Characterizing Rapid Fluctuations of Resting State Functional Connectivity in Demyelinating, Neurodegenerative, and Psychiatric Conditions: From Static to Time-Varying Analysis

Functional magnetic resonance imaging (fMRI) at resting state (RS) has been widely used to characterize the main brain networks. Functional connectivity (FC) has been mostly assessed assuming that FC is static across the whole fMRI examination. However, FC is highly variable at a very fast time-scale, as demonstrated by neurophysiological techniques. Time-varying functional connectivity (TVC) is a novel approach that allows capturing reoccurring patterns of interaction among functional brain networks. Aim of this review is to provide a description of the methods currently used to assess TVC on RS fMRI data, and to summarize the main results of studies applying TVC in healthy controls and patients with multiple sclerosis (MS). An overview of the main results obtained in neurodegenerative and psychiatric conditions is also provided. The most popular TVC approach is based on the so-called “sliding windows,” in which the RS fMRI acquisition is divided in small temporal segments (windows). A window of fixed length is shifted over RS fMRI time courses, and data within each window are used to calculate FC and its variability over time. Sliding windows can be combined with clustering techniques to identify recurring FC states or used to assess global TVC properties of large-scale functional networks or specific brain regions. TVC studies have used heterogeneous methodologies so far. Despite this, similar results have been obtained across investigations. In healthy subjects, the default-mode network (DMN) exhibited the highest degree of connectivity dynamism. In MS patients, abnormal global TVC properties and TVC strengths were found mainly in sensorimotor, DMN and salience networks, and were associated with more severe structural MRI damage and with more severe physical and cognitive disability. Conversely, abnormal TVC measures of the temporal network were correlated with better cognitive performances and less severe fatigue. In patients with neurodegenerative and psychiatric conditions, TVC abnormalities of the DMN, attention and executive networks were associated to more severe clinical manifestations. TVC helps to provide novel insights into fundamental properties of functional networks, and improves the understanding of brain reorganization mechanisms. Future technical advances might help to clarify TVC association with disease prognosis and response to treatment.

Functional magnetic resonance imaging (fMRI) at resting state (RS) has been widely used to characterize the main brain networks. Functional connectivity (FC) has been mostly assessed assuming that FC is static across the whole fMRI examination. However, FC is highly variable at a very fast time-scale, as demonstrated by neurophysiological techniques. Time-varying functional connectivity (TVC) is a novel approach that allows capturing reoccurring patterns of interaction among functional brain networks. Aim of this review is to provide a description of the methods currently used to assess TVC on RS fMRI data, and to summarize the main results of studies applying TVC in healthy controls and patients with multiple sclerosis (MS). An overview of the main results obtained in neurodegenerative and psychiatric conditions is also provided. The most popular TVC approach is based on the so-called "sliding windows," in which the RS fMRI acquisition is divided in small temporal segments (windows). A window of fixed length is shifted over RS fMRI time courses, and data within each window are used to calculate FC and its variability over time. Sliding windows can be combined with clustering techniques to identify recurring FC states or used to assess global TVC properties of large-scale functional networks or specific brain regions. TVC studies have used heterogeneous methodologies so far. Despite this, similar results have been obtained across investigations. In healthy subjects, the default-mode network (DMN) exhibited the highest degree of connectivity dynamism. In MS patients, abnormal global TVC properties and TVC strengths were found mainly in sensorimotor, DMN and salience networks, and were associated with more severe structural MRI damage and with more severe physical and cognitive disability. Conversely, abnormal TVC measures of the temporal network were correlated with better cognitive performances and less severe fatigue. In patients with neurodegenerative and psychiatric conditions, TVC abnormalities of the DMN, attention and executive networks were associated to more

INTRODUCTION
The human brain at resting state (RS) exhibits highly structured spontaneous fluctuations in functional magnetic resonance imaging (fMRI) data, which reflect the underlying network architecture (Biswal et al., 2010). RS functional connectivity (FC) captures the temporal associations between such fluctuations, and has been successfully used to characterize the main networks of the brain and map abnormalities of functional network architecture occurring in different neurological conditions. In healthy controls, RS FC strength was found to be associated to age, with RS fluctuations being strongest in adulthood and lowest in children and elderly (Mak et al., 2017). A dependency of connectivity from sex (Biswal et al., 2010;Mak et al., 2017), as well as from cognitive, emotional, and behavioral variables was also detected (Kelly et al., 2012).
Multiple sclerosis (MS) is an inflammatory and neurodegenerative disease of the central nervous system leading to a progressive increase over time of clinical disability and cognitive impairment (Filippi et al., 2017(Filippi et al., , 2018. Reorganization of brain functional networks in MS has been shown from the first RS fMRI studies (Lowe et al., 2002(Lowe et al., , 2008Rocca et al., 2010;Roosendaal et al., 2010), which is thought to limit the clinical consequences of widespread tissue damage (Filippi et al., 2013a;Sbardella et al., 2015). Cortical reorganization has been demonstrated to be variable across the different stages of the disease, and a progressive exhaustion or inefficiency of the adaptive properties of the cerebral cortex is likely to be among the factors responsible for the worsening of clinical disability (Rocca et al., 2010(Rocca et al., , 2018Roosendaal et al., 2010;Loitfelder et al., 2011). In neurodegenerative conditions, RS FC studies showed a progressive and gradual spreading of connectivity changes from a target brain network, reflecting specific behavioral and cognitive dysfunctions (Zhou et al., 2017). In psychiatric diseases, disruption of fronto-parietal network connectivity seems to be the common fingerprint across distinct forms of pathology (Baker et al., 2019).
However, current understanding of the role of functional abnormalities in neurological and psychiatric disorders is still incomplete, mostly due to inconsistencies in the findings from several studies. Specifically, in MS some investigations found trends toward lower RS FC vs. healthy controls in the defaultmode (Rocca et al., 2010(Rocca et al., , 2012(Rocca et al., , 2018Bonavita et al., 2011), sensorimotor (Rocca et al., 2018) and subcortical (Liu et al., 2011;Rocca et al., 2018) networks, while in other studies the opposite trends were observed (Roosendaal et al., 2010;Tona et al., 2014;Schoonheim et al., 2015). Similarly, even if RS FC abnormalities were principally located in the core regions hit by pathology, a certain variability of brain areas involved by RS FC changes was detected in neurodegenerative and psychiatric conditions (Busatto, 2013;Weiner et al., 2017).
The wide spectrum of clinical characteristics of MS patients has been considered as one of the main causes for the discrepancies described in RS fMRI literature (Filippi et al., 2013a;Sbardella et al., 2015). However, technical factors might also bias connectivity estimation, including scanner-related signal instabilities, an inappropriate control of confounding covariates, and the application of analysis methods based on inaccurate assumptions.
For instance, one of the main assumptions of classical RS FC assessment methods is that connectivity is static across the entire fMRI examination, e.g., it can be assessed by calculating the mean correlation between whole-length RS fMRI time series (Biswal et al., 2010). However, as widely evident by neurophysiological techniques, brain FC is highly variable at a very fast time-scale. The functioning human brain during any state of wakefulness repeatedly changes between different combinations of cognitive, sensorimotor, attentional, emotional, auditory, and visual-related tasks. Notably, the majority of brain regions experience continuous functional changes even during sleep (Tagliazucchi and van Someren, 2017). Thus, studying time-varying RS FC patterns is likely to shed light not only on physiological processes occurring in healthy subjects, but also to understand clinical manifestations of different neurological and psychiatric conditions. In fact, clinical symptoms associated to these diseases are likely to depend not only from damage to specific brain regions, but also from delayed (or abnormal) communication between brain areas. The study of the temporal reconfigurations of FC occurring within RS fMRI sessions has been defined as time-varying functional connectivity (TVC) (Hutchison et al., 2013;Calhoun et al., 2014;. The main goal of this review is to summarize the main results obtained using TVC in healthy and diseased populations. A particular focus is given to studies of patients with MS; however, the main findings of investigations performed in neurodegenerative and psychiatric conditions are also reported. The review is structured as follows: in section Methods Used to Assess Time-Varying Functional Connectivity, we present the main approaches developed to investigate TVC using fMRI data, with a main emphasis on the methods applied to study MS patients. Then, we summarize the results obtained applying these methods in healthy controls (section Application of TVC to Healthy Subjects) and in patients with MS (section Application of TVC Techniques to MS). An overview of the results derived from other neurological and psychiatric conditions is also given (section Application of Time-Varying FC Techniques to Psychiatric and Other Neurological Diseases). In the final part (section Current Limitations and Future Directions), current TVC methodological limitations are discussed and possible future developments are presented.

METHODS USED TO ASSESS TIME-VARYING FUNCTIONAL CONNECTIVITY
Several analysis strategies have been applied so far to quantify temporal variations of blood oxygenation level dependent (BOLD) signal fluctuations (Hutchison et al., 2013;. Some strategies aim at capturing variations in interregional associations between pairs of brain areas (Sakoglu et al., 2010;Allen et al., 2014), while others try to detect changing patterns of temporal synchrony at a multivariate level, e.g., considering all brain regions at once (Tagliazucchi et al., 2012;Liu and Duyn, 2013). One of the most popular methods for TVC analysis, which is based on the use of the so-called "sliding windows" (Sakoglu et al., 2010;Allen et al., 2014), belongs to the first category, since it relies on the calculation of a series of pairwise correlation coefficients over small shifting segments of fMRI time series.
Despite the great variability of available pipelines, TVC analysis usually requires the performance of the following steps: (1) selection of a set of regions of interest (ROIs) in the brain; (2) assessment of time-varying correlations among the selected ROIs; and (3) extraction of features quantifying connectivity changes over time, as described in details in the next paragraphs.

Selection of Regions of Interest for TVC Analysis
It is important to properly identify the ROIs (which may be areas of the brain, or even entire functional networks) that will be included in TVC analysis. Several factors can influence the choice of ROIs: spatial resolution, the use of a priori hypotheses or datadriven strategies, and the rationale of the experiment, which may focus on selected functional circuits or on the whole brain.
The large majority of studies assessing TVC in MS patients mainly relied on the use of a priori atlases, such as the Automatic Anatomical Labeling (AAL) (Tzourio-Mazoyer et al., 2002) or the Desikan (Desikan et al., 2006) cortical atlas (Leonardi et al., 2013;Lin et al., 2018;van Geest et al., 2018a,b). Some studies built ad-hoc ROIs centered in critical nodes of large-scale brain networks (Bosma et al., 2018). However, a widely used approach in previous literature consists in a ROI data-driven selection through independent component analysis (ICA; Rocca et al., 2010Rocca et al., , 2019Sakoglu et al., 2010;Filippi et al., 2013b;Allen et al., 2014;Damaraju et al., 2014;Yang et al., 2014;Zalesky et al., 2014;Bisecco et al., 2018;Castellazzi et al., 2018;d'Ambrosio et al., 2019) (Figure 1). The broad application of ICA in previous TVC studies can be explained by the flexibility of this approach, which allows to extract ROIs at different spatial resolution according to ICA dimensionality, to perfectly fit the data (avoiding non-linear registrations with a priori atlases, which may be challenging in diseased populations) and to reduce the impact of physiological and motion-related noise.
Since ROIs identified by "static" a priori atlases may not reflect significant connectivity variations occurring within brain regions at short time scales (Ryyppo et al., 2018), recent studies have suggested that incorporating information of time-varying connectivity between neighboring voxels to parcellate the brain may improve accuracy of TVC analyses Ryyppo et al., 2018).

Sliding Window Analysis
The most popular strategy used to examine time-varying correlations between RS fMRI time series relies on the use of sliding windows (Sakoglu et al., 2010;Allen et al., 2014). In this approach, a time window of fixed length is selected, and correlations between pairs of fMRI time series are calculated using data within that window. Then, the window is shifted in time by a certain number of time points, and correlations are re-assessed on the new data. This procedure results in a series of pair-wise correlation matrices that describe the time-resolved behavior of connectivity over the entire duration of the fMRI experiment ; Figure 1).
The choice of an appropriate length for sliding windows is crucial: too short time segments may introduce spurious fluctuations associated with intrinsic fMRI signal instability, while with increased window size TVC estimation may become too similar to the classic static FC (Leonardi and Van De Ville, 2015;. Different validation analyses recommended to set window length around 30-60 s (or the equivalent time expressed as repetition times, TRs), demonstrating consistent reproducibility of the obtained results Damaraju et al., 2014;Rashid et al., 2014Rashid et al., , 2016Zalesky et al., 2014;Leonardi and Van De Ville, 2015;Qin et al., 2015;Zalesky and Breakspear, 2015;Choe et al., 2017;Zhang C. et al., 2018).
Once sliding windows correlation matrices have been produced, different strategies can be applied to extract features describing connectivity reorganization through time inside the data (Leonardi et al., 2013;Allen et al., 2014;Miller et al., 2016), as described in details in section Extraction of Features Quantifying Time-Varying Connectivity.

Beyond Sliding-Window Analysis
A variety of approaches alternative to sliding windows have been developed to quantify TVC in fMRI data . For instance, time-frequency decomposition has been used to represent correlations between two fMRI time series in the joint time and frequency domain (Chang and Glover, 2010;Yaesoubi et al., 2015a; Figure 1). Point-process analysis allowed to detect recurring patterns of co-activation between brain regions from a small fraction of the total scans of a RS fMRI experiment (Tagliazucchi et al., 2012;Liu and Duyn, 2013). Phase coherence connectivity has been proposed to calculate RS FC at each recorded fMRI time point (Deco and Kringelbach, 2016).
In MS studies, two alternative methods to sliding windows have been applied. One study (Bosma et al., 2018) used dynamic FIGURE 1 | Schematic representation of the post-processing steps used in the assessment of time-varying functional connectivity (TVC). Top row: selection of regions of interest for TVC analysis, which can be done using data-driven approaches (e.g., independent component analysis, A) or using a priori atlases (B). Middle row: assessment of time-varying correlations between fMRI time series. The most popular approach consists in using a sliding-window analysis (C); alternative approaches, such as time-frequency analysis (D) or point-process analysis (E) have been also proposed. Bottom row: extraction of features quantifying connectivity changes over time, which can be done using several techniques, such as graph theory (F), k-means clustering to estimate recurring TVC states (G), or fuzzy meta-state analysis (H). ICA, independent component analysis; AAL, automatic anatomical labeling; ACC, anterior cingulate cortex; CAP, co-activation pattern; MCC, middle cingulate cortex; PCC, posterior cingulate cortex; SPG, superior parietal gyrus; MFG, middle frontal gyrus; R, right; L, left.
conditional correlations (DCC) to quantify TVC. DCC were originally proposed to study fluctuations of financial time series (Engle, 2002) and subsequently adapted to neuroimaging data to quantify time-varying variances and correlations between multivariate RS fMRI time series (Lindquist et al., 2014). DCC overcome some limitations intrinsic to sliding-window techniques, since they do not depend from any arbitrary window length and do not give the same weight to all time points within the window, ignoring older observations. Moreover, DCC are not easily confused by changes of correlation occurring in fMRI time series merely due to random noise (Lindquist et al., 2014).
Another study (Zhou et al., 2016) quantified connectivity reorganization over time using brain entropy (BEN). Entropy is a statistical and physical index that measures irregularity of a time-varying system (Sandler, 2006). In RS fMRI data, voxel-wise assessments of entropy were performed by calculating sample entropy, defined as the negative logarithm of the probability that if two time series of length m have a correlation < r, then two time series of length m+1 also have a correlation < r. A higher entropy indicates increased randomness of a system, meaning that the time-varying system activity is less predictable and less organized (Wang et al., 2014).

Extraction of Features Quantifying Time-Varying Connectivity
Sliding-window (or alternative) techniques produce a large amount of correlation data, calculated on several time segments. Some features have then to be extracted from this big data mass, to summarize to what extent functional relationships reorganize through time. The simplest summary TVC statistic is standard deviation (or variance) of sliding-window correlation time series (Sakoglu et al., 2010;Choe et al., 2017) or of DCC time series (Lindquist et al., 2014;Bosma et al., 2018). The mean TVC (Huang et al., 2019) or the sum of absolute differences in pairwise connectivity between consecutive windows have also been used as a summary TVC measure (van Geest et al., 2018a,b). Another interesting metric assessing temporal variability of BOLD fluctuations is the so-called amplitude of low-frequency fluctuations functional connectivity (ALFF-FC) (Shen et al., 2016), which sums up the spectral content of low-frequency RS fluctuations through consecutive sliding windows.
Flexibility metrics quantifying time-varying global and regional network properties were also calculated using a graph theory framework , as described in section Graph Theoretical Analysis. More complex strategies rely on the identification of connectivity patterns that reoccur over time during the course of the experiment. Reoccurring RS FC patterns, often called "states, " can be determined using clustering techniques , principal component analysis (Leonardi et al., 2013) or tensor decomposition (Mokhtari et al., 2018a), as detailed in section Definition of Reoccurring Connectivity States. Finally, approaches overcoming a rigid data decomposition into "fixed" connectivity states have been recently proposed , as described in detail in section Fuzzy Meta-State Analysis.

Graph Theoretical Analysis
Graph theory analysis can be applied to series of matrices derived from sliding-window analysis (Figure 1). Besides the classical network metrics (Rubinov and Sporns, 2010), which can be quantified as a function of time (Fukushima and Sporns, 2018), more specific metrics can be used to assess time-varying network structure. For instance, network power measures the summed values of TVC pairs in all windows, and density estimates how dense, on average, connections are over time. Specific timeresolved network features include network variation, which describes how different are connectivity values between two adjacent windows, flexibility of homologous, non-homologous and intra-hemispheric connections, which quantify connectivity differences between two consecutive windows for the specified type of connections , or the Fiedler value, which summarizes how well-connected a network is . Recently, novel approaches have been proposed to improve modeling of brain network TVC using graph theory (Khambhati et al., 2018). Such modeling strategies aim to assess time-varying patterns of connectivity (e.g., dynamic community detection or non-negative matrix factorization), time-varying patterns of activity, or a combination of both. A detailed review of these methods is reported in Khambhati et al. (2018).

Definition of Reoccurring Connectivity States
One of the most diffuse approaches used to identify reoccurring FC states from sliding-window matrices is based on hardclustering algorithms , such as the kmeans algorithm . In this approach, data are partitioned into different connectivity states by maximizing a cluster validity index, which describes the betweencluster/within-cluster distance ratio. In this way, identified recurring connectivity states have a minimal degree of overlap . The amount of time spent in each recurring state (dwell time) and the number of transitions between states can be calculated and compared between groups (Figure 1). Between-group comparisons can be also performed on pair-wise TVC strengths within each detected state .
Other ways to identify FC states from sliding-window data rely on principal component analysis (PCA) (Leonardi et al., 2013) or tensor decomposition (Mokhtari et al., 2018a). PCA is able to decompose sliding-window matrices into patterns of correlated connectivities (called "eigenconnectivities") between brain regions. Each eigenconnectivity pattern is characterized by a "contribution" (which can be thought as the equivalent of dwell times for k-means clustering analysis), which varies over time across subjects. Between-group comparisons of such contributions may allow to characterize TVC abnormalities in patients' populations (Leonardi et al., 2013). Similarly, tensor decomposition (Mokhtari et al., 2018a) is able to decompose sliding-window connectivity matrices in a set of components, each with an associated weight, which explain the majority of data content.

Fuzzy Meta-State Analysis
In hard-clustering analysis, windowed correlation matrices are forced to fit into determined TVC recurring states. However, the existence of just one state at each time point may be a too rigid assumption. A more flexible approach is to consider the possibility that multiple states might be represented to varying degrees at the same time point. The contribution of each state for a specific time is characterized by a vector that is called a "meta-state" . Four different measures of neural reorganization over time can be associated to such meta-states and can be calculated for each study subject: (1) the total number of distinct meta-states that a subject assumes during the experiment; (2) the number of changes between distinct meta-states; (3) the range of meta-states occupied in the n-dimensional meta-state space during the entire RS fMRI experiment; and (4) the total distance traveled in the n-dimensional state space (Figure 1).

Main TVC Findings in Healthy Subjects
The results of the main studies assessing TVC in healthy controls are summarized in Table 1.
In healthy subjects, it was always possible to identify a certain number of recurring connectivity configurations (from 3 to 12, depending on the method applied and on RS fMRI sequence Frontiers in Neuroscience | www.frontiersin.org     (Chang and Glover, 2010;Liu and Duyn, 2013;Lindquist et al., 2014;Zalesky et al., 2014) and when considering connections between the DMN and other crucial cognitive networks (Chang and Glover, 2010;Liu and Duyn, 2013;Allen et al., 2014;Marusak et al., 2017;Vidaurre et al., 2018). High dynamism was also observed in multimodal brain regions, involved in high-order emotional and cognitive processing (Yang et al., 2014;Zalesky et al., 2014;Marusak et al., 2017;Vidaurre et al., 2018). Such quick temporal reconfigurations may be required to facilitate transient psychological states between different brain functions (starting, maintenance or conclusion of the different attentional, cognitive, and executive tasks). Conversely, networks involved in sensory and motor processing showed more "static" connectivity profiles Zalesky et al., 2014). TVC was also useful to characterize age-and sex-related features. For instance, it was shown that children have higher TVC between the DMN and other subnetworks than young adults, but that young adults have stronger TVC than children among sensorimotor, executive control, and auditory networks (Cai B. et al., 2018). Moreover, variability of TVC among cognitive networks increased with age (Marusak et al., 2017). Overall, these results suggest that maturation is associated with a higher flexibility of functional connections. More discrepancies were found when analyzing sex-related characteristics of connectivity dynamics (Yaesoubi et al., 2015a,b;Nini et al., 2017). While some studies found no differences in TVC configurations between males and females (Yaesoubi et al., 2015b), other studies found that connectivity configurations were different between genders (Yaesoubi et al., 2015a,b;Nini et al., 2017): males showed a higher connectivity flexibility than females in the amygdala, hippocampus, fusiform, and temporal gyrus, whereas the opposite trend was found in the middle cingulate cortex, thalamus, precuneus, and some temporaloccipital regions (Nini et al., 2017).
TVC constitutes a complex and novel methodology. Studies from healthy controls also served to test how reliable and reproducible TVC results were across scanning sessions. This was the goal of some recent investigations Smith et al., 2018;Zhang C. et al., 2018), which found that TVC metrics were reliable across days  and had an overall good reproducibility Smith et al., 2018;Zhang C. et al., 2018), even if lower than that of the corresponding static FC metrics (Zhang C. et al., 2018). The highest reliability was found for intra-network connections in the DMN, fronto-parietal, sensorimotor, and occipital networks (Zhang C. et al., 2018).
To better investigate the intrinsic nature of TVC states and their electrophysiological correlates, simultaneously acquired electroencephalography (EEG) and RS fMRI data were analyzed and concurrent temporal variations were assessed (Allen et al., 2017). Results indicated that connectivity states detected by TVC analysis correspond to neuro-electric brain activity with distinct spectral signatures. Moreover, eyes open/eyes closed conditions show some common and some different connectivity patterns, with a greater integration within sensory systems, as well as reduced modularity and increased global efficiency, in the eyes-closed compared to the eyes-open condition (Allen et al., 2017). These results integrate and complete previous EGG/RS fMRI studies, which showed a variable TVC configuration between wakefulness and different stages of sleep (Tagliazucchi et al., 2013), with temporal memory and long-range temporal dependencies decreasing from wakefulness to deep non-rapid eye movement sleep.

Correlations Between TVC and Behavioral/Neuropsychological Performances in Healthy Subjects
To date, correlations between TVC measures and cognitive performances in healthy controls have been evaluated by one study (Cabral et al., 2017), which found that worse cognitive performance in healthy elderly was associated with a lower permanence in a TVC state characterized by strong, positive connectivity. These results suggest that a more static pattern of TVC may characterize poor vs. good performers.
Another study  analyzed the correlation between TVC and scores obtained at questionnaires of subjective well-being, and found that subjects with higher well-being scores spent less time in low cross-network and strong within-network connectivity states. The total number of transitions between states was also higher in subjects with high well-being scores, suggesting a more efficient transfer of information between networks in this group . Finally, two studies assessing the relationship between TVC and mindfulness in healthy adults (Lim et al., 2018) and children (Marusak et al., 2018) had similar conclusions, showing that high-mindfulness subjects spent more time in highly-connected states and switched more frequently between states than low-mindfulness subjects, suggesting a more efficient and flexible connectivity in the first group.

Main TVC Findings in MS Patients
The main studies assessing TVC abnormalities in MS patients are summarized in Table 2. As it is evident from this table, TVC methodologies applied in MS investigations were quite heterogeneous. Despite this, results of different studies share some common points.
First of all, networks showing the greatest amount of TVC abnormalities in MS patients in comparison to healthy subjects were the DMN, salience, executive and sensorimotor networks (Leonardi et al., 2013;Zhou et al., 2016;Bosma et al., 2018;Lin et al., 2018;d'Ambrosio et al., 2019;Rocca et al., 2019).
The regional pattern of TVC abnormalities was quite complex, and regions involved by TVC changes were variable across studies, probably depending from the used TVC approach and patients' clinical characteristics. The analysis of eigenconnectivity patterns helped to identify the presence of stronger TVC in parietal regions and weaker TVC in frontal/subcortical regions  in relapsing-remitting MS patients with mild to moderate disability compared to healthy controls (Leonardi et al., 2013). These patients also showed more frequently strong connections in temporal and parietal (angular gyrus) regions as well as weaker connections in motor and amygdalar regions vs. control subjects (Leonardi et al., 2013). Another study assessing TVC abnormalities in relapsing-remitting MS patients with mild disability found an increased BEN (corresponding to an increased connectivity disorganization) of regions involved in motor, executive and spatial coordination, as well as reduced BEN in memory brain areas (including temporal and hippocampal cortices) and relay areas as the cerebellum or the brainstem compared to healthy subjects (Zhou et al., 2016). A recent study using DCC to quantify TVC (Bosma et al., 2018) confirmed the results obtained by Leonardi et al. (2013), and found an increased BOLD signal variability in posterior regions of the DMN in MS patients vs. controls. The same study also found an increased TVC between the salience network and the ascending nociceptive pathway. Conversely, divergent results were obtained by Lin et al. (2018), who showed an overall reduction of network variation in MS patients compared to healthy controls, suggesting a globally more "static" FC configuration, but at the same time found an increased flexibility of interhemispheric connections, which was interpreted as a compensatory mechanism for the decreased global connectivity. A complex pattern of increased and decreased TVC was also shown by d'Ambrosio et al.
(2019), who found, in a multicenter study, a selective TVC increase between subcortical and visual/cognitive networks, and a TVC decrease between subcortical and sensorimotor networks in relapsing-remitting MS patients compared to healthy controls (d'Ambrosio et al., 2019).
Specific investigations of crucial systems involved in cognitive functions were performed by Van Geest et al., who studied TVC of the hippocampal network van Geest et al., 2018a,b) and of the DMN van Geest et al., 2018a,b) and by Huang et al., who investigated the attention network (Huang et al., 2019). Overall, hippocampal and DMN TVC were not different between MS patients and control subjects; however, including TVC measures in multivariate statistical models contributed to explain the performance of MS patients at visuospatial memory van Geest et al., 2018a,b) and information processing speed van Geest et al., 2018a,b) tasks. Huang et al. detected a complex pattern of TVC abnormalities, which was characterized by a TVC decrease within the dorsal and ventral attention networks, as well as TVC increase between the same networks (Huang et al., 2019).
Changes in TVC at the earliest stages of the disease have been rarely assessed, but interesting results have been observed. Patients with clinically isolated syndrome (CIS) suggestive of MS exhibited, early after the first demyelinating attack, reduced TVC in the functional networks more affected by the clinical onset, compared to healthy controls (Rocca et al., 2019). These patients also showed, in the first 2 years after the clinical event, a progressive increase over time of TVC strength, mainly between the DMN and sensorimotor/visual/cognitive networks, combined with a progressive increase over time of global fuzzy meta-state dynamism (Rocca et al., 2019).
Overall, these results suggest that, at the beginning of the disease, TVC dysfunctions have a specific correspondence with clinical symptoms. Then, a progressive increase of TVC oscillations occurs, probably trying to compensate disease-related damage. This initial phase seems to be followed by a loss of coordination and flexibility among brain regions in MS patients (Leonardi et al., 2013;Zhou et al., 2016;Lin et al., 2018;d'Ambrosio et al., 2019), which may be compensated by local increased fluctuations between specific areas van Geest et al., 2018a,b).
Recent studies tried to investigate TVC changes in MS populations affected by specific clinical manifestations. In details, cognitive impairment in patients with MS was associated to reduced TVC between subcortical and DMN areas, as well as to reduced global dynamism, compared to cognitively preserved patients (d'Ambrosio et al., 2019). Patients with MS suffering from neuropathic pain expressed selectively reduced TVC strength in the salience-descending nociceptive circuit (Bosma et al., 2018), whereas in patients without such neuropathic pain, TVC strength was increased in the same network (Bosma et al., 2018).
The large majority of the above-mentioned studies assessed TVC changes in relapsing-remitting MS patients, while detailed investigations of TVC abnormalities occurring in progressive MS phenotypes or over the course of the disease, are still missing.

Correlations Between TVC and Clinical, Neuropsychological, and Structural MRI Variables in MS
Different correlation analyses have been performed in MS patients, in order to understand the possible association between TVC abnormalities and motor and cognitive performances, as well as with specific clinical symptoms such as fatigue.
A higher expanded disability status scale (EDSS) score, reflecting more severe clinical disability, was found to be correlated with increased BEN in the bilateral supplementary motor area and in the right precentral operculum (Zhou et al., 2016), as well as with a more rigid (less fluid) global connectivity in MS patients . Conversely, other studies failed to show significant associations between TVC abnormalities and disability, probably because of the relatively low sample size and/or a narrow EDSS range (Leonardi et al., 2013).
Several correlations have been detected between TVC abnormalities and MS patients' cognitive performances. In particular, better scores in tests involving executive control functions and processing speed ability were correlated with a higher global network dynamism . Similar findings were shown by van Geest et al. (2018a), who found that a higher dynamism in the DMN during an information processing speed task vs. a resting state condition was associated with better information processing speed performances. These results are in agreement with the reduced network dynamics observed in cognitively impaired vs. preserved MS patients (d'Ambrosio et al., 2019). Conversely, a lower hippocampal TVC contributed to explain, at least partially, better verbal learning, visuospatial learning, and memory performances (van Geest et al., 2018b).
Lower fatigue was associated with reduced TVC in the parahippocampal gyrus, right posterior cerebellum, and brainstem (Zhou et al., 2016). Pain interference has been associated with increased TVC in the posterior cingulate cortex, an associative region involved in the salience and nociceptive networks and DMN (Bosma et al., 2018).
A few studies investigated the relationship between TVC and white matter lesions or MS-related structural damage. Decreased TVC between parietal and fronto-temporal regions of the attention network was associated with an higher lesion load in relapsing-remitting MS patients (Huang et al., 2019). A significant association has been demonstrated between reduced global dynamism in cognitively impaired MS patients and brain atrophy (d'Ambrosio et al., 2019), as well as between increased TVC and diffuse microstructural damage in relapsing-remitting MS patients, quantified as a higher mean diffusivity on diffusion-tensor MRI (Zhou et al., 2016). At the earliest stages of MS, a progressive increase of TVC over 2 years of follow-up was associated with a lower white matter lesion volume change over the same period of time (Rocca et al., 2019).
In summary, in MS patients, abnormal global TVC properties of the sensorimotor, DMN and salience networks were associated with more severe tissue damage at structural MRI, more severe clinical disability, worse cognitive performance and pain interference, evidencing a maladaptive neuronal response to direct disease-related damage. Conversely, abnormal TVC properties of the temporal network and relay areas as the cerebellum and brainstem were correlated with better cognitive performances and less severe fatigue, suggesting a compensatory role of TVC changes.

Main Findings in Psychiatric and Other Neurological Diseases
The main studies discussed in this section are summarized in Table 3.
In neurological disorders, TVC abnormalities have been mainly observed in areas directly affected by the disease. For example, subcortical stroke and mild traumatic brain injury patients showed TVC abnormalities in sensorimotor networks Vergara et al., 2018). Patients with myoclonic/frontal lobe epilepsy showed reduced TVC mainly in frontal and parietal brain regions, whereas patients with temporal lobe epilepsy experienced TVC decrease mainly in temporal regions (Ridley et al., 2017;Klugah-Brown et al., 2018;Wang et al., 2018). Generalized epilepsy was related to TVC strength changes mainly in the DMN and cognitive networks . Patients suffering from Alzheimer's disease had reduced regional (nodal) TVC (Alderson et al., 2018) and alterations in inter-network TVC of the anterior and posterior regions of the DMN (Jones et al., 2012;Quevenco et al., 2017), the frontal cortex and temporal areas (Jie et al., 2018). Patients with Parkinson's disease showed TVC changes mainly in sensorimotor, executive, cognitive , visual, and DMN areas (Diez-Cirarda et al., 2018), combined with reduced global and nodal TVC Diez-Cirarda et al., 2018).

Correlations Between TVC and Clinical, Neuropsychological, and Structural MRI Variables in Psychiatric and Other Neurological Diseases
In schizophrenia patients, reduced global time-resolved graph metrics have been related to structural disease-related damage (Yu et al., 2015), while abnormalities in TVC of auditory brain regions have been correlated with the presence of auditory hallucinations . Hallucinations were also         correlated with a more rigid, reduced global dynamism Mennigen et al., 2018). Autistic behavior and diagnosis were associated with longer dwell times in a globally disconnected state (Rashid et al., 2018a). In patients with temporal lobe epilepsy, recurring states characterized by high inter-network TVC expressed reduced dwell time and correlated with an early seizure onset (Klugah-Brown et al., 2018). Interestingly, reduced TVC in the ictal irritative zone was associated to an intracranial EEG connectivity increase in the same epileptic region in alpha, beta and gamma bands (Ridley et al., 2017). In patients with Alzheimer's disease, TVC abnormalities between the anterior and the posterior DMN areas correlated with poorer episodic memory performance (Quevenco et al., 2017), while reductions in global TVC were associated with microstructural tissue damage (Alderson et al., 2018). In patients with Parkinson's disease, TVC abnormalities in the DMN have been associated with memory performance (Engels et al., 2018), while TVC alterations in the putamen were associated with clinical disability .
At this moment, TVC approaches are applicable only at a group level. However, some preliminary investigations have successfully used TVC abnormalities to classify schizophrenia patients from bipolar patients and/or healthy controls Rashid et al., 2016), suggesting a future application of TVC at an individual level.

CURRENT LIMITATIONS AND FUTURE DIRECTIONS
The field of TVC is relatively new: all main technical developments have been achieved in the last 9 years. Nonetheless, in such short period of time TVC has provided greater insights into fundamental properties of functional networks, and has improved knowledge of the pathophysiological brain reorganization occurring in MS and other neurological and psychiatric diseases.
However, TVC methodology presents some inherent limitations that are likely to be overcome in the next future. Further investigations are also needed to better understand the physiological meaning of TVC fluctuations and their electrophysiological correlates.

How Reliably Can Time-Varying Fluctuations Be Detected From fMRI Data?
One of the main pitfalls of current TVC analysis approaches consists in the fact that the mere presence of signal fluctuations in an fMRI time series is often taken as an evidence of TVC (Hindriks et al., 2016). This might be not necessarily true: FC values fluctuating over time might be observed just because of noise, or statistical uncertainty. Several measures have been employed to test the effective presence of FC variability in fMRI time series, including variance (Sakoglu et al., 2010), standard deviation (Chang and Glover, 2010), kurtosis (Laumann et al., 2017), or more complex, non-linear measures (Zalesky et al., 2014). Usually, these metrics are compared between real fMRI data and simulated data, constructed ad-hoc to have a static FC. If the test is significant, the null hypothesis of stationarity can be rejected, and TVC can be considered to be effectively present in the data.
Results of studies assessing evidence of TVC in RS fMRI data were quite disappointing, showing that the power of TVC detection in typical 10-min RS fMRI acquisitions was relatively low (Leonardi and Van De Ville, 2015;Hindriks et al., 2016;Zhang C. et al., 2018). Solutions to improve the likelihood of detecting TVC might be the choice of appropriate lengths for sliding windows (Leonardi and Van De Ville, 2015) or the concatenation of more RS fMRI sessions (Hindriks et al., 2016). On the other hand, it is possible that measures used to test the hypothesis of dynamism so far might be not fully appropriate . Indeed, novel wavelet-based metrics  seem to be more sensitive to capture non-stationarities present in real RS fMRI data.

Improving Temporal Resolution of fMRI Acquisitions
Results of TVC also depend upon the temporal resolution used to acquire fMRI data. TVC studies usually investigate modifications in RS FC occurring within seconds, by using fMRI volumes acquired with TRs ranging from 1 s to 3 s (Chang and Glover, 2010;Allen et al., 2017;Cabral et al., 2017;Nini et al., 2017;Yaesoubi et al., 2017b;Marusak et al., 2018). Investigations performed on RS fMRI data acquired with a higher sampling rate, e.g., thanks to the use of simultaneous multislice imaging techniques, may be more powerful in detecting changing connectivity reconfigurations over time . Also, the use of ultra-fast fMRI acquisition techniques, such as inverse imaging (Lin et al., 2006), generalized inverse imaging (Boyacioglu and Barth, 2013), or multi-slab echovolumar imaging (Posse et al., 2013), might constitute an important improvement for TVC. Ultra-fast fMRI allows to acquire a single functional volume covering the whole brain in <300 ms, resembling the results of magnetoencephalography studies (Asslander et al., 2013). Therefore, fMRI scans acquired with ultra-fast techniques do not include physiological aliasing and allow the detection of more accurate BOLD signal responses to neural activity. Seminal studies already showed that ultra-fast fMRI significantly enhanced the sensitivity of mapping RS FC dynamics (Posse et al., 2013).

Improving TVC Pre-and Post-processing
Regardless of the analysis method, the signal-to-noise ratio of the BOLD signal in RS fMRI is low, especially in small temporal segments (Handwerker et al., 2012). Non-neural processes contaminating RS fMRI time series can affect TVC estimates (Hutchison et al., 2013;Murphy et al., 2013;. These confounds often include the effects of motion, cardiac and respiratory activity, and fluctuations in arterial CO 2 concentration (Hutchison et al., 2013;Murphy et al., 2013;Nikolaou et al., 2016;Glomb et al., 2018). Global signal regression (GSR) may be useful to better denoise RS fMRI time series (Murphy and Fox, 2017); however, it was shown to slightly reduce reliability of the estimated TVC connectivity states . Moreover, the impact of GSR was spatially heterogeneous across brain regions and was dependent from the amount of global signal magnitude across windows (Xu et al., 2018). As such, caution is suggested in applying GSR to slidingwindow correlation analyses, and a control of subjects' mental fluctuations during RS fMRI scanning is recommended (Xu et al., 2018). By applying accurate pre-processing steps on the fMRI data, the rate of artifacts present in the TVC analyses will be minimized (Murphy et al., 2013), thus increasing the quality of the observed findings.
Improvements can still be done not only to pre-processing of RS fMRI time series, but also to TVC post-processing, e.g., by implementing new, accurate methods to estimate changing connectivity over time. Recent papers proposed new approaches to analyse TVC, which aim at capturing change points of connectivity in functional correlation matrices (Cribben et al., 2012;Jeong et al., 2016;Kundu et al., 2018). Other studies introduced tensor-based multilayer community detection algorithms, which are able to describe how organization of functional networks evolves over time (Al-Sharoa et al., 2019). All these methods might be useful to complement TVC information obtained by using more standard, state-of-art methods, such as sliding-window analysis. Finally, improvements can still be done in statistical thresholding strategies. TVC assessment relies on the use of a massive amount of pairwise correlations, stored in series of connectivity matrices, and the best way to perform a proper adjustment for multiple comparisons is still an open issue. Traditional methods of correction for multiple comparisons (Friston et al., 1994) may be too conservative and may suppress all significant results; therefore, different approaches of adjustment for multiple comparisons might be more suitable. For instance, network-based statistic (NBS, Zalesky et al., 2010) was proposed as an alternative method of multiple comparison correction in studies using graph theory, which suffers of similar drawbacks as TVC. NBS has been rarely applied in TVC studies (Diez-Cirarda et al., 2018), probably because the process to construct "components" is not straightforward when connectivity matrices change over time. Future studies investigating new strategies of adjustment for multiple comparisons may propose new solutions for this issue.

Functional Interpretation of TVC Findings
Although some studies have tried to provide a functional interpretation of TVC output, several questions remain to be answered by future work. Preliminary data found some degree of correspondence between EEG rhythms and TVC frequency content (Allen et al., 2017) and hypothesized that some of the TVC states observed in healthy subjects, especially at the end of RS fMRI sessions, might be related to drowsiness or light sleep (Allen et al., , 2017. A preliminary study assessing contemporary TVC and EEG registrations confirmed the presence of connectivity changes over different phases of sleep, with long-range temporal dependencies becoming weaker during deep sleep (Tagliazucchi et al., 2013). Still, it is not clear why larger TVC oscillations have been registered in functional networks at the beginning of RS fMRI sessions (Allen et al., , 2017. Theories hypothesizing the brain functional "anticipation" (e.g., brain predisposition to switch quickly between different psychological states; Zalesky et al., 2014) might partially explain why more "specialized" functional networks (sensorimotor, auditory, visual) express more constant TVC behavior, while more complex, multi-modal networks express more dynamism. On the other hand, constant TVC oscillations in sensorimotor, auditory and visual networks might only reflect their lower activity during RS (Syed et al., 2017).
From this perspective, additional multi-modal studies integrating information from imaging and electrophysiological modalities are necessary for a better comprehension of the neural origin, mechanisms and function of temporal FC variations, as well as of the physiological meaning of TVC states.

CONCLUSIONS
The analysis of time-varying FC has contributed to provide significant information on intrinsic brain functional organization, both in healthy and diseased conditions, which complements data produced by static FC approaches. TVC seems to be an intrinsic property of the brain with a neural origin, although some open questions still remain about the correct interpretation of TVC output.
In MS patients, TVC helped to better understand the pathophysiological functional reorganization occurring in the brain, with a peculiar involvement of the DMN, salience, sensorimotor, and fronto-temporal networks. TVC abnormalities were partially correlated with more severe tissue damage and more severe clinical disability, while more extensive correlations were found with abnormal cognitive performances. In patients with neurodegenerative and psychiatric conditions, TVC abnormalities of the DMN, attention and executive networks were also associated to stronger clinical manifestations. Overall, these results suggest a maladaptive neuronal response to disease-related damage.
There are still several unmet needs in neurological and psychiatric conditions that TVC analysis may help to address. First, TVC may be useful to identify multi-modal regions, crucial for functional network plasticity, which may constitute possible targets for motor and cognitive neurorehabilitation protocols, as well as for symptomatic or disease-modifying treatments. Second, trajectories of TVC changes over time during the disease course need to be better defined, both in MS and in psychiatric/other neurodegenerative disorders. This may be the topic of future longitudinal studies, or of cross-sectional studies enrolling patients at different disease phases. Finally, it is still unclear whether TVC abnormalities may have a prognostic value on future disease course. The collection of clinical data at medium-or longterm follow-up may allow to define whether some TVC abnormalities are associated with a more favorable/worse disease prognosis.

AUTHOR CONTRIBUTIONS
MH wrote the first draft of the manuscript. PV, MF, and MR drafted/revised all sections of the manuscript. MR contributed to the study concept and acted as study supervisor. All authors contributed intellectually to manuscript revision, read, and approved the submitted version.