline was mainly attributed to the loss of local information processing capacity (mean cluster coefficient). The aforementioned studies indicate that functional brain networks are vulnerable to neurodegeneration phenomena, which affect brain networks through a multi-level way. The proposed dancing intervention seems to improve functional network performance as estimated through small-world property by reducing the average length of characteristic paths. This implies minimization of functional disconnection patterns and better transmission of the information across network nodes. This finding is of particular importance since Alzheimer's dementia is regarded as a disconnection syndrome resulting thus in longer characteristic path lengths (Stam et al., 2007; Frantzidis et al., 2014a). The specific hypothesis is supported by previous studies which linked longer path lengths and preserved cluster coefficient with a loss of complexity and a less optimal organization (Stam et al., 2007). Among the first findings of the network neuroscience were that increases of characteristic path length induce functional degradation of brain network and reduced interaction among brain nodes (Sporns and Zwi, 2004) which was also correlated with cognitive decline (Stam et al., 2007). So, the present study results regarding global network analysis (increase of small world property due to preservation of cluster coefficient and reduction of characteristic path length) may be attributed to neuroplasticity findings due to the dance intervention which seems robust for promoting active and healthy aging.

We also investigated whether the improvement of physical activity may be correlated with neuroplasticity effects. The study demonstrated positive correlations among the smallworld property and the Back Scratch (r = 0.301, p = 0.05) and Chair Sit and Reach (r = 0.385, p = 0.01) domains of the Fullerton Senior Fitness test. These two tests quantified the flexibility of the upper and the lower body respectively. Flexibility is regarded as an index of functional status of the elderly and could be benefited from physical training (Heyn et al., 2004). Dance training induces greater improvement on trunk flexibility and hip-joint mobility (Sofianidis et al., 2009). According to that study, traditional Greek dancing seems to be an ideal candidate for improving flexibility since it involves activities of single-limb standing, while the dancer's body should be flexed/extended, and many steps require forward/backward leaning (Sofianidis et al., 2009). However, previous studies have not investigated how such an intervention affects the functional organization of brain networks. Positive correlations imply that flexibility improvement may be induced by a more efficient function of the mature human brain as quantified by the small-world property. This finding is novel since it identified for first time a positive relationship among flexibility increase and global network performance.

Although, positive correlation of physical activity with global network characteristics is a novel finding which has not been earlier demonstrated to the best of our knowledge, it does not provide information about the mechanism that traditional Greek dance induces neuroplasticity. Aiming to further investigate this issue we performed local network analysis by investigating functional cartography (Guimerà and Nunes Amaral, 2005) and hub identification (Seo et al., 2013; Frantzidis et al., 2014a). This analysis was performed in order to define the functional node hierarchy both on the entire set of nodes and on specific resting-state (Default Mode, Fronto-Parietal and Executive) networks according to (Voss et al., 2010). We computed the BC as a generic hub index, PC as an index of how well a node is connected with nodes belonging to other modules and within-module z-score which quantifies the connection strength of a given node with the other nodes of its own module. The previous finding of positively correlating flexibility improvement with better functional organization (small-world increase) was further validated by the negative correlation observed among the Back Scratch and the PC for the nodes of the FPN (r = −0.32, p = 0.039). FPN seems to be greatly affected by the proposed intervention since there were also statistically significant changes in BC, PC and within module z-score for that network. FPN functioning was linked with goal-directed attention (Uddin, 2015), motor planning and execution (Hsu et al., 2017). It was also shown that fitness gains were associated with greater activation of the FPN and better executive functioning (Colcombe et al., 2004). Combining the previous neuroimaging findings with our results regarding the re-organization of the fronto-parietal and executive resting state network, we may conclude that traditional Greek dance is a potentially robust approach for improving the flexibility of senior citizens. This was achieved by inducing neuroplasticity effects on the modular architecture of resting-state networks known to be affected by increased physical activity (Colcombe et al., 2004; Voss et al., 2010). These networks were also previously demonstrated to govern complex motor planning and executive functioning resulting thus in improved flexibility which is essential during dancing. Neuroplasticity was also evident and for the DMN, as denoted by PC changes, but its effect was more attenuated in comparison to the aforementioned networks.

The aim of the present study was to examine whether a program of traditional Greek dance could induce neuroplasticity effects on the mature human brain. Dancing involves learning of motor skills through physical practice and observation (Kirsch et al., 2018), while it combines cognitive and physical training (Rehfeld et al., 2018). Although it involves a moderate level of physical activity, poses several sensorimotor and cognitive challenges. It also facilitates social interaction enriched with emotional stimulation (Kattenstroth et al., 2013). So, dancing is regarded as a promising intervention for promoting neuroplasticity, healthy and active aging (Merom et al., 2016). The pioneering work of Kattenstroth et al. (2013) demonstrated positive effects on posture, reaction time, cognition, motor performance and quality of life, when compared with a control group. Despite their novelty, the aforementioned studies did not investigated changes in terms of functional or structural neuroimaging, even though dance often involves the activation of premotor, parietal and occipito-temporal cortices (Vogt and Thomaschke, 2007). However, these brain regions are known to be affected by aging mechanisms, resulting thus in poor performance during complex motor sequences like those required for dancing (Kirsch et al., 2018). Our hypothesis was that an intervention based on traditional Greek dance would challenge mental, cognitive and physical abilities of senior citizens. This would be achieved through music stimuli stemming from their youth and requiring execution of complex motor sequences. During the choreography performance, dancers need to co-ordinate their body posture and movement according to the time-varying music characteristics. This would be achieved by improving their balance, flexibility and strength through the functional re-organization of the mature human cortex. Neuroplasticity effects were evident at a global network level denoted by faster information flow among distant brain regions. This may imply the optimized performance of specific neural networks governing elaborate cognitive processing, balance and entire body co-ordination (Voss et al., 2010). Our results validated previous studies that compared a challenging 6-month dance training program with a conventional fitness intervention of equal intensity (Rehfeld et al., 2018). Although both interventions increased physical fitness, dancing was associated with greater brain volume increase and with an increase in plasma BDNF level (Rehfeld et al., 2018). Longer duration of dance intervention was associated with greater activation of the left hippocampal volume and with improved balance performance (Rehfeld et al., 2017). These studies employed advanced neuroimaging techniques for localizing either structural (Rehfeld et al., 2017, 2018) or functional (Kirsch et al., 2018) changes associated with dancing. Our approach aims to provide an evaluation framework, quantifying dynamical interactions among cortical regions based on contemporary mathematical tools derived from graph theory. So, we focused not only on how the dance training changed the macroscopic functional organization of the human brain, but we also investigated local changes on the default-mode, fronto-parietal and executive resting state networks.

Although our results and their interpretations shed light into the neuroplasticity induced by dancing interventions, they should be considered in the context of several limitations. Firstly, the analysis employed an EEG device with 57 active electrodes. Similar EEG recording settings have been used successfully in previous studies (Styliadis et al., 2015; Klados et al., 2016), nonetheless one should take into account in the interpretation of the results the low spatial resolution of such a setting. High density EEG systems with 128 channels may identify with greater accuracy the cortical activations obtained as outcome measures from similar intervention procedures. Moreover, our brain network analysis was based on a generic head anatomy modeling which could not deal with inter-subject variability and therefore may be vulnerable to errors due to atrophy, life style factors, tissue conductivity or injuries. Given that the local network analysis was based on ROIs defined by a previous neuroimaging study (Voss et al., 2010), estimation of cortical activity for these ROIs may be prune to methodological limitations and results should be interpreted with caution. However, our results identifying a major re-organization of fronto-parietal and ENs are in line with a plethora of previous neuroimaging studies (Colcombe et al., 2004; Voss et al., 2010; Hsu et al., 2017; Kirsch et al., 2018). Future studies may also avoid using thresholding of the connectivity values and employ weighted networks. An important aspect of future research will be the validation of our results with resting state data derived from other recording modalities such as magnetoencephalography (MEG) or fMRI. Another possible limitation is the relatively small sample size of participants we employed. Although, power analysis indicated that the proper number of participants is close to the actual one (48 vs. 44), larger cohorts is expected to provide more definitive results regarding the neuroplastic changes due to the dance intervention. We should also mention that the present study enrolled only Greek senior citizens and lacks diversity concerning ethnic background. Our future aim is to investigate the generalizability of the intervention to other sample cohorts across the world and to compare it with other traditional dance types. Finally, future studies should investigate whether the beneficial effects of dancing are dependent from demographic factors (age, gender), education, cognitive reserve, socioeconomic status, etc.

To sum up, the present study investigated whether traditional Greek dance training would induce neuroplasticity effects on senior citizens. It employed network neuroscience in order to investigate the beneficial role of dancing on both global and local cortical level. To the best of our knowledge, it is the first attempt that demonstrated improved functional performance on cortical level through increased small world-property. This finding was mainly attributed to faster information flow and more accurate information integration among distant cortical regions (smaller characteristic path length). Neuroplasticity was more evident on specific modules such as the fronto-parietal and executive resting state networks known for being responsible for attention, motor planning and execution. The present computational approach seems to provide an integrative framework for evaluating the intervention impact on groups of senior citizens and to be robust for quantifying neuroplasticity effects on multi-level aspects of the functional cortical networks.

#### AUTHOR CONTRIBUTIONS

VZ designed and implemented the dance program, collected the data and prepared the initial draft of the manuscript. CF collected and analyzed the EEG data, guided the analysis, wrote EEG methodological parts, discussed the results and revised the manuscript. ER conceived the analysis of data. EP revised the manuscript. SD guided the study. PB co-guided the study and revised the manuscript.

## FUNDING

This work has been partially funded by the European Union's Horizon 2020 research and innovation programme under grant agreement No 643555 Ubiquitous iNteroperable Care for Ageing People (http://www.uncap.eu).

#### REFERENCES


networks in mild Alzheimer's disease and amnestic mild cognitive impairment: an eeg study using relative wavelet entropy (RWE). Front. Aging Neurosci. 6:224. doi: 10.3389/fnagi.2014.00224


hippocampal plasticity and balance abilities in healthy seniors. Front. Hum. Neurosci. 11:305. doi: 10.3389/fnhum.2017.00305


graph theoretical studies of brain networks. Neurobiol. Aging 34, 2023–2036. doi: 10.1016/j.neurobiolaging.2013.02.020


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Zilidou, Frantzidis, Romanopoulou, Paraskevopoulos, Douka and Bamidis. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Greek Traditional Dances: A Way to Support Intellectual, Psychological, and Motor Functions in Senior Citizens at Risk of Neurodegeneration

#### Styliani Douka<sup>1</sup> , Vasiliki I. Zilidou1,2 \*, Olympia Lilou<sup>1</sup> and Magda Tsolaki<sup>3</sup>

<sup>1</sup> Laboratory of Sports, Tourism and Recreation Management, School of Physical Education and Sport Science, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>2</sup> Laboratory of Medical Physics, Medical School, Aristotle University of Thessaloniki, Thessaloniki, Greece, <sup>3</sup> Department of Neurology, Medical School, Aristotle University of Thessaloniki, Thessaloniki, Greece

One of the major problems that elderly people are facing is dementia. For scientist's dementia is a medical, social and economic problem, as it has been characterized as the epidemic of the 21st century. Prevention and treatment in the initial stages of dementia are essential, and community awareness and specialization of health professionals are required, with the aim of early and valid diagnosis of the disease. Activities are recommended to the senior citizens to improve their physical and mental health. Dance has been suggested as an appropriate recreational activity for the elderly that brings functional adjustments to the various systems of the body, psychological benefits, and makes exercise to seem interesting and entertaining as it combines the performance of multiple animations with musical accompaniment. A Greek traditional dance program was performed where our sample consisted of 30 healthy elderly and 30 with Mild Cognitive Impairment – MCI. It lasted 24 weeks, two times a week for 60 min. Specific traditional dances from all over Greece were selected. The dances were of a moderate intensity at the beginning with a gradual increase in intensity, according to the age and physical abilities of the participants. The results showed a significant improvement in: attention (S4viac-Healthy: z = −3.085, p = 0.002; MCI: z = −3.695, p < 0.001, S4viti-Healthy: z = −2.800, p = 0.005; MCI: z = −3.538, p < 0.001), anxiety (Healthy: z = −2.042, p = 0.041; MCI: z = −2.168, p = 0.030), verbal fluency for MCI (Verflx: t = −2.396, df = 29, p = 0.023, Verfls: t = −3.619, df = 29, p = 0.001, Verfmo: t = −3.295, df = 29, p = 0.003) and in executive functions (FUCAS: z = –2.168, p = 0.030). Significant improvement also showed in physical condition (Arm curl– Healthy: z = –3.253, p = 0.001; MCI: z = −3.308, p = 0.001, Chair stand – Healthy: t = –3.232, df = 29, p = 0.003; MCI: t = −2.242, df = 29, p = 0.033, Back scratch– Healthy: z = −1.946, p = 0.052; MCI: z = −2.845, p = 0.004, 2 min step– Healthy:

#### Edited by:

Ashok Kumar, University of Florida, United States

#### Reviewed by:

Denis Gris, Université de Sherbrooke, Canada Maria Velana, Universitätsklinikum Ulm, Germany

> \*Correspondence: Vasiliki I. Zilidou vickyzilidou@gmail.com

Received: 15 June 2018 Accepted: 10 January 2019 Published: 25 January 2019

#### Citation:

Douka S, Zilidou VI, Lilou O and Tsolaki M (2019) Greek Traditional Dances: A Way to Support Intellectual, Psychological, and Motor Functions in Senior Citizens at Risk of Neurodegeneration. Front. Aging Neurosci. 11:6. doi: 10.3389/fnagi.2019.00006

**565**

z = –2.325, p = 0.020; MCI: z = −2.625, p = 0.009, FootUpandGo– Healthy: z = −4.289, p < 0.001; MCI: z = −3.137, p = 0.002, Sit and Reach: z = −3.082, p = 0.002, Balance on One leg: z = −3.301, p = 0.001) and Quality of life (Healthy: z = −1.937, p = 0.053; MCI: z = −2.130, p = 0.033). This study proves that dancing not only improves the cognitive and physical condition of the elderly but also contributes to a better quality of life.

Keywords: Greek traditional dances, dementia, quality of life, physical health, mental health

#### INTRODUCTION

People living with dementia have poor access to appropriate healthcare, even in most high-income country settings, where only around 50% of people living with dementia receive a diagnosis. In low and middle-income countries, less than 10% of cases are diagnosed. As populations age due to increasing life expectancy, the number of people with dementia is increasing. We estimate that there were 46.8 million people worldwide living with dementia in 2015 and this number will reach 131.5 million in 2050 (World Alzheimer Report 2016; Prince et al., 2016). In Greece, there are more than 200,000 patients with dementia and this figure is expected to exceed to 600,000 patients by 2050, while the annual cost of dementia in Greece is now approaching six billion euros (Alzheimer Athens).

The term "dementia" is generic and refers to a complex group of changes with known or unknown etiology, which occur with widespread disruption of cognitive abilities and social functions of the individual. "Dementia can be reversible or irreversible, with rapid or slow progression, and characterized by multiple deficits of cognitive functions or almost exclusive disorder of emotion, initiative and personality" (Gorelick et al., 2011). The most common type of dementia that mainly occurs in the elderly is Alzheimer's disease type (AD). The rapid increase in dementia ranges from about 2–3% among people aged 70–75 years and from 20–25% among people aged 85 and over (Ferri et al., 2005). The most serious and early cognitive problem in AD is memory loss. This loss is gradual and occurs within the limits of a normal level of consciousness, without any other central nervous system disorder that could explain these symptoms.

Mild Cognitive Impairment (MCI), is an emerging term that encompasses the clinical stage between normal cognitive status and dementia. It is a condition considered to be a transition between normal mental changes due to age and early clinical signs of dementia (Petersen, 2004). Its features, applications, and definitions are controversial. The MCI is now focusing on studies of natural history, biological markers and on the prevention of AD. The stage of the MCI is probably the best stage at which we could intervene with preventive strategies. Despite the conflict, progress has been made in determining the risk factors for progression from MCI to dementia. Now, treatments in order to prevent the development of AD are focusing on the MCI as a treatment group, and neurologists will increasingly be called upon to do this diagnosis. This interest is motivated by patient requests for prognosis and treatment. Therapists-neurologists have at their disposal a wealth of research information, though it is illustrated by the lack of practical suggestions for patient management (Chertkow et al., 2008). MCI is associated with an increasing risk of developing dementia. Patients with this pattern of early deficits develop dementia at a rate of 10–15% per year, while the rate for healthy controls is only 1–2% per year. However, data on the prevalence of MCI and its rate of conversion to dementia vary widely, depending on the different determinants applied.

Also, the functionality is a key area affected by aging. The functional capacity is considered to be an important part of health and wellness. The lack of mobility is the major reason that older people have problems with functionality. During aging, there are some problems in the musculoskeletal system and joints. Strength and muscle mass decrease over time. Physical activity, especially strength training, is very important action to come up against this situation (Keller and Engelhardt, 2013). There exists a small decrease in muscle strength up to 40–50 years and a 30–40% decrease in 70 to 80 years. This reduction is due to sarcopenia, which occurs more in elderly women than in men (Cruz-Jentoft et al., 2010). Genetic factors and lifestyle, such as reduction of physical activity, smoking and the use of alcoholic beverages can contribute also to sarcopenia and dementia. Muscular weakness is associated with increased risk of falls (Soriano et al., 2007; Leveille et al., 2009), resulting in possible fractures (Aniansson et al., 1984; Carpintero et al., 2014).

Physical activity stimulates the physiological functions of the body and can contribute to the stabilization of a good level of cognitive functions making the elderly more energetic. Exercise improves physical health, behavior, mental state, communication and functionality in the elderly with cognitive impairment, especially exercise that is for durability, agility, muscle strength and balance (Garber et al., 2011). The activities proposed for participation by the elderly should lead to the improvement or maintenance of physical, spiritual and mental health (Kim, 2009).

In international literature, dance has been suggested as an appropriate recreational activity for the older adults. Dance is a physical activity that causes functional changes in various systems of the human body. Previous studies have shown that elderly who dance at regular intervals have significant benefits as better balance, stability, flexibility and cognitive status than other elderly who do not dance on a regular interval (Kattenstroth et al., 2011). Dance, also offers psychological benefits and can also make the exercise more interesting and entertaining, as it combines the execution of multiple kinetic tasks with music accompaniment. For elderly, dancing is a pleasure, is exercise capacity, companionship, mental balance, wellness, coordination and muscle tone (Mullen et al., 2012). Music is still an important component of pleasure, as individuals enjoy it and

express themselves through it. The rhythmic music, improves the coordination of gait and proprioceptive movement control in people with neuromuscular and skeletal disorders and leads to increased mobility and stability. Mild dance activity can prevent the risk of high blood pressure, of diabetes and the cardiovascular diseases (Gordon et al., 2004). It also helps to prevent falls and loss of bone density (Minne, 2005), improves the flexibility of the joints, especially the lower limbs, as all the muscle groups are exercised through a combination of slow and fast steps. Along with the well-being, it activates the muscles, accelerates the cardiac resistance and blood circulation, increases the burns and thus affects the metabolism, increases the maximum oxygen intake, improves the myocardial contractility, increases the frequency of breathing (Xerakia and Kalogerakou, 2000). It also requires simultaneous operation of both cerebral hemispheres, while at the same time activates kinesthetic, logical, musical and emotional processes. For this reason, dancing as a physical activity helps at a rate of 76% the risk reduction in dementia. The standard steps and specific figures does not help much. Creativity is the special component in dance that offers more results (Powers, 2010).

Greek traditional dances are an activity which offering pleasure, entertainment, education and characterized by diversity, complexity, since the combinations of lower and upper limb movements dominate and differ in intensity and in movements from other types of dance. Apart from the entertainment they offer, they are also classified as an aerobic activity that causes a burden but as part of the physiological adjustments (Galanou, 2003). Also classified as an aerobic leisure activity offering a variety of intensity and rhythm, as there is a pleasant climate during the practice (Pitsi, 2005).

In recent years, there has been growing interest in studying the quality of life. The "quality of life" is a concept with a broad scope, that is something that makes difficult its measure and its integration into the scientific study (Fallowfield, 2007). It includes epidemiological, biomedical, functional, economic and cultural approaches, as well as personal preferences, perceptions and experiences. International organizations such as the Organization of the United Nations (UN) and the World Health Organization (WHO) recognize the importance of quality of life through various declarations and conventions. Participating in properly organized exercise programs, physical education and physical activity programs contributes to positive self-esteem and high self-assessment, factors that lead to the adoption of appropriate and desirable attitudes and behaviors, greatly ensure physical well-being and mental health (Landers and Arent, 2001; Hamill et al., 2011).

The positive contribution of physical exercise to quality of life is well documented, as participation in physical activities and systematic exercise help to enhance mental well-being, increase positive mood, seek pleasurable and intense experiences, improve health and control stress, both in healthy and in clinical populations (Theodorakis, 2010). Continuously new research and work proves that exercise and participation in physical activities are associated with better performance of cognitive functions, self-esteem and self-confidence, reduction of anxiety and depression, mental well-being and an improvement in quality of life.

The aim of this study is to demonstrate the importance of Greek traditional dances in improving both the cognitive and physical health of the senior citizens. Dance is an enjoyable type of aerobic exercise that can cause various changes in the human body. We investigated if the Greek traditional dance can be an important tool for enhancing health status of senior citizens and simultaneously to improve their quality of life. Furthermore, we investigated if dance may delay the beginning of a cognitive impairment or dementia.

### MATERIALS AND METHODS

### Subjects

The sample consisted of elderly people (n = 60) who were selfserving, had good functional and emotional state and normal or non-normal cognitive status. The subjects were divided into two groups depending on their diagnosis. More precisely, thirty participants (n1 = 30) were healthy seniors with median age of 65.50 years [Interquartile range (IQR) = (62.00, 68.00)] and median education of 13 years [IQR = (8.75, 16.25)] while thirty participants (n2 = 30) had a diagnosis of mild cognitive impairment (MCI). The MCI participants had a median age of 67.50 years [IQR = (63.00, 70.00)] and median education of 6 years [IQR = (6.00, 8.25)]. Participants did not participate in other Greek Traditional Dances programs or at any other cognitive rehabilitation programs. The intervention took place at the Greek Association of Alzheimer Disease and Relative Disorders (Alzheimer Hellas) and at the Day Care Centers of Municipality of Thessaloniki It lasted 24 weeks with a frequency of two times per week in sessions of 60 min.

Inclusion criteria were age ≥60 years, senior citizens with mild cognitive impairment, agreement of a medical doctor and time commitment to the dance protocol. Exclusion criteria were concurrent participation in another study, hypertension, heart and respiratory failure, uncorrectable vision problems, inability to participate at 80% of the hours of the program. The training program was provided at no cost and participants received no compensation. At the beginning of the program and at the end of this, a neuropsychological evaluation was performed by a psychologist to assess the cognitive, functional, and behavioral status of each participant. The fitness and functional capacity evaluated by a fitness instructor and their quality of life was assessed through appropriate questionnaires. The required descriptions were given for the purpose of this research and written consent was requested from the senior citizens to participate. Ethical and Scientific Committee of GAARD approved the protocol of this study.

#### Outcome Measures Psychological Evaluation

The neuropsychological assessment was performed before the intervention (initial assessment) and after the intervention of dance (final assessment). Fifteen (15) different tests were used, which examine all cognitive functions (memory,

reason, judgment, abstract thinking, complex skills, attention, concentration, orientation, audiovisual perception), activities of daily living behavioral problems and quality of life. The tests that were selected are the following with a reference to what they evaluate: Mini Mental State Examination (MMSE), it is a screening instrument to separate patients with cognitive impairment from those without it (Fountoulakis et al., 2000), Clinical Dementia Rating (CDR), characterize six domains of cognitive and functional performance (Morris, 1993), Functional Cognitive Assessment Scale (FUCAS), assesses executive function in daily life activities directly in patients with dementia (Kounti et al., 2006), Functional Rating Scale for Symptoms of Dementia (FRSSD), assesses the daily functionality (Hutton, 1990), Instrumental Activities of Daily Living (IADL), assesses the independent living skills (Theotoka et al., 2007), Test of Every Day Attention (TEA), assesses the attention (Robertson et al., 1996), Trail-making Test (TMT), assesses the executive function (Vlahou and Kosmidis, 2002), Rey–Osterreith Complex Figure Test (ROCF), assesses the memory (Rey and Osterrieth, 1993), Rey Auditory Verbal Learning Test (RAVLT) and Rivermead Behavioral Memory Test (RBMT), assesses the memory (Efkildes et al., 2002), Verbal Fluency Test (VFT), assesses the cognitive function (Kosmidis et al., 2004), Neuropsychiatric Inventory (NPI), assesses the range of neuropsychiatric symptoms (Politis et al., 2004), Geriatric Depression Scale (GDS), assesses the depression (Fountoulakis et al., 1999), Quality of Life in Alzheimer's Disease (QOL-AD), assesses the quality of life (Logsdon et al., 1999), Beck Anxiety Inventory (BAI), record the anxiety (Beck and Steer, 1988). These tests have been selected on the basis of the validity and reliability and the existence of norms for the Greek population (Kosmidis, 2008; Tsolaki and Kounti, 2010).

#### Physical Evaluation

In order to assess their physical condition and functional capacity, the Body Mass Index (BMI) was first calculated and then the Senior Fitness Fullerton Test was used, which consists of six tests and evaluates the flexibility of low back and hamstrings, the functional capacity of individuals through the strength of the lower limbs and the dynamic balance, the speed, agility and balance during movement (Jones and Rikli, 2002). In addition, their static equilibrium was evaluated through the Flamingo test (Barabas et al., 1996), the length of time spent on one leg was calculated, the strength of the strong hand was recorded with the use of the dynamometer (Saehan Corp., Masan, South Korea) and the jumping ability (vertical jump) was evaluated using OptoJump system (Microgate, Bolzano, Italy). Specifically, the tests used are: Chair stand, 8 FootUpandGo, Back Scratch, Arm Curl, Chair Sit and Reach, 2 Min Step, Balance One leg, Hand Grip Strength, and Jumping ability.

#### Quality of Life Evaluation

To evaluate the quality of life, the questionnaire developed by the WHO was used, the WHOQOL, which aims to promote an intercultural Quality of Life assessment system and the use of this questionnaire in the wider health sector. It includes 26 questions and is divided into four thematic sections (Skevington et al., 2004) where the relevant questions address: (a) physical health; (b) mental health; (c) social relations; and (d) the environment. It also includes two questions, which offer an overall assessment of Quality of Life and Health Status (Ginnieri-Kokkosi et al., 2003). The results with the highest values are an indication of a better quality of life. In general, the multifaceted Quality of Life is examined, as well as a general state of health.

#### Selection of Greek Traditional Dances

The traditional dances selected from all over Greece. The design of the program held by dividing the dances into three categories depending on the complexity and number of steps, the rate of intensity (slow speed) and the position-movement of the hands. Dances were also classified into three categories: mild, moderate and high intensity. Most of them were in moderate intensity, with progressive and increasing intensity, indicative of the age and physical abilities of the participants.

### Statistical Analysis

#### Demographics

We planned comparisons between the independent variables age and education level between groups, respectively. Initially, demographic data were tested for normality assumption between groups (Healthy, MCI) using visual inspection of histograms, normal Q-Q plots and boxplots, in terms of Skewness and Kurtosis as well as using the normality tests (Shapiro and Wilk, 1965; Razali and Wah, 2011). If the independent variable was approximately normally distributed in both groups, differences between groups were explored using parametric methods (independent samplest-tests). However, if the normality assumption was not met, non-parametric analysis (Mann– Whitney U-Test) was followed. Additionally, the possible association between and the gender (male, female) and the group (Healthy, MCI) was investigated by means of Chi-squared test.

The participants' demographic information was described in tables in terms of mean (standard deviation) or median, interquartile range, respectively, depending on the normality assumption. More precisely, when normality assumption was met, the mean (standard deviation) was used whereas in nonnormally distributed variables, median and interquartile range was depicted.

Statistical analysis was performed using the IBM SPSS Statistics (Version 20) and the level of significance was set at p < 0.05.

#### Data

Tasks examining the neuropsychological and somatometric state as well as the quality of life of the participants were performed both before and after the intervention in both groups (Healthy and MCI participants). As assumptions for a Mixed Model Analysis of Variance (or Split-plot ANOVA) were not fulfilled, an alternative analysis was performed. Differences in scores collected from the neuropsychological and somatometric assessment at the two-time points (after training – before training scores) were computed and then tested for normality. The within-group changes, after grouping our data by diagnosis, were explored using either paired t-test or Wilcoxon signed-rank test depending

on normality assumption of score differences at the twotime points. Additionally, the between-group differences were explored comparing score differences between the two groups using either independent samples t-test or Mann–Whitney U-test based on normality assumption of score differences. The methodology used has been published elsewhere (Cramer, 1998; Cramer and Howitt, 2004; Doane and Seward, 2011; Arvanitidou-Vagiona and Xaidits, 2013; Athanasiou et al., 2017; Pandria et al., 2018).

#### RESULTS

#### Demographics

Both variables age and education were not approximately normally distributed for both groups (Healthy, MCI).

Planned comparisons between groups revealed that the age did not significantly vary between healthy and MCI participants (U = 333.500, p = 0.084) whereas MCI individuals seem to have significantly fewer educational years compared to healthy participants (U = 183.500, p < 0.001) (**Table 1**). The proportion of male/female (6/24) participants were equal for both groups and as such no significant association was found between gender and group (χ <sup>2</sup> = 0.000, df = 1, p = 1.000).

#### Neuropsychological Data

The performance of healthy and MCI participants significantly changed at the subtests of TEA test, S4viac (Healthy: z = −3.085, p = 0.002; MCI: z = −3.695, p < 0.001) and S4viti (Healthy: z = −2.800, p = 0.005; MCI: z = −3.538, p < 0.001). More precisely, both healthy and MCI participants showed significant improvement in S4viac test [Healthy – Before training: 9.00 (5.00, 10.00); After training: 10.00 (9.75, 10.00); MCI – Before training: 6.50 (4.00, 10.00); After training: 10.00 (8.00, 10.00)]. However, a significant decrease was observed in S4viti test for both groups [Healthy – Before training: 6.32 (5.20, 9.47); After training: 5.36 (4.36, 6.50); MCI – Before training: 8.58 (6.03, 11.70); After training: 6.39 (5.36, 7.72)].

Additionally, significant decreases were found in RBMT1 and RBMT2 tasks for both groups [RBMT1: Healthy – Before training: 14.00 (11.75, 15.00); After training: 12.00 (9.13, 15.00); z = −3.176, p = 0.001; MCI – Before training: 11.00 (9.00, 13.00); After training: 8.00 (5.88, 10.00); z = −3.811, p < 0.001; RBMT2: Healthy – Before training: 12.25 (10.00, 15.00); After training: 12.00 (7.88, 15.00); z = −1.986, p = 0.047; MCI – Before training: 10.00 (6.00, 13.00); After training: 6.25 (4.00, 10.00); z = −3.580, p < 0.001]. Moreover, anxiety levels have found to

TABLE 1 | Demographic data as age and education of healthy and MCI participants.


be considerably altered, as measured by the BAI test (Healthy: z = −2.042, p = 0.041; MCI: z = −2.168, p = 0.030). More precisely, MCI individuals showed improvement in anxiety levels based on the scores in BAI test when comparing their scores in two-time conditions [Before training: 7.50 (3.00, 12.25); After training: 4.50 (2.00, 10.25)] whereas anxiety levels in healthy participants were increased [Before training: 2.50 (1.00, 6.50); After training: 4.00 (1.00, 8.50)]. In contrary to this, when comparing the MCI participants' scores in two-time conditions at PSS test they showed significant increase in their scores [Before training: 6.63 (6.75); After training: 9.73 (5.92); t = −2.168, df = 29, p = 0.024].

Healthy individuals seem to improve their immediate memory and delayed recall as they scored higher at RAV test [Before training: 39.77 (9.36); After training: 42.87 (10.08); t = 2.095, df = 29, p = 0.045] in the post-intervention screening (**Figure 1**). On the other hand, MCI participants benefited the most from dancing to verbal fluency scoring higher at tasks Verflx [Before training: 7.97 (3.10); After training: 9.33 (3.19); t = −2.396, df = 29, p = 0.023], Verfls [Before training: 8.03 (3.08); After training: 10.00 (3.33); t = −3.619, df = 29, p = 0.001] and Verfmo [Before training: 8.21 (3.02); After training: 9.54 (2.99); t = −3.295, df = 29, p = 0.003].

Considerable deviations in the performance of MCI individuals at S1map1 [Before training: 25.00 (19.75, 30.25); After training: 21.50 (16.75, 29.00); z = −2.153, p = 0.031] and S1map2 [Before training: 41.87 (8.11); After training: 37.67 (9.57); t = 2.508, df = 29, p = 0.018] tasks have observed. Marginally significant difference was found in scores of MCI group at NPI task (z = −1.912, p = 0.056) whereas a statistically significant decrease was reported in their functionality based on FUCAS test [Before training: 42.00 (42.00, 46.00); After training: 44.50 (42.00, 46.00); z = −2.168, p = 0.030] (**Figure 2**).

Furthermore, we performed comparisons of score differences in two-time conditions (post-pre-scores) between groups resulting in significant interactions time × group at FRSSD [Healthy: 0.9 (2.56); MCI: −0.93 (2.64); t = 2.729, df = 58, p = 0.008], RBMT2 [Healthy: −1.00 (–2.63, 0.25); MCI: −2.00 (–5.00, 0.00); U = 312.500, p = 0.041] and BAI [Healthy: 0.00 (–0.25, 4.00); MCI: −2.00 (–5.00, 1.00);U = 262.000, p = 0.005] tests (**Figure 3**). Based on the aforementioned results, MCI participants showed greater improvement in FRSSD and BAI tests.

#### Somatometric Data

Both groups showed considerable improvement after dancing in their strength of both upper [Arm curl task: Healthy – Before training: 27.00 (22.00, 28.25); After training: 28.00 (25.75, 30.25); z = −3.253, p = 0.001; MCI – Before training: 24.00 (21.50, 25.00); After training: 25.50 (24.00, 28.00); z = −3.308, p = 0.001] and lower limbs [Chair stand task: Healthy – Before training: 16.97 (4.80); After training: 18.57 (4.76); t = −3.232, df = 29, p = 0.003; MCI – Before training: 15.73 (4.02); After training: 16.97 (2.46); t = −2.242, df = 29, p = 0.033] as well as in the flexibility of the shoulder belt [Back scratch task: Healthy – Before training: −3.00 (–13.75, 2.00); After training: −2.50 (−14.25, 5.00); z = −1.946, p = 0.052; MCI – Before training: −13.00

FIGURE 1 | Significant design. differences in the performance of Healthy participants when comparing tests' scores in two-time conditions. Asterisk indicates the p-values that reached statistical significance (p < 0.05).

(−20.00, 2.25); After training: −8.00 (–16.25, 4.00); z = −2.845, p = 0.004]. Additionally, the intervention seems to promote gains in aerobic capacity [2-min step: Healthy – Before training: 96.00 (82.75, 115.25); After training: 102.00 (81.50, 125.00); z = −2.325, p = 0.020; MCI – Before training: 90.50 (81.75, 104.50); After training: 99.00 (85.50, 113.50); z = −2.625, p = 0.009] and movement coordination [FootUpandGo task: Healthy – Before training: 4.89 (4.36, 5.78); After training: 4.41 (4.18, 5.16); z = −4.289, p < 0.001; MCI – Before training: 4.99 (4.53, 5.70); After training: 4.73 (4.13, 5.14); z = −3.137, p = 0.002] for both health and MCI individuals (**Figures 4**, **5**). Moreover, healthy participants improved their suppleness of back and the lower femoral back along with their balance achieving higher scores at Sit and Reach task [Before training: 2.00 (–0.25, 5.00); After training: 4.00 (1.50, 9.25); z = −3.082, p = 0.002] and Balance on One leg task [Before training: 37.28 (13.46, 57.30); After training: 45.76 (19.54, 67.06); z = −3.301, p = 0.001], respectively. On the other hand, MCI individuals marginally altered their performance at the Handgrip task [Before training: 23.13 (9.27); After training: 24.27 (8.54); t = 2.014, df = 29, p = 0.053].

FIGURE 4 | Healthy participants improved their performance in most of the somatometric tests. Asterisk indicates the p-values that reached statistical significance (p < 0.05).

Planned comparisons of score differences in two-time conditions between groups revealed a marginally significant interaction time × group in Sit and Reach task [Healthy: 2.50 (0.75, 6.00); MCI: 1.00 (–1.25, 3.25); U = 322.000, p = 0.057] (**Figure 6**).

### Quality of Life Parameters

Dance seems to promote generally significant gains in the quality of life and health status for both healthy [Before training: 62.63 (57.25, 71.25); After training: 64.88 (62.13, 73.56); z = −1.937, p = 0.053] and MCI [Before training: 61.25 (54.44, 70.13); After training: 65.00 (57.06, 70.50); z = −2.130, p = 0.033] individuals (**Figure 7**). Moreover, healthy participants due to the intervention improved their interaction with the environment [Before training: 72.00 (61.25, 81.00); After training: 75.00 (69.00, 81.00); z = −2.062, p = 0.039].

#### DISCUSSION

In this research, it is observed that the dance intervention has presented significant benefits to mental and physical health in healthy elderly and in elderly with MCI, also in their quality of life. Their performance significantly changed in the assay examining the daily attention (selective attention, sustained attention, stirring of attention and execution of dual work in visual and auditory attention). Specifically, in two sub-tests, significant statistical results were observed. It seems that the effect of dance was positive in this particular test. The test requires in terms of the participant not only speed but also good visual competence. Consequently, the effect of dance seems to be particularly beneficial to participants, increasing their alertness and improving their visual perception. In addition, intervention has altered the performance of visual and audio information, and the level of anxiety was found to have changed significantly. In particular, Mild Cognitive Impairment (MCI) individuals showed an improvement in their anxiety levels based on their twotime scores, while healthy participants increased stress levels. Healthy individuals appear to have improved their immediate and delay memory as they were scored higher in the postintervention test. This increase indicates, that after the dance intervention participants were able to improve their memory levels. The fact that dance intervention has had positive effects on memory improvement is considered to be particularly important

FIGURE 6 | Score differences in two-time conditions between groups showed a marginally significant interaction time × group in Sit and Reach task.

because the most common form of dementia is AD, which basically decreases the brain hippocampus resulting in memory difficulties. The speech fluency seems to be dwindling as dementia progresses, so patients cannot even say a word in advanced stages of the disease. In the final count, the number of words recalled in each sub-test of the verbal option was significantly larger than the original one. In our research, it was observed that the dance had a positive effect on the healthy group as it improved their performance, but the effects were more admirable in the MCI participants. This shows that dance has produced positive results, which helped them to function and think faster. All of these findings, are based on the effectiveness of dance in elderly people and in particular in their physical and mental health, as shown by other studies (Kim et al., 2011; Kattenstroth et al., 2013).

The criterion for the separation of MCI from dementia is the ability to resolve daily activities, something that patients with MCI can achieve with a potentially relatively slower pace than normal elderly, but they can achieve it with positive results, while patients with dementia, even at baseline, seems to be unable to do so. After the dance intervention, the price seems to have increased. This could be explained in two ways: (a) Probably the increase in units in the test is a normal increase, as we know that dementia is an evolving disease and therefore over time it is expected that the patient will have more difficulty and (b) dance can maintain the functionality of the patients which would statistically worsen if they did not take the dance intervention. As shown in Hwang and Braun (2015), survey, adding physical activity to one's life is an effective method of preventing, controlling, and alleviating some health conditions. Studies have demonstrated that physical activity has positive effects on depression, anxiety, dementia, heart failure, stroke, cognition and sleep. The harmful effects resulting from physical inactivity and the positive effects of physical activity suggest that further efforts are needed to encourage physical activity, with an emphasis on populations at high risk for inactivity. Maintaining the functional capacity of the elderly at satisfactory

levels, is something that leads them to an independent and quality lifestyle, while reducing the risk of various diseases. In our research, the two groups showed a significant improvement after the intervention of the dance in their strength both in the upper and low part of the body, as well as in the flexibility. Moreover, the dance seems to promote significant benefits to their aerobic capacity and coordination of movements, both in healthy and in MCI individuals.

In addition, healthy participants improved their flexibility in the lower back and in femoral back as well as their equilibrium by achieving higher scores. Both teams have improved their functional capacity and body balance, especially the skills that related to day-to-day activities such as luggage or shopping and they have gained more confidence and independence as physical strength and energy allow them to engage in more activities with less fatigue.

Controlling the balance and maintaining the strength of the lower limbs, are considered important in order to reduce the episodes of drastic falls (Buchner et al., 1997). Many investigations have reported the importance of the ankle joint in the entire human body's mechanics, particularly in sensitive groups such as the elderly, in whom the muscles around the joint seem to be more affected by old age. As the strength of the dorsal flexors of the ankle seems to be affected by aging, especially in the elderly with a history of falls, it is likely that the strengthen of this particular muscle group improves the control of the elderly, resulting in falls reduction. Balance is an important functional capacity that influences significantly the ability of human to perform daily activities for their survival, such as maintaining a stable posture, the straight movement from one position to another and maintain the upright posture (Islam et al., 2004).

Quality of life refers not only to one or some particular external features, but rather expresses an existential state of the individual. Its assessment methodology should focus on identifying those factors that have a particular focus on subjective judgment and quality of life assessment. The results of our research showed statistically significant effects in the overall quality of life and the general health status of participants in the Greek traditional dances program. Dance generally seems to promote significant gains in quality of life and health status for both the healthy and for the MCI subjects. In addition, healthy participants due to dance intervention improved their interaction with the environment. It appeared that they began to acquire a sense of security with regard to external dangers, they have also started to engage in various recreational activities in their spare time and to use the means of transport more easily. Cruz-Ferreira et al. (2015), presented the positive effects of dance in different dimensions of functioning and the potential to contribute to healthy aging. This could be related to the integrated mobilization of physical, cognitive, and social skills promoted by creative dance. Also, Sivvas et al. (2015) at his study, shown that dancing helps in many ways to preserve and improve human health, as far as physical health is concerned – as it maintained the physical state in good level, but also concerning mental health – by minimizing stress and depression. Finally, social health also proved to be positively affected as the factors that prevent an individual from socialization were reduced.

Generally, our findings are in line with Teri et al. (2008) study that for long-term participation of dementia patients in exercise programs, it is necessary to be able to perform them, be fun and enjoyable. Thus, the pleasant environment, the effective communication, the feeling of security as well as the entertaining character of the exercise programs, strongly consent to the regular participation of patients in them.

In Greece than in many other countries, there is a "living tradition," i.e., in the villages living tradition continues, evolves and sometimes consciously experienced. This is an element of the inextricable relationship of the Greeks with their traditions that are lost in the depths of the ages. Traditional dances are deeply rooted in the culture of Greece, expressing local history, traditions and customs, express culture. By learning the dances of an area, someone could learn at the same time its peculiarities, history, geographic stories and myths. It is a necessity for the Greek people to remain in tradition, with the result that local dances have a strong expression of emotions but also creativity. We could investigate dances from different countries in a next study in order to evaluate the results that will arise.

## CONCLUSION

The participation of the elderly in activities helps them to maintain their physical status, but also gives them the opportunity to interact with other people of all ages. This interaction, removes the feeling of loneliness, which stimulates their psychological state. Also, improves their self-esteem as they realize that they can participate in new skills. Dance as exercise, increases body resistance, helps them to maintain proper posture, stimulates the muscular system and improves physical fitness. Combined with music, helps to express emotions, combat stress and improves mental health. With the repetition of the steps, the movement of the hands and their combined function helps to improve the cognitive functions.

Furthermore, dance seems to be particularly important to protect against dementia and to slow the progression of the disease. Socialization contributes to the maintenance of positive psychology, which can be a protective shield for dementia, and to protect patients with MCI from potential depression, which adds to the situation of a patient with mental problems. It is true that nowadays, the non-pharmaceutical interventions for the treatment of dementia play a particularly important role on the world stage, and our research confirms the studies that have been established so far which mention the importance of dance.

## FUTURE DIRECTIONS

The results of our research suggest that the intervention of Greek traditional dances in elderly is effective for their physical and mental health. Future research should focus on an additional diagnostic test such as neuropsychological recordings through an electroencephalography that records the brain's functions and in particular its electrical activity. After the end of interventional assessment will identify any changes that occur in brain activity. Generally, the electroencephalography is a very useful method, because it can non-invasively and painlessly give a complete view of brain function or even brain disorders.

#### AUTHOR CONTRIBUTIONS

fnagi-11-00006 January 23, 2019 Time: 17:10 # 10

VZ designed and implemented the dance program, collected and analyzed the data, prepared the initial draft of the manuscript, guided the analysis, and revised the manuscript. OL implemented the dance program, collected the data, and revised the manuscript. MT guided the study. SD co-guided the study.

#### REFERENCES


#### FUNDING

This work was partly supported by the project "Augmentation of the Support of Patients suffering from Alzheimer's Disease and their caregivers (ASPAD/2875)," which is materialized by the Special Account of the Research Committee at Aristotle University of Thessaloniki. The project was funded by the European Union (European Social Fund) and the Ministry of Education, Lifelong Learning and Religious Affairs in the context of the National Strategic Reference Framework (NSRF, 2007– 2013).



Osterrieth's "The complex figure copy test" (J. Corwin & F. W. Bylsma, Trans.). Clin. Neuropsychol. 7, 3–21.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Douka, Zilidou, Lilou and Tsolaki. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

digital media

of impactful research

article's readership