Characterizing resting‐state networks in Parkinson’s disease: A multi‐aspect functional connectivity study

Abstract Purpose Resting‐state functional magnetic resonance imaging (Rs‐fMRI) can be used to investigate the alteration of resting‐state brain networks (RSNs) in patients with Parkinson's disease (PD) when compared with healthy controls (HCs). The aim of this study was to identify the differences between individual RSNs and reveal the most important discriminatory characteristic of RSNs between the HCs and PDs. Methods This study used Rs‐fMRI data of 23 patients with PD and 18 HCs. Group independent component analysis (ICA) was performed, and 23 components were extracted by spatially overlapping the components with a template RSN. The extracted components were used in the following three methods to compare RSNs of PD patients and HCs: (1) a subject‐specific score based on group RSNs and a dual‐regression approach (namely RSN scores); (2) voxel‐wise comparison of the RSNs in the PD patient and HC groups using a nonparametric permutation test; and (3) a hierarchical clustering analysis of RSNs in the PD patient and HC groups. Results The results of RSN scores showed a significant decrease in connectivity in seven ICs in patients with PD compared with HCs, and this decrease was particularly striking on the lateral and medial posterior occipital cortices. The results of hierarchical clustering of the RSNs revealed that the cluster of the default mode network breaks down into the three other clusters in PD patients. Conclusion We found various characteristics of the alteration of the RSNs in PD patients compared with HCs. Our results suggest that different characteristics of RSNs provide insights into the biological mechanism of PD.

occurs in 77% of people with PD, which causes changes in the neural connections of the cortical and subcortical structures such as the basal ganglia, thalamus, and frontal cortices (Mimura, 2007). For example, functional connections between the posterior and anterior hubs of the default mode network in Parkinson's patients affect the deterioration of cognitive functions (Weil et al., 2019). Moreover, an emotional task fMRI study demonstrated abnormal emotional valence in subcortical limbic structures in PD (Bell et al., 2019).
To explain some of the motor and nonmotor deficits found in patients with PD, the abnormal pattern of spontaneous activity and disrupted connectivity of the brain networks can be captured by fMRI at rest (Brooks & Pavese, 2011). Resting-state fMRI (rs-fMRI) connectivity analysis has great potential in investigating the underlying mechanism of PD. The rs-fMRI reports the grid of brain areas that individually activated but functionally connected, known as resting-state networks (RSNs). The RSNs reflect spontaneous neural activity signals between correlated temporal regions of the brain (Rektorova, 2014). There are mainly two large RSNs in PD: salience network (SN) and default mode network (DMN) (Bressler & Menon, 2010;Buckner et al., 2013;Raichle, 2015;Yuan et al., 2016).
Previous studies have shown that alteration of the RSNs causes different emotional, cognitive, and behavioral impairment in neurodegenerative diseases (Lebedev et al., 2014) such as PD (Ghasemi & Foroutannia, 2019;Ghasemi et al., 2021;Wu et al., 2009) and other neurological and psychiatric diseases mainly Alzheimer's disease (Badhwar et al., 2017), dementia (Gratwicke et al., 2020), depression (Greicius et al., 2007), and schizophrenia (Whitfield-Gabrieli et al., 2009). RSNs are usually identified by the seed-based method (Biswal, 2012;Fernández-Seara et al., 2015;Tahmasian et al., 2015) or a data-driven method based on independent components analysis (ICA) (Beckmann et al., 2005;Erhardt et al., 2011). In the following, some related findings regarding RSN connection changes in PD will be mentioned.  found diminished functional connectivity (FC) of the right medial temporal lobe and the bilateral inferior parietal cortex within the DMN in PD patients compared with healthy controls (HCs) . Disbrow et al. (2014) assessed the DMN and executive RSNs in PD patients and HCs and found that neuropsychological investigations were not significantly different in the executive RSNs between two groups, although there was a reduced FC in the DMN of PD patients (Disbrow et al., 2014). Tahmasian et al. reported aberrant FC in bilateral inferior parietal lobule and the supramarginal gyrus in PD patients and found that these regions formed an interconnected network in PD patients, mainly with the DMN (Tahmasian et al., 2017). In another article, functional connections of the SN increase in group comparisons PD and HC (Navalpotro-Gomez et al., 2020). Tessitore's et al. also examined the salience, central executive, and DMN and showed that there was an increase in connectivity in the SN and DMN, as well as reduced connectivity in the central executive networks (Tessitore et al., 2017).
Previous studies have shown changes in resting brain network connectivity in PD patients, and limited studies have examined the disruption of the main RSNs in PD. what areas do RSNs break down when they are disrupted and disintegrated in PD? To answer these questions, we compared the connectivity of within and between RSNs among PD patients and HCs to investigate alterations of the brain network in PD. We sought to address the interindividual variability of the brain network in PD by proposing a subject-specific score-named an RSN score. After that, we investigated a voxel-wise comparison between extracted RSNs by randomizing method and a hierarchical clustering analysis between the RSNs based on connectivity measures in the PD patient and HC groups.

| Overall procedure
The overall procedures of this study are shown in Figure 1. After preprocessing the rs-fMRI data, independent component (IC) group extraction and statistical tests were performed. Twenty-three RSNs were separated from other noisy components by overlapping the spatial maps of ICs and resting-state template reference networks. After applying the dual-regression analysis and finding subject-specific IC maps, an RSN score was calculated for each IC of all subjects. Then, the RSN scores of all components were statistically compared among the two groups using the nonparametric Kruskal-Wallis test. Next, a comparison of one-by-one individual RSN maps was made between the two groups via randomization.
Basic network modeling was performed using the IC's time series using FSLNets (http://fsl.fmrib.ox.ac.uk/fsl/fslwi ki/FSLNets; RRID: SCR_002823) for PD and HC groups based on full and partial correlation measures. Finally, an additional structural MRI analysis was carried out with the FSL-VBM tool to investigate voxel-wise differences in the volume and topography of the gray matter in both the PD patient and HC groups.

| Patients
Data for this study were randomly selected from the Parkinson's Progression Markers Initiative database (PPMI; RRID: SCR_006431) (Marek et al., 2011

| Structural MRI analysis
In order to verify the typical pattern of atrophy in PD patients, we evaluated gray matter (GM) volume differences between patients with PD and HCs. The voxel-based morphometric analysis was performed on structural MRI data using the FSL-VBM toolbox (FSL, RRID: SCR_002823) (Ashburner & Friston, 2000). First, the skullstripped structural images of the brain were extracted, and the brain tissues were segmented (Zhang et al., 2001). The resulting gray matter volume images were normalized to the standard Montreal Neurological Institute (MNI) 152 using FNIRT. Voxel-wise nonparametric statistical test (n = 10,000) was applied in FSL to identify significantly different (p < .05; family-wise error [FWE]-corrected) GM voxels among the two groups using the threshold-free cluster enhancement (TFCE) technique (Smith & Nichols, 2009).
Moreover, data were temporally high-pass filtered at 0.01 Hz.

| Group independent components analysis
We generated RSNs using FSL's Melodic tool version 3.14 (Beckmann & Smith, 2004). The concatenated multiple fMRI datasets were decomposed using ICA to identify large-scale patterns of functional connectivity in the subject population. The IC maps were thresholded using the false-discovery rate at p < .05 (Beckmann & Smith, 2004).
The dataset was decomposed into 120 components, including artifactual and desired components. For extracting RSNs, we spatially correlated all IC maps to a set of 20 reference resting-state template networks (Laird et al., 2011) with threshold 0.3 using the "fslcc" tool in FSL. A list of brain regions associated with the resting-state template networks is presented in Table 2. We name these 20 desired regions as extracted RSNs (interested ICs) hereafter.

| Group difference in rsn score
After extraction of the group-specific RSNs by ICA, a dual-regression approach was used to identify subject-specific spatial maps and associated temporal dynamics for each subject. We defined "RSN score" for each individual IC as the mean value of spatial regression maps (parameter estimation) across all IC voxels. The RSN scores of all ICs were compared among the PD patient and HC groups using the nonparametric Kruskal-Wallis test to find ICs with significantly different scores (p < .05).

| Voxel-wise network analysis
A voxel-wise statistical test using a randomized approach was performed to compare the subject-specific maps of patients with PD and HCs for each IC (Nichols & Holmes, 2002). The general linear model (GLM) matrix, nonparametric permutation testing (10,000 permutations), and the "fslstats," "fslmaths," and cluster tools in FSL were used to find significant differences in the spatial maps of the PD patient and HC groups (Nickerson et al., 2017;Reineberg et al., 2015).

| Network modeling
The time-domain dependency of the RSNs makes a network model of connections. The strength of the connections between different regions leads to a hierarchical network. Brain network modeling based on hierarchical clustering reorders the ICs and brings together highly correlated components to form larger-scale networks.
Hierarchical clustering analysis was performed using the FSLNets tool for the PD patient and HC groups.

| Structural analysis
The results of our structural MRI analysis using VBM-FSL revealed no significant difference in the gray matter of PD patients and HCs (p > .33).

| Group RSN analysis
We identified 120 group ICs from 43 subjects. IC maps were thresholded at a level of 0.7 (threshold IC maps [signal > noise] = 0.7).
By overlapping these 120 ICs with the rest network templates, 97 ICs were associated with artifacts, and 23 ICs were identified as brain components ( Figure 2). It is noteworthy that some brain areas were associated with more than one IC (e.g., the superior and middle frontal gyri in IC 4 and 12, and the cerebellum in IC 33, 42, and 44). The anatomical locations of ICs were identified using the Harvard-Oxford cortical and subcortical atlases and are listed in Table 3.

| RSN scores in PD and HC
We extracted subject-specific RSN scores and examined the group differences as defined in Section 2.7. The RSN scores of 7 ICs in patients with PD were significantly smaller than that in HCs (p < .05). Statistical results and boxplot are shown in Table 4, and Figure 3, respectively. It is noteworthy that the average RSN scores of almost all ICs in patients with PD were smaller than that in HCs.

| Voxel-wise differences
Based on methods described in Section 2.8, we found that individual differences between PD and HC were associated with some voxels in 2 regions: the occipital pole and cerebellum ( Figure 4). The size and location of significant clusters identified by the dual-regression analysis are listed in Table 5.

| Connectivity clustering in two groups
We compiled the matrix of full correlation and partial correlation of and IC 41-marked as A in Figure 5. On the other hand, the partial correlation between them was not very strong-marked A′ in TA B L E 2 A list of brain regions associated with the resting-state template networks used in this research Figure 5-indicating indirect functional connectivity between these nodes. This is true for IC 8 and IC 38, marked as B and B′ in Figure 5.
The symmetric full and partial correlation connections were observed between IC 41 and IC 12-marked as C and C′ in Figure 5- as H and H′ in the top panel), while this connection became completely disconnected in PD (marked as a, and a′ in the bottom panel).
In addition, the direct connection between IC 54 and IC 33 was negative in HCs (marked as I in top panel), while this connection became positive in the PD patient group (marked as b in bottom panel).
The changes in the connection strength resulted in new hieratical clustering in the PD patient group, and therefore, the RSNs were grouped in 3 clusters in PD. Alteration of the RSN clustering from HCs to PD patients is shown in Figure 6, in which HCs have a cluster corresponding to the DMN, and this cluster dissolved into three other clusters in PD patients.

| D ISCUSS I ON
In our study, we examined changes in functional brain connectivity architecture on a whole brain and network level in patients with PD. Previous studies have reported that the connectivity of brain regions at rest was affected by PD (Baggio, Segura, Sala-Llonch,

TA B L E 3
The regions of extracted group RSNs. The peak coordinates of maximum intensity, cluster size, the mean, and standard deviation of the extracted RSNs (ICs) are listed. The peak coordinates are given in MNI space, and anatomical information is represented according to Harvard-Oxford Atlas. The maximum value reflects the peak value of the IC map  , 2015;Manes et al., 2018;Wu et al., 2009). In our study, various analyses were performed to investigate different aspects of the alteration of the whole-brain RSNs.
Large-scale disruption of the brain network in PD was evaluated by the intraaspects and interaspects of RSN connectivity associated with PD. While the majority of previous studies have investigated the alteration of functional connectivity in PD by using seed-based analysis, we employed a data-driven approach based on group ICA to extract and evaluate RSNs.
We identified 23 RSNs in the PD patient and HC groups, which were composed of several subnetworks. Previous studies also extracted RSNs using group ICA (Onu et al., 2015;Yao et al., 2014). Overall, other investigators have also found F I G U R E 3 Boxplot of the RSN scores of 7 ICs that were significantly different among PD patients and HCs. Note that the RSN scores of patients with PD are significantly smaller than those in HCs, which indicates a reduction in brain connections in PD F I G U R E 4 Voxel-wise spatial regression analysis to compare subjectspecific maps of patients with PD and HCs. Voxels that were significantly different in the two groups are overlaid with a blue color on the MNI standard atlas

TA B L E 5
The result of the voxel-wise comparison of the subject-specific spatial IC maps. Significant clusters identified by randomizing analysis are listed. Anatomical locations of the clusters, maximum, mean, and standard deviation of intensity, and coordinate (in MNI space) of the maximum intensity are listed a widespread and significant reduction in connectivity in PD patients compared with HCs at rest in different brain areas (Dubbelink et al., 2014).
The RSN score used in the current study has a number of advantages. First, the RSN scores were extracted from the whole-brain rs-fMRI dataset using an ICA and dual-regression method, without F I G U R E 5 The correlation matrix was extracted from the time series associated with 23 RSN components in HCs (top panel) and patients with PD (bottom panel). Each row or column is a set of correlations between a network node (IC) and all other nodes (ICs). The lower and upper triangular matrices in each panel represent the full and partial correlations, respectively. According to the strength of correlations, the matrix of the nodes was arranged and hierarchical clustering was made F I G U R E 6 Clusters of ICs in patients with PD and HCs. Note that 4 clusters in HCs were reduced to 3 clusters in patients with PD. The cluster with green color in HCs contains IC 29,76,39,54,and 51 and is related to the DMN. This cluster no longer exists as a separate cluster in patients with PD any a priori assumption on the pathophysiology of PD. This finding is in contrast to previous studies where parameters were extracted using predefined masks of regions. Second, this score can be used in statistical approaches and machine-learning algorithms to distinguish patients with PD from other groups.
We used the ICA dual-regression approach for extracting subject-specific spatial RSNs and deriving the functional connectivity network of group RSNs. Once the RSNs were estimated, their spatial distributions were compared voxel-wise among the two groups (Littow et al., 2010;Yu et al., 2015). We also considered both direct and indirect functional communication patterns between distinct RSNs. Based on dualregression analysis, we investigated the alteration of the brain network in PD compared with HCs and found disintegration and disruption of the RSNs, and in particular the DMN, in PD patients.
Functional connectivity between RSNs in our analyses indicated four main clusters of RSNs in HCs (top panel in Figure 5) and three main clusters in patients with PD (bottom panel in Figure 5 down).
Our results for healthy subjects are in agreement with previous studies that reported multiple clusters in the motor, executive, and visual networks in these subjects (Onu et al., 2015;Smith et al., 2013). As shown in Figure 6, the 4 clusters among HCs were reduced to 3 clusters in PD. resting-state network analysis is choosing and comparing with the template references of resting network and the limit number of predefined RSNs. However, in some regions, such as the cerebellum, we got more than one network.

| CON CLUS ION
In the current study, various analyses were performed to investigate different aspects of the alteration of the whole-brain resting-state networks. While the majority of the previous studies have investigated the alteration of the functional connectivity in PD using seedbased analysis, we employed a data-driven approach based on group ICA to extract and evaluate resting-state networks. Based on dualregression analysis, changes in the network structure, as well as the connectivity network, were analyzed. Functional communication patterns between distinct RSNs were examined. We found disintegration and disruption of the DMN in PD.

ACK N OWLED G M ENTS
The authors wish to thank Andrew J. Gienapp (Neuroscience Institute, Le Bonheur Children's Hospital and Department of Neurosurgery, University of Tennessee Health Science Center, Memphis, TN, United States) for editing the manuscript.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest. has supervised the project and provided access to crucial research components.

I N FO R M E D CO N S E NT
Informed consent was obtained from all individual participants included in this study.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/brb3.2101.

DATA AVA I L A B I L I T Y S TAT E M E N T
This clinical study will be conducted according to the protocol and in compliance with Good Clinical Practice (GCP), with the Declaration