# DYNAMIC FUNCTIONAL CONNECTIVITY IN NEUROPSYCHIATRIC DISORDERS: METHODS AND APPLICATIONS

EDITED BY : Wenbin Guo, Feng Liu and Zaixu Cui PUBLISHED IN : Frontiers in Neuroscience, Frontiers in Psychiatry and Frontiers in Neurology

#### 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 978-2-88966-195-4 DOI 10.3389/978-2-88966-195-4

### 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

## DYNAMIC FUNCTIONAL CONNECTIVITY IN NEUROPSYCHIATRIC DISORDERS: METHODS AND APPLICATIONS

Topic Editors: Wenbin Guo, Central South University, China Feng Liu, Tianjin Medical University General Hospital, China Zaixu Cui, University of Pennsylvania, United States

Citation: Guo, W., Liu, F., Cui, Z., eds. (2020). Dynamic Functional Connectivity in Neuropsychiatric Disorders: Methods and Applications. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88966-195-4

# Table of Contents

*04 Editorial: Dynamic Functional Connectivity in Neuropsychiatric Disorders: Methods and Applications*

Xiaoya Fu, Feng Liu, Zaixu Cui and Wenbin Guo


Liyuan Li, YiFeng Wang, Liangkai Ye, Wang Chen, Xinju Huang, Qian Cui, Zongling He, Dongfeng Liu and Huafu Chen

*25 Reduced Dynamic Interactions Within Intrinsic Functional Brain Networks in Early Blind Patients*

Xianglin Li, Ailing Wang, Junhai Xu, Zhenbo Sun, Jikai Xia, Peiyuan Wang, Bin Wang, Ming Zhang and Jie Tian


Xuesong Li, Yuhui Xiong, Simin Liu, Rongsong Zhou, Zhangxuan Hu, Yan Tong, Le He, Zhendong Niu, Yu Ma and Hua Guo

*54 Increased Temporal Dynamics of Intrinsic Brain Activity in Sensory and Perceptual Network of Schizophrenia*

Youxue Zhang, Gang Guo and Yuan Tian

*62 Dynamic Alterations of Spontaneous Neural Activity in Parkinson's Disease: A Resting-State fMRI Study* Chao Zhang, Binru Dou, Jiali Wang, Kai Xu, Haiyan Zhang, Muhammad Umair Sami, Chunfeng Hu, Yutao Rong, Qihua Xiao, Nan Chen and Kuncheng Li

# Editorial: Dynamic Functional Connectivity in Neuropsychiatric Disorders: Methods and Applications

Xiaoya Fu1,2, Feng Liu<sup>3</sup> , Zaixu Cui <sup>4</sup> and Wenbin Guo1,2 \*

*<sup>1</sup> Department of Psychiatry, The Second Xiangya Hospital of Central South University, Changsha, China, <sup>2</sup> National Clinical Research Center on Mental Disorders, Changsha, China, <sup>3</sup> Department of Radiology, Tianjin Medical University General Hospital, Tianjin, China, <sup>4</sup> Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, United States*

Keywords: functional magnetic resonance imaging, dynamic functional connectivity, neuropsychiatric disorders, cognitive deficits, regional homogeneity, amplitude of low-frequency fluctuations

**Editorial on the Research Topic**

### **Dynamic Functional Connectivity in Neuropsychiatric Disorders: Methods and Applications**

Resting-state functional magnetic resonance imaging (RS-fMRI), a non-invasive measurement of spontaneous brain activity, has greatly broadened our understanding of neural substrate underlying neuropsychiatric disorders over the last several decades. Since Biswal et al. discovered synchronized brain activity in different brain areas even without any tasks or stimuli (Biswal, 2012), numerous studies have investigated resting-state coupling (i.e., functional connectivity, FC) between different brain areas in neuropsychiatric disorders (Guo et al., 2015; Zhu et al., 2018).

FC can be defined as a temporal correlation of blood-oxygen-level dependent (BOLD) signal between spatially distributed brain regions (Biswal et al., 1997). Most previous RS-fMRI studies assumed that FC was constant throughout the observation period of task-free experiments (Hutchison et al., 2013). Recently, several studies have demonstrated the feasibility of dynamic methods in characterization of functional brain changes, such as dynamic FC (dFC) investigated by the sliding-window method, which provide novel insights into underlying neural activity (Chang and Glover, 2010; Liu et al., 2017; Duan et al., 2019). However, window size, window stepsize, and window type are open areas of research and important parameters to capture the resting-state FC dynamics. Sliding-window and time-frequency analyses are the two frequently used dynamic functional analyses (Hutchison et al., 2013). Apart from dFC, dynamic amplitude of low frequency fluctuations (dALFF) and dynamic regional homogeneity (dReHo) are also widely used (Deng et al., 2016; Fu et al., 2018). Both static and dynamic functional metrics provide great insight into understanding functional deficits of neuropsychiatric disorders (Biswal, 2012; Hutchison et al., 2013). Therefore, deep and detailed understanding of the method and application of dynamic functional metrics in neuropsychiatric disorders is critical.

This special issue focuses on the recent developments in dynamic functional analyses and their applications in neuropsychiatric disorders. A total of 7 articles were included in this Research Topic.

### POSTTRAUMATIC STRESS DISORDER (PTSD)

Fu et al. applied dReHo and dFC to investigate both local and large-scale functional coupling in patients with PTSD. Results indicated increased dReHo in the left precuneus in patients with PTSD. Also, the left precuneus exhibited increased dFC with the left insula and decreased dFC with the left inferior parietal lobe and right precuneus, suggesting that the left precuneus might be critical for the pathophysiology of PTSD.

### Edited and reviewed by:

*Vince D. Calhoun, Georgia State University, United States*

\*Correspondence: *Wenbin Guo guowenbin76@csu.edu.cn*

#### Specialty section:

*This article was submitted to Brain Imaging Methods, a section of the journal Frontiers in Neuroscience*

Received: *26 December 2019* Accepted: *20 March 2020* Published: *20 April 2020*

#### Citation:

*Fu X, Liu F, Cui Z and Guo W (2020) Editorial: Dynamic Functional Connectivity in Neuropsychiatric Disorders: Methods and Applications. Front. Neurosci. 14:332. doi: 10.3389/fnins.2020.00332*

**4**

## GENERALIZED ANXIETY DISORDER (GAD)

Brain signal variability (BSV) is a method to measure the temporal variability of standard variation of BOLD signal, which reflects capacity of transition between brain states and processing various external stimuli. Li L. et al. evaluated the changes of BSV in patients with GAD and found that extensive brain regions exhibited decreased BSV in patients with GAD compared to healthy controls (HCs), suggesting that the brain of patients with GAD may be in a less flexible state compared to HCs.

### EARLY BLIND

Dynamic causal modeling (DCM) is an approach to measure causal functional interactions among neuronal populations, i.e., effective connectivity. Li, Wang et al. used spectral DCM to investigate whether early visual deprivation had an impact on the dynamic causal interactions among regions within the default mode network, salience network, and dorsal attention network in patients with early blind. Abnormal patterns of effective connectivity within all these three networks were found in patients with early blind compared to HCs, which might imply the effect of early sensory deprivation on brain plasticity.

### PARKINSON'S DISEASE (PD)

Li, Xiong et al. recruited 62 participants with PD and demonstrated that dynamic nodal efficiency measurement, which was calculated from RS-fMRI brain network and sliding-window analysis, could be used to predict the severity level of PD after drug therapy. Hippocampus, post-central gyrus, cingulate gyrus, and orbital gyrus were the contributed regions for the prediction. This study offered an example of using RS-fMRI data to predict the treatment effect in patients with PD.

Another study on PD by Zhang C. et al. used dALFF to explore the feasibility of differentiating patients with PD from HCs. Increased coefficient of variation in the left precuneus was observed in patients with PD. Moreover, coefficient of variation of dALFF in the left precuneus was positively correlated with disease duration in the patients. These findings were likely to provide a new direction for diagnosis of PD.

### SUBACUTE STROKE

Chen et al. explored abnormal dynamic characteristics in patients with subacute stroke. Results of both dALFF and dReHo showed

### REFERENCES


significant intergroup differences of regional brain activity. Fugl-Meyer assessment, an index for evaluating the degree of motor deficit, exhibited a positive correlation with dALFF variability in supplementary motor area (SMA) and a negative correlation with dReHo variability in ipsilesional middle frontal gyrus (MFG). The receiver operating characteristic analysis suggested that dALFF in SMA and dReHo in ipsilesional MFG might be potential markers to distinguish patients with subacute stroke from HCs. Therefore, dALFF and dReHo have the potential for evaluating the motor function in patients with subacute stroke.

### SCHIZOPHRENIA

Zhang Y. et al. focused dReHo and dynamic fALFF to investigate abnormal dynamic local functional activity in schizophrenia. Results revealed deficits in the sensory and perception functional networks and a positive relationship between dReHo of the thalamus and the severity of symptoms in the patients, which highlighted the importance of the sensorimotor networks in the physiopathology of schizophrenia.

Taken together, all studies in this special issue suggested progress in the methodology and application of dynamic functional properties in neuropsychiatric disorders. These advances would promote better understanding in the temporal evolution of brain functional activity and provide valuable insight into the development of objective neuro-biomarker of neuropsychiatric disorders. Clinicians will be benefit from this topic in regard to theoretical, experimental and clinical questions related to the nature and origins of dFC in neuropsychiatric disorders.

### AUTHOR CONTRIBUTIONS

WG has contributed to the conception and design of the editorial. XF has not contributed to the conception or design of the editorial, yet contributed considerably to the writing. FL and ZC contributed to the revision of the editorial.

### FUNDING

This study was supported by grants from the National Key R&D Program of China (Grant No. 2016YFC1307100), the National Natural Science Foundation of China (Grant No. 81771447), and the Natural Science Foundation of Tianjin (Grant No. 18JCQNJC10900).


subdivisions in treatment-naive, first-episode schizophrenia. Schizophr. Bull. doi: 10.1093/schbul/sbz087. [Epub ahead of print].


generalized tonic-clonic seizure: dynamic FNC in IGE-GTCS. Hum. Brain Mapp. 38, 957–973. doi: 10.1002/hbm.23430

Zhu, F., Liu, F., Guo, W., Chen, J., Su, Q., Zhang, Z., et al. (2018). Disrupted asymmetry of inter- and intra-hemispheric functional connectivity in patients with drug-naive, first-episode schizophrenia and their unaffected siblings. EBiomedicine 36, 429–435. doi: 10.1016/j.ebiom.2018.09.012

**Conflict of Interest:** 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 © 2020 Fu, Liu, Cui and Guo. 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.

# Dynamic Alterations in Spontaneous Neural Activity in Multiple Brain Networks in Subacute Stroke Patients: A Resting-State fMRI Study

Jing Chen<sup>1</sup>† , Dalong Sun<sup>2</sup>† , Yonghui Shi<sup>1</sup> , Wei Jin<sup>1</sup> , Yanbin Wang<sup>3</sup> , Qian Xi<sup>3</sup> and Chuancheng Ren1,4 \*

<sup>1</sup> Department of Neurology, Shanghai Fifth People's Hospital, Fudan University, Shanghai, China, <sup>2</sup> Division of Gastroenterology, Department of Internal Medicine, Zhongshan Hospital, Fudan University, Shanghai, China, <sup>3</sup> Department of Radiology, Shanghai East Hospital, Tongji University, Shanghai, China, <sup>4</sup> Department of Neurology, Shanghai East Hospital, Tongji University, Shanghai, China

#### Edited by:

Zaixu Cui, University of Pennsylvania, United States

#### Reviewed by:

Yang Yu, Second Affiliated Hospital, Zhejiang University School of Medicine, China Songran Yang, Sun Yat-sen Memorial Hospital, China

#### \*Correspondence:

Chuancheng Ren rccfns17@sina.com †These authors have contributed equally to this work

#### Specialty section:

This article was submitted to Brain Imaging Methods, a section of the journal Frontiers in Neuroscience

Received: 28 August 2018 Accepted: 11 December 2018 Published: 07 January 2019

### Citation:

Chen J, Sun D, Shi Y, Jin W, Wang Y, Xi Q and Ren C (2019) Dynamic Alterations in Spontaneous Neural Activity in Multiple Brain Networks in Subacute Stroke Patients: A Resting-State fMRI Study. Front. Neurosci. 12:994. doi: 10.3389/fnins.2018.00994 Objective: To examine whether subacute stroke patients would exhibit abnormal dynamic characteristics of brain activity relative to healthy controls (HC) and to investigate whether the altered dynamic regional indexes were associated with clinical behavior in stroke patients.

Methods: The dynamic amplitude of low-frequency fluctuations (dALFF) and dynamic regional homogeneity (dReHo) in 42 subacute stroke patients and 55 healthy controls were compared. Correlation analyses between dALFF and dReHo in regions showing significant intergroup differences and clinical scores (i.e., the National Institutes of Health Stroke Scale, Fugl-Meyer assessment and lesion volume size) were conducted in stroke patients. Receiver operating characteristic (ROC) curve analysis was used to determine the potential value of altered dynamic regional indexes to identify stroke patients.

Results: Significantly dALFF in the bilateral cerebellum posterior lobe (CPL), ipsilesional superior parietal lobe, ipsilesional inferior temporal gyrus (ITG), the midline supplementary motor area (SMA), ipsilesional putamen and lentiform nucleus were detected in stroke patients compared to HC. Relative to the HC group, the stroke patients showed significant differences in dReHo in the contralesional rectal gyrus, contralesional ITG, contralesional pons, ipsilesional middle frontal gyrus (MFG). Significant correlations between dALFF variability in midline SMA and Fugl-Meyer assessment (FMA) scores or between dReHo variability in the ipsilesional MFG and FMA scores were detected in stroke patients. Furthermore, the ROC curve revealed that dynamic ALFF at SMA and ReHo at ipsilesional MFG might have the potential to distinguish stroke patients.

Conclusion: The pattern of intrinsic brain activity variability is altered in stroke patients compared with HC, and dynamic ALFF/ReHo might be potential tools to assess stroke patients' motor function.

Keywords: stroke, dynamic intrinsic brain activity, resting-state fMRI, amplitude of low-frequency fluctuations, regional homogeneity

**7**

## INTRODUCTION

fnins-12-00994 December 28, 2018 Time: 16:45 # 2

Stroke is the most common cause leading to varying degrees of neurological dysfunction with a very high likelihood of long-term disability (Liu et al., 2011; Yang et al., 2013). Movement disorders are the major common conditions of stroke-induced disability, and motor functional recovery remains highly variable. Although the exact mechanism of motor deficits and motor recovery are still under investigation, recent advances in neuroimaging have expanded our understanding. Resting-state functional magnetic resonance imaging (fMRI), which is operationally defined as taskindependent spatiotemporal correlations within functionally related regions of the brain (Biswal et al., 1995), has been extensively used to delineate neural function abnormalities in stroke patients.

Resting-state fMRI measures spontaneous brain activity in low-frequency fluctuations which can be reflected by the blood oxygen level dependent (BOLD) signal. An increasingly large body of resting-state fMRI studies in stroke patients has focused on the characteristics of within-region or interregion functional connectivity, such as connections within motor networks or between motor networks and non-motor networks (Wang et al., 2010; Grefkes and Fink, 2014; Wu et al., 2015). However, few studies have examined regional brain activities in patients with stroke. Neural regional properties are crucial for a better understanding of the neurophysiological and neuropathological conditions, such as regional abnormal energy consumption suggesting excessive or decreased resting metabolic rates (Raichle, 2006; Fox and Raichle, 2007). Currently, one of the methods to measure regional properties of the BOLD signal is the amplitude of low-frequency fluctuations (ALFF), which measures the signal strength in low-frequency oscillations of spontaneous neural activity (Zang et al., 2007). The ALFF is correlated with field potential activity in local brain regions (Logothetis et al., 2001), and the amplitude of oscillations can be applied as an index to examine alterations in neural function (Mohamed et al., 2004). Another approach is regional homogeneity (ReHo), which reflects the statistical similarity of local neural activity among spatially adjacent regions (Zang et al., 2004). These two approaches have been widely adopted for evaluating local neural function in neurologic disorders and neuropsychiatric diseases (Qiu et al., 2011; Li et al., 2012; Liu et al., 2012, 2013).

It has been reported that ALFF or ReHo were altered under resting conditions in stroke patients with movement disorders (Skidmore et al., 2013; Tsai et al., 2014) and that the ALFF value or ReHo value in certain brain regions were associated with the severity of motor deficits (Liu et al., 2015; Zhu et al., 2015). However, the aforementioned investigations of regional brain activities assumed that the BOLD signal is stationary during the entire fMRI scan, ignoring the characteristics of dynamic changes of brain spontaneous activity over time. Indeed, evidence has accumulated that brain responds to internal or external stimuli by dynamic integration or adjustment over multiple time scales (Abrams et al., 2013; Yin et al., 2013). Fortunately, the dynamic nature of brain activity may be detected by task manipulations using methods such as electroencephalography and can also be informed by the lower temporal resolution of resting-state fMRI (Calhoun et al., 2008). In recent years, sliding window approaches to functional connectivity have effectively examined abnormal brain function in stroke (Duncan and Small, 2017). Nevertheless, it is not enough to merely focus on timevarying dynamic functional connectivity, since evidence from neuroimaging techniques of high spatiotemporal resolution has verified that local brain activity itself exists with substantial fluctuations (Liao et al., 2015; Fu et al., 2017), and until now, no study explored the dynamic characteristics of local brain activity indexes in stroke patients. These dynamic local approaches are expected to explore the variability of the oscillation amplitudes and regional synchronization of spontaneous brain activity and to advance our understanding of brain function by identifying specific pathophysiological function signatures and our ability to decipher the neural underpinnings of normal or abnormal human behaviors.

Hence, the present study applied resting-state fMRI to investigate whether subacute stroke patients would exhibit abnormal dynamic characteristics of spontaneous brain activity by calculating regional indexes, ALFF and ReHo, compared with healthy controls (HC). Furthermore, another goal was to explore whether the altered dynamic ALFF (dALFF) and dynamic ReHo (dReHo) were correlated with the clinical behavior of the stroke patients. In the current study, we included subacute stroke patients for two reasons. First, the condition of stroke patients at the subacute stage is relatively stable than acute stroke patients, and the patients' compliance is relatively high, facilitating the smooth progress of the current study. Second, we enrolled stroke patients within 1–3 weeks after symptom onset, and this period is well within the recovery window. A period of dramatic changes in functional and structural reorganization may provide more information of spontaneous neural activity. We hypothesized that variability of regional brain activity was altered in patients with stroke-induced motor deficits compared with HC and that dynamic regional indexes in certain regions detected to be associated with the Fugl-Meyer assessment (FMA) scores could provide more information for evaluating of motor function in stroke patients.

### MATERIALS AND METHODS

### Subjects

This study was part of an integrated stroke and rehabilitation project at Shanghai 5th People's Hospital affiliated with Fudan University and was approved by the local ethical committee of Shanghai 5th People's Hospital affiliated with Fudan University. Written informed consent was obtained from all subjects before participating according to the Declaration of Helsinki. A total of 45 subacute stroke patients were recruited. Additionally, 55 HC, who were right-handed and matched for age, gender and education, were recruited from the local communities. The inclusion criteria for stroke patients were as follows: (1) they were aged 40–80 years; (2) it was a first-onset stroke with a single lesion in right-side subcortical regions as verified by diffusion-weighted imaging (DWI); (3) they were examined within 1–3 weeks after stroke symptom onset; (4) they were clinical evidence of a motor

deficit based on neurological examination; and (5) they were right-handed before the stroke. The exclusion criteria for both stroke patients and HC were the presence of any of the following: (1) other brain abnormalities, or clinically significant or unstable medical diseases; (2) unconsciousness, cognitive impairment, or cooperation difficulties; (3) patients with use of medications that could affect motor examination, such as antipsychotics and antiepileptics; (4) patients with cerebellar lesions; and (5) contraindications for MRI scanning. For all stroke patients, the right hemisphere corresponded to the ipsilesional hemisphere. The National Institutes of Health Stroke Scale (NIHSS) and Mini-Mental State Examination (MMSE) were used to evaluate neurological function impairment and cognitive conditions. FMA for upper and lower extremities was applied to evaluate the degree of motor deficit. These clinical behavior scores were collected on the same day as fMRI data acquisition.

### Data Acquisition

All resting-state fMRI data were acquired using a Philips Achieva 3.0 T MR scanner (Philips Medical Systems, Best, Netherlands). Tight but comfortable foam pads and earplugs were used to reduce head motion and scanner noise. Resting-state fMRI was collected using an echo-planar (EPI) sequence with the following scan parameters: repetition time (TR) = 2000 ms; repetition echo time (TE) = 30 ms; flip angle (FA) = 90◦ ; field of view (FOV) = 220 mm × 220 mm; voxel size = 3 mm × 3 mm × 3 mm; matrix = 64 × 64; slice thickness = 3 mm; gap = 1 mm; interleaved transversal slices = 38; and number of volumes = 180. High-resolution sagittal T1-weighted images were acquired using a 3D magnetization prepared rapid gradient echo (MPRAGE) sequence: TR = 8.0 ms; TE = 3.7 ms; FA = 12◦ ; FOV = 256 mm × 256 mm; voxel size = 1 mm × 1 mm × 1 mm; matrix = 256 × 256; slice thickness = 1 mm; and slices = 180. During scanning, all participants were instructed to remain awake, keep their eyes closed, and stay motionless without thinking of anything in particular.

The lesion location of each patient was determined by an experienced neuroradiologist on T1-weighted MRI images. We manually outlined the lesion profiles on T1-weighted MRI images slice by slice using the software MRIcron<sup>1</sup> and generated a lesion mask for each patient. After spatial normalization to Montreal Neurological Institute (MNI) space, all the patients' lesion masks overlapped. We then averaged the individual lesion masks and overlaid them with a template to create the lesion overlap map shown in **Figure 1**.

### Preprocessing of Resting-State fMRI Data

The preprocessing of resting-state fMRI data was performed using the Data Processing Assistant for Resting-State fMRI (DPARSF) version 4.0<sup>2</sup> . The first 10 volumes of each participant were deleted to allow the signal to reach equilibrium and the subjects to adapt to the environment. The remaining 170 volumes were corrected for acquisition time delay between slices. Realignment was conducted to correct head motion. The participants with head motion of >2.0 mm in maximum displacement or >2.0◦ rotation in angular motion were excluded from the study. The mean framewise displacement (FD) was computed by averaging the FD of each subject across the time points, and no significant differences were found between stroke patients and HC (p = 0.235). In addition, each subject's mean FD was included in all group-level analyses as a covariate to further control the head move effect. Subsequently, the structural image was coregistered to the mean functional image after the motion correction, and the transformed structural image was segmented into gray matter, white matter, and cerebrospinal fluid. Then, the segmented images were normalized to MNI space using Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra (DARTEL) algorithm (Ashburner, 2007). Next, the motion-corrected functional volumes were normalized to the MNI space using the normalization parameters for their respective structure images and resampled into a voxel size of 3 mm × 3 mm × 3 mm. Nuisance covariates (24 head motion parameters, cerebrospinal fluid signal, white matter signal and linear trend) were regressed out. Given that it is still a controversy of removing the global signal (Murphy et al., 2009); we did not regress out the global signal. For the ReHo calculation, an additional processing step was that the regressed functional images were temporally bandpass filtered (0.01–0.08 Hz) to reduce low frequency drift and high-frequency noise.

### dALFF and dReHo Analysis

Dynamic regional metrics analysis was performed using Temporal Dynamic Analysis (TDA) toolkits based on DPABI

<sup>2</sup>http://rfmri.org/DPARSF/

<sup>1</sup>http://www.mccauslandcenter.sc.edu/mricro/mricron/

(Yan et al., 2017). Sliding window-based analysis, which is sensitive in detecting time-dependent variations (Hindriks et al., 2016; Liu F. et al., 2017; Yip et al., 2017), was applied to examine the dALFF or dReHo variability over the whole brain. In the sliding window analysis, a temporal window of certain size and shape is chosen, and ALFF and ReHo within that window are calculated. Ideally, the window size should be small enough to detect potentially transient signals, and yet large enough to analyze the lowest frequencies of interest in the signals (Sakoglu et al., 2010). Previous work of sliding window connectivity have applied a sliding window length of as small as 10 s (Thompson et al., 2013) and as long 180 s (Gonzalez-Castillo et al., 2015). In this work, a moderate-length sliding window of 32 TR (64 s) and a shifting step size of one TR (2 s) were used to simultaneously maximize statistical power within the window and also maximize statistical power for cross-level analyses (Allen et al., 2014). The remaining 170 time points after removing the first 10 time points for each individual were segmented into 139 windows in total. In each sliding window, ALFF and ReHo were calculated. For ALFF, the time series was first converted to the frequency domain using a fast Fourier transform, and then the ALFF value of a given voxel was obtained by calculating and summing the square root of the power spectrum between 0.01 and 0.08 Hz. For ReHo, the Kendall's coefficient of concordance (KCC) of the time course of every 27 nearest neighboring voxels was calculated (Zang et al., 2004). The standard deviation (SD) of ALFF values and ReHo values at each voxel across 139 windows was calculated to assess the variability of ALFF and ReHo. To reduce the global effects of variability across subjects, the dALFF and dReHo of each voxel were divided by the global mean dALFF and dReHo values within a gray matter mask, respectively. Finally, the mean normalized dALFF and dReHo maps were spatially smoothed with an isotropic Gaussian kernel of 4 mm full-width-at-half-maximum (FWHM).

### Statistical Analysis

A general linear model (GLM) was used in a voxel-wise manner to compare group differences between the stroke group and HC group in dALFF and dReHo with age, gender, educational level, MMSE and mean FD as covariates. Multiple comparisons were corrected using a voxel-level familywise error rate (FWE) method with corrected p < 0.05 (cluster size ≥50 voxels).

The Shapiro–Wilk statistic was first used to test for normality, and then group comparisons of clinical measures were performed using two-sample t-tests for continuous data and Pearson's chi-squared test for categorical data. Partial correlation analyses were conducted in stroke patients between the clinical measures (NIHSS scores, FMA scores and lesion volume size) and the mean dALFF/dReHo value of each cluster showing significant group differences between stroke group and HC group. The age, gender, educational level, MMSE, mean FD, illness duration and intravenous thrombolysis (IVT) were also considered as covariates. As the correlation analyses were exploratory in nature, the significance levels were set at uncorrected p < 0.05. Furthermore, receiver operating characteristic (ROC) curve analysis was performed to examine the potential value of altered dynamic ALFF or ReHo values in clusters showing significant correlations with clinical behaviors in stroke patients. The optimal cut-off between sensitivity and specificity was determined by maximizing the Youden's index J (J = sensitivity + specificity − 1). A two-tailed p-value of 0.05 was considered statistically significant for the analyses conducted with SPSS version 19.0 statistical software (IBM Corporation, Armonk, NY, United States).

### RESULTS

### Clinical Data

Data obtained from three stroke patients were excluded because of excessive head motion during scanning. Demographic and clinical characteristics of 42 patients with right hemisphere stroke (21 men; mean age, 57.86 ± 11.17 years) and 55 HC (29 men; mean age, 56.73 ± 10.21 years) are listed in **Table 1**. No significant differences in were found (p > 0.05) in gender, age, education level, high risk factors (hypertension, diabetes, hyperlipidemia and atrial fibrillation) and MMSE scores between the stroke group and HC group. The mean lesion volume of stroke patients was 6.10 ± 8.55 cc<sup>3</sup> . Among them, four patients received IVT therapy. Of the 42 patients, 15 had corona radiate lesions, 15 had internal capsule lesions, 10 had basal ganglia lesions, and two had thalamus lesions. The lesion overlaps of stroke patients are shown in **Figure 1**.

TABLE 1 | Demographic and clinical data.


Data represent n (%) or mean ± SD. MMSE, the Mini-Mental State Examination; NIHSS, the National Institutes of Health Stroke Scale; FMA-total, Fugl-Meyer assessment for upper and lower extremities; IVT, intravenous thrombolysis.

## Differences in Dynamic ALFF and Dynamic ReHo

fnins-12-00994 December 28, 2018 Time: 16:45 # 5

The significant differences in dALFF and dReHo between the stroke group and HC group are shown in **Figures 2A,B**, respectively. Compared with HC, significantly increased dALFF variability in the contralesional cerebellum posterior lobe (CPL), ipsilesional superior parietal lobe, ipsilesional inferior temporal gyrus (ITG), ipsilesional CPL and cerebellum tonsil and decreased dALFF variability in the midline supplementary motor area (SMA), ipsilesional putamen and lentiform nucleus were detected in stroke patients (**Table 2** and **Figures 2A,C**). Relative to the dReHo variability of subjects in the HC group, these stroke patients showed a significant increase in contralesional rectal gyrus, contralesional ITG, contralesional pons and a significant decrease in ipsilesional middle frontal gyrus (MFG) (**Table 3** and **Figures 2B,D**).

### Correlational Analysis

A significant positive correlation was detected between the FMA scores and dALFF variability in the SMA (r = 0.347, p = 0.035, uncorrected; **Figure 3A**), a significant negative correlation between the FMA scores and dReHo variability in the midline ipsilesional MFG was found in stroke patients (r = −0.462, p = 0.004, uncorrected; **Figure 3B**). No other significant correlations between NIHSS, size of lesion volume and dynamic regional indexes were observed in the stroke group.

### ROC Analysis

As shown above, significant correlations were detected between the FMA scores and dynamic ALFF variability in SMA or dynamic ReHo variability in ipsilesional MFG, which proposed that the dynamic ALFF/ReHo in these regions might be utilized to differentiate the stroke patients from healthy persons. To verify this possibility, mean dALFF/dReHo values in the SMA or ipsilesional MFG were extracted. Then, ROC analysis was performed to investigate this possibility. The results demonstrated that the area under the curves (AUC) of SMA and ipsilesional MFG were 0.965 and 0.911, respectively (**Figure 4**), which suggested that dynamic ALFF in SMA and ReHo values in the ipsilesional MFG might have the potential to distinguish the stroke patients from healthy subjects. Further diagnostic analysis showed that the sensitivity and specificity were relatively high (**Table 4**).

#### TABLE 2 | Brain regions with significant differences in dynamic ALFF between groups.


ALFF, amplitude of low-frequency fluctuation; HC, healthy controls; BA, Brodmann's area; MNI, Montreal Neurological Institute; IL, ipsilesional; CL, contralesional.

TABLE 3 | Brain regions with significant differences in dynamic ReHo between groups.


ReHo, regional homogeneity; HC, healthy controls; BA, Brodmann's area; MNI, Montreal Neurological Institute; IL, ipsilesional; CL, contralesional.

FIGURE 3 | (A) The dALFF variability in the SMA was positively correlated with the FMA scores of the stroke patients (r = 0.347, p = 0.035, uncorrected). (B) The dReHo variability in the ipsilesional MFG was negatively correlated with the FMA scores of the stroke patients (r = –0.462, p = 0.004, uncorrected). dALFF, dynamic amplitude of low-frequency fluctuation; dReHo, dynamic regional homogeneity; FMA-total, Fugl-Meyer assessment for upper and lower extremities; SMA, supplementary motor area; MFG, middle frontal gyrus; IL, ipsilesional.

### DISCUSSION

In the present study, dynamic regional brain activity between stroke patients with motor deficits and healthy controls was examined using resting-state fMRI. To the best of our knowledge, there is no prior study using a TDA approach to detect strokerelated brain activity changes in humans. Given that young adult stroke could be different both in causes and in outcomes (Edwards et al., 2018) and the role of educational level in recovery should not be ignored (Hillis and Tippett, 2014). Therefore, we used age and educational level as covariates in intergroup difference analyses and correlation analyses. Moreover, patients were enrolled in this current study within 1–3 weeks after stroke symptom onset, during which period imaging data and behavior performance were obtained. This period is well within the recovery window for stroke (Venketasubramanian et al., 2017).

healthy subjects. ALFF, amplitude of low-frequency fluctuation; ReHo, regional homogeneity; SMA, supplementary motor area; MFG, middle frontal gyrus; IL, ipsilesional; AUC, area under curve.

Additionally, medication, especially IVT therapy, can largely improve patients' functional outcomes for hyperacute ischemic stroke (Ferrigno et al., 2018). Hence, we also considered patients' illness duration and IVT use status as covariates in correlation analyses to reduce the confounding effects.

Differences in dALFF variability were observed between the stroke group and HC group in the contralesional CPL, ipsilesional superior parietal lobe, ipsilesional ITG, ipsilesional CPL, cerebellum tonsil, the midline SMA, ipsilesional putamen and lentiform nucleus, while differences of dReHo variability in contralesional rectal gyrus, contralesional ITG, contralesional pons and MFG distinguished stroke patients from HC. In addition, relationships were observed between the FMA scores and dynamic ALFF or ReHo variability in SMA or ipsilesional MFG in stroke patients. Further ROC analyses suggested that dynamic ALFF in SMA or ReHo in ipsilesional MFG had the potential to distinguish the patients of subacute stroke from healthy subjects.

Our findings were in line with previous studies that investigated stroke patients with movement disorders by functional connectivity, structural connectivity or regional metrics, such as ALFF or ReHo (Skidmore et al., 2013; Tsai et al., 2014; Zhu et al., 2015). The sensorimotor network (SMN), which is critical for voluntary movement, connects primary motor cortex function with SMA function (Damoiseaux et al., 2006; Chang et al., 2013). The SMA is a part of the primate cerebral cortex that contributes to the control of movement. Neurons in the SMA project directly to the spinal cord and may play a role in the direct control of movement (Nudo and Masterton, 1990). It has been reported that SMA plays roles in the postural stabilization of the body, the coordination of both sides of the body such as during bimanual action, the control of movements that are internally generated rather than triggered by sensory events, and the control of sequences of movements (Shima and Tanji, 1998; Cunnington et al., 2003; Zhang et al., 2012). Stroke-induced disturbance of intrinsic neural activity, which may be due to a complex cascade of events that are associated with structural reorganization and axonal sprouting as demonstrated by tract-tracing studies in animal models of stroke (Carmichael et al., 2001; Dancause et al., 2005), impedes brain network function of voluntary motion. Based on prior studies, we speculate that altered dALFF and dReHo variability in the SMN (e.g., in the superior parietal lobe and SMA) of stroke patients with motor deficits might occur as a compensatory mechanism and might be a significant factor in the reorganization and integration of resting-state functional networks at the subacute stage.

We observed decreased dReHo variability in the ipsilesional MFG, which belongs to the default mode network (DMN). This finding may indicate that stroke-related brain activity changes not only occurred in motor-related areas but also in non-motor regions. The DMN plays a pivotal role in "resting" brain activity, which is involved in sustaining attention, self-consciousness and exhibiting self-control (Andrews et al., 2011; Yuan et al., 2018). In stroke patients, alterations in brain activity of the DMN may be associated with advanced neural function of cognitive and emotional control (Liu F. et al., 2017). Evidence has shown that the DMN regulates consciousness, processes emotionally salient stimuli, and coordinates the interactions of cognitive function and emotional processing (Soddu et al., 2012; Zhang et al., 2017). In the current study, we demonstrated that the stroke group showed significantly decreased ReHo variability in the ipsilesional MFG, indicating that the synchronous neural activity was also disrupted in the ipsilesional MFG in stroke patients.

The increased regional brain activity in contralesional pons and bilateral cerebellum CPL, ipsilesional cerebellum tonsil may


<sup>∗</sup>By this optimal cut-off point, the dynamic ALFF value of SMA could correctly classify 40 of 42 patients and 50 of 55 healthy subjects, resulted in a sensitivity of 95.2% and a specificity of 90.9.2%. The means of other optimal cut-off point was similar. MNI, Montreal Neurological Institute; AUC, areas under the curves; dALFF, dynamic amplitude of low-frequency fluctuation; dReHo, dynamic regional homogeneity; SMA, supplementary motor area; MFG, middle frontal gyrus; IL, ipsilesional.

result from the dysfunction of the cerebro-ponto-cerebellar circuit and act as a compensatory response (Lu et al., 2011). The excessive ReHo variability of the prefrontal cortex (the rectal gyrus) may compensate for the deficits in motor function in stroke patients. However, this concept needs to be confirmed. Moreover, rectal gyrus is considered as a part of affective network (AN). Study has reported that abnormal regional neural activity was observed mainly in component of DMN, SMN, cerebellar lobes (CPL) and AN in stroke patients with emotional abnormality (Liu T. et al., 2017). Hence, we tend to speculate that the dysfunction of the volitional system may lead to the disorder of emotional system in stroke patients.

We also found decreased dALFF variability in the ipsilesional putamen and contralesional lentiform nucleus in stroke patients. Considered the entrance to the basal ganglia, the lentiform nucleus receives the input from cerebral cortex, and there was partial stroke lesion overlap observed in the present study and shown in **Figure 1**. The primary roles of these regions are to regulate movements at various stages, such as during motor preparation and motor execution. Meanwhile, they play important roles in motor learning, which can be considered a sensory feedback of frontal mediated goal parameters and posterior-mediated motor programs; there are anatomical links that exist between these motor-execution and frontal-parietal motor control systems (Ween, 2008). Previous human and nonhuman studies of stroke models have shown that decreased ALFF values were found in the core of stroke lesions (Liu et al., 2007; Skidmore et al., 2013). The mechanism may involve abrupt decreases of the blood flow resulting from cytotoxic swelling, calcium overload and membrane ion pump failure in core regions of stroke lesions (Badaut et al., 2011). Based on the previous studies and this present study, we infer that deceased or vanished oxygen consumption and intrinsic brain activity may contribute to less variation and low ALFF values.

In addition, we also detected excessive intrinsic brain activity variability reflected both in dALFF in the ipsilesional ITG and in dReHo in the contralesional ITG of the stroke patients. The ITG belongs to the higher levels of the ventral stream of visual processing and is related to the representation of complex object features. A previous study found that the visual network was activated in recovery from sensorimotor stroke, and limb movement critically relies on visual guidance (Archer et al., 2016). The mechanism underlying the abnormal brain activity of contralesional ITG is presently poorly understood.

The current study has several limitations. First, although we identified significant differences between the stroke patients and the HCs, the sample size of 42 stroke patients analyzed in the current study was somewhat lacking in statistical power, and large sample size studies are needed for further demonstration. Second, it was reported that corticospinal tract lesion load was a significant predictor of motor deficit (Zhu et al., 2010; Feng et al., 2015). However, in the current study, we focused on investigating whether subcortical stroke patients would exhibit abnormal dynamic characteristics of brain activity relative to healthy controls and to investigate whether the altered dynamic regional indexes were associated with clinical behavior in stroke patients. Hence, here, we did not give much thought to the characteristics of diffusion-tensor imaging of the patients. Third, the correlation analyses cannot pass the FDR or Bonferroni correction. Larger sample size will be necessary to confirm the current results in the future studies. Fourth, physiological noise of cardiac and respiratory cycles was not monitored during the MRI scanning, which may influence brain activity alterations. It is possible that alterations in network dynamics may reflect changes in brain state, since few constraints were imposed on a participant's cognitive processes during the scanning. Finally, although the correlation analyses between dynamic regional indexes in regions showing significant group differences and size of lesion volume revealed no significant correlation, the heterogeneous clinical characteristics, such as lesion location, stroke severity and size of lesion volume, exhibited a relatively large variation across subjects and should be taken into consideration when interpreting the results.

### CONCLUSION

The pattern of intrinsic brain activity variability in multiple brain networks is altered in stroke patients with motor deficits compared with healthy controls. The alterations of dynamic brain activity in the SMN and DMN were correlated with the degree of motor functional impairment. Resting-state fMRI dynamic regional indexes might be potential tools to assess stroke patients' motor function. Future studies will be needed to clarify the underlying mechanisms of alterations in the dynamic regional metrics after stroke.

### AUTHOR CONTRIBUTIONS

JC contributed to the experiments, data analysis and writing of the manuscript. DS contributed to performing the experiments, and writing and revising the manuscript. YS contributed to the data collection. WJ designed the experiments and revised the manuscript. YW contributed to the data analysis and manuscript revision. QX contributed to the manuscript revision. CR is the guarantor of this study and had complete access to all data in the study. All authors are accountable for the contents of this research.

### FUNDING

The study was supported by the National Natural Science Foundation of China (81571277).

### ACKNOWLEDGMENTS

The authors thank all volunteers who participated in the study and the staff of the Med-X Research Institute and School of Biomedical Engineering Shanghai Jiao Tong University in Shanghai, China for their selfless and valuable assistance.

### REFERENCES

fnins-12-00994 December 28, 2018 Time: 16:45 # 9


brief, whole-brain functional connectivity patterns. Proc. Natl. Acad. Sci. U.S.A. 112, 8762–8767. doi: 10.1073/pnas.1501242112


**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 Chen, Sun, Shi, Jin, Wang, Xi and Ren. 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.

fnins-12-00994 December 28, 2018 Time: 16:45 # 10

# Altered Brain Signal Variability in Patients With Generalized Anxiety Disorder

Liyuan Li 1,2, YiFeng Wang1,2, Liangkai Ye1,2, Wang Chen1,2, Xinju Huang1,2, Qian Cui 1,3 , Zongling He1,2,4, Dongfeng Liu1,2 and Huafu Chen1,2 \*

*<sup>1</sup> The Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, University of Electronic Science and Technology of China, Chengdu, China, <sup>2</sup> School of Life Science and Technology, Center for Information in BioMedicine, University of Electronic Science and Technology of China, Chengdu, China, <sup>3</sup> School of Public Affairs and Administration, University of Electronic Science and Technology of China, Chengdu, China, <sup>4</sup> Mental Health Center, The Fourth People's Hospital of Chengdu, Sichuan, China*

Generalized anxiety disorder (GAD) is characterized by a chronic, continuous symptom of worry and exaggerated startle response. Although functional abnormality in GAD has been widely studied using functional magnetic resonance imaging (fMRI), the dynamic signatures of GAD are not fully understood. As a vital index of brain function, brain signal variability (BSV) reflects the capacity of state transition of neural activities. In this study, we recruited 47 patients with GAD and 38 healthy controls (HCs) to investigate whether or not BSV is altered in patients with GAD by measuring the standard deviation of fMRI signal of each voxel. We found that patients with GAD exhibited decreased BSV in widespread regions including the visual network, sensorimotor network, frontoparietal network, limbic system, and thalamus, indicating an inflexible brain state transfer pattern in these systems. Furthermore, the correlation between BSV and trait anxiety score was prone to be positive in patients with GAD but negative in HCs. The opposite relationships between BSV and anxiety level in the two groups indicate that the brain with moderate anxiety level may stay in the most stable rather than in the flexible state. As the first study of BSV in GAD, we revealed extensively decreased BSV in patients with GAD similar to that in other mental disorders but with a non-linear relationship between BSV and anxiety level indicating a novel neurodynamic mechanism of the anxious brain.

### Edited by:

*Feng Liu, Tianjin Medical University General Hospital, China*

### Reviewed by:

*Feng Kong, Shaanxi Normal University, China Sufang Li, Johns Hopkins University, United States*

> \*Correspondence: *Huafu Chen chenhf@uestc.edu.cn*

#### Specialty section:

*This article was submitted to Neuroimaging and Stimulation, a section of the journal Frontiers in Psychiatry*

Received: *17 October 2018* Accepted: *06 February 2019* Published: *04 March 2019*

#### Citation:

*Li L, Wang Y, Ye L, Chen W, Huang X, Cui Q, He Z, Liu D and Chen H (2019) Altered Brain Signal Variability in Patients With Generalized Anxiety Disorder. Front. Psychiatry 10:84. doi: 10.3389/fpsyt.2019.00084* Keywords: brain signal variability, fMRI, generalized anxiety disorder, neurodynamics, non-linear relationship

### INTRODUCTION

Generalized anxiety disorder (GAD) is one of the most prevalent mental disorders characterized by exaggerated startle response and chronic, pervasive, and intrusive worry (1, 2). Based on static analysis methods, many functional magnetic resonance imaging (fMRI) studies have found aberrant brain activation related to cognition and emotion functions in GAD (3, 4). Inefficient intrinsic brain activity associated with integration of interoceptive and somesthetic functions has also been found in anxiety disorders (5–7). These static analysis methods for brain activity or activation have provided abundant evidence for us to understand the neural mechanism of GAD. However, clinically effective biomarker is still lacking. Recently, a mass of studies have shown that dynamic brain activity can provide novel information of neural characteristics for various neural

**17**

disorders (8–10). Whether the dynamic brain activity analysis can provide insightful information about the neural mechanism of GAD, however, is unknown.

Human brain activity is naturally variable (11). In previous years, fMRI research had regressed blood oxygen level-dependent (BOLD) signal variance as measurement-related or other confounds (12, 13). However, researchers found that the "noise" variance in data is an important feature of brain function in the recent 10 years (14, 15). In a neuroimaging time series, BSV measures the magnitude of variability from moment to moment (16). The forms of BSV include variance (17) and mean square successive differences (18), especially standard deviation (SD) (14, 15). As the next frontier in brain mapping, the brain signal variability (BSV) reflects the capacity of state transition of neural activities and dynamic range of brain functional systems (16). BSV has been suggested to be an excellent proxy of the characteristics of neural dynamics, cognitive performance, and brain disorders (14, 19–21). Great BSV has been suggested to be associated with increased ability to transfer between brain states (22) and to process varying and unexpected external stimuli (16, 23). Measured with the SD of brain signal, the BSV is more powerful than mean brain signal in predicting neural aging (14). Recently, the quadratic change in lifespan BSV trajectory has been further uncovered (24). Furthermore, a number of studies have demonstrated abnormal BSV in schizophrenia, attention deficit hyperactivity disorder, autism, and patients with disorders of consciousness, reflecting the dynamical dysfunction of neural activities in mental disorders (25, 26). Specifically, the non-linear dynamics of brain signal over a range of temporal scales are mainly decreased compared with those in healthy controls (HCs) (27). The hypothesis of "unhealthy brain is less variable than healthy brain" has been demonstrated in various clinical populations, such as dementia, untreated patients with schizophrenia, autism, and mesial temporal lobe epilepsy (16, 27, 28). In other words, many findings support Pool's opinion that "healthy brain is a chaotic brain" (11).

The anxious brain was viewed as an inflexible system, grounded in poor inhibition (29). In patients with GAD, a reduced capacity to inhibit cognitive (worry), behavioral (avoidance), and accompanying physiological manifestations was associated with cognitive rigidity and inflexibility (30). By using the mean-based methods, the core symptom, worry, which predominantly reflected a stimulus-independent mental processing also leads to the inflexible functional brain configurations in the prefrontal cortex, cingulate gyrus, and amygdala (31). However, the respective spatial patterns profiled by the SD-based method (like BSV) and mean-based method were highly different (14). To profile the flexibility of brain, BSV is a brand new effective index. Therefore, we hypothesized that patients with GAD may show decreased BSV compared with HCs in more related core regions.

In this study, we investigated the altered BSV in patients with GAD and healthy participants. To perform comprehensive comparisons, the relationship between BSV and anxiety level was also observed. Notably, this work is the first study of altered BSV in patients with GAD. Therefore, the BSV may be able to provide new insights into understanding the neural dynamics of GAD.

### METHODS AND MATERIALS

### Participants

Forty-seven patients with GAD were recruited from the mental health center of Chengdu, Sichuan, China. All patients were determined by consensus of two experienced psychiatrists by using the Structured Clinical Interview for DSM-IV (patient edition) (32). Clinical states of the patients were evaluated using the Hamilton anxiety scale (HAMA). Data from one patient was deleted because the BSV is extremely lower than the others (< mean-5 SD). The exclusion criteria included schizophrenia, mental retardation, or personality disorder, history of loss of consciousness, substance abuse, and serious medical or neurological illness. The HC group was composed of 38 age-, gender-, education-, mean frame-wise displacement (FD) (33) matched healthy participants. The Structured Clinical Interview for DSM-IV non-patient version was employed to ensure the absence of psychiatric illnesses in the HCs. The HCs did not finish the HAMA scale test because the test was only obtained in hospital by two well-trained psychiatrists. None of the HCs had a history of serious medical or neuropsychiatric illness or a family history of major psychiatric or neurological illness in their first-degree relatives. All participants (including GADs and HCs) finished the Chinese vision of Trait Anxiety Inventory (TAI) questionnaire (34), which is often used in clinical application in the diagnostic work-up of mental disorders and has good validity and reliability (35, 36).

Ultimately, 47 patients with GAD and 38 HC were included in the study (**Table 1**). Written informed consent, approved by the research ethical committee of the School of Life Science and Technology at University of Electronic Science and Technology of China, was obtained from each participant.

### Data Acquisition

MRI data were acquired using a 3.0T GE 750 scanner (General Electric, Fairfield, Connecticut, USA) equipped with high-speed gradients. An 8-channel prototype quadrature birdcage head coil fitted with foam padding was applied to minimize the head motion. Ear plugs were used to minimize the scanner noise. Participants were instructed to simply rest with their eyes closed, minds relaxed, awake, and motionless. Functional images were acquired using a gradient-recalled echo-planar imaging (EPI) sequence. The parameters were as follows: repetition time/echo time = 2,000 ms/30 ms, 90◦ flip angle, bandwidth = 250 Hz/pixel, 43 axial slices (3.2 mm slice thickness without gap), 64 × 64 matrix, and 22 cm field of view. For each participant, 255 volumes were obtained.

### Data Preprocessing

Functional images were preprocessed using the Data Processing Assistant for resting-state fMRI (DPARSF 2.2, http://www. restfmri.net/forum/DPARSF). The first five volumes were discarded to ensure signal equilibrium and for the participants to familiarize themselves with the scanning environment. The



*SD, standard deviation; GAD, generalized anxiety disorder; HCs, healthy controls; HAMA, Hamilton Anxiety Rating Scale; TAI, Spielberger Trait Anxiety Inventory. a Chi-square test; b Independent-sample t-test.*

remaining 250 images were slice-time corrected, spatially aligned, and spatially normalized to the Montreal Neurological Institute (MNI) EPI template and resampled to 3 × 3 × 3 mm<sup>3</sup> voxels. After signal detrending, the images were spatially smoothed (8 mm full width at half maximum Gaussian kernel). Afterwards, Friston 24 head motion parameters, white matter signal, and cerebrospinal fluid signal were further regressed out. Finally, signal was filtered at 0.04–0.07 Hz because of the less noise contamination within this frequency range (37). The frame-wise displacement (FD) was used to represent instantaneous head motion. The mean FD of each participant was <0.5 mm.

### Brain Signal Variability

Firstly, before calculating the BSV, we performed the temporal normalization for each voxel during the entire time of 500 s. The purpose of this step is to eliminate the contamination of the mean signal (14). Secondly, the SD which is simply the square root transformation of variance, was calculated in each voxel cross time series by using a custom-built function in MATLAB (The MathWorks, Inc.). According to Garrett's study, the SD of BOLD signal is temporal variability called BSV (14). The Anatomical Automatic Labeling 90 template was transformed to a binary mask and used to constrain the calculation in the gray matter (21).

### Statistical Analysis

Two-sample t-test was used to assess the difference of BSV between the GAD and HC groups. Multiple-comparison correction was performed on the contrast brain map via the false discovery rate approach (p < 0.05). Pearson correlations were calculated between the BSV and TAI score in the two groups, correspondingly, with age, sex, education, and mean FD as control variables. The correlation analysis was performed under the regions with significantly different BSVs between the two groups. We adopted the cross-voxel correlation (21, 38) to evaluate the spatial correlation between the correlation maps of the GAD and HC groups. As suggested by Liang et al., the two 3D maps of correlation coefficient were first transformed into columns and then transformed into z-score by minus mean then divided by SD. Pearson correlation was finally computed between these two columns of data. Then, we built a linear mixed-effects model with State-Trait Anxiety Inventory (STAI) score as a factor to show the differences in these correlations between BSV and STAI score in the two groups.

### RESULTS

### Decreased BSV in Patients With GAD

Compared with the HC group, patients with GAD show decreased BSV across the widespread brain regions (**Figure 1A**). Significantly low BSV is primarily located in the visual network, sensorimotor network, frontoparietal network, limbic system, and thalamus (**Figure 1B** and **Table 2**).

### Correlations Between BSV and Anxiety Severity in GAD and HC

As shown in **Figure 2A**, the correlation between BSV and TAI score in the HC group is mainly negative, whereas that in the GAD group is mainly positive. **Figures 2B,C** show the correlation map for the HC and GAD groups, respectively. The correlation coefficients in the HC group and GAD group are negatively related to each other (r = −0.285, p = 4.12e−274). In other words, if the correlation between BSV and TAI score in the HC group is negative in one voxel, that correlation in the GAD group is positive in that voxel. The differences of correlations between BSV and STAI score in the two groups are also mainly located in the visual network, sensorimotor network, and frontoparietal network (**Figure 3**).

### DISCUSSION

To the best of our knowledge, this work is the first study on the abnormal BSV in patients with GAD. In accordance with other studies in mental disorders, we observed widely decreased BSV in patients with GAD compared with HCs, suggesting multiple deficits in neural systems of GAD. Interestingly, we observed opposite distributions of linear relationship between TAI score and BSV in the GAD and HC groups, indicating a non-linear relationship between anxiety level and BSV. In other words, the chaotic brain might be not always the healthy one.

### Decreased Temporal Variability in Patients With GAD

In line with reduced BSV in other mental disorders, the current study revealed widely decreased BSV in patients with GAD, arguing that reduced BSV is a general characteristic of mental



*BA, Brodmann's area; MNI, Montreal Neurological Institute; X, Y, Z, coordinates of primary peak locations in the MNI space; T statistical value of peak voxel showing BSV decreased. All the clusters survived p* < *0.05, FDR corrected, and a minimum cluster size of 50 voxels. Degree of freedom*= *78.*

disorders. Recent studies have shown that low amplitude of low-frequency fluctuation (ALFF) in the left supplementary motor area, right middle occipital gyrus, cerebellum, prefrontallimbic system, and thalamus is associated with high-trait anxiety (32, 39, 40). Because ALFF is the square root of power (41) and power is equivalent to BSV (21). The BSV may reflect similar neural mechanisms to ALFF but in a nonlinear way (21). Specifically, decreased BSV is associated with small potential kinetic energy to handle external environmental demands (16, 22) and low ability to transfer between different brain states (i.e., rest state and task state) (42). In patients with anxiety disorder, the functional deficit of environmental detection and the inflexible functional brain configurations has been found (31, 43). Therefore, the widely decreased BSV in patients with GAD may be associated with limited ability to adapt different external conditions and inflexible pattern of information transfer.

Low BSV in GAD is mainly located in the visual cortex, somatosensory cortex, anterior cingulate cortex (ACC), and thalamus. Multiple studies have documented deficits in the visual cortex in different anxiety disorders (44–46). Visual cortex plays a core role in the social brain network during visual and emotion processing (47). Visual cortex is also involved in the processing of fear generalization (48), which may greatly contribute to the GAD psychopathology (49, 50). The postcentral gyrus, as a critical substrate of interoceptive processing (51), is responsible for receiving, integrating, and interpreting most of the sensory information transmitted by the thalamus (52, 53). As shown by many studies, the thalamus is also strongly involved in interoceptive awareness (54–57). Abnormal connectivity between the postcentral gyrus and thalamus has been found in panic disorder, which may be associated with the typical symptoms (e.g., the extreme feeling of heartbeat) of panic disorder

respectively; (B) The r map of HC group; (C) The r map of GAD group. L: left; R: right.

(6). In general, as a hub of the interoceptive network, the ACC participates in mediating visceromotor activity and has projections into motor systems (56). Decreased functional connections between the ACC and thalamus was negatively correlated with HAMA in GAD, which may cause some somatic disturbed symptoms, such as rapid heart rate, low skin conductance, and difficulty breathing (58). Above all, the reduction of variability in these brain areas may be related to biased perception in processing exogenous and endogenous information (6, 59), which is implicated in the GAD pathological mechanism.

Reduced BSV in patients with GAD is also found in the dorsolateral prefrontal cortex and inferior parietal lobe, which are the key brain regions of the frontoparietal network (FPN). The deficiencies in FPN have been found in individuals with high-trait anxiety who have poor cognitive control and are easily distracted by emotional distractors from external environment (60, 61). Individuals with high anxiety generally require additional attentional control even in the absence of threat-related stimuli (62). The FPN is also a flexible hub during brain state transition (63). Low variability in the FPN may be associated with low ability to transfer between different brain states (42). Therefore, low BSV in the FPN may be associated with inflexible information to transform patterns facing the external world in patients with GAD.

### Opposite Relationships Between BSV and Anxiety in the GAD and HC Groups

The opposite correlations between the BSV and TAI score in the HC group and in the GAD group has been found in the present study, indicating that the moderate anxious brain has small BSV. Moderate anxiety is associated with the best performance (64–66), and BSV has been demonstrated to be closely related to behavior performance (19, 24). In certain brain regions, maintaining great signal stability is also a critical marker of good task performance (22). In other words, the direction of correlation between BSV and behavioral performance depends on both cognition and brain region (21, 24, 67). Therefore, the opposite relationships between BSV and anxiety level may indicate that (1) moderate anxiety is associated with the most stable brain state and the best performance, and (2) increased BSV in subjects with high anxiety and low anxiety may be related to distinctive brain states, leading to different performances. Considering that (1) the dopaminergic system is closely associated with both anxiety and BSV (68, 69), (2) the non-linear relationship between dopaminergic system and behavioral performance (70), and (3) the quadratic relationship between anxiety and behavioral performance (66), the opposite BSV-anxiety relationship may reflect different neural dynamical configurations mediated by the dopaminergic system in patients with GAD and healthy people. Furthermore, the negative correlation between the HC group and GAD group suggests that the neuropsychological association in HC and patients with GAD may be driven by the same mechanism rather than independent from each other. This hypothesis deserves further investigations.

Although increased anxious levels in GADs showed increased BSV which indicated the great capacity of detection, the mechanism of this seemingly increasing ability can be totally different from healthy participants. In patients with GAD, the excessive activation is an apparent clinical symptom (1, 2), which might be associated with the high detection. In addition, this conclusion still needs further research. Meanwhile, the great BSV in some regions can be a compensation mechanism of an inefficient brain (14).

Overall, the opposite correlations between the BSV and TAI score in two groups indicate close relationships among dopaminergic system, behavioral performance, anxiety level, and BSV. These relationships can be described by the U law. Therefore, BSV is not always linearly related to brain health (24). These results provide valuable insights into understanding the relationship between BSV and health.

### LIMITATIONS

Some limitations remain in this study. First, the sample is relatively small (n=85), impeding the uncovering of a strong correlation between BSV and anxiety level. Second, the HAMA score is lacking in the HC group because these participants were tested at the university rather than at the hospital, hindering the full investigation of clinical relevance of BSV. Third, some undiagnosed comorbid psychological disorders may exist in this sample, which may mix the true relationship between BSV and anxiety. A large sample with accurate diagnosis and complete scale collection is needed in future studies.

### CONCLUSION

In summary, the decreased BSV in patients with GAD and different neuropsychological relationships in patients with GAD and HCs may reveal a novel neurodynamic mechanism, suggesting that the chaotic brain is not always the healthy one.

### DATA AVAILABILITY

The datasets generated for this study are available on request to the corresponding author.

### ETHICS STATEMENT

All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975 and the applicable revisions at the time of investigation. Informed consent was obtained from all patients included in the study.

### AUTHOR CONTRIBUTIONS

LL, YW, and HC designed the study. LY, WC, DL, and HX wrote the computing scripts. QC and ZH interviewed all patients by using the DSM-IV. LL managed the literature searches and analyses, acquired, and analyzed the data, and wrote the first manuscript draft. All authors contributed and have approved the final manuscript. The authors thank all subjects participating in this study.

### FUNDING

This work was supported by the National Natural Science Foundation of China (31600930, 81771919, and 61533006), the Fundamental Research Funds for the Central Universities (ZYGX2016KYQD120), the Sichuan Science and Technology Program (2018TJPT0016), the Science Foundation of Ministry of Education of China (14XJC190003), the Scientific Research Project of Sichuan Medical Association (S15012), and the Youth Innovation Project of Sichuan Provincial Medical Association (Q14014).

### ACKNOWLEDGMENTS

The authors also thank the two reviewers for their suggestions and comments.

### REFERENCES


**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 Li, Wang, Ye, Chen, Huang, Cui, He, Liu and Chen. 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.

# Reduced Dynamic Interactions Within Intrinsic Functional Brain Networks in Early Blind Patients

Xianglin Li1,2, Ailing Wang<sup>3</sup> , Junhai Xu<sup>4</sup> , Zhenbo Sun<sup>2</sup> , Jikai Xia<sup>5</sup> , Peiyuan Wang<sup>5</sup> , Bin Wang<sup>2</sup> \*, Ming Zhang<sup>1</sup> \* and Jie Tian1,6 \*

<sup>1</sup> Department of Medical Imaging, The First Affiliated Hospital of Xi'an Jiaotong University, Xi'an, China, <sup>2</sup> Medical Imaging Research Institute, Binzhou Medical University, Yantai, China, <sup>3</sup> Department of Clinical Laboratory, Yantai Affiliated Hospital of Binzhou Medical University, Yantai, China, <sup>4</sup> Tianjin Key Laboratory of Cognitive Computing and Application, School of Artificial Intelligence, College of Intelligence and Computing, Tianjin University, Tianjin, China, <sup>5</sup> Department of Radiology, Yantai Affiliated Hospital of Binzhou Medical University, Yantai, China, <sup>6</sup> School of Life Sciences and Technology, Xidian University, Xi'an, China

#### Edited by:

Feng Liu, Tianjin Medical University General Hospital, China

#### Reviewed by:

Yifeng Wang, University of Electronic Science and Technology of China, China Mengting Liu, University of Southern California, United States

\*Correspondence:

Bin Wang binwang001@aliyun.com Ming Zhang zhangming01@mail.xjtu.edu.cn Jie Tian jie.tian@ia.ac.cn

#### Specialty section:

This article was submitted to Brain Imaging Methods, a section of the journal Frontiers in Neuroscience

Received: 28 December 2018 Accepted: 07 March 2019 Published: 22 March 2019

#### Citation:

Li X, Wang A, Xu J, Sun Z, Xia J, Wang P, Wang B, Zhang M and Tian J (2019) Reduced Dynamic Interactions Within Intrinsic Functional Brain Networks in Early Blind Patients. Front. Neurosci. 13:268. doi: 10.3389/fnins.2019.00268 Neuroimaging studies in early blind (EB) patients have shown altered connections or brain networks. However, it remains unclear how the causal relationships are disrupted within intrinsic brain networks. In our study, we used spectral dynamic causal modeling (DCM) to estimate the causal interactions using resting-state data in a group of 20 EB patients and 20 healthy controls (HC). Coupling parameters in specific regions were estimated, including the medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), and inferior parietal lobule (IPC) in the default mode network (DMN); dorsal anterior cingulate cortex (dACC) and bilateral anterior insulae (AI) in the salience network (SN), and bilateral frontal eye fields (FEF) and superior parietal lobes (SPL) within the dorsal attention network (DAN). Statistical analyses found that all endogenous connections and the connections from the mPFC to bilateral IPCs in EB patients were significantly reduced within the DMN, and the effective connectivity from the PCC and lIPC to the mPFC, and from the mPFC to the PCC were enhanced. For the SN, all significant connections in EB patients were significantly decreased, except the intrinsic right AI connections. Within the DAN, more significant effective connections were observed to be reduced between the EB and HC groups, while only the connections from the right SPL to the left SPL and the intrinsic connection in the left SPL were significantly enhanced. Furthermore, discovery of more decreased effective connections in the EB subjects suggested that the disrupted causal interactions between specific regions are responsive to the compensatory brain plasticity in early deprivation.

Keywords: early blindness, effective connectivity, spectral dynamic causal modeling, intrinsic brain networks, brain plasticity

### INTRODUCTION

The recognition of objects could be manipulated through the coordinated cross-modal interactions of different modalities, such as vision, touch and audition (Amedi et al., 2005; Dormal et al., 2018). Deprivation of one sensory modality could give us a chance to explore the plasticity changes of the cognitive functions (Jiang et al., 2015). Two hypotheses have been raised to explain the plasticity in early blindness. The first hypothesis indicated a kind of maladjustment caused by the

early blindness. When the sensory information is manipulated, there is a decrease in processing capacities by the early visual deprivation (Jiang et al., 2015). The other hypothesis is based on the compensatory explanation, in which the blind patients exhibit a superior ability in retained sensory modalities (Pascual-Leone et al., 2005). The early visual deprivation has been demonstrated to cause the structural and functional remodeling or reorganization in both intact and deprived sensory cortices (Bauer et al., 2017, 2018; Hou et al., 2017). Furthermore, impaired cognitive performances have been discovery in specific brain regions or networks (Dormal et al., 2018; Manescu et al., 2018; Vercillo et al., 2018). The baseline metabolism and blood flow in specific cortices have been suggested to show a significant increase in early blind (EB) patients (Viski et al., 2016). Functional correlations were observed between the subregions of the visual cortex in the retinotopic pattern in EB subjects at rest (Bock et al., 2015). Furthermore, EB patients also exhibit disrupted functional connections and stronger parietal and auditory networks compared with sighted subjects (Boldt et al., 2014; Hou et al., 2017; Abboud and Cohen, 2019). Therefore, EB patients showed a compensatory pattern in the primary sensory networks, while the other brain networks related to cognition might be disrupted due to the early deprivation.

The resting-state technique on functional magnetic resonance imaging (fMRI) has been proved to be powerful in revealing the abnormalities of intrinsic functional connections. The default mode network (DMN) is the most prominent system in the resting state (Mccormick and Telzer, 2018; Prestel et al., 2018), which consists of the medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), and bilateral inferior parietal lobule (IPL), and some brain areas in the temporal lobe (Sharaev et al., 2016). By applying advanced brain network analyses to the resting-state data, abnormalities of the functional integration in EB patients have been examined in many studies, reflecting the statistical dependencies between distant brain regions (Heine et al., 2015; Sabbah et al., 2016). However, correlation parameters can only calculate the statistical dependency between two regions, while the causal interactions cannot be estimated. Moreover, DMN regions have been observed to be hyperactivated in many cognitive tasks, like autobiographical information retrieval, mind-wandering, emotional processing, and spontaneous cognition (Turkheimer et al., 2015; Hu et al., 2017).

In addition to the DMN, two other distinct brain networks (ventral and dorsal) play a vital role in EB subjects in the allocation of attention (Jimenez et al., 2016). The ventral network consists of the temporoparietal junction, anterior cingulate cortex (ACC), and anterior insulae (AI), which has been suggested to be responsible for stimulus-driven attention (Jimenez et al., 2016; Liu et al., 2017; Wang et al., 2018). The other dorsal network, constituted of the lateral frontal eye field (FEF) as well as the bilateral superior parietal lobes (SPL), is correlated with the voluntary, sustained orienting of attention (Corbetta and Shulman, 2011; Tsvetanov et al., 2016). Recent studies have highlighted the dysfunction in the ventral network in subjects with disorders, but not in the dorsal network (White et al., 2013; Wynn et al., 2015). The ventral attention network can also be called the "salience network" (SN). Pathophysiological studies have demonstrated that EB subjects with dysfunctions could result in the incorrect assigning of salience (Palaniyappan and Liddle, 2012). Therefore, the application of the resting-state fMRI approach can be potentially powerful in the assessment of attentional deficits in EB patients when we intend to examine the relative contributions of these two networks.

Dynamic causal modeling (DCM) is an appropriate approach to examine the causal influence, which is designed to calculate the effective connectivity changes underlying human responses based on Bayesian analysis. When one specific model is constructed (including specific brain regions, directed connections, and modulations), the parameters are estimated based on the observed data and model structures. Dynamic causal modeling is model-based, and shows great advantages compared to the functional connectivity or data-driven effective connectivity like the Granger causality, as DCM allows different hypotheses to be tested, which could capture the functional brain architectures corresponding to a specific hypothesis (Friston et al., 2017). Many studies have been reported by applying DCM to fMRI data and Magnetoencephalography/electroencephalography (MEG/EEG) (Chahine et al., 2017; Yang et al., 2017; He and Johnson, 2018; Van De Steen et al., 2019). Recently, a new version of DCMspectral DCM was developed to estimate the intrinsic effective connections from resting-state data using the cross spectra of the signals through a deterministic model (Friston et al., 2014). The cross-spectra can be considered as a more complete measure of functional connections (Park et al., 2018). Spectral DCM renders the model essentially in a deterministic way to get rid of the heavy load in estimating the random fluctuations in neural states. Therefore, spectral DCM makes DCM in restingstate fMRI slightly simpler and does not require a bilinear term accounting for condition-specific effects (Razi et al., 2015). Furthermore, spectral DCM is intended to simply compare the endogenous coupling between different groups (healthy controls vs. patients), so it only estimates the time-invariant parameters of the cross spectra. As the frequency domain is used in estimating the effective connectivity, the high computational efficiency and stable estimation make spectral DCM significantly powerful to compare the couplings and directionality in the intrinsic brain networks between groups of subjects, like patients and healthy cohorts (Li et al., 2017). Therefore, this is the reason for choosing the spectral DCM approach in this study.

Previous studies have suggested that the EB subjects showed a disruption of functional connectivity both in the visual cortex and other brain regions located in other cognitive brain networks (Liu et al., 2017; Dormal et al., 2018; Abboud and Cohen, 2019). Early deprivation could trigger the functional preference for the selective auditory recruitment in the cognitive brain networks (Dormal et al., 2018) and a reduction of the interhemispheric functional connectivity in the cognitive regions (Hou et al., 2017). Although the disruption of the functional interactions among the regions from the visual cortex and the cognitive brain networks, like the DMN and SN, it remains unclear how the directed interactions between these regions changed in the EB subjects. The purpose of our study is to examine the changes in the

effective connectivity within three high-order cognitive networks between EB patients and healthy controls (HC) by estimating the corresponding parameters with a newly developed spectral DCM. In this study, we had one hypothesis: early deprivation could affect the dynamic interactions between regions within three brain networks. Furthermore, the strength of the effective connectivity between regions in all three networks might be decreased due to the early deprivation. The spectral DCM was used to explore the causal neuronal influences within the intrinsic functional networks (DMN, SN and DAN) in a sample of 20 EB patients and 20 matched HC. Firstly, an independent component analysis was made to the resting-state functional images to generate the DMN, SN and DAN components. Then, the spectral DCM was employed to compute the effective connectivity within DMN, SN, and DAN of EB patients and HC. A Bayesian model selection procedure was adopted to determine the optimal DCM model at the group level. Statistical analyses were further conducted to compare the differences in the effective connectivity parameters within DMN, SN, and DAN between EB and HC subjects.

### MATERIALS AND METHODS

### Participants

Twenty EB subjects and 20 sighted subjects participated in this study. The mean age was 22.3 ± 1.4 years for the blind group, and 20.7 ± 1.2 years for the sighted group; there was no significant difference between both groups. The blindness of these subjects was evaluated and diagnosed by two professional ophthalmologists from Yantai Affiliated Hospital of Binzhou Medical University based on the retinal pathology. All the EB subjects had the same categories of blindness, mainly caused by cataracts and retinal pigment degeneration (Cataract: 8; Retinal pigment degeneration: 12). **Table 1** shows the detailed information for all EB subjects. All subjects were right-handed, and no subject had neurological problems except for the visual deprivation. After hearing a detailed explanation on the study, all subjects gave written informed consent. This study was carried out in accordance with the recommendations of Institutional Review Board of Binzhou Medical University. The protocol was approved by the Institutional Review Board of Binzhou Medical University. All subjects gave written informed consent in accordance with the Declaration of Helsinki after hearing detailed explanation about the study.

TABLE 1 | Demographic information for the early blind (EB) groups.


Onset of visual deprivation means the age of the EB subject sight loss.

### Data Acquisition

The functional images were scanned using a 3.0 T Siemens Skyra scanner with a 32-channel head coil at Yantai Affiliated Hospital of Binzhou Medical University. Participants were instructed to relax and keep their eyes closed without thinking about anything in particular. A high-resolution structural image was collected using a T1 weighted 3D MPRAGE sequence (repetition time (TR) = 1900 ms, echo time (TE) = 2.52 ms, inversion time (TI) = 1100 ms, voxel size = 1 × 1 × 1 mm<sup>3</sup> , matrix size = 256 × 256, flip angle (FA) = 90◦ ). A gradient-echo planar imaging (EPI) sequence (TR = 2000 ms, TE = 30 ms, voxel size = 3.1 × 3.1 × 4.0 mm<sup>3</sup> , matrix size = 64 × 64, slices = 33, slices thickness = 4 mm, slices gap = 0.6 mm, FA = 9◦ ) was used for functional data collection. Earplugs and foam pads were used to reduce the scanner noise and head motion.

### Data Preprocessing

SPM12 was used for the data preprocessing<sup>1</sup> . For each participant, the first 10 functional images were first removed to allow for participants' adaptation to the environment and equilibration effects. The remaining data were processed by a slice timing correction and head motion correction by a realign analysis. No participant was excluded, as all participants' head movements were not larger than 1.5 mm and 1 degree. The high-resolution structural image was coregistered with the functional images and subsequently segmented into the gray matter, white matter and cerebrospinal fluid (CSF). The generated spatial parameters from the segmentation procedure were then applied to spatially normalize the realigned images to 3 × 3 × 3 mm<sup>3</sup> in the Montreal Neurological Institute (MNI) space. Finally, the functional images were smoothed with a full-width at halfmaximum Gaussian filter (FWHM = 4 mm) to attenuate spatial noise. Several variances were regressed out with the temporal derivatives with the linear regression, including six head motion parameters and averaged signals from the CSF and white matter. Specifically, the global signal regression was not performed in this study, as global signals are thought to be irrelevant to non-neural noise and should not be regressed out (Chen et al., 2012).

### Group Independent Component Analysis (Group ICA)

To extract the regions for the subsequent DCM analysis, the DMN, SN, and DAN components were first identified for each subject. Group ICA was performed to decompose the resting state functional images into spatially independent components (ICs) using the GIFT software<sup>2</sup> after data preprocessing. First, the number of the optimal components were estimated using the minimum description length criteria (MDL) (Li et al., 2007). The number of components in our experiment were 41. Therefore 41 independent spatial components were produced using the infomax algorithm. For each participant, the corresponding spatial maps were generated through a back-reconstruction step.

The DMN, SN, and DAN templates were generated by the WFU PickAtlas Tool 3.0 to identify the DMN, SN, and DAN

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

<sup>2</sup>http://icatb.sourceforge.net/

components (Calhoun et al., 2008). In our study, we generated the DMN template by including the posterior cingulate cortex, bilateral precuneus, superior medial frontal cortex, and inferior parietal lobe. A multiple spatial regression analysis between the template and ICs was then conducted to sort all the spatial components. By comparing the regression coefficients, the component that fit best with the template with the greatest one was identified as the DMN component. The one-sample t-test was used to generate a group-mean pattern (p < 0.001, uncorrected, cluster size = 10 voxels). For the SN and DAN components, the same procedure was performed by creating the SN and DAN templates. To keep the consistency of the locations of each ROI between two groups, the Euclidean distance was also calculated between the individual ROI and group-mean ROI.

### Definition of Region of Interest (ROI)

For each subject, all the specific ROIs were selected based on the spatial map of the individual component representing three brain networks. Based on the regions identified above, four ROIs were defined for DMN: the mPFC, PCC, and bilateral parietal cortex (lIPC and rIPC). We defined three regions: dorsal cingulate cortex (dACC), left anterior (lAI), and right anterior insulae (rAI) for the salience network. Additionally, the DAN included the bilateral frontal eye fields (lFEF, rFEF) and bilateral superior parietal lobes (lSPL, rSPL). The detailed information of ROIs is shown in **Table 2**.

All the subject-specific ROIs were defined as spheres with a radius of 8 mm centered at the local maximum from the identified DMN, DAN, and SN components for each subject. We also included a white matter mask to remove the influence of the white matter. This procedure would ensure the further DCM analysis was conducted on the consistent regions identified as one functionally connected network for each subject. Finally, time series were extracted from all ROIs as the residuals of the general

TABLE 2 | Coordinates of the regions of interest used in spectral dynamic causal modeling (DCM) analysis.


mPFC, medial prefrontal cortex; IPC, inferior parietal lobule; PCC, posterior cingulate cortex; dACC, dorsal cingulate cortex; AI, anterior insulae; FEF, frontal eye field; SPL, superior parietal lobe; DMN, default mode network; SN, salience network; DAN, dorsal attention network; l, left; r, right.

linear model, which was constructed by the following regressors: the head motion parameters, cosine basis functions to model the aliased respiratory and cardiac signals, one constant regressor to model the baseline and a high-pass filter of 1/128 Hz to regress the ultraslow fluctuations.

## Spectral Dynamic Causal Modeling

To estimate the effective connectivity parameters in specific regions, the spectral DCM analysis was conducted for the defined nodes of all three networks using DCM12 in SPM12. The spectral DCM is designed to estimate the intrinsic effective connectivity from resting state fMRI images with the crossspectra of the signals. The cross spectral in the frequency domain could be considered as a more complete measure of functional connectivity (Park et al., 2018). In theory, the spectral DCM is distinct from stochastic DCM, which can estimate the coupling parameters among coupled populations of neurons in the frequency domain. It calculates the spectral measure by using a neuronally plausible power-law model among measured responses, and the time-invariant parameters. It models the observed functional connectivity between nodes with their second-order statistics instead of the neural signals under a deterministic assumption. In other words, spectral DCM could estimate the covariance of the random fluctuations from previous experiences to create the complex cross spectra among measured responses (Razi et al., 2015).

The expected cross spectra can be generated by the following model:

$$\lg\_\mathbb{Y}(\omega,\theta) = |K(\omega)|^2 \lg\_\mathbb{V}(\omega,\theta) + \lg\_\mathbb{e}(\omega,\theta)$$

Here, gy(ω, θ) represents the predicted cross spectra that can be estimated. K(ω) is a function of the effective connection, which is the Fourier transform of the system's Volterra kernel.

For each participant, a fully connected model was constructed with reciprocal connectivity between any pair of all ROIs by a Bayesian network discovery analysis (Friston et al., 2011). For resting-state fMRI data, no exogenous input was added to the model construction, then we made a parameter estimation for all models. Unlike the stochastic DCM, the convolution kernel representation in the spectral DCM was transformed to a spectral representation in the frequency domain. An estimation procedure was applied to characterize the spectral densities over frequencies for all estimated parameters within the DMN, SN, and DAN. Bayesian model selection (BMS) was finally performed with a post hoc optimization method to estimate the optimal model with the best balance between complexity and accuracy. The BMS was conducted on both groups separately, as we assumed both groups did not share the same model structure. The corresponding parameters for the best model were also estimated.

### Statistical Analysis

After the parameter estimation was completed for the optimal model, one-sample t-tests were conducted to examine whether the effective connectivity between regions was significant for both EB and HC groups. Multiple comparison corrections were made on the results of one-sample t-tests at p < 0.05 by

applying the false discovery rate (FDR) procedure. To examine the abnormalities of the effective connectivity within all three networks between EB and HC cohorts, the coupling parameters from both groups were tested using the two-sample t-tests. A correction for FDR was used to determine the results of two-sample t-test at p < 0.05.

### RESULTS

### Independent Component Analysis

We first decomposed the separated spatial patterns by conducting the group ICA on the resting-state functional images. Fortyone independent components were generated according to the MDL criterion, and the generated components were sorted by their correlations with the DMN, SN, and DAN templates. The independent components that showed a best fit with all three templates were identified as the DMN, SN, and DAN components. Subject-specific nodes were defined as 8-mm spheres centered at the peak values of all ROIs for each subject. **Figure 1** shows the profiles of the DMN, SN, and DAN components. Based on the group ICA analysis, four ROIs were defined for DMN, including the mPFC, PCC, bilateral IPC; three ROIs for SN: dACC, lAI, and rAI, and four bilateral ROIs for DAN: lEFF, rFEF, lSPL, and rSPL. The coordinates of all selected ROIs are shown in **Table 2**. The statistical analysis on the Euclidean distance suggested there was no significant difference in the Euclidean distance between two groups, indicating the consistency of the locations of the subject-specific ROI.

## DCM Analysis and Effective Connectivity of DMN, SN, and DAN

The time series of all ROIs were first extracted for each subject. For illustration, we extracted the principal eigenvariates from the corresponding time series of the selected regions within the DMN, as shown in **Figure 2**. The generalized fitting procedure

FIGURE 1 | Functional brain networks chosen from the independent component analysis. (A) represents the default mode network (DMN); (B) represents the salience network; (C) represents the dorsal attention network. The red color regions correspond to the threshold z-value.

suggested that the observed spontaneous fMRI signals can be modeled by the conditional expectations in the neuronal activations. For each participant, a Bayesian model reduction procedure was used to search for the optimal model from the model space. The detailed results of the estimation are shown in **Figure 3**. Finally, 6536 models were constructed with all possible combinations using four ROIs. The log-evidence of all models for DMN is shown in **Figure 3A**, suggesting a model with more connections shared larger evidence. In this study, we found that the fully connected model was optimal with the most significant efficiency. The meant that our resting-state fMRI data could be stimulated by the fully connected model. There was a deduction of the log-evidence when any connection was removed.

The same DCM analyses were conducted for SN and DAN. As expected for DMN, the best models were the fully connected models for SN and DAN. In the EB and HC groups, the coupling parameters of the optimal model for each network were calculated by the Bayes model average analysis. **Figures 3B–D** shows the estimated fixed effects and the posterior probabilities for DMN, SN, and DAN in one typical subject. The Bayes model average parameters were then estimated for all subjects, as shown in **Figure 4**. Most of the coupling parameters within the DMN for EB and HC groups were observed to be significant (p < 0.05, df = 19), except the connections from PCC and mPFC to the right IPC and from bilateral IPC to PCC in the HC group; the connection from the left IPC to the other three regions and from the mPFC to the left IPC in the EB subjects, as shown in **Table 3**. All the effective connections within the SN were found to be significant in both the EB and NC groups (p < 0.05, df = 19). For the DAN, no significance was found except in one effective connection (left SPL→ left FEF) in the HC group, and in two connections (left SPL→ left FEF and left FEF→ right FEF). **Tables 4**, **5** showed the details of the average parameters of SN and DAN.

### Alteration of the Effective Connectivity Within DMN, SN, and DAN

After the parameters within the DMN, SN, and DAN were estimated, the two-sample t-tests analysis between the EB and HC groups showed that EB subjects showed significantly enhanced effective connectivity from PCC and lIPC to mPFC, and from mPFC to PCC (p < 0.05, df = 18). Besides all endogenous connections, the effective connectivity from the mPFC to bilateral IPC were significantly decreased (**Figure 5A**). For the SN, almost all the significant effective connections between the EB and NC groups were reduced, except the intrinsic connection in the right AI including the connections from the dACC to bilateral AI, right AI to dACC, and effective connections from the left AI to right AI (p < 0.05, df = 18; **Figure 5B**). As shown in **Figure 5C**, more significant effective connections were found to be reduced within the DAN between EB and HC groups, which included the connections from the left FEF to right FEF, and left SPL, the connections from the right FEF to all the other three nodes, and the intrinsic connections in the rSPL (p < 0.05, df = 18) Only the connections from the right SPL and itself to the left SPL were observed to be enhanced. **Figure 5** shows the significantly enhanced and reduced effective connections within the DMN, SN, and DAN between EB and HC subjects.

fnins-13-00268 March 20, 2019 Time: 18:3 # 6

## DISCUSSION

In this study, the spectral DCM was used to compute the effective connections within the DMN, SN and DAN with resting-state fMRI data. The intrinsic brain networks were identified using the group ICA, and several nodes were defined for further DCM analyses. The fully connected model was found to be the optimal explanation for all three networks for our functional data. The one-sample t-test analysis suggested that most effective connections within all three networks were significant in both the EB and HC groups. Two-sample t-test analyses found that reduced coupling parameters of the effective connectivity for each network were observed by comparing the EB patients to the HC subjects, suggesting that there was a disruption of the effective integration within the resting-state brain networks in EB subjects.

### Effective Connectivity Analysis Within DMN

Increasing attention has been paid to the functional interactions within the DMN in normal and disordered people (Fang et al., 2016; Van Den Heuvel and Thomason, 2016; Khalili-Mahani et al., 2017). Directed functional interactions within the DMN regions have been explored using a series of approaches, such as the partial coherence analysis (Silfverhuth et al., 2011), Granger causality analysis (GCA) (Zhou et al., 2011) and Bayesian networks (Wu et al., 2011). One latest study found a reduction of interhemispheric functional connectivity in early blindness using the resting-state fMRI (Hou et al., 2017), suggesting the disruption of the functional connections by early deprivation. However, the DCM method was distinct from the above approaches in theory, because the DCM was model-based by constructing a hemodynamic model to estimate the hidden neuronal states (Li et al., 2012). In our study, the model evidence suggests that the optimal models for all three networks were the whole-connected model. For the EB subjects, there were bidirectional connections between the bilateral IPC and PCC, mPFC and left IPC, as well as the mPFC and right IPC, which was consistent with previous causal studies on the effective connectivity within the DMN (Zhou et al., 2011; Xu et al., 2017). We also discovered directed connections between the PCC and mPFC, which are both functionally and anatomically connected, and play a crucial role in the DMN.

Inconsistent findings on the coupling between regions within the DMN have been reported in many studies. Using the stochastic DCM, our recent study found an influence from the PCC on the mPFC (Xu et al., 2017). Some other studies using GCA showed an mPFC to PCC effective connection. Inconsistent with our findings, bilateral IPC were observed to be capable of modulating mPFC and PCC (Di and Biswal, 2014). Furthermore, reciprocal connections between bilateral IPC were observed in the present study, as suggested in another fMRI study (Li et al., 2012). In our study, bilateral IPC showed a slight functional asymmetry and exerted causal influences on the PCC and mPFC, but not vice versa. Therefore we can assume that bilateral IPCs possess a modulating or driving role, as confirmed by a previous study (Di and Biswal, 2014). However, some inconsistent results were reported. Using GCA, the directed connection from the left IPC to mPFC was observed for EB subjects, whereas there were symmetrical connections from the bilateral IPC to PCC, as well as the connectivity from the mPFC to bilateral IPCs (Jiao et al., 2011). This discrepancy may be caused by the smaller sample

TABLE 3 | Bayes model average parameters of the DMN for healthy controls (HCs) and EBs.


The averaged parameters represented the effective connectivity between regions within the DMN. The positive value between from node A to node B showed a positive modulation, suggesting that node A caused an increase in the rate of change of node B's signal and vice versa. mPFC, medial prefrontal cortex; IPC, inferior parietal lobule; PCC, posterior cingulate cortex; DMN, default mode network; l, left; r, right. <sup>∗</sup>means p < 0.05; ∗∗means p < 0.001.

TABLE 4 | Bayes model average of the salience network for HCs and EBs.


The averaged parameters represented the effective connectivity between regions within the salience network. The positive value between from node A to node B showed a positive modulation, suggesting that node A caused an increase in the rate of change of node B's signal and vice versa. dACC, dorsal cingulate cortex; AI, anterior insulae; SN, salience network; l, left; r, right. <sup>∗</sup>means p < 0.05; ∗∗means p < 0.001.

TABLE 5 | Bayes model average parameters of the dorsal attention network for HCs and EBs.


The averaged parameters represented the effective connectivity between regions within the dorsal attention network. The positive value between from node A to node B showed a positive modulation, suggesting that node A caused an increase in the rate of change of node B's signal and vice versa. FEF, frontal eye field; SPL, superior parietal lobe; DAN, dorsal attention network; l, left; r, right. <sup>∗</sup>means p < 0.05; ∗∗means p < 0.001.

sizes in previous studies, which lead to several connections not being significant.

In addition, the approach of the spectral DCM in this study is crucial, as it can elucidate the neuronal connectivity abnormalities underlying the functional correlations of fMRI signals and provide more insight into the organization of the functional networks (Birnbaum and Weinberger, 2013). Our DCM analysis suggested that the mPFC appeared to be a zonal area, as there were separate efferent and afferent connections between the mPFC and other regions, which was of particular interest given its role in personal evaluation, choice behavior and reality monitoring (Rushworth et al., 2004; Metzak et al., 2015). The function of the mPFC has been proposed to manipulate the correlations between learning associations between different actions, context and adaptive responses (Euston et al., 2012). Moreover, one functional connectivity study has further demonstrated that signals in the mPFC may be responsible for the signal integration between mPFC and other brain regions, such as the posterior superior temporal cortex and AI (Hare et al., 2010). Using a multimodal MR-imaging technique, decreased functional connectivity were observed between the parietal and frontal areas (Bauer et al., 2017), while this reduced effective connectivity were also discovered in the DMN in the EB subjects.

### Effective Connectivity Analysis Within Other Functional Brain Networks

The left AI has played a critical role not only in the semantic and language network (Binder and Desai, 2011), but has also been considered as a "supramodal convergence zone" (Binder et al., 2009; Binder and Desai, 2011), which is involved in the association network and information integration from several sensory modalities. Decreased effective connection between right AI and other regions were observed in our study, as shown in **Figure 5**. The right AI has been confirmed to be a functional hub in coordinating brain networks to produce adaptive behavior, especially the modulation in the DMN and DAN (Wang et al., 2018), as suggested in our study. A weakened functional connection in the early deprivation subject was discovered between the subregions of AI (Nomi et al., 2016; Liu et al., 2017), as the reduced connection from left AI to right AI in our study. Abnormalities within the IPL have been associated with working memory dysfunction as well as impairments in cognitive insight and reality perception (Guo et al., 2014; Lee et al., 2015; Chahine et al., 2017). A recent research based on connectivity-based parcellation of the left inferior parietal lobule suggested that this region was functionally connected to areas more strongly involved in the higher level of social cognitive and language processes, as opposed to more rostro-ventral areas of the IPL that appeared to be associated with the lower functionality level (Bzdok et al., 2016). Our findings suggest that the specific abnormalities of the AI and general alterations within the IPL may be mediated by inhibitory interactions from other regions of the left dorsal attention network. A recent study in EB subjects also suggested that early visual deprivation can selectively reshape the functional architecture of the salience network (Liu et al., 2017). Our study found that most of the effective connections were decreased in the EB subjects, suggesting a corresponding reorganization of the human brain due to the early deprivation.

There are several limitations to our study. First, our analysis does not allow us to conclusively determine the functional

significance of these correlations. This study displayed low average positive symptomatology which possibly precluded us from finding additional correlations and, more importantly, our resting-state experiments were not accompanied by the working memory, executive functioning or language-related behavioral paradigms, so that we could obtain more functionally specific correlations. Second, the sample size of both the EB patients and normal control groups was relatively small and moderate. More qualified subjects should be recruited in future studies to increase the reliability of our observations and the effect of the causal interactions in the DMN, SN, and DAN. This study showed preliminary findings and further studies are needed to retest our discoveries in other blind cohorts. Finally, we did not examine the changes of the effective connectivity within all the brain networks, especially the primary sensory networks like the visual network. The purpose of this study is to see if there were disruptions in the high-order cognitive networks that are associated with the cognitive behaviors in the EB subjects. And the Group ICA could not separate the visual cortex into small parcellations. In future studies, we could examine the dynamic changes of EB subjects using other techniques that could parcellate the visual cortex. In this study, we tested the hypothesis that there was a disruption of the effective connectivity in the cognitive networks. The cognitive networks may functionally connected as a whole, thus the future direction will focus on the causal interactions between the cognitive networks besides the directed connections between regions.

### CONCLUSION

In this study, fully connected models were identified to be the optimal model for the effective connections within the intrinsic

### REFERENCES


functional brain networks in EB subjects. Statistical analyses suggested significant differences in the effective connectivity within the DMN, SN, and DAN between EB and HC groups. More reduced effective connections within three networks by comparing the EB to HC groups indicated that the interactions within the high-order brain networks got greatly suppressed due to the early deprivation. Our analysis additionally revealed the neural connective abnormalities that may underlie alterations of the intrinsic brain network in EB subjects.

### DATA AVAILABILITY

The datasets generated for this study are available on request to the corresponding author.

### AUTHOR CONTRIBUTIONS

MZ and JT designed the experiments. XL, AW, JiX, and PW performed the experiments. XL, JuX, and ZS analyzed the data. XL, JuX, and ZS wrote the manuscript. BW, MZ, and JT revised the manuscript.

### FUNDING

This work was supported by National Natural Science Foundation of China (No. 61703302), Shandong Provincial Natural Science Foundation of China (No. ZR2015HM081), Project of Shandong Province Higher Educational Science and Technology Program (J15LL01), and Shandong Provincial Key Research and Development Project (No. 2018GSF118156).

experience. J. Neurosci. 35, 12366–12382. doi: 10.1523/JNEUROSCI.4715-14. 2015


blind and sighted individuals. J. Cogn. Neurosci. 30, 86–106. doi: 10.1162/jocn\_ a\_01186


fnins-13-00268 March 20, 2019 Time: 18:3 # 11


**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 Li, Wang, Xu, Sun, Xia, Wang, Wang, Zhang and Tian. 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.

fnins-13-00268 March 20, 2019 Time: 18:3 # 12

# Altered Local and Large-Scale Dynamic Functional Connectivity Variability in Posttraumatic Stress Disorder: A Resting-State fMRI Study

*Shishun Fu1,2, Xiaofen Ma2, Yunfan Wu2, Zhigang Bai3, Yin Yi1,2, Mengchen Liu1,2, Zhihong Lan1,2, Kelei Hua1,2, Shumei Huang2,4, Meng Li2 and Guihua Jiang2,1\**

*1 The Second School of Clinical Medicine, Southern Medical University, Guangzhou, China, 2 The Department of Medical Imaging of Guangdong Second Provincial General Hospital, Guangzhou, China, 3 The Department of Medical Imaging of Affiliated Hospital, Inner Mongolia University for Nationalities, Hohhot, China, 4 Guangdong Medical University, Dongguan, China*

### *Edited by:*

*Feng Liu, Tianjin Medical University General Hospital, China*

### *Reviewed by:*

*Meiling Li, Harvard Medical School, United States Xin Di, New Jersey Institute of Technology, United States*

> *\*Correspondence: Guihua Jiang jianggh@gd2h.org.cn*

#### *Specialty section:*

*This article was submitted to Neuroimaging and Stimulation, a section of the journal Frontiers in Psychiatry*

*Received: 23 October 2018 Accepted: 28 March 2019 Published: 12 April 2019*

#### *Citation:*

*Fu S, Ma X, Wu Y, Bai Z, Yi Y, Liu M, Lan Z, Hua K, Huang S, Li M and Jiang G (2019) Altered Local and Large-Scale Dynamic Functional Connectivity Variability in Posttraumatic Stress Disorder: A Resting-State fMRI Study. Front. Psychiatry 10:234. doi: 10.3389/fpsyt.2019.00234*

Posttraumatic stress disorder (PTSD) is a psychiatric condition that can emerge after exposure to an exceedingly traumatic event. Previous neuroimaging studies have indicated that PTSD is characterized by aberrant resting-state functional connectivity (FC). However, few existing studies on PTSD have examined dynamic changes in resting-state FC related to network formation, interaction, and dissolution over time. In this study, we compared the dynamic resting-state local and large-scale FC between PTSD patients (*n* = 22) and healthy controls (HC; *n* = 22; conducted as standard deviation in resting-state local and large-scale FC over a series of sliding windows). Local dynamic FC was examined by calculating the dynamic regional homogeneity (dReHo), and large-scale dynamic FC (dFC) was investigated between regions with significant dReHo group differences. For the PTSD patients, we also investigated the relationship between symptom severity and dFC/ dReHo. Our results showed that PTSD patients were characterized by I) increased dynamic (more variable) dReHo in left precuneus (PCu); II) increased dynamic (more variable) dFC between the left PCu and left insula; and III) decreased dFC between left PCu and left inferior parietal lobe (IPL), and decreased dFC between left PCu and right PCu. However, there is no significant correlation between the clinical indicators and dReHo/dFC after the family-wise-error (FWE) correction. These findings provided the initial evidence that PTSD is characterized by aberrant patterns of fluctuating communication within brain system such as the default mode network (DMN) and among different brain systems such as the salience network and the DMN.

Keywords: posttraumatic stress disorder, resting-state functional magnetic resonance imaging, dynamic functional connectivity, regional homogeneity, default mode network

### INTRODUCTION

Posttraumatic stress disorder (PTSD) is a psychiatric condition that can emerge after exposure to an exceedingly traumatic event (1). In the general population, PTSD occurs most commonly after traffic accidents and affects 10%–32% of those involved within 12 months after the event (2). Symptoms of PTSD include intrusive memories, hypervigilance, insomnia, and emotional numbing (3). Previous studies have indicated that PTSD patients exhibited abnormal interactions among the brain systems (4, 5). For example, Zhang et al. found that the dorsolateral prefrontal cortex showed increased resting-state functional connectivity (FC) with the visual cortex, suggesting that the disrupted frontaloccipital system may be associated with the dysfunction of visual information processing (5). One of robustly identifiable networks is the default mode network (DMN) (6), which is involved in processing self-relevant stimuli (7, 8). The dysfunction of the DMN in PTSD patients may indicate impaired self-generated thoughts and autobiographical memory during rest (9).

One effective approach for exploring brain communication is through the analysis of resting-state fMRI studies (10). Recent resting-state studies in both animals and humans have revealed the dynamic nature of the spatiotemporal organization of blood oxygen level-dependent (BOLD) signals (11–13). Due to unconstrained mental activity, the resting state even shows more dynamic features than in task-stimuli studies (14). A recent study of dynamic FC network indicated that the static FC represented average connectivity across different dynamic states during the whole scanning period; it may not be sensitive enough to detect the alteration of neurofluctuations (15). In order to investigate the dynamic features of inter-regional BOLD signal fluctuations over temporal scales, the sliding window analysis of dynamic FC (dFC) was developed. This approach measured the variety correlations among discrete (large-scale) brain regions (16) using a short, sliding temporal window. Kaiser et al. found that the resting-state dFC revealed the interactions among networks or subnetworks over time (17). Early studies suggested that the dFC can be associated with the changes in arousal (18) and vigilance (19) since hypervigilance and hyperarousal are two typical symptoms of PTSD. We proposed to use dFC to investigate the characteristic features of PTSD in the resting state. Moreover, since changes in brain network topology are associated with those in local brain activity (20), it was reasonable for us to measure both the large-scale and local dynamic FC in our study.

Regional homogeneity (ReHo) is one of the commonly used algorithms in measuring local FC (21–23). ReHo is a reliable measurement technique and robust against noise in the fast imaging sequence data (24). A prior animal study has suggested an association between ReHo variability and different states of neural activity (25). A recent study of dReHo using the sliding-window approach also indicated that brain regions with high dReHo fluctuation tended to be functional hubs in brain systems (26). A resting-state study has shown that the gene variants affected dReHo in attention-deficit/hyperactivity disorder (27). These findings introduced the clinical potential of dReHo analysis.

The investigation of dFC and dReHo in the resting state may provide new insight into the aberrant brain connectivity in PTSD. Previous studies investigating major depression (28), schizophrenia (29, 30), and bipolar disorder (31) showed abnormal dFC and dReHo under the resting state of these patients, and all of these psychiatry researches found aberrant dFC or aberrant dynamic local activities in the DMN. A possible explanation for these abnormalities is the dynamic nature of the DMN, which exhibits dynamic interactions with a number of other brain systems in the resting state (32). Kaiser et al. (17) indicated that the investigation of altered dynamic activity in areas of the DMN may be important in understanding the pathophysiology of psychiatric disorders.

Although there are few studies available that have focused on the dynamic brain activity in PTSD, prior static studies have suggested that the symptoms of PTSD are associated with the DMN. Mounting evidence has indicated that PTSD is associated with aberrant DMN connectivity (33–35). A static FC (sFC) study suggested that the aberrant activities in the DMN can be a predictor of the symptom severity of PTSD (8). A previous study also approved that the static fMRI data can be used to discriminate the PTSD from HC by using the multilevel parametric classification approach (36). A recent study compared the accuracy of sFC to the accuracy of dFC in classifying PTSD patients and HC (37). The results showed that the peak classification accuracy of dFC reached 94.2%, while the peak classification accuracy of sFC was 86.7%; this research indicated that the temporal dFC is a better predictor than sFC of the diagnostic features of PTSD. Additionally, this study indicated that, in comparison with the HC, the PTSD patients were characterized by decreased temporal variability of brain connectivity. Preti et al. also indicated that PTSD patients often stay trapped in one state and exhibited a decreased dFC in comparison with HC subjects (38). All of these studies indicated that the aberrant connectivity variability of brain networks is vital in the investigation of the neurophysiological mechanism of PTSD.

In order to explore the characteristic resting-state temporal variability of PTSD, we decided to measure both large-scale and local dynamic FC. Based on previous findings, we hypothesized that PTSD patients would exhibit altered dReHo in regions within the DMN. We also expected regions with dReHo alterations to show aberrant dFC and the connectivity measures to be associated with subjects' symptomatology.

### METHOD

### Subjects

Permission to undertake this study was granted by the ethics committee of Guangdong Second Provincial General Hospital. In January and February 2017, we recruited 30 trauma-exposed subjects from a serious highway traffic accident in Guangdong province. Prior to the examination, none of the patients had undergone any psychotherapy. The inclusion criteria for the PTSD patients were as follows: I) age >18 years; II) right-hand dominance; III) no preexisting psychiatric disorders or physical conditions as determined by a structural clinical interview using the *Diagnostic and Statistical Manual of Mental Disorders*, 4th edition (DSM-IV); IV) no psychiatric medications or substance abuse; V) no MR imaging contraindications; VI) no head trauma or neurologic disorders; VII) fulfills the criteria of DSM-IV and has a Clinical-Administered PTSD Scale (CAPS) score >40; and VIII) not pregnant or nursing. After considering the strict requirements, eight subjects were excluded, five of them for failing to obtain the CAPS score >40. Twenty-two demographically matched healthy controls (HCs) were recruited for this study. The inclusion criteria for HCs were as follows: I) age >18 years; II) right-hand dominance; III) no preexisting psychiatric disorders or physical conditions as determined by a structural clinical interview using the DSM-IV; IV) no psychiatric medications or substance abuse; V) no MR imaging contraindications; and VI) not pregnant or nursing. Each participant provided written informed consent, which was obtained prior to the MRI scanning.

### Assessment of Mental Status

PTSD diagnosis was determined following the DSM-IV diagnostic criteria. Before undergoing resting-state MRI, all PTSD patients were screened with CAPS (39) in order to estimate the intensity and frequency of the symptoms. In addition, emotion assessments were conducted of all participants, including the Self-rating Anxiety Scale (SAS) (40) and the Selfrating Depression (SDS) (41), in order to evaluate the emotional status. A further Structured Clinical Interview for DSM-IV was also performed to evaluate psychiatric disorder comorbidities.

### Magnetic Resonance Imaging Data Acquisition

Each of the participants underwent a resting-state MRI in a 3.0-T MR imager (Ingenia; Philips, Best, The Netherlands) equipped with a 32-channel head coil at the Department of Medical Imaging in Guangdong Second Provincial General Hospital. A diagnostic T1-weighted image and a T2 fluid attenuated inversion recovery (T2-FLAIR) image were taken to exclude participants with brain lesions. The resting-state fMRI data were acquired using gradient echo-planar imaging (EPI) with the following parameters: repetition time (TR)/echo time (TE) = 2,000 ms/30 ms; matrix = 64 × 64; field-of-view = 230 mm × 230 mm; flip angle = 90; slice thickness = 3.6 mm, 0.6-mm gap; interleaved scanning; 38 transverse slices covering the whole brain at all 240 volumes were acquired for each participant within 480 s; each volume was aligned along the anterior–posterior commissure. Each participant was instructed to lie still and to avoid falling asleep or thinking of anything in particular during MR scanning.

### Resting-State Functional Magnetic Resonance Imaging Data Preprocess

Standard preprocessing of the functional images was performed with the DPARSF 4.3 Advanced Edition (http://rfmri.org/ DPARSF) and the SPM12 package (www.fil.ion.ucl.ac.uk/spm) based on MATLAB (Mathworks, Inc., Natick, MA, USA). The first 10 volumes of each dataset were discarded for signal equilibration. The remaining data were performed using slice timing correction and realignment and co-registered with the anatomical scan. The co-registered T1-weighted images were segmented into gray matter, white matter, and cerebrospinal fluid. And then the functional images were normalized into the Montreal Neurological Institute (MNI) space with a voxel size of 3 × 3 × 3 mm3 . The head movement parameters were obtained from the realignment steps in the DPARSF. We took the mean FD Jenkinson (42) as the head motion TABLE 1 | Demographic and clinical data.


*Demographic data are presented as mean ± SD. PTSD, posttraumatic stress disorder; HC, healthy control; CAPS, Clinical-Administrated PTSD Scale; SAS, Selfrating Anxiety Scale; SDS, Self-rating Depression Scale. The P-value was obtained by the two-sample t test.*

reference standard. We eliminated the subjects with motion (mean FD Jenkinson) greater than 2 × standard deviation (SD) above the group mean motion as recommended in a previous study (43). No subject was eliminated in this step. There was no significant difference in head motion between the PTSD patients and the HC (see **Table 1**). Linear detrending processing was conducted to remove the linear signal drift. Individual-level regression analysis was conducted to minimize the influence of head motion (Friston 24 model), white matter signal noise, and cerebrospinal fluid signal noise. A temporal band-pass filter (0.0167–0.10 Hz) was applied to the data to remove the physical noise and any frequencies for which the period was shorter than that of a single sliding window (44). We performed spatial smoothing with a 6-mm full-width at halfmaximum (FWHM) kernel before performing the dReHo group analysis. As for the dFC, we performed the spatial smoothing with a 6-mm FWHM kernel before the linear detrending and nuisance signals regression, and band-pass filtering. Considering the size of FWHM Gaussian kernel might affect the results of dReHo/dFC analysis (45, 46), we used 4- and 8-mm FWHM Gaussian kernel to test the consistency of our results (45, 46) (see **Supplementary Figures 1** and **2**).

### Computation of dReHo and dFC

ReHo calculation: The ReHo algorithm measures voxel-wise short-distance FC with Kendall's coefficient of concordance (23) using the following formula:

$$W = \frac{\sum\_{i=1}^{N} R\_i^2 - N\overline{R}^2}{\frac{1}{12}K^2(N^3 - N)}$$

where *W* is the Kendall's coefficient of concordance among the given voxels, *N* denotes the length of the time series, *K =* 27 is the size of the voxel cluster containing 3 × 3 × 3 adjacent voxels, *Ri* denotes the summation of the rankings of the BOLD signal amplitude of all *K* voxels at the *i*th time point, and *R* is the mean of *Ri* .

To compute the dReHo for these data, the time course was segmented into 60-s Hamming windows (30 dynamics). By sliding the onset of each window by 2 dynamics (4 s), for a total of 101 overlapping windows in the first level analysis, the dReHo was estimated by using the calculated SD of the ReHo through the windows at each voxel, yielding a set of ReHo maps for each participant.

Two-sample *t*-test with head motion parameters (mean FD Jenkinson values), age, and sex as covariates was performed to test the difference in dReHo maps between the PTSD patients and HC at each voxel. Multiple comparisons correction was performed with Gaussian random field (GRF) theory at the cluster level (minimum *z* > 3.54; cluster significance: *p* < 0.05, two-tailed GRF corrected).

To determine whether the dReHo metrics were associated with clinical indicators, we performed general linear models with the clinical indicators (CAPS, SAS, SDS) and mean dReHo values from clusters with significant group differences as independent variables, and head motion parameters, age, and sex as covariates. The correlation analysis was accomplished with the SPSS software with a significance threshold of *p* < 0.05 (uncorrected).

Voxel-wise seed-based FC analyses were performed using the DPARSF 4.3. We employed the aberrant dReHo region, which we calculated above, as a seed region. Then we used the slidingwindow approach as we have used in the dReHo calculation; the time course was segmented into 60-s Hamming windows by sliding the onset of each window by 2 dynamics, for a total of 101 overlapping windows in the first level analysis. Within each sliding window, the whole brain FC maps for the seed region were computed as the Fisher *z* transformed Pearson correlation coefficient between the averaged time course of all voxels in the seed and the time course of all other voxels in the whole brain, yielding a set of sliding window *z*FC maps for each participant. The dFC was estimated by calculating the standard deviation in *z*FC values through windows at each voxel.

Two-sample *t*-test with head motion parameters (mean FD Jenkinson values), age, and sex as covariates was performed to investigate the difference of dFC values between the PTSD group and the HC group at each voxel. Multiple comparisons were performed with GRF correction at the cluster level (minimum *z* > 3.29; cluster significance: *p* < 0.05, two-tailed GRF corrected).

To explore the relationship between dFC metrics and clinical indicators, we performed general linear models with the clinical indicators and mean dFC values from clusters with significant group differences as independent variables, and head motion parameters, age, and sex as covariates. The correlation analysis was accomplished with SPSS software with significance threshold of *p* < 0.05 (uncorrected).

### RESULTS

The demographic and clinical data are summarized in **Table 1**. Inconsistent with our prediction, compared with the HC, the PTSD patients exhibited an increased dReHo (more variability) in the left posterior cingulate cortex (PCC)/precuneus (PCu) (**Figure 1**). We also found a decreased dFC (less variability) between the left and right PCu, and the left inferior parietal lobe (IPL)/angular gyrus (AG), but increased dFC between the PCC and the left insula (**Table 2** and **Figure 2**).

Next, we conducted analyses to test the association between the clinical indicators (CAPS, SAS, SDS) of PTSD and dReHo within the PTSD group. There is no significant correlation between the clinical indicators and dReHo/dFC after FWE correction (see **Supplementary Figure 4**).

### DISCUSSION

In this study, by using resting-state dReHo analysis, we determined the local aberrant variability in the left PCu. The PCu is a key hub in the DMN of human brain (47). Aberrant restingstate temporal dynamic brain activities were found in the dReHo and large-scale dFC of specific brain regions, which were mainly located in the posterior DMN (pDMN) and the primary region of the salience network (SN). These observations provide new insights into the aberrant brain activities in PTSD.

This study revealed a significantly increased dReHo (more variability) in left PCu, suggesting that the neurofluctuation of the left PCu is unstable in the PTSD patients in the resting state. The PCu is a key hub of the pDMN (47–49) and is considered to


*L, left; R, right; MNI, Montreal Neurological Institute; dFC, dynamic functional connectivity; AAL, anatomical automatic labeling; PCu, precuneus; IPL, inferior parietal lobe.*

be involved in self-referential processing (7) and autobiographic memory (50). In the resting state, the DMN exhibits dynamic interaction with a number of brain systems, such as the frontal– parietal control network and the dorsal attention network (32). Although we found an aberrant local connectivity variety in the PCu, the evidence of aberrant DMN connectivity in PTSD patients is not entirely persuasive. In order to explore the dynamic interactions among the DMN and other brain networks, we employed the left PCu as the seed region and carried out the dFC of the whole brain.

Several studies have identified aberrant activities in the DMN of PTSD patients (6, 34, 35). In the present study, using seedbased dFC, we identified lower-variability regions located in the right PCu and the left IPL. These results suggest that compared with the normal controls, the PTSD group exhibits decreased dFC (less variability) within the pDMN. We suspected that this restrained neurofluctuation within the pDMN represents decreased regulation of the self-referential processing. Previous studies using the independent component analysis identified the aberrant pDMN in PTSD patients (4, 51). Furthermore, Zhang et al. found a decreased intranetwork connectivity within the pDMN by measuring resting-state sFC. These researches suggested that the decreased FC in pDMN was associated with the dysfunction of evaluation of the self-related events in PTSD. The interaction between DMN and the executive control systems, which includes dorsal lateral middle frontal cortex and the IPL (52), is essential in regulating the self-generated thought (32). In the present study, a deceased dFC was found between the left PCu and the left IPL. Therefore, we inferred that the deceased dFC suggested a discrete FC state between the DMN and the executive control system, which may induce the dysregulation of self-referential processing. Previous studies indicated that the IPL is engaged in mediating visuospatial processing (53), which is critical when dealing with life-threatening events (54). Additionally, previous study indicated that the IPL is a vulnerable brain region to the neurotoxic effects of stress (55). Therefore, we suspected that the aberrant dFC between the PCu and the IPL might be a potential biomarker of PTSD.

In the present study, the left insula was the only region in the brain that exhibited significant positive dynamic correlation with the left PCu in the PTSD group. The insula is a key hub in the SN and is thought to be involved in the detection of personally salient internal and external stimuli that guide behaviors in order to maintain equilibrium (56). In addition, the insula is thought to be involved in mediation of the "switching" between activation of the DMN and the central executive network (CEN) to direct appropriate behavioral responses to the salience stimuli (57). Therefore, we suspected that the positive dynamic correlation between the left PCu and left insula might suggest an excessive interaction between the SN and the DMN. Previous studies using the graph theory approach identified dysregulation in three intrinsic brain networks (1, 58). Lei et al. found a disequilibrium among the CEN, DMN, and SN and suggested that the SN was crucial to the PTSD symptoms (58). Previous resting-state sFC study also revealed an increased correlation between the DMN seed region [PCC and ventromedial prefrontal cortex (vmPFC)] and SN (insula and precentral sulcus) (59). They found a positive correlation between the PTSD symptom severity and the vmPFC-precentral sulcus FC values. Our results provided an additional piece of evidence that, compared with the HC, the PTSD patients exhibited more variable connectivity between the DMN and the SN.

There are some limitations in the present study that should be highlighted. Firstly, little information is available on the meaning of the resting-state dFC in neurocognitive functioning. For example, it remains unknown whether the abnormal dynamic activities in the resting state are intrinsic properties or

### REFERENCES


are affected by the present-moment cognitive activities (17, 60). As the number of resting-state dFC studies grows, we may gain a better understanding of these properties and their relation to the psychopathology. Secondly, since the dFC based on the sliding-window approach is composed of a few time points, the dynamic analysis is particularly sensitive to the physiological noise (61). Although, we did not denoised the physiological noise individually, we denoised the physiological noise in the preprocessing steps and group-level test and we also chose a relatively large window size in order to diminish the adverse effects of physiological noise. Thirdly, we only examined significant differences in regions exhibiting abnormal dynamic activity to focusing on the dynamic pattern related to PTSD; further exploration of static results is needed in our future works. Fourthly, the correlation results did not survive the FWE correction, so further exploration of the abnormal dynamic patterns and CAPS subscales is needed in our future large sample research to evaluate the relations between the dynamic patterns and specific clinical symptoms severity, such as intrusive memory and flash back.

In conclusion, this resting-state dFC (combine the dReHo and dFC) study provided evidence that the PTSD patients exhibited aberrant dReHo and dFC in comparison with the HC. Decreased variability within the DMN may suggest dysfunction of self-referential processing in PTSD patients, while increased variability between the insula and PCu may suggest dysregulation between the DMN and the SN.

### ETHICS STATEMENT

Permission to undertake this study was granted by the ethics committee of Guangdong Second Provincial General Hospital.

### AUTHOR CONTRIBUTIONS

SF designed the experiment. SF and XM carried out the experiment. ZB collected and sorted out the data. YW, MLiu, YY, ZL, SH, MLi, and KH helped on data management and processing. SF and GJ wrote the manuscript.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2019.00234/ full#supplementary-material

vehicle accidents. *Psychopathology* (2009) 42:236–41. doi: 10.1159/ 000218521


independent component analysis. *Cell Physiol Biochem* (2018) 48:99–110. doi: 10.1159/000491666


preprocessing, scan duration, imaging resolution and computational space. *NeuroImage* (2013) 65:374–86. doi: 10.1016/j.neuroimage.2012.10.017


**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 Fu, Ma, Wu, Bai, Yi, Liu, Lan, Hua, Huang, Li and Jiang. 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.*

# Predicting the Post-therapy Severity Level (UPDRS-III) of Patients With Parkinson's Disease After Drug Therapy by Using the Dynamic Connectivity Efficiency of fMRI

Xuesong Li <sup>1</sup> , Yuhui Xiong<sup>2</sup> , Simin Liu<sup>2</sup> , Rongsong Zhou<sup>3</sup> , Zhangxuan Hu<sup>2</sup> , Yan Tong<sup>4</sup> , Le He<sup>2</sup> , Zhendong Niu<sup>1</sup> , Yu Ma<sup>3</sup> \* and Hua Guo<sup>2</sup>

*<sup>1</sup> School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China, <sup>2</sup> Department of Biomedical Engineering, Center for Biomedical Imaging Research, School of Medicine, Tsinghua University, Beijing, China, <sup>3</sup> Department of Neurosurgery, Tsinghua University Yuquan Hospital, Beijing, China, <sup>4</sup> Nuffield Department of Clinical Neurosciences, Wellcome Centre for Integrative Neuroimaging, FMRIB, University of Oxford, Oxford, United Kingdom*

#### Edited by:

*Feng Liu, Tianjin Medical University General Hospital, China*

#### Reviewed by:

*Bing Zhang, Nanjing Drum Tower Hospital, China Hui-Jie Li, Chinese Academy of Sciences, China Heng Chen, Guizhou University, China*

> \*Correspondence: *Yu Ma mayu@tsinghua.edu.cn*

#### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *01 February 2019* Accepted: *06 June 2019* Published: *02 July 2019*

#### Citation:

*Li X, Xiong Y, Liu S, Zhou R, Hu Z, Tong Y, He L, Niu Z, Ma Y and Guo H (2019) Predicting the Post-therapy Severity Level (UPDRS-III) of Patients With Parkinson's Disease After Drug Therapy by Using the Dynamic Connectivity Efficiency of fMRI. Front. Neurol. 10:668. doi: 10.3389/fneur.2019.00668* Parkinson's disease (PD) is a multi-systemic disease in the brain arising from the dysfunction of several neural networks. The diagnosis and treatment of PD have gained more attention for clinical researchers. While there have been many fMRI studies about functional topological changes of PD patients, whether the dynamic changes of functional connectivity can predict the drug therapy effect is still unclear. The primary objective of this study was to assess whether large-scale functional efficiency changes of topological network are detectable in PD patients, and to explore whether the severity level (UPDRS-III) after drug treatment can be predicted by the pre-treatment resting-state fMRI (rs-fMRI). Here, we recruited 62 Parkinson's disease patients and calculated the dynamic nodal efficiency networks based on rs-fMRI. With connectome-based predictive models using the least absolute shrinkage and selection operator, we demonstrated that the dynamic nodal efficiency properties predict drug therapy effect well. The contributed regions for the prediction include hippocampus, post-central gyrus, cingulate gyrus, and orbital gyrus. Specifically, the connections between hippocampus and cingulate gyrus, hippocampus and insular gyrus, insular gyrus, and orbital gyrus are positively related to the recovery (post-therapy severity level) after drug therapy. The analysis of these connection features may provide important information for clinical treatment of PD patients.

Keywords: fMRI, dynamic nodal efficiency, Parkinson's disease, drug treatment, prediction of post-therapy severity level

### INTRODUCTION

Parkinson's disease (PD) is one of the most common neurodegenerative disorders. It is clinically characterized by some specific motor symptoms, including rigidity, slowness of movement, tremor at rest, bradykinesia, and postural instability and some other non-motor symptoms such as cognitive deficits, impaired olfaction, emotional problems (1, 2). PD can be considered as a

**45**

multi-systemic disease in the brain arising from dysfunction in several neural networks (3–5). The motor and cognitive impairments in PD have been related to abnormal functional connectivity and disrupted network integration in the brain (6–8).

Several graph theoretic studies revealed an abnormal topological organization of functional brain networks in PD patients. Specifically, Skidmore et al. combined fMRI and graph analysis to find a smaller global efficiency of brain networks in advanced PD patients (9). Wei et al. found that PD had significantly decreased efficiency in the cortico-basal ganglia motor pathway (10). In addition, Dubbelink et al. using magnetoencephalography and graph theory, reported that impaired local network efficiency and network decentralization are very early features of PD that continue to progress over time, along with reductions in global efficiency (6). In summary, the graph theory provides a powerful and general framework to characterize brain connectivity at global and local levels, and offers a collection of metrics that can quantify the segregation and integration of information within functional networks among the brain regions. However, most of the previous studies did not consider the important dynamic properties of FC over time, such as the dynamic nodal efficiency; instead, FC was usually assumed to be constant during the rs-fMRI experiment (8).

The graph theory-based approach applied to dynamic FC show that the variability in brain network may also provide important information on the underlying nature of neurodegeneration. In a study by Yu et.al., the reduced variability of local and global network efficiency was detected in a patient with schizophrenia (11). In a more recent PD study, dynamic topological properties of brain networks can characterize the underlying nature of Parkinson's disease and correlate with clinical features (8).

The dynamic property of fMRI can enrich the graph theory. We wonder if the dynamic nodal efficiency (dnE) can be used to predict the recovery effect after drug therapy (i.e., posttherapy severity level) of PD patients. If possible, it may provide useful guidance information for drug therapy. Connectomebased predictive modeling is a recently developed data-driven method for identifying the relationship between functional brain connectivity and the behavioral and cognitive variables of interest, and then predicting the behavior of patients (12–14). The predictive modeling procedure has been applied to analyze connectivity, such as attention control and temperament trait (14, 15). Its core idea is the cooperative analysis of the relationship between behavior and FC, finally finding the strong functional networks that are correlated to the behavior with statistical significance. It provides an effective way to explore the correlation between altered topological properties and clinical indexes of interest.

In the present study, we used rs-fMRI and sliding-window analysis to build the individual dnE network by computing each nodal efficiency of each sliding-window and predicted the post-therapy severity level of PD patients. The global efficiency is chosen to calculate the dnE, since it may reveal more PD properties than local efficiency, as indicated by TABLE 1 | Participant demographic and clinical characteristics.


*Values are given as mean and SD. MoCA, Montreal Cognitive Assessment; BDI-II, Beck Depression Inventory-II; UPDRS-III, Unified Parkinson Disease Rating Scale III.*

Kim et al. whose study showed a significant difference in global efficiency between PD and the healthy control, but not in local efficiency (8). Specifically, we proposed a rigorous crossvalidated prediction framework incorporating feature selection and regression techniques, to predict the drug therapy effect of levodopa (the most commonly used drug in PD treatment), which is evaluated by Unified Parkinson Disease Rating Scale III (UPDRS-III) (16) scores, using the rs-fMRI data from 62 PD patients. We aim to investigate the possibility of predicting individual after-therapy UPDRS-III scores using whole-brain dnE network. The post-therapy UPDRS-III scores for certain patients was estimated, and the potentially important connections that contribute to the recovery degree were predicted by the rs-fMRI data.

### MATERIALS AND METHODS

### Subjects

Sixty-two subjects (mean age, 58.5 ± 10.1 years; 31 females and 31 male patients) were recruited from Tsinghua University Yuquan Hospital, Beijing, China. Patients were told to stop taking drugs 12 h before the rs-fMRI scan (before therapy). Patients diagnosed with PD based on the UK Brain Bank criteria (17) were enrolled. Exclusion criteria includes a history of psychiatric or neurological disease other than PD, other major medical diseases, head injury, alcohol/drug dependency/abuse (8). Disease severity of each patient was evaluated by the UPDRS-III (16) scores given by an experienced specialist after taking levodopa, including the medication-on and medication-off states. These PD patients took different doses of levodopa, according to a widely used guidance (18) for each patient. None of them have taken other medicines. Details of the demographic information can be found in **Table 1**. All participants signed the informed consent form before the experiment. This research was approved by the Ethics Committee of Tsinghua University Yuquan Hospital.

### MR IMAGE ACQUISITION

All data were collected on a 3T Philips Achieva MRI scanner (Philips Healthcare, Best, The Netherlands) with a 32-channel head coil. Head motion was controlled by fixing the head during scanning. Resting-state blood-oxygenation-level dependent (BOLD) signals were collected with following imaging parameters: 35 axial slices; repetition time (TR) = 2,000 ms; echo time (TE) = 30 ms; flip angle (FA) =90◦ ; slice thickness =4.0 mm; slice gap =0.8 mm; acquisition matrix = 64 × 64; field of view = 224 × 224 mm<sup>2</sup> . All the PD patients have only experienced one rs-fMRI scan, which was carried out before taking levodopa. During the scan, the participants were instructed to keep their eyes closed, relax their minds, and remain as motionless as possible but not to fall asleep. The rs-fMRI scan with 240 dynamic scans lasted for 8 min. High-resolution T1-weighted structural images in coronal view were acquired with slice thickness of 1 mm without slice gap. Other sequence parameters were: TR/TE = 7.64/3.73 ms, FOV = 256 × 256 mm<sup>2</sup> (acquisition matrix = 256 × 256 × 160).

### Data Processing and Network Analysis

The pre-processing of rs-fMRI data was conducted using the SPM12 (http://www.fil.ion.ucl.ac.uk/spm) and GRETNA (19) software. The first four scans were discarded to allow for magnetization equilibration. Four subjects with the mean framewise displacement value exceeding the maximum displacement of 1 mm were excluded from either the above demographic information or subsequent data analysis. Data were realigned to the first volume to correct for head movement. A 0.01– 0.10 Hz band-pass was used to reduce the effects of low frequency drift and high-frequency physiological noises. The nuisance signal regression (24-parameter head motion profiles, global signal, CSF signal, and WM signal) was performed. Data were spatially smoothed with a 4 mm full-width at halfmaximum Gaussian kernel. In order to perform group analysis, the first scan of fMRI time series was co-registered to the same participant's T1-weighted images. The transformed T1 structural images were normalized to the Montreal Neurological Institute (MNI) template space, using the voxel size of 3 × 3 × 3 mm<sup>3</sup> .

The flowchart of the subsequent data processing is shown in **Figure 1**. The GRETNA software was used to construct the whole-brain networks for each sliding-window (19). The human Brainnetome Atlas (http://atlas.brainnetome.org/) was applied to obtain 246 brain regions (i.e., nodes, with 123 in each hemisphere), including 210 cortical and 36 subcortical regions (20). The sliding-window approach was used to explore the timevarying changes of FC. The window was slided by 2 s along the 240 dyanmic scans (480 s). We chose the window size of 50 time points for the trade-off between the accuracy of capturing state transitions accurately and the number of overall state transitions (21), resulting in 191 consecutive windows across the entire scan. For each sliding window for a participant, the nodal efficiency was computed, resulting in a 191 nodal efficiency curve. For each patient, the dnE matrix (246 × 246) were calculated by computing the inter-node pearson correlation of the 191-time-point dynamic efficiency curve. The value of each element in the dynamic nodeal efficiency matrix ranges between −1 and 1.

### Prediction Model

The least absolute shrinkage and selection operator (LASSO) method (14) was performed to select the features and build the model.

The least absolute shrinkage and selection operator (LASSO) is a regression analysis method that performs both variable selection and regularization in order to enhance the prediction accuracy and interpretability of the statistical model it produces. The object function is as below:

$$\min\_{\boldsymbol{\mathfrak{B}}} \left\lVert \mathbf{y} - \mathbf{x} \boldsymbol{\mathfrak{B}} \right\rVert\_2^2 + \frac{1}{2} \lambda \left\lVert \boldsymbol{\mathfrak{B}} \right\rVert\_1$$

where **x** denotes the dnE matrix and **y** donates the actual UPDRS-III scores. The objective of the whole regression process is to solve the matrix β by minimizing the loss function. This L1-norm regularization typically sets most coefficients to zero and retains one random feature among the correlated ones, the selection of parameter λ is a trade-off between the prediction error and L1-norm regularization we used λ = 0.08 in this study.

The prediction model was chosen to depict the correlation between the connectome-based feature and the UPDRS-III score. Considering the size of the dataset, it is not convincing if we only use part of the training-validate-test dataset. Therefore, the Leave-One-Out-Cross-Validation (LOOCV) was used to maximize the loss function (22, 23). In LOOCV, N-1 (N is number of subjects, N = 62) samples were used as training data and the remaining samples were used as validation data. The left subjects were used as the input to the training model which was derived with inner training data, generating the estimated UPDRS-III scores. This loop was repeated N times to test all subjects. Each time, the predicted UPDRS-III scores for the leftone-out subjects, the identified FCs, and their corresponding weights in the training model were obtained. By pulling all testing subjects across N loops together, we obtained the prediction results for all subjects. Thus, there were N regression models of the same type with the same parameters for learning and predicting different data. The prediction performance was assessed by the Pearson correlation (with Bonferroni correction) between the model predicted UPDRS-III scores and the actual scores. Permutation test (1,000 times) was carried out to access the significance (24). Mean Absolute Error (MAE) were used to measure the magnitude of the error between the predicted and the actual UPDRS-III scores.

### FEATURE IDENTIFICATION

Since we applied a cross-validation strategy to estimate the UPDRS-III scores, in each iteration, slightly different connections were selected. The relative weights for all selected connections were determined by averaging the regression weights of all loops. For better interpretation and visualization (14, 15), we grouped the 246 FC nodes into 24 relatively large brain regions defined by the Brainnetome atlas (20), and estimated the inter-region contributing power by averaging the weights of all FCs connecting between two specific macroscale regions.

FIGURE 1 | The prediction and validation flowchart incorporating feature selection and regression analysis. (A) shows the detailed steps of the data preprocessing including parcellation, efficiency network computing, feature selection, regression model, and the final feature verification. (B) is the related information from the image preprocessing to feature identified.

## RESULTS

### Clinical Data

Firstly, we compared the UPDRS-III scores before and after levodopa therapy ("medication-off " and "medication-on," respectively). The paired t-test shows significant difference (p < 0.0001) of the UPDRS-III scores between medication-off (44.1 ± 12.0) and -on (22.2 ± 11.8), which demonstrates the efficacy of levodopa.

## Feature Selection for Medication-off and Medication-on

The mean contributing weights of whole-brain dnE network are shown in **Figure 2**. For the medication-off status, MedioVentral Occipital Cortex (MVOcC), are the important regions. For the medication-on status, the important regions are frontal regions including the Inferior Frontal Gyrus (IFG), Middle Frontal Gyrus (MFG), Superior Frontal Gyrus (SFG) and Orbital Gyrus (OrG).

### Prediction Performance

The dnE network based prediction models achieved significant correlation between the predicted and the true UPDRS-III scores of either medication-off or -on for the 62 PD patients (**Figure 3**). Specifically, Pearson correlation of r = 0.54 (p = 4.56 × 10−<sup>6</sup> , MAE = 9.49) and r = 0.65 (p = 8.06 × 10−<sup>9</sup> , MAE = 7.52) were obtained for medication-off and –on, respectively. All the results passed Bonferroni corrections for the multiple comparisons. The p-value of the permutation test is 0.004 and 0.001 for medicationoff and –on, respectively.

### Connections Identified

The relative weights of dnE network connecting between each pair of the 24 anatomically defined macro-scale areas are displayed in **Figure 4**. The identified features in the dnE network of either medication-off or -on include the negative connections (blue) and positive connections (red). The width of the inter-node lines represents the strength of the connections. For the negative connections, stronger connectivity reflects

their lobe locations. dnE, dynamic nodal efficiency.

lower disease severity thus better recovery after drug therapy. The positive connection case reflects the contrary. Specifically, for medication-off, dnE network connections show more contributing power between Middle Temporal Gyrus (MTG) and STG, Postcentral Gyrus (PoG) and Superior Parietal Lobule (25). The stronger the connections, the better the recovery of PD.

macro-scale regions in the Brainnetome Atlas, and a bigger circle represents a higher predictive weight. Names of 24 macroscale regions were colored according to

There is a negative correlation between the strength of some connections and the UPDRS-III scores, such as the connections between Precuneus (Pcun) and Orbital Gyrus (OrG), Inferior Parietal Lobule (IPL), and lateral Occipital Cortex (LOcC), IPL, and Fusiform Gyrus (FuG). The stronger the connections, the worse the recovery of PD.

For medication-on, in terms of the feature analysis of predicting recovery effect after the drug therapy, the contributing power is mainly concentrated on hippocampus (Hipp), PoG, Pcun, Cingulate Gyrus (CG), Insular Gyrus (INS), and OrG. Particularly, the features of the connections between Hipp and CG, INS, Pcun, OrG, respectively, have much influence on the recovery after therapy. The stronger the connections, the better the recovery. For other regions, such as MFG and MTG, the stronger the connection between MFG and Inferior Temporal Gyrus (ITG) or the connection between MTG and SPL, the worse the recovery of PD. The prediction efficacy of each region is shown in **Figure 4C**. The results are from the regression model and are normalized to the range of 0 to 1.

### DISCUSSION

Through analyzing the efficiency of correlation networks using rs-fMRI, the present study investigated the important features of connections that are correlated to the post-therapy disease severity (UPDRS-III scores) after taking levodopa in the 62 PD patients. Studying the effect of drug therapy is an important topic in PD research. Finding what influences the effect of drug therapy is of great significance. This work predicted the correlation between the dnE network and the actual effect of drug therapy by training a regression model. The major findings of the present study are as follows: (1) The connection efficiency of networks based on rs-fMRI can effectively depict the severity of PD (UPDRS-III scores), and further predict the recovery effect after drug therapy. (2) The Hipp region is an important area that indicates drug therapy effect in the dnE network. (3) Increased cortical functional connectivity from ITG and MTG has a negative effect on the recovery. As such, these findings provide new evidence that the rs-fMRI network connectivity strengthening or weakening within key functional networks in dnE network plays an important pathophysiological role in the recovery of PD patients.

The FC analysis of the brain network has revealed that the brain is organized according to a highly efficient small-world topology, combining a high level of segregation (local efficiency) with a high level of global integration (global efficiency) (26). Most of the previous studies did not consider the important dynamic properties of FC, as functional connectivity was assumed to be constant during rs-fMRI scanning. However, dynamic FC may yield novel insights into brain function and dysfunction (8). The sliding window approach is commonly used for examining dynamics in resting-state FC, resulting in quantification of the time-varying behavior of a chosen metric. In this study, we selected the nodal global efficiency (i.e., nodal efficiency) as the metric, then obtained the time-varying behavior of each nodal global efficiency. Global efficiency is a network attribute that quantifies how easily information can be exchanged over the network. It provides information on the communication efficiency of a network as a whole, with higher values indicating more efficient information transmission through the whole brain. We further calculated the inter-node correlation of the dnE, to further reveal the synchronization of the dynamic property of FC between two regions.

The prediction model analyses demonstrated that some specific subnetworks with decreased connectivity are correlated with the recovery effect after drug therapy. The regions mainly include parietal lobe, insular lobe, limbic lobe, and Hipp. The connections between these regions are directly positively correlated to the recovery after drug therapy: the stronger the connections, the better the recovery (lower UPDRS-III scores). Nonetheless, there are several key pathways in the SPL, ITG, MFG, MTG showed negative influence on the recovery of PD.

Previous studies showed that the decreased functional connectivity of the temporal cortex is related to the disease progression of PD (27, 28). We drew the similar conclusion here, especially for the connection between MTG and STG. In addition, we found that a distributed set of regions in the

FIGURE 4 | The identified features in the dnE network between medication-off (A) and medication-on (B), respectively, including the negative connections (NC) represented by blue and the positive connections (PC) represented by red, respectively. The width of the inter-node lines represents the strength of connections. For the negative connections, stronger connectivity reflects smaller disease severity and better recovery after drug therapy. The positive connection case is on the contrary. The prediction efficacy of each node for medication-on is shown in (C). The results are from the regression model and are normalized to the range of 0 to 1. As shown in the circle plots, the 246 FC nodes (inner circle) are grouped into 24 macro-scale brain regions (outer brain representations) that are anatomically defined by the Brainnetome atlas. Specifically, nodes incorporated in each of 24 macro-scale brain areas are plotted with different colors, which delineate their corresponding anatomy locations in the outer brain representations.

frontal lobe, temporal lobe, occipital lobe, and parietal lobe showed decreased inter-node correlation of dnE in these patients. The decreased correlation is related to the high post-therapy UPDRS-III scores of PD patients. It reveals that these regions are also related to the disease progression of PD. Among these important connections exist, such as the connection between the paracentral cortex and OrG. Some other studies have also reported that connections among these regions is related to cognitive decline (29).

This study shows that the Hipp region, which was previously reported to influence dementia (30), is also an important area to indicate drug therapy effect. Previous study also shows that there is a high correlation between PD and dementia, i.e., PD with dementia (31). This study reveals that individualized recovery effect after drug therapy can be influenced by the functional connections between Hipp and other areas, on which enough attention should be paid before therapy.

The selected features (**Figure 2**) are not necessarily useful for the following prediction. For example, the frontal area appears to be important in the feature selection, but the Hipp area is the dominant feature for prediction. Therefore, how to improve feature selection needs to be further studied. Reducing feature dimension while maintaining all of the connection information might need to be carefully considered.

There are several limitations of this study. First, the number of subjects was relatively small to draw firm conclusions. Second, the dynamic property of FC we investigated in this study is only the commonly used global efficiency. It should be noted that choosing the global efficiency does not mean less importance of other network properties. Other dynamic properties, such as local efficiency, may provide additional information and are worthwhile to be further investigated in future studies. Third, the long-term follow-up study should be carried out in the future to follow the outcome of patients.

### CONCLUSION

In this study we studied a sample of 62 PD patients and calculated the dnE based on rs-fMRI. With connectome-based predictive

### REFERENCES


models using LASSO, we demonstrated that the dnE properties can successfully predict the post-therapy severity level of PD patients after taking levodopa. The contributed regions for the prediction include hippocampus, post-central gyrus, cingulate gyrus, and orbital gyrus. Specifically, the connections between hippocampus and cingulate gyrus, hippocampus and insular gyrus, insular gyrus and orbital gyrus are positively related to the recovery after drug therapy. The analysis of these connectivity features can provide guidance information for clinical therapy in PD patients.

### ETHICS STATEMENT

This study was carried out in accordance with The Nuremberg Code, World Medical Association Declaration of Helsinki and International Ethical Guidelines for Biomedical, Ethics Committee of Tsinghua University Yuquan Hospital. Research Involving Human Subjects, and 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 Tsinghua University Yuquan Hospital.

### AUTHOR CONTRIBUTIONS

XL contributed to the experiments, data analysis, and writing of the manuscript. YX and SL contributed to performing the experiments and revising the manuscript. RZ and LH contributed to the data collection. ZH and YT revised the manuscript. ZN contributed to the manuscript revision. YM and HG are the guarantors of this study and had complete access to all data in the study. All authors are accountable for the contents of this research.

### FUNDING

This work was supported by the Beijing Institute of Technology Research Fund Program for Young Scholars and the National Natural Science Foundation of China (Grant No. 61801026).

3. Brooks DJ, Pavese N. Imaging biomarkers in Parkinson's disease. Progr Neurobiol. (2011) 95:614–28. doi: 10.1016/j.pneurobio.2011. 08.009

4. Luo C, Song W, Chen Q, Zheng Z, Chen K, Cao B, et al. Reduced functional connectivity in early-stage drug-naive Parkinson's disease: a resting-state fMRI study. Neurobiol Aging. (2014) 35:431–41. doi: 10.1016/j.neurobiolaging.2013.08.018


**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 Li, Xiong, Liu, Zhou, Hu, Tong, He, Niu, Ma and Guo. 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.

# Increased Temporal Dynamics of Intrinsic Brain Activity in Sensory and Perceptual Network of Schizophrenia

*Youxue Zhang1\*, Gang Guo1 and Yuan Tian2*

*1 School of Psychology, Chengdu Normal University, Chengdu, China, 2 School of Foreign Languages, Chengdu Normal University, Chengdu, China*

Schizophrenic subject is thought as a self-disorder patient related with abnormal brain functional network. It has been hypothesized that self-disorder is associated with the deficient functional integration of multisensory body signals in schizophrenic subjects. To further verify this assumption, 53 chronic schizophrenic subjects and 67 healthy subjects were included in this study and underwent resting-state functional magnetic resonance imaging. The data-driven methods, whole-brain temporal variability of fractional amplitude of low-frequency fluctuations and regional homogeneity (ReHo), were used to investigate dynamic local functional connectivity and dynamic local functional activity changes in schizophrenic subjects. Patients with schizophrenia exhibited increased temporal variability ReHo and fractional amplitude of low-frequency fluctuations across time windows within sensory and perception network (such as occipital gyrus, precentral and postcentral gyri, superior temporal gyrus, and thalamus). Critically, the increased dynamic ReHo of thalamus is significantly correlated with positive and total symptom of schizophrenic subjects. Our findings revealed that deficit in sensory and perception functional networks might contribute to neural physiopathology of self-disorder in schizophrenic subjects.

Keywords: schizophrenia, functional connectivity, temporal variability, self-disorder, sensory and perceptual network

### INTRODUCTION

About 1% of the whole adult population suffer from schizophrenia, which is one of the costliest mental disorders. Schizophrenic subject is typically considered as a self-disorder (1). Self-disorder could be associated with several positive symptoms. The major point of schizophrenic subjects' positive symptom is unable to efficiently distinguish self and others. This symptom would lead to a worse deficit that the schizophrenic patients could not confirm their actions and thoughts are related to external information or stimulation. Importantly, in schizophrenia, the symptoms related to self-disorder have been considered a crucial factor to identify whether the psychiatric patient is schizophrenic or not (2).

There are many neuroimaging studies that have been employed in investigating the neuropathological mechanism of schizophrenia (3–5). Although many functional connectivity studies of schizophrenia focused on the abnormal long-range functional connectivity among spatially distributed brain regions (6, 7), few studies paid attention on local functional information of blood oxygen level dependence and functional interaction between spatially adjacent regions (8, 9). Thus, to quantify local or short-range functional connectivity in human brain, several measures were commonly employed in neuroimaging

#### *Edited by:*

*Zaixu Cui, University of Pennsylvania, United States*

#### *Reviewed by:*

*Jin Liu, Beijing Normal University, China Nan Zhang, University of Science and Technology of China, China*

### *\*Correspondence:*

*Youxue Zhang yx\_zhang06@163.com*

#### *Specialty section:*

*This article was submitted to Neuroimaging and Stimulation, a section of the journal Frontiers in Psychiatry*

*Received: 02 December 2018 Accepted: 19 June 2019 Published: 12 July 2019*

#### *Citation:*

*Zhang Y, Guo G and Tian Y (2019) Increased Temporal Dynamics of Intrinsic Brain Activity in Sensory and Perceptual Network of Schizophrenia. Front. Psychiatry 10:484. doi: 10.3389/fpsyt.2019.00484*

**54**

studies, including regional homogeneity (ReHo) (10), local power of blood oxygen level dependence [low-frequency fluctuations (fALFF)] (11), and functional connectivity strength (12) derived from resting-state functional magnetic resonance imaging (fMRI). Several studies have reported that there are significant relationships between static ReHo/fALFF and several factors, such as age, gender, and intelligence in healthy subjects (13, 14). These findings have revealed that the static local neural activity and short-range functional connectivity have been linked with the physiological and psychological factors in human brain.

In schizophrenia, multi-site resting-state fMRI study has shown that schizophrenic subjects exhibited decreased static fALFF in cuneus, middle temporal gyrus, and posterior cingulate cortex compared with healthy subjects (15). Guo et al. has also found that the schizophrenic patients showed both decreased static fALFF in the posterior cingulate cortex and decreased gray matter volume in medial prefrontal cortex, indicating that the changes of brain function and anatomy within default model network might contribute separately to the pathophysiology of schizophrenia (16). Besides, recent studies have indicated that schizophrenic patients have shown reduced static functional connectivity density in primary sensory network of schizophrenia and decreased static ReHo in visual and sensorimotor networks compared with healthy controls (17). Furthermore, the symptomatology (e.g., auditory hallucinations) in schizophrenia has been proved to be related to abnormal multisensory static functional connectivity (18). In conclusion, the deficit static functional connectivity of sensory and perceptual systems may potentially contribute to physiopathology of schizophrenia. While these studies have implicitly revealed that functional connectivity is a stable characteristic across the entire resting scan period, recent studies have indicated that functional connectivity is not stationary and changes over time (19, 20).

Assessing brain dynamic functional connectivity from resting-state fMRI has advanced our knowledge of the brain (21). Specifically, a recent neuroimaging study has stated that functional connectivity variability seems to be a reliable feature, partly dependent on functional relationships among distributed brain regions (22). Dynamic functional connectivity analysis could provide a novel method to sensitively capture the abnormal functional connectivity related with psychiatric disorders (23– 26). The results of dynamic functional connectivity analyses also revealed transient states of dysconnectivity in schizophrenia (27, 28), which support and expand current knowledge regarding dysconnectivity in schizophrenia. Moreover, a recent study demonstrated that the feature of dynamic functional connectivity significantly outperforms the static connectivity in classification analysis (29). These findings reveal that static functional analysis may obscure important dynamic features of network behavior.

During recent years, few studies have focused on altered local temporal variability of functional activity or short-range functional connectivity in schizophrenia, which could reveal information that is not from static functional connectivity (30). Thus, we sought to determine whether altered temporal variability of regional neural activity was associated with symptom of schizophrenia in this study. The dynamic neural activity analysis used in this study includes dynamic ReHo and fALFF, which allow us to identify voxel-level dynamic functional alterations in schizophrenia compared with healthy subjects. On the basis of previous results about abnormal static functional connectivity in primary motor and perception networks, we hypothesize that abnormal dynamic neural activity in schizophrenia would locate in primary perceptual systems, such as primary sensory-motor cortex and related visual and thalamus regions. In addition, schizophrenic subjects are expected to show significant association between altered variability of these network and symptom of schizophrenic subjects.

### MATERIALS AND METHODS

### Subjects Selection and Schizophrenic Patients' Clinical Symptoms

Fifty-three chronic schizophrenic subjects and 67 healthy controls are included in this study. Related resting-state fMRI data are collected from the Center for Biomedical Research Excellence. The patients with schizophrenia are diagnosed according to Diagnostic and Statistical Manual of Mental Disorders, 4th Edition, diagnostic. The psychiatric symptom severity is measured using positive and negative syndrome scale (PANSS) assessment. Healthy subjects are also recruited, those who do not have schizophrenia and not exhibiting Axis I symptoms. These research procedures were in accordance with institutional review boards of the USA. Written informed consent was obtained from each subject before the study. Details of demographic characteristics of both groups are shown in **Table 1**.

### Data Acquisition and Image Preprocessing

Functional imaging scan was performed on a 3T MRI scanner (Siemens Trio). Resting-state functional image are collected with single-shot full k-space echo-planar imaging (EPI) (repetition time = 2,000 ms, echo time = 29 ms, number of slices = 32, slice thickness = 3 mm, matrix size: 64 × 64, flip angle = 7°, field of view = 256 × 256 mm2 ). Subjects underwent 6-min scan. A total of 180 volumes of EPI images were obtained.

The preprocessing steps of functional image were performed using commonly processing steps [Data Processing and Analysis

TABLE 1 | Dataset (The Center for Biomedical Research Excellence, chronic).


*Indicated values are shown as mean ± standard deviation. PANSS, positive and negative symptom scale; FD, Framewise displacement.*

*aIndicates the p values from the comparison analysis (Chi-square test).*

*bIndicates the p values from the comparison analysis (two-sample t-test).*

of Brain Imaging (DPABI) (31), http://rfmri.org/dpabi] and briefly described here. First, temporal and spatial corrections were performed, including slice time and head motion correction, furthermore normalized (voxel size: 3 mm) into EPI template. Any subjects who had a maximum translation in any of the cardinal directions larger than 3 mm or a maximum rotation larger than 3° were excluded from subsequence analysis. Moreover, framewise displacement (FD) was evaluated in two groups as suggested by Power et al. (32). Second, detrending analysis was performed on the normalized data to minimize the effect of linear trend. Third, several nuisance signals were regressed out from functional image through linear regression analysis. The nuisance signals include six motion parameters and their first temporal derivative, white matter and cerebrospinal fluid signals. In this study, the global signal was not removed from the functional image (33, 34).

### Temporal Variability Analysis

Two widely used approaches, including ReHo and fALFF, were used to measure voxel-level functional maps (35). We calculated dynamic ReHo and fALFF through sliding window analysis (**Figure 1A**). Based on the "rule of thumb," which is 1/fmin of data should be equal or less than the length of window (36), the whole-run time series of each voxel was segmented into 50 TR windows and sliding the onset of these windows by one TR. Then, within each window, we calculated ReHo and fALFF at each voxel in whole-brain mask.

In the ReHo analysis, the frequency band passing (0.01–0.08 Hz) was done on fMRI data. Then, Kendall's W value was calculated for each voxel, between the time series of the target voxels and the series of their nearest voxels (26 voxels) in the whole-brain mask (10). In the fALFF analysis, fALFF is defined as the percentage of the power within the low-frequency range (0.01–0.08 Hz) in total power of whole frequency range (0–0.25 Hz) (11).

Across *n* window, we calculated the coefficient of variation (CV) maps of ReHo and fALFF for each subject. We define the CV of a voxel *k* as:

$$CV\_k = \frac{\sqrt{\sum\_{t=1}^{n} \left(\omega\_t - \omega\_{\text{mean}}\right)^2 / n}}{\omega\_{\text{mean}}}$$

where *xt* is ReHo or fALFF score of voxel *k* over time window *t*, *t =* 1,2,…,*n*; *xmean* is mean score of *xt* across time window *t*. Finally, individual voxel-wise ReHo and fALFF CV maps were

window was systematically shifted by one TR, and the corresponding fALFF and ReHo were computed. Then, the temporal variability of the dynamic fALFF and ReHo were defined as the CV maps across the sliding windows. The pattern of temporal variability of the fALFF (B) and ReHo (C) of the schizophrenic subjects/ healthy controls were shown.

standardized by dividing the whole-brain mean values and, furthermore, spatially smoothed (6-mm full width at half maximum of the Gaussian kernel). Then, two-sample t-tests were performed for ReHo and fALFF CV maps, respectively (DPABI, http://rfmri.org/dpabi), between schizophrenic and healthy subjects with age, gender, education level, handedness, and FD as covariates, with a statistical significance level corrected by false discovery rate (*p* < 0.05).

### Correlations With Pathological Factors

We assessed the association between the score of clinical score and significant changes of temporal variability in regional functional measurements in patients with schizophrenia. We extract the mean CV score from the peak voxel and its nearest voxels (26 voxels) for each significant cluster. Then, the partial correlation analysis was performed between ReHo and fALFF CV scores and patients' PANSS scores with age, gender, education level, handedness, medication dosage, and FD as covariates (*p* < 0.05).

### Validation Analysis

Recent fMRI study has indicated that sliding window-based dynamic functional connectivity could be largely explained by head motion (37). Patient is chronic schizophrenic subjects in this study. The antipsychotic treatment might have an effect on dynamic local neural activity of schizophrenic subjects. Thus, we preformed the validation analysis to investigate the influence of these factors on dynamic temporal variability of regional functional measurements in schizophrenic subjects.

First, spike-regression-based scrubbing was performed to take into account transient head motion (38, 39). We defined the "bad points" with high FD (above 0.5 mm) and their adjacent time points (1 back and 2 forward) for each subject. These "bad points" were modeled as separate regressor in the nuisance regression models in the preprocessing analysis. Then, for new preprocessed fMRI data, we reevaluated the temporal variability of ReHo and fALFF through sliding window analysis. Twosample t-tests were also performed between two groups with age, gender, education level, handedness, and FD as covariates. Second, to take account of antipsychotic treatment, we calculated the relationship between altered temporal variability of fALFF/ ReHo and medication dosage in schizophrenia group by using correlation analysis (*p* < 0.05).

### RESULTS

### Temporal Variability of fALFF/ReHo Between Schizophrenic and Healthy Groups

Temporal variability of dynamic fALFF and ReHo were shown at each voxel for each subject (**Figures 1B**, **C**) with the BrainNet viewer (http://www.nitrc.org/projects/bnv/) (40). The variability of these dynamic local neural activity displayed a nonuniform spatial distribution across the brain. The lowest variability was located in the limbic system. The largest variability was mainly located in the heteromodal association region, including the temporal–parietal junction, prefrontal and posteromedial cortex. The primary sensory and visual cortices showed a moderate level of variability. Furthermore, using two-sample *t* test, schizophrenic subjects showed increased temporal variability in both dynamic fALFF and ReHo compared with healthy controls (**Table 2**, **Figure 2**) with the DPABI viewer (41). Within temporal variability of fALFF, increased dynamic fALFF were observed in thalamus, super temporal gyrus, precentral/postcentral gyrus, and lingual gyrus in schizophrenic subjects. Similar increased dynamic ReHo were also being found in patients with schizophrenia, including super temporal gyrus, thalamus, postcentral gyrus, middle cingulum cortex, and cuneus. Furthermore, these findings were observed by using spike-regression-based scrubbing procedure (**SFigure 1**).

### Correlations With Pathological Factors

We observed positive correlation between PANSS scores and the increased temporal variability of ReHo in schizophrenic subjects: PANSS-positive score and thalamus within basal ganglia network (BGN) (*r* = 0.317, *p* = 0.021, **Figure 3A**) and PANSStotal score and thalamus within BGN (*r* = 0.369, *p* = 0.006, **Figure 3B**). The relationship was observed between PANSS-total score and thalamus within BGN by using spike-regression-based scrubbing procedure (**SFigure 2**). Moreover, no other significant correlations were found between the altered temporal variability of fALFF/ReHo and medication dosage in schizophrenia group.

### DISCUSSION

This study has presented some new insights in alterations of dynamic temporal variability of ReHo and fALFF in schizophrenia through sliding window analysis. Consistent with our hypothesis, increased dynamic temporal variability of ReHo and fALFF were observed in sensory and perceptual networks in schizophrenic subjects. Critically, the psychiatric symptom analysis has indicated that increased temporal variability of ReHo showed significantly positive relationship with the positive symptoms of schizophrenic subjects. These findings provide evidence that there is deficient temporal variability of local neural activity in low-level perceptual processing in schizophrenic subjects.

While these are well known about the abnormal higher-order brain function in schizophrenia, such as memory and cognitive (42, 43), neuroimaging studies have also documented some basic sensory processing deficits in schizophrenic subjects. The perceptual deficits have been increasingly observed in the sensory networks, including primary motor and visual regions (44, 45). A recent study has also revealed that schizophrenic subjects has shown increased resting-state functional connectivity variability in sensory and perceptual networks (46). Most of these locations were in line with the meta-analysis' results of schizophrenia (47). Increased variability of local neural activity of sensorimotor regions might reflect the deficits in the integration of multisensory stimuli in schizophrenia (48). Moreover, enhanced dynamic local neural activity might indicate that the abnormal bottom–up processing is associated with the pathological mechanism of schizophrenia (46). In this study, we observed increased temporal variability of TABLE 2 | Significant increased dynamic fALFF and ReHo in schizophrenic subjects.


ReHo and fALFF in sensory and perceptual system across time windows in schizophrenic subjects. These increased local temporal variabilities might provide some new evidences to support deficient dynamic neural activity in primary sensorimotor, as well as the abnormal dynamic bottom–up processing in schizophrenia.

Furthermore, schizophrenic patients could be commonly considered as a self-disorder with abnormal functional network (49). Recent studies have revealed that the processing and integration of multisensory bodily signals underlay a coherent self-experience in healthy controls (50, 51). In the "rubber-hand illusion" experiment, Botvinick and Cohen pointed out that the subjects would have true self-experience when they saw the fake hand was stroked, synchronous individual's unseen hand (50). Disturbances in self-experiences were also reported by Ehrsson; they found that visual perception was not match with proprioceptive information (51). These studies have provided the evidence that the sense of self-experiences depend on multisensory information that arose from the body, such as proprioceptive, spatial, and temporal sensorimotor signals. In schizophrenia, the deficits of visual and motor networks appear to be related to self-disorder (46, 52, 53). Besides, the neurobiological model of self-disorder has also indicated that deficient sense of self in schizophrenia is largely related to the abnormal multisensory signals integration from body and external stimuli (54, 55). Thalamus is a very crucial key role in gating and in integrating multisensory and cognitive information

in human brain. Thus, previous studies have indicated that the altered static function of the thalamus is an important feature related to the schizophrenic subjects' self-disorder symptom (56, 57). In this study, we found increased temporal variability of ReHo and fALFF in primary visual and somatosensory area in schizophrenic patients. These increased dynamic neural activity across time may be related to the high interaction within regional sensorimotor functional network in schizophrenia. Increased temporal variability of thalamus was also observed in schizophrenic subjects, which may suggest that abnormal dynamic functional integration across time in schizophrenia exists between multisensory regions and higher order cognitive functional system. A significant relationship was observed between increased dynamic ReHo of thalamus and PANSSpositive score. These findings indicated that schizophrenic subjects have altered dynamic local functional connectivity and local dynamic neural activity in thalamus regions. Moreover, increased local dynamic functional connectivity of the thalamus maybe related with a positive symptom of schizophrenic subjects. Therefore, the abnormal dynamic local neural activity within the visual, sensorimotor, and thalamus areas might provide more evidence about abnormal self-processing in schizophrenia.

While our results provide a new insight of dynamic functional activity for understanding the self-disorder in schizophrenia, several main methodological points of this study should be further addressed. First, dynamic temporal

FIGURE 3 | The relationship between altered temporal variability of dynamic ReHo and PANSS scores. (A) The positive association is observed between increased CV score of thalamus region and PANSS-positive score in schizophrenic subjects. (B) The PANSS-total score was also positively related with CV score of thalamus region in patients.

variability of ReHo and fALFF were calculated through sliding window correlation analysis. The size of window length is one parameter that does not have formal consensus, although we selected it based on the frequency of preprocessed data. Second, the patient we chose is chronic schizophrenic subjects. The antipsychotic treatment might have an effect on dynamic local

neural activity of patients. We should validate our findings in the first-episode schizophrenic subjects in further study. Third, self-experience assessment is not included in the current study. We should measure it and investigate the relationship between self-experience score and static/dynamic local neural activity in schizophrenic subjects.

### CONCLUSION

In conclusion, this study has combined resting-state fMRI and dynamic functional analysis. Our findings have revealed an increased temporal variability of ReHo and fALFF in primary visual and sensorimotor networks, as well as in the thalamus in schizophrenia patients. It has been showed that the increased dynamic neural activity of the thalamus was significantly related with a positive symptom of schizophrenic subjects. Thus, our findings might have potential interpretation for the neural physiopathology of self-disorder in schizophrenia.

### AUTHOR CONTRIBUTIONS

YZ made a substantial contribution to the conception and drafting and revising of the article. YZ, GG, and YT made a

### REFERENCES


substantial contribution to the analysis and interpretation of the data, and gave final approval of the version to be published.

### FUNDING

This scientific work was supported by grants from the National Nature Science Foundation of China (grant number: 81801775) and Advanced Talents Introduction Program of Chengdu Normal University (YJRC2017-4).

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2019.00484/ full#supplementary-material


tonic-clonic seizure. *Hum Brain Mapp* (2017) 38(2):957–73. doi: 10.1002/ hbm.23430


**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 Zhang, Guo and Tian. 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.*

# Dynamic Alterations of Spontaneous Neural Activity in Parkinson's Disease: A Resting-State fMRI Study

Chao Zhang1†, Binru Dou1†, Jiali Wang<sup>1</sup> , Kai Xu<sup>1</sup> \*, Haiyan Zhang<sup>2</sup> , Muhammad Umair Sami <sup>1</sup> , Chunfeng Hu<sup>1</sup> , Yutao Rong<sup>1</sup> , Qihua Xiao<sup>3</sup> , Nan Chen<sup>4</sup> and Kuncheng Li <sup>4</sup>

*<sup>1</sup> Department of Radiology, Affiliated Hospital of Xuzhou Medical University, Xuzhou, China, <sup>2</sup> Department of Radiology, Affiliated 2 Hospital of Xuzhou Medical University, Xuzhou, China, <sup>3</sup> Department of Neurology, Affiliated Hospital of Xuzhou Medical University, Xuzhou, China, <sup>4</sup> Department of Radiology, Xuanwu Hospital, Capital Medical University, Beijing, China*

Objective: To investigate the dynamic amplitude of low-frequency fluctuations (dALFFs) in patients with Parkinson's disease (PD) and healthy controls (HCs) and further explore whether dALFF can be used to test the feasibility of differentiating PD from HCs.

#### Edited by:

*Wenbin Guo, Central South University, China*

### Reviewed by:

*Xiuqin Jia, Capital Medical University, China Tianmei Si, Peking University Sixth Hospital, China Li Wang, Capital Medical University, China*

> \*Correspondence: *Kai Xu xukaixz@163.com*

*†These authors have contributed equally to this work*

#### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *07 June 2019* Accepted: *17 September 2019* Published: *01 October 2019*

#### Citation:

*Zhang C, Dou B, Wang J, Xu K, Zhang H, Sami MU, Hu C, Rong Y, Xiao Q, Chen N and Li K (2019) Dynamic Alterations of Spontaneous Neural Activity in Parkinson's Disease: A Resting-State fMRI Study. Front. Neurol. 10:1052. doi: 10.3389/fneur.2019.01052* Methods: Twenty-eight patients with PD and 28 demographically matched HCs underwent resting-state functional magnetic resonance imaging (rs-fMRI) scans and neuropsychological tests. A dynamic method was used to calculate the dALFFs of rs-fMRI data obtained from all subjects. The dALFF alterations were compared between the PD and HC groups, and the correlations between dALFF variability and disease duration/neuropsychological tests were further calculated. Then, the statistical differences in dALFF between both groups were selected as classification features to help distinguish patients with PD from HCs through a linear support vector machine (SVM) classifier. The classifier performance was assessed using a permutation test (repeated 5,000 times).

Results: Significantly increased dALFF was detected in the left precuneus in patients with PD compared to HCs, and dALFF variability in this region was positively correlated with disease duration. Our results show that 80.36% (*p* < 0.001) subjects were correctly classified based on the SVM classifier by using the leave-one-out cross-validation method.

Conclusion: Patients with PD exhibited abnormal dynamic brain activity in the left precuneus, and the dALFF variability could distinguish PD from HCs with high accuracy. Our results showed novel insights into the pathophysiological mechanisms of PD.

Keywords: Parkinson's disease, resting-state fMRI, dynamic brain activity, amplitude of low-frequency fluctuations, support vector machine

## INTRODUCTION

Parkinson's disease (PD) is a common neurodegenerative disorder characterized by progressive impairment of motor function and widespread non-motor symptoms, which affects patients' quality of life and is, hence, a significant social burden (1–3). At present, the pathophysiological mechanism of PD is not fully understood, and it is still a neuroimaging challenge to form a

**62**

definitive diagnosis at the early stage of the disease (1). Magnetic resonance imaging (MRI) has made great contributions in the clinical evaluation of PD (4, 5). Conventional MRI has been used to exclude secondary parkinsonism caused by neoplasms, vascular parkinsonism, and multiple sclerosis among others. The common imaging features of primary PD include iron deposition and substantia nigra atrophy (6, 7). However, several new reports have revealed that the cerebral region is widely involved in patients with PD. Therefore, it is important to explore novel imaging features that could help effectively identify PD.

In recent years, advanced neuroimaging techniques have allowed us to noninvasively explore the nature of the human brain in an efficient manner (8–10). Resting-state functional MRI (rs-fMRI) is an established tool to investigate the intrinsic neuronal activity of the human brain by measuring the amplitude of spontaneous low-frequency fluctuations (ALFFs) of bloodoxygen-level dependent (BOLD) signals (11–13). ALFF has been widely used to investigate regional brain activity in neurological diseases. Abnormal ALFFs in PD have been detected in extensive brain regions and act as an important characteristic related to subtypes of motor symptom or comorbidities (14). When compared with healthy controls (HCs), patients with tremordominant PD exhibited increased ALFF in the right cerebellar posterior lobe, while those with PD with postural instability/gait difficulty exhibited decreased ALFF in the bilateral putamen and cerebellar posterior lobe (15). PD patients with depression had significantly lower ALFF in the prefrontal cortex and anterior cingulated cortex than PD patients without depression (16). PD patients with apathy showed lower ALFF in the left orbital middle frontal gyrus and bilateral superior frontal gyrus (17) than PD patients without apathy. Further, PD patients with visual hallucinations showed lower ALFF in both lingual gyrii/cunei and greater ALFF in the temporo-parietal regions, medial temporal gyrus, and cerebellum than those with non-visual hallucinations and HCs (18). By measuring the local spontaneous brain activities, these studies provided satisfactory evidence that widespread cerebral regions were involved in PD, which greatly contributed to the understanding the neurobiological foundations of such disorders. However, the results of these studies were limited in that their focus on abnormal brain activities in PD with different motor/non-motor symptoms was based solely on group-level analysis; thus far, to our best knowledge, no study has used these abnormal features to distinguish PD with specific symptoms at an individual level.

Previous studies on aberrant static ALFF in PD are inconsistent as they report both lower and higher local brain activity in various cerebral areas. However, a static state analysis of ALFF ignores the dynamic characteristics of brain activity during the whole scanning period. Evidences from rs-fMRI that employ a sliding window approach have effectively detected dynamic functional connectivity features with higher sensitivity than the static state method (12, 19, 20). Recent studies reported that dynamic rs-fMRI analysis strategy not only made good contributions to human-brain exploration but also played an important role in studying the pathogenesis of schizophrenia (21–23). Nevertheless, few reports have focused on time-varying local spontaneous neuronal activity in PD, as ALFF itself exists with substantial fluctuations either (24, 25). Furthermore, support vector machine (SVM) is one of the most widely used supervised machine-learning approaches that can enable individual-level classification and prediction with high accuracy (26, 27). Uddin et al. (28) applied SVM to distinguish patients with autism from normal individuals with a classifier accuracy of over 80%. Accordingly, SVM has been proposed as an effective tool for diagnostic application in the clinic.

To date, no studies have evaluated the combined effects of dynamic ALFF and supervised machine-learning approaches on PD. Based on previous findings that a static ALFF has greater heterogeneity in PD and the proven property of dynamic methods, we hypothesized that dynamic intrinsic local spontaneous neuronal activity will show greater variability in subjects with PD than HCs, and that it may be another powerful index of rs-fMRI in exploring the underlying mechanisms of PD. We aimed to identify the cerebral regions that displayed abnormal dynamic local neuronal activity based on the voxel level of the whole brain. Furthermore, we applied SVM to observe whether the aberrant dALFF could be used as a feature to distinguish PD from HCs. This study may improve our understanding of the potential pathophysiological mechanism of PD. Moreover, we hope that this research can contribute to the clinical diagnosis of PD at an individual level.

### MATERIALS AND METHODS

### Subjects

This study was carried out in accordance with the tenets of the Helsinki Declaration and approved by the local ethics committee of Xuzhou Affiliated Hospital, Xuzhou Medical University. Written informed consent was obtained from all subjects before participation in the study.

The project used a convenience sample of 31 hospitalized patients with PD who met the UK Bank diagnostic criteria for PD (29). All patients underwent neuropsychological tests such as the Mini Mental Status Examination (MMSE) and the Montreal Cognitive Assessment (MoCA) and motor impairment evaluation including the Unified Parkinson's Disease Rating Scale (UPDRS) and the Hoehn and Yahr (H-Y) stage when patients were off medication. The included patients had no history of mental illness or other neurological diseases. The exclusion criteria of the participants were MRI-confirmed brain abnormalities such as trauma, stroke, tumor, and infection and contraindications to MRI such as claustrophobia and implanted metal devices. In addition, the subjects with a history of drug and/or alcohol abuse and syncope were also excluded. All patients underwent routine treatment, and none of them received any other relevant interventions. Thirty-two age- and sex-matched healthy volunteers were included as HCs. Three patients and four HCs were excluded because of head motion artifacts. Finally, 28 patients with PD (15 male and 13 female, 59.17 ± 9.72 years old) and 28 HCs (14 male and 14 female, 58.18 ± 6.46 years old) were included for analysis. There were no significant intergroup differences with respect to age and sex (**Table 1**). All patients underwent functional MRI scanning when they were off medication; all the HCs also underwent the same protocol

#### TABLE 1 | Demographics and clinical data.


*PD, Parkinson's disease; HCs, healthy controls; M, male; F, female; UPDRS-III, unified Parkinson's disease rating scale; H-Y, Hoehn and Yahr disability scale; MoCA, Montreal Cognitive Assessment; MMSE, Mini-Mental Status Examination.*

*Data are presented as the range and mean* ± *SD.*

#*The p-value was obtained using a chi-square test.*

\**The p-value was obtained by a general linear mode analysis.*

for MRI scanning and neuropsychological tests. All subjects were right-handed.

### MRI Data Acquisition

All participants were scanned in a 3.0 Tesla MRI scanner (GE Medical Systems, Signa HD, Waukesha, WI) with an eightchannel head coil. During the scan, comfortable foam pads were used to stabilize the head of each subject to minimize head motion, and all subjects wore earplugs to reduce the noise from the MRI machine. Then, an echo-planar imaging sequence was employed to acquire resting BOLD images. The parameters of the protocol are as follows: time of repetition, 2,000 ms; time of echo, 30 ms; field of view, 220 mm × 220 mm; slice thickness, 3 mm; slice gap, 1 mm; voxel size, 3.4 mm × 3.4 mm × 4.0 mm; number of slices, 36; flip angle, 90◦ ; and total volume of each subject, 185.

### Preprocessing of rs-fMRI Data

The rs-fMRI data preprocessing were carried out using data Processing and Analysis for (Resting-State) Brain Imaging (DPABI) (http://www.rfmri.org/dpabi) (30). The first 10 time points of each subject were excluded to stabilize the status and allow participants to adapt to the scanning condition. Slice timing was carried out on the remaining 175 volumes to correct time differences. Realignment was performed to correct head motion by using a Friston-24 model for individual-level correction, and any subject with a head maximum displacement >2 mm, maximum rotation >2.0◦ , or mean framewise displacement (FD) >0.3 was excluded. In our study, mean FD was set as a covariate for further group-level statistics to minimize the potential influences of head motion. Several covariates such as the Friston 24 head-motion parameters, cerebrospinal fluid signal, and white matter signal were regressed. Then, the processed volumes were normalized to the standard Montreal Neurological Institute (MNI) echo planar imaging (EPI) template with a voxel size of 3 mm × 3 mm × 3 mm. Finally, functional volumes were smoothed with 6-mm full width at half maximum. We did not carry out global signal regression of our data given that there is still some controversy regarding removal of the global signal (31–33).

### Dynamic ALFF Analysis

The analysis of dynamic amplitude of low-frequency fluctuations (dALFF) was carried out using Temporal Dynamic Analysis (TDA) toolkits based on DPABI (34). Before dALFF calculation, functional volumes were bandpass filtered (0.01–0.08 Hz) to minimize the influences of low-frequency drifts and fluctuations of the signal. The sliding window is an important parameter to capture dynamic spontaneous neural activities, and the proper window length is critical for dynamic analysis. Too small a window length may not allow robust estimation of dynamic changes, and too long a window length may not be able to detect dynamic activity. Previous studies provided the range of the appropriate window length as 10–75 TR, step = 1 TR) (12, 35). To maximize the statistical power, a moderate sliding window length of 50 TR (step = 1 TR) was selected. The post-processed 175 volumes of each subject were segmented into 126 windows in all. The ALFF was calculated in each sliding window. The standard deviation (SD) of ALFF values of each voxel across 126 windows was further calculated to assess the variability of ALFF. We also calculated the static ALFF containing the whole sliding window.

### Statistical Analysis

Two-sample t-test was used to observe intergroup differences in age and MoCA/MMSE scores. Sex-based group difference was evaluated using the chi-square test. A general linear mode (GLM) with age, sex, and mean FD as covariates was used to compare the difference of dALFF/ALFF between the PD and HC groups. Multiple comparisons were corrected using Gaussian Random-Field (GRF) method (voxel level, p < 0.001; cluster level, p < 0.05).

Partial correlation analysis was calculated between dALFF variability and disease duration/MoCA/MMSE/UPDRS/H-Y with age and sex as covariates (p < 0.05). All statistical analyses were performed using SPSS version 16 (SPSS Inc., Chicago, IL, USA).

### Support Vector Machine Analysis

The intergroup dALFF difference was used as the classification feature in this study. We then trained the SVMs by providing labeled observations, for which the classification results were known. To overcome the limitations of our samples, the leaveone-out cross-validation (LOOCV) method was applied to estimate the generalization ability of our classifier. To verify the ability of the validation strategy, we also made a classification comparison by introducing 10-fold cross-validations. Then, the total accuracy, sensitivity, and specificity were obtained to assess classifier performance.

A permutation test was used to evaluate the statistical significance of this classification accuracy (36). The permutation test was repeated 5,000 times, and during each time, the classifier randomly reallocated labels of PD and HC to the training subjects and repeated the entire classification process. The p-value was obtained after the total permutation was accomplished.

### RESULTS

### Demographics and Clinical Data

The details of age, sex, and MoCA/MMSE scores are listed in **Table 1**. The results showed no significant difference in age (p = 0.652), sex (p = 0.789), and MMSE (p = 0.831) between the PD and HC groups. However, the MoCA score of the PD group was significantly lower than that of the HCs (p < 0.05).

### Differences in ALFF/Dynamic ALFF and Correlational Analysis

The intergroup differences in dALFF are shown in **Figure 1** and **Tables 1**, **2**. Compared with HCs, significantly increased coefficient of variation (CV) of dALFF was noted in the left precuneus of PD patients (p < 0.001). In addition, we found that the CV of dALFF was positively correlated with disease duration (p < 0.001, r = 0.800) (**Figure 1**), and no significant correlation was found between dALFF variability and MoCA/MMSE/UPDRS/H-Y scores (**Supplementary Material**). There were no significant intergroup differences in ALFF.

### Classification Results

Classification results are shown in **Figure 2**. The accuracy of linear SVM classifier using LOOCV achieved an accuracy of 80.36%, sensitivity of 85.71%, and specificity of 75% (p < 0.001, non-parametric permutation approach). The receiver operating characteristic (ROC) curve of the classifier was 0.82. A 10-fold validation was also employed in our study to verify the reliability of the classification method, which generated a classifier accuracy of 71.43%.

### DISCUSSION

Upon literature review, we observed that only few studies employed a TDA method to explore the neural-activity characteristics of PD. The present study showed the following findings: (1) the dALFF of patients with PD compared to HCs was notably different in the left precuneus; (2) a significant correlation between CV of dALFF in the left precuneus and the course of the disease was found in PD; (3) dALFF in the left precuneus showed high accuracy in distinguishing between patients with PD and HCs.

Prior studies have noted the importance of cerebral static local neural activities in PD (4, 14, 15). To our best knowledge, dynamic changes in spontaneous neural activity has been very


FIGURE 1 | Increased CV of dALFF in the left precuneus displayed in coronal (A), sagittal (B), transverse (C), and three-dimensional view (D). Box plots with Whiskers (min–max) show the CV values in the left precuneus of the two groups (E), and scatterplots show the relationship between the CV in the precuneus of the PD group and the disease duration (F). CV, coefficient of variation; dALFF, dynamic spontaneous low-frequency fluctuation.

validation methods in PD groups, respectively.

poorly researched. The rs-fMRI analysis was based on the hypothesis that brain activity was in a stationary state during the entire scanning period; thus, the dynamic characteristics could not be identified. Dynamic algorithm was proven to represent the time-dependent characteristics of brain activity under the given scanning period. A recent study captured abnormal dALFF/ Regional Homogeneity (ReHo) in stroke patients by using TDA and compared it with HCs; they further found that variability in brain activity could be used to evaluate patients' motor function (24). Liu et al. (12) found abnormal functional network connection (FNC) through dynamic instead of static state. Dynamic FNC was significantly correlated with the frequency of epileptic seizures and the course of the disease. Furthermore, dynamic FNC could distinguish patients with idiopathic generalized epilepsy with generalized tonic–clonic seizures from controls with an accuracy of 77.91% through linear SVM classifier (p < 0.001). Dynamic algorithm showed the capability to characterize neural activity of the human brain by identifying specific function signatures.

Our dynamic algorithm showed that the dALFF in the left precuneus in PD was notably different than HCs. The result was partially consistent with previous reports that the left precuneus was a key cerebral region in patients with PD. Precuneus, which mainly constitutes the medial and posterior part of the parietal lobe, contributes to motor and cognitive tasks, and has been reported as displaying the highest resting metabolic rate among all cerebral regions (37). Interestingly, the precuneus consumed 35% more glucose than other brain regions in the resting brain (38). Perfusion single photon-emission computed tomography (SPECT) and [18F]fluorodeoxyglucose positron emission tomography ([18F]FDG-PET) proved that the precuneus was the most remarkable area of hypometabolism in the posterior cortical region (39). Another [18F]FDG-PETbased study on PD found that the metabolic capability of the left precuneus decreased with disease progression (40). Similarly, a more recent research using the arterial spin labeling (ASL) technique showed that cerebral blood flow (CBF) in the left precuneus significantly reduced in the PD group when compared with HCs (41). These functional imaging studies supported the view that the left precuneus might be more prone to attack in neuropsychiatric disorders. Additional rs-fMRI studies showed that the left precuneus was closely associated with motor and non-motor symptoms in PD. Hu et al. (42) found that increased local brain activity in the left precuneus was related to the Hamilton Depression Rating Scale score, by using static brain activity analysis. Thibes et al. (43) used a brain connection algorithm and showed that the left precuneus was a critical node connecting with specific cerebral regions in PD. In addition, morphological changes of the left precuneus were also found in PD with cognitive impairment and isolated apathy through voxelbased morphometry (44, 45). Therefore, the left precuneus is undoubtedly an important and vulnerable structure in patients with PD.

In our study, the higher variability of dynamic local brain activity level in the left precuneus was positively correlated with the course of PD. This meant that the degree of variation was significantly increased with an extended disease course, which reflected the increased or decreased brain activity at different sliding windows during the whole scanning period. These findings revealed a localized brain function impairment over time in PD. However, the abnormal dALFF did not correlate with UPDRS/H-Y scores in the present findings, likely because the heterogeneous motor symptoms in PD were associated with integration of multiple cerebral region function, rather than being determined by a single brain region impairment (46, 47). Unlike previous reports, the present study did not find significant intergroup differences with respect to ALFF, either because the sample size in our study was relatively small or because the result was not powerful enough to pass the multiple comparison correction of the present statistical methods. In fact, the dynamic features were concealed under the static analysis that represented a measure of the average amplitude of local activity across different scanning time points within the whole scan (12). Thus, static rs-fMRI may not be as sensitive as dynamic analysis to detect neural-activity changes. Our study indicated that dynamic analysis could completely unearth information of brain activity. In addition, the present result suggested that the left precuneus was an important structure involved in PD, and higher dALFF in this region was a promising imaging marker reflecting the disease duration. Besides, our findings did not show a correlation between dALFF and MoCA/MMSE tests. This may be because the MoCA and MMSE scales were mainly appropriate for cognitive screening, and our study lacked detailed assessment of cognitive performance compared to previous studies (48).

The imaging diagnosis of PD remains a challenge even now, as a confirmed diagnosis in most patients is still made depending on the clinical symptoms (49). An assessment of the iron content and volume of substantia nigra may be useful indicators to identify PD and evaluate the disease progression (6, 7). However, this approach has not been widely applied in the clinical management of patients with PD. Previous reports demonstrated that SVM was a powerful tool utilizing imaging features to distinguish PD patients from HCs. In our study, we tested the inter-group difference of dALFF in the left precuneus as a classification feature to discriminate PD from HCs through a linear SVM classifier. The accuracy of this classification was 80.36% when an LOOCV method was employed (non-parametric permutation correction, p < 0.001). Further, to compare the performance of SVM using LOOCV, a nested 10-fold cross-validation method was used to assess the classifier's performance; the accuracy was 71.43%. These findings showed that SVM could achieve better classification capability with LOOCV, and the results also provided evidence that patients with PD could be distinguished from HCs at the individual level when using dALFF variation in the left precuneus. These results support the hypothesis that the dALFF could identify individual PD patients.

Our study has some limitations. First, all patients were on medication. Although patients underwent fMRI scanning while they were off medication, the effects of the long-term treatment could not be completely ruled out. Second, the classification power based on the 28 PD patients was still not strong enough, and we just used SVM in the same sample to testify the classification accuracy. Third, the patients did not undergo comprehensive cognitive scales testing, which could have prevented a more accurate detection of cognitive performance. Future research should include a larger sample size and another independent test sample should be recruited for testifying classification accuracy.

## CONCLUSION

To our best knowledge, this is the first study to attempt to investigate the dynamic spontaneous neural activities in patients with PD. Our results provided evidence that dynamic analysis was more sensitive to detect alteration of brain activity than a static method. In addition, the CV of dALFF was found to be correlated with the course of the disease, which may ultimately contribute to identifying PD at the individual level. Thus, our results provide novel insights on the pathophysiological mechanisms of PD.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## ETHICS STATEMENT

The studies involving human participants were reviewed and approved by Ethical committee of Xuzhou Affiliated Hospital. The patients/participants provided their written informed consent to participate in this study.

## AUTHOR CONTRIBUTIONS

CZ and BD contributed equally to this work for the conception/design of the study, the acquisition, analysis, interpretation of data, drafting of the manuscript, final approval of the version to be published, and agreement to be accountable for all aspects of the research. JW, CH, YR, HZ, and MS were responsible for data analysis, drafting of the manuscript, final approval of the version to be published, and agreement to be accountable for all aspects of the research. QX, NC, and KL were responsible for revision of the manuscript, final approval of the version to be published, and agreement to be accountable for all aspects of the work. KX was responsible for design of the study, revision of the work, final approval of the version to be published, and agreement to be accountable for all aspects of the work.

### FUNDING

This study was supported by the National Natural Science Foundation of China (No. 81871339; No. 81271556) and Xuzhou Science and Technology Project (No. KC18051).

### REFERENCES


### ACKNOWLEDGMENTS

We thank all the volunteers who participated in this research.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2019.01052/full#supplementary-material

instability/gait difficulty subtypes of Parkinson's disease: a resting-state fMRI study. CNS Neurosci Ther. (2015) 21:855–66. doi: 10.1111/cns.12464


**Conflict of Interest:** 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 Zhang, Dou, Wang, Xu, Zhang, Sami, Hu, Rong, Xiao, Chen 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) 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