Abnormal brain functional network dynamics in obsessive–compulsive disorder patients and their unaffected first‐degree relatives

Abstract We utilized dynamic functional network connectivity (dFNC) analysis to compare participants with obsessive–compulsive disorder (OCD) with their unaffected first‐degree relative (UFDR) and healthy controls (HC). Resting state fMRI was performed on 46 OCD, 24 UFDR, and 49 HCs, along with clinical assessments. dFNC analyses revealed two distinct connectivity states: a less frequent, integrated state characterized by the predominance of between‐network connections (State I), and a more frequent, segregated state with strong within‐network connections (State II). OCD patients spent more time in State II and less time in State I than HC, as measured by fractional windows and mean dwell time. Time in each state for the UFDR were intermediate between OCD patients and HC. Within the OCD group, fractional windows of time spent in State I was positively correlated with OCD symptoms (as measured by the obsessive compulsive inventory‐revised [OCI‐R], r = .343, p<.05, FDR correction) and time in State II was negatively correlated with symptoms (r = −.343, p<.05, FDR correction). Within each state we also examined connectivity within and between established intrinsic connectivity networks, and found that UFDR were similar to the OCD group in State I, but more similar to the HC groups in State II. The similarities between OCD and UFDR groups in temporal properties and State I connectivity indicate that these features may reflect the endophenotype for OCD. These results indicate that the temporal dynamics of functional connectivity could be a useful biomarker to identify those at risk.


| INTRODUCTION
Obsessive-compulsive disorder (OCD) is a group of neuropsychiatric diseases with obsessive thinking and compulsive behavior as the main clinical manifestations. OCD is common in the general population (2.5-3%) (Robbins, Vaghi, & Banca, 2019). Family members of OCD patients are at a higher risk for OCD compared with the general population (Nestadt, Grados, & Samuels, 2010), indicating a high genetic risk.
In recent years, the development of neuroimaging technology has provided a new way to explore the pathophysiological mechanisms underlying obsessive-compulsive disorder. Particularly, resting-state functional magnetic resonance imaging (rs-fMRI) has attracted attention due to a number of advantages: it is noninvasive, easy to perform, can be repeated, and avoids individual differences in the execution of tasks that might complicated the use of task-based fMRI (Barkhof, Haller, & Rombouts, 2014). Numerous rs-fMRI studies have identified abnormalities of the cortico-striato-thalamo-cortical circuit (CSTC) as a common characteristic in OCD patients (Calzà et al., 2019;Jung et al., 2013;Posner et al., 2014;van den Heuvel et al., 2016;Zhao et al., 2019). However, previous neuroimaging studies have suggested that abnormalities are not limited to the CSTC circuit and other regions (Anticevic et al., 2014;Hou et al., 2014;Milad et al., 2013), but also can be seen in cortical brain network connectivity (Fan et al., 2017a;Gürsel, Avram, Sorg, Brandl, & Koch, 2018;Shin et al., 2014).
Most previous rs-fMRI studies on OCD patients have investigated FC patterns as a static phenomenon. Recent studies have found that FC varies over time (Calhoun, Miller, Pearlson, & Adali, 2014) and such temporal fluctuations can be captured by dynamic functional network connectivity (dFNC) methods, providing greater insight into the fundamental properties of brain networks (Hutchison et al., 2013). Studies of a variety of psychiatric disorders have revealed that abnormal dFNC characteristics (Espinoza et al., 2019), including in autism spectrum disorder (ASD) (de Lacy, Doherty, King, Rachakonda, & Calhoun, 2017), schizophrenia (Rabany et al., 2019), and major depression . Previous dFNC studies of OCD patients have been limited in many ways. Gürsel and colleagues performed group-based independent component and sliding time window analyses to investigate dFNC alterations (Gürsel et al., 2020). They focused on a subset of networks (default mode network, frontoparietal network, and salience network) and did not examine whole brain connectivity patterns. Liu and colleagues examined firstepisode and treatment-naive patients with obsessive-compulsive disorder (OCD) (Liu et al., 2020) and did not examine patients undergoing treatment. Neither study examined nonaffected relatives in order to examine whether dFNC patterns might serve to identify endophenotypes for OCD.
Endophenotypes have been defined as "measurable components unseen by the unaided eye along the pathway between disease and distal genotype" (Gottesman & Gould, 2003). Endophenotypes can serve as a more direct indicator of a genetic component of a disease than overt disease symptoms. Endophenotypes are often present in unaffected family members at a higher rate than in the general population. Therefore, data from unaffected relatives of those with OCD is critical for identifying common endophenotypes that can be used for diagnosis and treatment. Relatives of OCD patients are more likely to suffer from OCD than the general population (Gottesman & Gould, 2003;Pauls, 2008;Pauls, Abramovitch, Rauch, & Geller, 2014).
The concept of endophenotype (Gottesman & Shields, 1973) has proven useful in helping to bridge the gap between genetics and behavioral disease processes and has been widely used in the study of psychiatric illnesses including OCD, schizophrenia, attention deficit hyperactivity disorder (ADHD), and depression (Chamberlain et al., 2008;Chamberlain & Menzies, 2009;De Vries et al., 2014;Gottesman & Gould, 2003;Gould & Gottesman, 2006;Menzies et al., 2007;Peng et al., 2015;Shaw et al., 2015;Viswanath, Janardhan Reddy, Kumar, Kandavel, & Chandrashekar, 2009). The clinical relevance and potential biomarker utility of dFNC in particular is supported by clinical studies of schizophrenia (Du et al., 2016), autism (Yao et al., 2016) and Parkinson disease (Kim et al., 2017). A primary goal of our study was to examine whether dFNC properties might serve as biomarker for OCD by directly comparing OCD patients with their unaffected first-degree relatives as well as healthy controls with no family history of OCD.
We performed group ICA on rs-fMRI and a sliding-window analysis to compare dFNC in OCD patients, their unaffected first-degree relatives (UFDR) and healthy control participants (HC). We hypothesized that (a) OCD patients would show altered dFNC, compared with healthy controls. (b) Clinical features in OCD would correlate with altered dFNC temporal properties and (c) UFDR may show dFNC disruption similar to that found in OCD.

| Participants
We enrolled 48 OCD patients, 24 UFDR, and 49 HC. All participants gave informed consent according to the institutional research and ethics committee of the Guangzhou Psychiatric Hospital. The groups were matched on age (range between 18 and 50) and gender. All participants were right-handed. OCD patients and their UFDR were recruited from the Guangzhou Psychiatric Hospital. HC were recruited through local and community advertisements. In order to maintain diagnostic consistency over time within our lab's OCD database, all patients were diagnosed according to DSM-IV criteria using the Structured Clinical Interview (SCID) for DSM-IV-TR Axis I disorders.
All participants were diagnosed by one experienced clinical psychiatrist and one experienced psychologist. All subjects gave written informed consent before participation.
Patients were excluded: (a) if they had a history of brain trauma or neurological disease; (b) if they had a history of alcohol or substance abuse. Twenty-six OCD patients were receiving treatment with selective serotonin reuptake inhibitors; all had been stable on their medication for at least 4 weeks. Details of the medications and dosages for OCD patients are provided in the supplementary materials (Table S1). Twenty-five patients had comorbidity with anxiety symptoms, and eight had depression symptoms. Comorbid anxious and depressive symptoms were not considered as an exclusion criterion, provided that OCD was the primary clinical diagnosis.
The exclusion criteria for UFDR and HC were same as those for the OCD patients. In addition, they were excluded if they reported any history of mental illness and/or treatment with any psychotropic medication as screened by using the SCID for DSM-IV-TR AXIS I disorders. If HC had a family history of any psychiatric disorders as defined by the DSM-IV, they were also excluded.

| Data acquisition and preprocessing
After preprocessing, two OCD patients were excluded due to head motion greater than 2 mm or 2 .

| Group independent component analysis
After data preprocessing, the resting state data was analyzed using spatial group independent component analysis (sGICA) as implemented in the GIFT software toolbox (GIFT v4.0a; http://icatb. sourceforge.net). Specific steps were as follows: (a) We used principal component analysis (PCA) to reduce the dimensionality of the data in two steps. We set the number of independent components to 100 in advance, process the data of single participants individually, and then connect all the participants' data into groups for processing. (b) The Infomax algorithm was used to estimate the independent components. It was repeated 100 times in ICASSO (Bell & Sejnowski, 1995) to ensure the stability of the results. (c) The independent components (including spatial maps and time series) for each subject were reconstructed in reverse and Fisher Z transformation was performed.
The data after Z transformation approximately obeyed the normal distribution with the mean value of the SD. (d) The Display GUI module in the GIFT toolbox was used to identified relevant network components. To identify which of the 100 ICs were meaningful, we chose those for which the peak activation coordinates were located primarily in gray matter and had low levels of spatial overlap with vascular, ventricular corresponding to artifacts (Allen et al., 2011). We further used Stanford functional ROIs (http://findlab.stanford.edu/ functional_ROIs.html) as templates to select ICs with high similarity to the templates. All ICs retained for analysis were located on gray matter, had low spatial overlap with cerebral ventricles and blood vessels, and had time courses dominated by low frequency signals (ratio of powers below 0.1 Hz to 0.15-0.25 Hz in spectrum) . A total of 39 meaningful ICs were identified according to these criteria. These ICs fell within the following functional networks: auditory (AUD), visual (VIS), sensorimotor (SMN), cognitive executive (CEN), default mode (DMN), and cerebellar (CB) networks ( Figure 1 and Table S2).
Additional postprocessing was applied to the time courses of the 39 meaningful independent components to remove remaining noise sources . Subject specific time courses were detrended and despiked using 3dDespike, then filtered using a fifthorder Butterworth low-pass filter with a high frequency cutoff of 0.15 Hz.

| dFNC analysis
The GIFT toolbox was applied to calculate dFNC through a sliding window analysis followed by k-means clustering. First, we used a sliding-window approach, in which a sliding time window of the 22-repetition time (TRs) method was applied to each participant, with a Gaussian window alpha value of 3 and a step between windows of 1 TR, resulting in 208 consecutive windows. We also analyzed the effect of different window lengths on the results. The results were highly consistent across a wide range of window sizes (18-26 TRs), suggesting that the identified altered dFNC was not caused by random artifacts related to window size (Figures S1-S4). To promote sparsity in the estimations, a penalty was imposed on the L1 norm of the precision matrix. We then applied a k-means clustering algorithm to the resulting 208 FC window FNC matrices for all the participants, which was iterated 150 times. The dFNC matrices of all participants were then clustered by using the k-means algorithm to assess the frequency and structure of the recurring FNC patterns. There are several different rules of thumb for determining the appropriate value of cluster number K Yang et al., 2021). In this study, we used the elbow criterion of the cluster validity index, and the optimal number of clusters was set as 2, which we refer to as State I and State II.
2.6 | Group differences in dynamic connectivity: Temporal properties and strength discovery rate (FDR) correction (Benjamini & Hochberg, 1995) between the three groups, applying a least significant difference post hoc test. The total number of connections taken into account by the multiple comparisons correction was 741.

| Clinical and neuropsychological data analysis
Statistical analyses were performed using SPSS 22.0 for Windows.

| Identification of states
Using the k-means clustering algorithm, we identified two different functional connectivity states that were recurrent throughout the rs-fMRI acquisition and across all participants, as shown in Figure 2.
State I was less frequent (27% of the scan duration) characterized by substantial between-network FC, particularly between ICs in the Auditory, Visual, Sensorimotor, and Default Mode networks. State II was more frequent (73% of scan duration) and was characterized by primarily within-network functional connectivity and minimal between-network functional connectivity.

| Within state connectivity differences between groups
We compared the strength of connections across three participant groups at each state by ANOVA (Table S4) In contrast with the comparison with the OCD group, in which differences were predominantly found in State II, when the UFDR Chi-square test was used to compare categorical variables across groups (OCD, UFDR, and HC). State I connectivity for OCD and UFDR that differs from State I connectivity in HC. In contrast, in State II UFDR were more similar to HC than to OCD: OCD differed from both HC and UFDR, whereas UFDR and HC showed relatively fewer differences when directly compared.

| Temporal properties of the dynamic states
We examined the proportion of time spent in each of the two states and frequency of switching to and from State I and State II. There was a significant group difference in the fraction of total time spent in each of the two states ( Figure 4a) (Figure 4c).  Table S3.

| DISCUSSION
This is the first study using dynamic functional network connectivity that compared obsessive-compulsive disorder patients with both their first-degree relatives and healthy controls. We identified two distinct connectivity states that were present across all three groups.

| Network connectivity within states
Our study provides additional insight into cortico-cortico network differences in OCD. Many previous studies focused on CSTC networks with the striatum as the core (Alexander, DeLong, & Strick, 1986). However, there is growing evidence that people with OCD have a wider range of brain network disorders (Fan et al., 2017b;Hou et al., 2012) We found within State I, generally characterized by between -network connectivity, that network connectivity between the DMN and other networks (SMN, VIS) was lower in OCD. The DMN has been related to the brain's monitoring of internal and external environment, emotional processing, creativity, self-reflection, and episodic memory extraction (Raichle, 2015). Our results are consistent with previous studies that also found lowered default network in functional activity and connectivity in OCD Shin et al., 2014). The influential "triple network model" in OCD (Anticevic et al., 2014;Beucke et al., 2013;Harrison et al., 2009) proposes that DMN is modulated by the salience network.
Our results were also consistent with Kwak and colleagues who indicated that not only within-DMN rs-FC but also functional connectivity between brain regions involved in the DMN were critical and that rs-FC features in somatosensory-motor, visual and auditory, and cinguloopercular networks were associated with clinical symptom severity improvement. (Kwak et al., 2020). We also found within State I that connectivity within motor and sensory networks (CB-AUD, CB-VIS, SMN-VIS) was greater in OCD compared with HC. The cerebellum plays an important role in cognition and emotion, in addition to motor function (D'Angelo & Casali, 2012). Previous research has associated OCD with abnormal cerebellar structure and function. At rest, OCD patients show lower spontaneous activity of the cerebellum (Hou et al., 2014), and the functional connection strength between cerebellum and the whole brain increases (Anticevic et al., 2014).
Within State II, OCD showed reduced between-network connectivity and increased within-network connectivity. This overall higher modularity is consistent with previous studies. Zhang et al. (2011) characterized networks using graph theory and found that OCD had abnormally higher clustering coefficient and shortest path length, both consistent with high modularity and network segregation. In contrast networks in HC controls were characterized by the small-world property, which has been shown to indicate an effective balance between modularization and decentralized information processing.

| Temporal dynamics of state transitions
We found that patients with OCD spent longer in State II, both overall (fraction window) and for each individual instance (dwell time), and less time in State I, than the HC group. These results reveal that OCD F I G U R E 5 Correlation between OCD symptoms (measured via the OCI-R) and the temporal properties for the OCD group. (a) Proportion of time (measured as fractional windows) in State I positively correlated with OCI-R score. (b) Proportion of time (measured as fractional windows) in State II negatively correlated with the OCI-R score patients were engaged across time in a brain configuration pattern characterized by a lack of between-network connections at the expense of strong within-network connections. Although there was no significant difference in number of transitions among the three groups, there was a trend for OCD to have fewer transitions than HC. These results are consistent with a study of schizophrenia and ASD, which found that both clinical groups displayed an increased time spent in a state of weak, intra-network connectivity (Rabany et al., 2019). In contrast, a study of Parkinson disease found that these patients showed more time in a State characterized by betweennetwork connections (Kim et al., 2017). In addition, the temporal properties were significantly associated with clinical features. The proportion of time in State I (fractional windows) was positively correlated with OCI-R scores and proportion of time in State II was negatively correlated with OCI-R scores, indicating that those with higher OCD severity spend more time in the highly segregated state. These findings emphasized the importance of dFNC studies, as it could reveal potential characteristic of OCD.
One previous study examined temporal properties of dFNC in OCD (Liu et al., 2020). It is hard to directly compare the studies because the data-driven algorithm extracted four different states in the Liu study between which the OCD group showed more frequent switching than controls, whereas our study the algorithm indicated two states with no significant difference in number of transitions. The two studies also differed in that Liu et al examined treatment naive unmedicated participants whereas in our study participants were receiving medication therapy. Future research will be needed to determine how these varying factors may affect functional connectivity in OCD.

| dFNC as an endophenotype for OCD
Our results show a number of similarities between OCD and their UFDR which may prove useful as a biomarker reflecting a shared endophenotype for OCD and UFDR. Previous studies have shown that dFNC may reflect various aspects of the neural system functional capacity (Deco, Jirsa, & McIntosh, 2011;Kucyi et al., 2016) and thus, may serve as a novel physiological biomarker of disease Hutchison et al., 2013 (Fan et al., 2016). With regard to within-state connectivity patterns we found evidence that State I connectivity may be an especially promising candidate to use as a biomarker of OCD. Overall OCD and UFDR showed greatest similarity in State I: there were no significant differences in connectivity between the groups in State I, whereas both OCD and UFDR showed multiple differences in connectivity when compared with HC. In contrast, in State II UFDR appeared to be more similar to HC than OCD. State I was the less frequent and highly integrated state, whereas State II was the more highly modular state.
Overall, then, UFDR show OCD like patterns of between-network integration in State I, but HC like patterns of modularity in State II.

| LIMITATIONS
Several limitations should be taken into account for the current study.
First, OCD patients were being treated with various types of therapy and medication. Therefore, it is possible that the results were influenced by these treatment conditions. However, we found that there were no significantly different in temporal properties of dFNC between medicated and unmedicated OCD patients (Table S5). Second, the sample size of UFDR was limited. Third, the relatively small sample size in this study and clinical heterogeneity of comorbidity across OCD patients, may have contributed variance to the study.
Future studies with larger sample size, OCD subgroups and drugnaive patients are needed to confirm our results. Fourth, the HC group scored higher on STAI than UFDR. This may be due to the fact that many of the controls were college students who participated in the experiment during the final exam period, thus leading to their high anxiety scores.

| CONCLUSION
We identified several dFNC differences in OCD and UFDR that may be useful for establishing biomarkers of an OCD endophenotype.
First, OCD patients had abnormal temporal properties which correlated with clinical features, which were shared with their UFDR. Second, we found similar connectivity patterns for UFDR and OCD within a dFNC state characterized by between-network integration.
These results provide new insights into the pathophysiology of OCD patients and indicate dFNC measures that could be used as biomarker to identify those at risk.

DATA AVAILABILITY STATEMENT
The original data are available from the first author on request.