Abstract
Generalized anxiety disorder (GAD) is a common anxiety disorder experiencing psychological and somatic symptoms. Here, we explored the link between the individual variation in functional connectome and anxiety symptoms, especially psychological and somatic dimensions, which remains unknown. In a sample of 118 GAD patients and matched 85 healthy controls (HCs), we used multivariate distance-based matrix regression to examine the relationship between resting-state functional connectivity (FC) and the severity of anxiety. We identified multiple hub regions belonging to salience network (SN) and default mode network (DMN) where dysconnectivity associated with anxiety symptoms (P < 0.05, false discovery rate [FDR]-corrected). Follow-up analyses revealed that patient’s psychological anxiety was dominated by the hyper-connectivity within DMN, whereas the somatic anxiety could be modulated by hyper-connectivity within SN and DMN. Moreover, hypo-connectivity between SN and DMN were related to both anxiety dimensions. Furthermore, GAD patients showed significant network-level FC changes compared with HCs (P < 0.01, FDR-corrected). Finally, we found the connectivity of DMN could predict the individual psychological symptom in an independent GAD sample. Together, our work emphasizes the potential dissociable roles of SN and DMN in the pathophysiology of GAD’s anxiety symptoms, which may be crucial in providing a promising neuroimaging biomarker for novel personalized treatment strategies.
Introduction
Generalized anxiety disorder (GAD) is a common mental disorder in the adult population with a combined lifetime prevalence of 3.7% (Wittchen and Hoyer 2001; Mackenzie et al. 2011; Ruscio et al. 2017), which is characterized by excessive and uncontrollable psychological and somatic symptoms (Association 2000; Mennin et al. 2004; Stein and Sareen 2015; DeMartini et al. 2019). Although psychological and somatic anxieties are closely related clinically, however, it could be possible that their underlying mechanisms are different. Accumulating evidence has shown that pharmacotherapy has differential treatment outcomes in GAD patients with somatic and/or psychological anxiety symptoms (Lydiard et al. 2010). For example, tricyclic antidepressants (TCAs), selective serotonin reuptake inhibitors (SSRIs), mixed serotonin-noradrenaline reuptake inhibitors (SNRIs), and 5-hydroxytryptamine 1A (5-HT1A) agonists exhibit greater efficacy in managing psychological anxiety symptoms than somatic anxiety symptoms (Pollack et al. 2001; Katz et al. 2002; Rickels et al. 2003; Meoni et al. 2004), while benzodiazepines, facilitating inhibitory GABAergic transmission, have been shown to be more efficacious in case of somatic anxiety symptoms (Rickels et al. 1988; Rickels et al. 1993). Thus, understanding the potential dissociable neural mechanisms of psychological and somatic anxiety has important implications for developing novel symptom-oriented therapeutic strategies.
To date, functional magnetic resonance imaging (fMRI) has proven to be very important in understanding large-scale brain functional organization and studying neural circuits related to GAD. Previous fMRI studies have reported abnormal activation of default mode network (DMN) and salience network (SN) hubs, including the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC) (Zhao et al. 2007; Buff et al. 2016), dorsal anterior cingulate cortex (Paulesu et al. 2010; Palm et al. 2011; Blair et al. 2012), supplementary motor area (SMA) (Wang et al. 2018), and insula (Buff et al. 2016; Cui et al. 2020), in patients with GAD. In addition, there is evidence showing that anxiety severity is inversely correlated with the functional connectivity (FC) between the PCC and the mPFC (Andreescu et al. 2014). However, some studies have reported the opposite FC pattern of the DMN. For example, FC between the PCC and mPFC has been found to be elevated in GAD patients (Rabany et al. 2017). More recent studies have shown that GAD patients exhibit disrupted insula-based FC, mainly in the mPFC and SMA (Cui et al. 2020), as well as significantly greater FC between the mPFC and the superior temporal gyrus (Xiong et al. 2020). Despite the growing body of evidence for network-level dysfunction and pathological roles of DMN and SN abnormalities in GAD, it remains unknown whether the individual functional connectome is associated with specific clinical symptoms, either psychological or somatic anxiety.
The resting-state functional connectome has emerged as a promising tool in investigating the intrinsic architecture of brain networks (He et al. 2019; Lu, Cui, et al. 2020; Lu, Li, et al. 2020). Notably, most studies have applied traditional hypothesis-driven seed-based analysis to identify disease-associated functional network changes across the experimental groups. However, the use of preselected brain seed regions and group averaging method limits the ability to capture the whole-brain functional architecture (Power et al. 2011) beyond these regions of interest and individual variation (Mueller et al. 2013) in network topography within clinical populations. To address these limitations, multivariate distance-based matrix regression (MDMR) (Zapala and Schork 2006, 2012; Shehzad et al. 2014), a more exploratory and entirely data-driven statistical procedure, has emerged as a promising framework to perform a connectome-wide association study of phenotypic variables. MDMR can search the whole-brain region for multivariate connectivity patterns that may vary with clinical symptoms across the level of individual brain networks without requiring a priori selection of the networks or a particular seed. Moreover, the advantages of the conventional seed-based approaches, like straightforward interpretability and visualization, can still be retained by MDMR. However, there is no study using MDMR approach in investigating the GAD-associated functional connectome network. More importantly, there is no neuroimaging-based evidence on the neural basis of psychological and somatic anxiety symptoms and their mechanistic relationships with individual functional connectome networks in patients with GAD.
Here, we adopted MDMR approach to locate associations between the complex individual functional connectome and the anxiety symptoms of GAD patients. After identifying the networks correlated with symptom-imaging, we attempted to investigate the specific connectivity patterns for psychological and somatic anxiety subscales and the altered network connectivity changes in patients with GAD. Finally, we used the observed connectivity patterns as the features in the support vector regression (SVR) model to predict specific anxiety symptoms of GAD in an independent validation set.
Materials and methods
Participants and clinical assessments
The study protocol was approved by the ethical committee of the University of Electronic Science and Technology of China (UESTC) and registered at ClinicalTrials.gov (Identifier: NCT02888509). All participants were provided information about the procedure and aims of the study, and provided their written informed consent prior to the start of the investigation. A total of 130 patients with GAD were recruited from the Clinical Hospital of Chengdu Brain Science Institute of UESTC after meeting the following inclusion criteria: (i) GAD diagnosed by 2 independent psychiatrists according to the criteria of the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV; Association 2000); (ii) age 18 years or older; (iii) Han Chinese ethnicity.
The anxiety symptoms of patients were assessed using the 14-item Hamilton Anxiety Scale (HAMA; Hamilton 1959) equipped with a specific subscale for psychological and somatic anxiety. Each item is rated on a 5-point Likert-type scale ranging from 0 to 4, with higher scores suggesting more severe anxiety (Shear et al. 2001). The psychological subscale includes items 1–6 and 14, which assesses the more subjective cognitive and affective symptoms of anxiety (e.g. anxious mood, tension, fears, insomnia, memory, or attention disorder). The somatic component (items 7–13) emphasizes features of GAD such as muscular symptoms, sensory symptoms, cardiovascular symptoms, respiratory symptoms, gastrointestinal symptoms, gastro-intestinal symptoms, and autonomic symptoms. The exclusion criteria were as follows: (i) any other major Axis I diagnosis; (ii) current pregnancy. It is worth noting that the pathological mechanisms of patients with GAD and those comorbid GAD with major depressive disorder may be different (Coplan et al. 2016; Choi et al. 2020), we excluded those patients with comorbidity of anxiety and depression to reduce the heterogeneity of the patients. In addition, the most patients with GAD (n = 109) included in this study were received medication treatment. Considering the confounding effects of medication, we used the strategy suggested by the previous studies (Phillips et al. 2008; Versace et al. 2008; Almeida et al. 2009) to calculate the total medication load index. Briefly, each medicine was transferred to equivalence dose of fluoxetine and coded as 1, 2, 3, or 4 based on previously developed criteria, with reference to the daily dose range and duration of the medication. Detailed information on the conversion between medication dosage and corresponding levels is shown in the tables in the Appendix of Sackeim (2001). Notably, there are 2 medications (escitalopram and duloxetine) do not included in the criteria, so they were coded as 1, 2, 3, or 4 according to the midpoint of the daily dose range recommended by the physician’s desk reference. We then obtained the total medication load for each patient by summing all medication codes for each medication category. To validate the reliability of results, we randomly divided all GAD patients into 2 independent subsets in a ratio of 3:1 according to the previous study’s procedure (Quah and Thwin 2003).
Eighty-five unrelated healthy controls (HCs) group-matched for age， sex, and education were recruited from the local community residents through advertisements. The HCs had not been diagnosed with any major Axis I disorders and had no family history of mental disorders. Any subjects suffering from severe medical condition, illegal drug or alcohol abuse/dependence were excluded.
Data acquisition and preprocessing
All participants were scanned using a 3 T GE DISCOVERY MR750 scanner (General Electric, Fairfield, Connecticut, USA). Resting-state fMRI data were acquired using a gradient-recalled echo-planar imaging (EPI) sequence (repetition time/echo time = 2,000/30 ms, flip angle = 90°, matrix size = 64 × 64, 43 axial slices, slice thickness = 3.2 mm with no gap, field of view = 240 × 240 mm2, and 255 volumes). During the 8.5-min resting state scan, participants were asked to keep eyes closed, relax, and not fall asleep. Resting-state fMRI images were preprocessed using the Data Processing Assistant for Resting-State fMRI (DPARSF v 4.4, http://www.restfmri.net) (Chao-Gan and Yu-Feng 2010) based on SPM12 (www.fil.ion.ucl.ac.uk/spm). The first 10 volumes were omitted to minimize the effects of scanner signal stabilization and to ensure the participants were familiar with the scanning environment. The remaining 245 images were slice-time corrected and head motion corrected (participants whose maximal head motion exceeding 2.5 mm displacement or 2.5° of rotation were discarded). Then spatially normalized to Montreal Neurological Institute echo-planar imaging template and resampled to 3 × 3 × 3 mm3 voxels. Confounding regression included Friston 24 motion parameter, cerebrospinal fluid, white matter signals, and global signal. Finally, time series were linearly detrended and band-pass filtered to retain frequencies between 0.01 and 0.08 Hz. Filtered images were smoothed by a Gaussian kernel of 8 mm3 full-width at half maximum. Gross head motion has sensitive effects on FC, so the mean frame-wise displacement (FD) was calculated to evaluate the head movement (Li et al. 2019; Lu, Cui, et al. 2020) and the largest mean FD of each participant was less than 0.25 mm.
Multivariate distance-based matrix regression
MDMR is a multivariate statistical method that could relate anxiety symptoms to whole-brain FC patterns mainly in 3 steps (Shehzad et al. 2014; Satterthwaite et al. 2016; Sharma et al. 2017; Elliott et al. 2018; Ling et al. 2019; Wang et al. 2020) (Fig. 1). Here, to make the analysis computationally tractable, we used the 264 putative functional area masks (Power et al. 2011) to extract time series of 264 nodes from a spherical region with 10 mm diameter (Elliott et al. 2018; Ye et al. 2019). First, the mean time series of a region of interest (ROI) of the 264 nodes was used to calculated a seed-based FC for each patient at the discovery set (n = 90). Second, a distance matrix between each pair of participant’s FC maps is computed to encode the multivariate similarity between each participant’s FC. Third, a pseudo-F value was generated to assess the relationship between the HAMA scores and the distance matrix while controlling the effects of covariates (e.g. age, sex, and education). This MDMR procedure identified ROI where anxiety score affected the overall pattern of connectivity. A nonparametric permutation test was calculated with 5,000 permutations in the MDMR analysis. Then a false discovery rate (FDR) correction with a threshold of 0.05 was applied to the permutation P-values to correct for multiple comparisons.
Visual representation of the main analytical process pipeline. (A) rsfMRI and HAMA scores were collected from participants. (B) For each node in 264 putative functional areas as determined by Power et al. (2011), MDMR was used to generate an F-statistic to assess the relationship between the subject-wise HAMA matrix and distance matrix at that node-level. (C) Cortical projection displayed 11 ROIs identified by MDMR. All ROIs were FDR corrected by using a threshold at P < 0.05. (D) ROIs returned by MDMR were used for follow-up seed-based analysis and (E) within- and between-network analysis. (F) GAD subgroup clustering analysis and between-group analysis. (G) Psychological/somatic scores predicted by SVR.
Follow-up seed-based analyses
In general, MDMR is used to identify certain seed regions where the whole-brain connectivity is significantly varied with the anxiety scores. However, it does not describe which specific connectivity pattern of these ROIs is driving this association. To visualize these spatial connectivity patterns and explore whether it was related to specific psychological or somatic anxiety, we conducted a post hoc seed-based analysis for each MDMR-returned ROI. Seed-based analyses were conducted by performing Fisher’s z-transformed Pearson’s correlations and regressing the z-transformed voxel-wise FC maps against anxiety scores (including the total HAMA scores and somatic and psychological subscales). To obtain 3 spatially correlated brain maps that reflect how ROI to the whole brain FC changes with 3 kinds of anxiety scores, here besides using the aforementioned covariates, we included somatic/psychological scores as covariates when computed the psychological-associated/somatic-associated brain map. Then we computed the z-scores of the spatially correlated brain maps. As the MDMR produced multiple seeds, we computed the averaged z-scores statistical maps at each voxel across the MDMR-based seeds belonging to one functional network to assist the visualization of such regionally consistent spatial patterns. Finally, we determined the corresponding connectivity network of each seed given the position of these MDMR-based seeds in the 7-network parcellation as determined by Yeo et al. (2011). It is worth noting that these follow-up analyses after MDMR do not constitute independent hypothesis tests, because the seeds were family wise error-controlled MDMR findings. Hence, these voxel-wise brain maps generated in follow-up analyses are uncorrected with P < 0.05.
Within- and between-network connectivity analyses
To summarize the pairwise interactions among these ROIs derived from MDMR, a ROI–ROI network framework diagram was constructed. Primarily, for each patient we computed the ROI to ROI FC matrix among the MDMR-identified ROIs and covariates as above. Then, one-sample t-tests were implemented to test the statistical significance of network connectivity, and FDR correction with a threshold of P < 0.05 was used. Next, the connectivity matrices were averaged across edges within the network. To examine the anxiety effects on the network connectivity, the Pearson’s correlations between anxiety symptoms scores (HAMA, psychological and somatic) and the averaged within-network or between-network connectivity were computed, respectively. Correction for multiple comparisons employed FDR with an adjusted threshold of P < 0.05.
GAD subgroup clustering and ablative analyses
Next, we aimed to verify the potential unique network connectivity patterns that allude to different underlying mechanisms of psychological and somatic anxiety symptoms. First, we took the psychological and somatic scores as input and divided GAD patients into 2 subgroups by k-means clustering algorithm (Kanungo et al. 2002; Likas et al. 2003). Then, we compared within-and between-networks FC between 2 GAD subgroups using ablative analysis. In particular, 2-sample t-tests were performed to evaluate the group differences after regressing the psychological and somatic effects on FC within and between networks, respectively. FDR correction was applied and the threshold for statistical significance was set at P < 0.05.
Between-group network connectivity analyses between GAD and HCs
Additional between-group analyses were conducted to examine whether there were significant connectivity differences between the patients with GAD and HCs. The ROI–ROI FC matrices within MDMR-selected ROIs were computed for HCs as same as above. Subsequently, 2-sample t-tests were performed between the matrices of 2 groups, and FDR correction with a threshold of P < 0.05 was used. Finally, the mean within- and between-network FC were calculated to show the differences between GAD and HCs.
Anxiety symptoms prediction analyses
To investigate a potential clinical application for FC-based prediction of anxiety symptoms in GAD, a support vector machine for regression in the Library for Support Vector Machines toolbox (Chang and Lin 2011) was used to predict the anxiety symptoms in patients at independent validation set (n = 28). The leave-one-out cross-validation approach was used to train a model and estimate patient’s symptom score. Briefly, we selected 27 patients at validation set as the training set and the remained one patient as test set, then we used training set to train the model and obtained the symptoms score of test set through this model. The psychological and somatic scores of each patient were predicted respectively using SN and DMN within network connectivity, as well as SN and DMN between network connectivity as features. Eventually, Pearson’s correlation between the observed and predicted symptoms scores was calculated. R2 and mean square error (MSE) were calculated to measure the prediction accuracies. The significance of the predictive model was assessed using 1,000 permutation tests with a threshold of P < 0.05.
Validation analysis of confounding factors
To validate whether the results were independent of the functional parcellation scheme, the 907 putative functional mask as determined by Yeo et al. (2011) was used to repeat the above analysis. In addition, to test whether the connectivity in MDMR-identified ROIs would be affected by other confounding variables including medication loads and motion parameters, we used Spearman’s rank to examine the potential correlations among within- and between-network connectivity and total medication load index/mean FD. A statistical threshold was set as P < 0.05.
Results
Demographic and clinical data
Twelve patients were excluded because of the poor quality of images or data missing. Therefore, the final patient sample size of this study was 118 patients. In order to validate the reliable and generalization ability of results, we randomly divided all patients into 2 independent subsets with a roughly 3:1 ratio. Ultimately, subset I consisted of 90 patients as the discovery set, and subset II has 28 patients as the validation set. There were no differences in age, sex, education, mean FD between patients of subset I and healthy controls (all P > 0.05). The demographic and clinical characteristics of the subset I and subset II GAD groups were statistically matched (all P > 0.05). The demographic and clinical data and of all subjects and comparisons are shown in Table 1. Correlation analyses showed that the total HAMA, psychic, and somatic scores highly correlated with each other (P < 0.05, FDR-corrected, Supplementary Fig. 1).
Characteristics of demographic information.
| Variables | GAD Discovery set (n = 90) | GAD Validation set (n = 28) | HCs (n = 85) | P1 value | P2 value |
|---|---|---|---|---|---|
| Sex (Male/Female) | 35/55 | 9/19 | 41/44 | 0.2125a | 0.5191a |
| Age (years) | 38.84 ± 12.34 | 44.00 ± 9.49 | 40.14 ± 12.35 | 0.5329b | 0.0222b |
| Education (years) | 11.77 ± 3.58 | 11.18 ± 4.08 | 12.51 ± 3.25 | 0.1057b | 0.3015b |
| Mean FD | 0.09 ± 0.04 | 0.07 ± 0.03 | 0.09 ± 0.04 | 0.4100b | 0.2472b |
| Total HAMA score | 24.40 ± 6.12 | 24.68 ± 5.53 | - | - | 0.8310c |
| Psychological subscale | 13.50 ± 3.22 | 13.43 ± 2.17 | - | - | 0.9128c |
| Somatic subscale | 10.90 ± 4.00 | 11.25 ± 4.43 | - | - | 0.6939c |
| Age of first onset (years) | 35.00 ± 12.00 | 38.68 ± 11.07 | - | - | 0.1249b |
| Duration of illness (months) | 44.23 ± 57.19 | 65.50 ± 69.09 | - | - | 0.0212b |
| Medication load index | 1.69 ± 0.73 | 1.82 ± 0.67 | - | - | 0.4697b |
| Variables | GAD Discovery set (n = 90) | GAD Validation set (n = 28) | HCs (n = 85) | P1 value | P2 value |
|---|---|---|---|---|---|
| Sex (Male/Female) | 35/55 | 9/19 | 41/44 | 0.2125a | 0.5191a |
| Age (years) | 38.84 ± 12.34 | 44.00 ± 9.49 | 40.14 ± 12.35 | 0.5329b | 0.0222b |
| Education (years) | 11.77 ± 3.58 | 11.18 ± 4.08 | 12.51 ± 3.25 | 0.1057b | 0.3015b |
| Mean FD | 0.09 ± 0.04 | 0.07 ± 0.03 | 0.09 ± 0.04 | 0.4100b | 0.2472b |
| Total HAMA score | 24.40 ± 6.12 | 24.68 ± 5.53 | - | - | 0.8310c |
| Psychological subscale | 13.50 ± 3.22 | 13.43 ± 2.17 | - | - | 0.9128c |
| Somatic subscale | 10.90 ± 4.00 | 11.25 ± 4.43 | - | - | 0.6939c |
| Age of first onset (years) | 35.00 ± 12.00 | 38.68 ± 11.07 | - | - | 0.1249b |
| Duration of illness (months) | 44.23 ± 57.19 | 65.50 ± 69.09 | - | - | 0.0212b |
| Medication load index | 1.69 ± 0.73 | 1.82 ± 0.67 | - | - | 0.4697b |
Notes: Values are mean ± SD. FD, frame-wise displacement; HAMA, Hamilton Anxiety Scale; SD, standard deviation. All P1 values were obtained by comparing GAD discovery set and healthy controls, P2 values were obtained by comparing GAD discovery set and validation set characteristics. If the variables were not normally distributed, we used a non-parametric analysis, if the variables were normally distributed, we used a parametric analysis.
aChi-square test.
bMann-Whitney U test.
cUnpaired t test. FD, mean frame-wise displacement.
Characteristics of demographic information.
| Variables | GAD Discovery set (n = 90) | GAD Validation set (n = 28) | HCs (n = 85) | P1 value | P2 value |
|---|---|---|---|---|---|
| Sex (Male/Female) | 35/55 | 9/19 | 41/44 | 0.2125a | 0.5191a |
| Age (years) | 38.84 ± 12.34 | 44.00 ± 9.49 | 40.14 ± 12.35 | 0.5329b | 0.0222b |
| Education (years) | 11.77 ± 3.58 | 11.18 ± 4.08 | 12.51 ± 3.25 | 0.1057b | 0.3015b |
| Mean FD | 0.09 ± 0.04 | 0.07 ± 0.03 | 0.09 ± 0.04 | 0.4100b | 0.2472b |
| Total HAMA score | 24.40 ± 6.12 | 24.68 ± 5.53 | - | - | 0.8310c |
| Psychological subscale | 13.50 ± 3.22 | 13.43 ± 2.17 | - | - | 0.9128c |
| Somatic subscale | 10.90 ± 4.00 | 11.25 ± 4.43 | - | - | 0.6939c |
| Age of first onset (years) | 35.00 ± 12.00 | 38.68 ± 11.07 | - | - | 0.1249b |
| Duration of illness (months) | 44.23 ± 57.19 | 65.50 ± 69.09 | - | - | 0.0212b |
| Medication load index | 1.69 ± 0.73 | 1.82 ± 0.67 | - | - | 0.4697b |
| Variables | GAD Discovery set (n = 90) | GAD Validation set (n = 28) | HCs (n = 85) | P1 value | P2 value |
|---|---|---|---|---|---|
| Sex (Male/Female) | 35/55 | 9/19 | 41/44 | 0.2125a | 0.5191a |
| Age (years) | 38.84 ± 12.34 | 44.00 ± 9.49 | 40.14 ± 12.35 | 0.5329b | 0.0222b |
| Education (years) | 11.77 ± 3.58 | 11.18 ± 4.08 | 12.51 ± 3.25 | 0.1057b | 0.3015b |
| Mean FD | 0.09 ± 0.04 | 0.07 ± 0.03 | 0.09 ± 0.04 | 0.4100b | 0.2472b |
| Total HAMA score | 24.40 ± 6.12 | 24.68 ± 5.53 | - | - | 0.8310c |
| Psychological subscale | 13.50 ± 3.22 | 13.43 ± 2.17 | - | - | 0.9128c |
| Somatic subscale | 10.90 ± 4.00 | 11.25 ± 4.43 | - | - | 0.6939c |
| Age of first onset (years) | 35.00 ± 12.00 | 38.68 ± 11.07 | - | - | 0.1249b |
| Duration of illness (months) | 44.23 ± 57.19 | 65.50 ± 69.09 | - | - | 0.0212b |
| Medication load index | 1.69 ± 0.73 | 1.82 ± 0.67 | - | - | 0.4697b |
Notes: Values are mean ± SD. FD, frame-wise displacement; HAMA, Hamilton Anxiety Scale; SD, standard deviation. All P1 values were obtained by comparing GAD discovery set and healthy controls, P2 values were obtained by comparing GAD discovery set and validation set characteristics. If the variables were not normally distributed, we used a non-parametric analysis, if the variables were normally distributed, we used a parametric analysis.
aChi-square test.
bMann-Whitney U test.
cUnpaired t test. FD, mean frame-wise displacement.
MDMR identifies anxiety-associated multiple brain regions
MDMR analysis revealed 11 brain ROIs where the whole-brain connectivity patterns were significantly varied with HAMA scores (Fig. 2A; P < 0.05, FDR-corrected). These regions included the left anterior insula and SMA, the bilateral lateral temporal lobe, and the right angular gyrus， which showed a very high degree of overlap with the SN (4 seeds) and DMN (7 seeds). The detailed information of these ROIs is presented in Supplementary Table 1.
![The MDMR results and follow-up seed-based connectivity analyses. (A) Data-driven MDMR analysis identified 4 ROIs within SN (green) and 7 ROIs within DMN (red), where the whole-brain connectivity was varied with anxiety severity (total HAMA scores) across GAD patients (n = 90). (B) Follow-up seed-based analyses revealed the spatial functional connectivity patterns associated with anxiety scores for each MDMR functional network. All clusters corrected with a voxel height of z > 1.64. (C) The relative contributions of 7 canonical intrinsic functional networks [as determined by Yeo et al. (2011)] to these mean seed-based connectivity patterns. VN, visual network; SMN, somatomotor network; DAN, dorsal attention network; LN, limbic network; FPN, frontoparietal network.](https://oup.silverchair-cdn.com/oup/backfile/Content_public/Journal/cercor/33/10/10.1093_cercor_bhac509/1/m_bhac509f2.jpeg?Expires=1748267910&Signature=2AiHpi8rkj9o3myBpw0siFJm5xElPM41e2P3~PTyjaxiLUyy5aM38E0ClBT68s6DXVdW4hn6UHoHBlC2mwavdumRpnOR8TJTlIuY7iY4UILtey9zvDeDpxifM1IKneug-NMlD-FDlEiHRenb1K9Q3Oj0JjNsSTPrnpbPoq-z8uqXFBYVPHdC0GN7xgTRa2sPWC3Zi36jRpqhiukSzbF5h0I~dnSfVBleJwDTtsQ11V1Or~QgLwtdHvMZ~HjmJsG8W3q3ltdhmny9SidPknU0G0iNr2az~J3G3Qzuejo~gP1RrURDKypDUeFCTipaHsp9PufYWaFSBVWA0~YZ30rJag__&Key-Pair-Id=APKAIE5G5CRDK6RD3PGA)
The MDMR results and follow-up seed-based connectivity analyses. (A) Data-driven MDMR analysis identified 4 ROIs within SN (green) and 7 ROIs within DMN (red), where the whole-brain connectivity was varied with anxiety severity (total HAMA scores) across GAD patients (n = 90). (B) Follow-up seed-based analyses revealed the spatial functional connectivity patterns associated with anxiety scores for each MDMR functional network. All clusters corrected with a voxel height of z > 1.64. (C) The relative contributions of 7 canonical intrinsic functional networks [as determined by Yeo et al. (2011)] to these mean seed-based connectivity patterns. VN, visual network; SMN, somatomotor network; DAN, dorsal attention network; LN, limbic network; FPN, frontoparietal network.
Seed-based analysis reveals the FC patterns driving MDMR
The follow-up seed-based analyses were used to determine the connectivity patterns that were associated with anxiety symptoms and to explore the relationship with specific anxiety subscores. We observed that anxiety symptoms, particularly the somatic symptoms, were positively correlated with SN-based connectivity to brain clusters including the anterior insula, supramarginal gyrus, precentral gyrus, SMA, and postcentral gyrus. Moreover, psychological symptoms were positively correlated with DMN-based connectivity to brain clusters including the PCC, dorsomedial and ventromedial prefrontal cortex, angular gyrus, lateral temporal lobe and parahippocampal gyrus (Fig. 2B). Validation analyses showed that age, sex, and education did not correlated with anxiety symptoms or MDMR network FC, suggesting the specificity of the association between MDMR network connectivity and anxiety symptoms (Supplementary Table 2). To aid the visualization of network connectivity patterns, we averaged the z-scores of mean connectivity maps associated with anxiety symptoms to functional network parcellation (Yeo et al. 2011). In general, these analyses demonstrated that the multivariate results from MDMR were driven by hyper-connectivity within the specific functional network. In particular, SN-based connectivity showed a very high degree of overlap with hubs of the SN and SMN, while DMN-based connectivity showed a very high degree of overlap with the DMN and limbic network (LN) (Fig. 2C).
Within- and between-network analyses delineate the association between network connectivity and anxiety symptoms
To summarize the pairwise interactions of these 2 networks, we performed network-level analysis within the MDMR-selected ROIs (Fig. 3A). This network-level analysis demonstrated that (1) somatic scores (r = 0.33, pFDR = 0.0024) were positively associated with connectivity within SN, (2) HAMA scores (r = 0.35, pFDR = 0.0018), psychological scores (r = 0.36, pFDR = 0.0012), and somatic scores (r = 0.25, pFDR = 0.0202) were positively associated with connectivity within DMN, (3) HAMA scores (r = −0.39, pFDR = 0.0012), psychological scores (r = −0.39, pFDR = 0.0009), and somatic scores (r = −0.34, pFDR = 0.0018) were all negatively related with decreasing connectivity between SN and DMN (Fig. 3B, all pFDR-corrected).
Dissociable patterns of within- and between-network connectivity linked to anxiety scores in GAD patients. (A) The layout of mean functional connectivity within a network identified by MDMR ROIs. One sample t-test was applied to map the dissociable patterns of network connectivity in GAD (P < 0.05, FDR-corrected). (B) The left column shows that the functional connectivity within/between SN and DMN was positively/negatively correlated with HAMA scores, respectively; the middle column displays the functional connectivity correlated with psychological scores; and the right column represents the functional connectivity correlated with somatic scores, all P values were obtained by FDR corrected.
GAD subgroup clustering and ablative analyses indicate the anxiety symptoms driven unique connectivity pattern
We obtained 2 GAD clusters through k-means clustering analysis, the patients in cluster 1 showed high load of both somatic and psychological symptoms, and the patients in cluster 2 showed relatively mild psychological and somatic symptoms (Fig. 4A). Before we eliminated the effect of psychological or somatic symptoms, we found significant within- and between-network connectivity differences between the 2 GAD subgroups (P < 0.05 or pFDR < 0.05). Followed ablative analysis showed that the GAD patients in cluster 1 exhibited significantly enhanced connectivity within SN (pFDR < 0.05) when the psychological effect were regressed out, whereas when the somatic effect were regressed out the GAD subgroups showed no group differences (Fig. 4B). These results indicated that GAD exhibited higher FC within SN may robustly related to somatic symptoms, while GAD exhibited higher FC within DMN and lower FC between SN and DMN may linked with both psychological and somatic symptoms.
MDMR network functional connectivity variation in GAD. (A) Clustering analysis. The left scatter diagram shows the distribution of 2 GAD clusters, and 2-sample t-test (P < 0.05, FDR-corrected) was used to explore the psychological and somatic score differences in 2 GAD subgroups. The right violin plot shows the significant between-group anxiety scores variances. (B) Anxiety-specific network connectivity differences between GAD subgroups. The left plot shows the significant between-group differences in functional connectivity in within- and between-network. The middle and right plot respectively shows the psychological and somatic anxiety specific connectivity comparisons between the 2 GAD subgroups in within- and between-network. The solid star represents FDR-corrected P < 0.05 and hollow star represents uncorrected P < 0.05. SN, salience network; DMN, default mode network.
Between-group analysis reveals the FC alterations in patients with GAD
The between-group analyses showed significant network connectivity differences within MDMR-selected ROIs between patients with GAD and HCs. Compared with HCs, GAD exhibited significantly enhanced FC within SN and DMN while decreased FC between SN and DMN (pFDR < 0.0001, Fig. 5).
Functional connectivity changes within MDMR network in patients with GAD. Between-group analysis exploring the connectivity differences between patients with GAD and HCs. A 2-sample t-test (P < 0.01, FDR-corrected) was used to investigate the between-group differences in functional connectivity within a network of the MDMR-selected seeds. The violin plot shows post hoc analysis of mean functional connectivity, which exhibited significant between-group differences in within- and between-network (P < 0.0001, FDR-corrected). SN, salience network; DMN, default mode network.
FC predicts anxiety symptoms in independent samples
To validate the generalization of the results, we used SVR and LOOCV to obtain the predictive models at the independent GAD validation samples. As an observation of the predictive power of network connectivity, 3 significant predictions were identified when (1) within DMN network-based connectivity (r = 0.73, P < 0.0001, R2 = 0.53, MSE = 2.19, pprem < 0.05) was used to predict the psychological scores, and between SN and DMN network-based connectivity was used to predict the psychological scores (r = 0.38, P = 0.0485, R2 = 0.14, MSE = 4.06, pprem < 0.05) and somatic scores (r = 0.54, P = 0.0029, R2 = 0.29, MSE = 29.2769, pprem < 0.05), respectively (Supplementary Fig. 2). However, within SN network-based connectivity could not predict any anxiety scores (all P > 0.05).
Validation analysis results
To validate the reliability of these results, we repeated analyses using the other 907 putative functional mask. Remarkably, the validation results were similar to the main results analyzed through 264 functional mask (Supplementary Fig. 3). In addition, correlation analysis between the network connectivity and medication load (Supplementary Fig. 4) or mean FD (Supplementary Fig. 5) revealed no correlation at the statistical threshold of P < 0.05, suggesting that the connectivity in MDMR-identified regions may not be confounded by medications and head motion.
GAD functional network association model
We summarized a network model to explain the changes within- and between-network along with psychological and somatic anxiety symptoms derived from our results (Fig. 6). Collectively, our findings suggest that the GAD’s psychological anxiety dimension is dominated by the higher positive FC within DMN, and patient’s somatic anxiety dimension is associated with both increase of FC within SN and DMN. In addition, the anti-correlation between SN and DMN was associated with both psychological and somatic anxiety symptoms. The FC between SN and DMN was decreased along with increasing anxiety symptoms.
Dissociable functional network model of GAD symptoms. This model schematically summarizes the main changes in behaviors associated with individual functional network connectivity. GAD’s psychological anxiety dimension is dominated by the higher positive FC within DMN, and patient’s somatic anxiety dimension is associated with both increase of FC within SN and DMN. Moreover, the negative functional connectivity between SN and DMN is related to both increasing psychological and somatic anxiety symptoms.
Discussion
In this study, we conducted a connectome-wide MDMR framework in patients with GAD, which revealed multivariate patterns of dysconnectivity among SN and DMN regions, including the insula, SMA, lateral temporal lobe, and the angular gyrus. Follow-up seed-based and network-level analyses showed that the somatic anxiety subscale was dominated by the hyper-connectivity within the SN, whereas both the hyper-connectivity within the DMN and the hypo-connectivity between these 2 networks were associated with both anxiety subscales. Additional between-group analyses demonstrated that patients with GAD showed significant alterations in SN and DMN within- and between-network connectivity compared with HCs. Moreover, we verified that the FC pattern of DMN could robustly predict the individual psychological anxiety symptom in an independent set. Collectively, these findings provide novel evidence for the dysconnectivity of SN and DMN as potential dissociable mechanisms underlying somatic and psychological anxiety dimensions for GAD.
Our MDMR analysis revealed multiple regions in which whole-brain connectivity patterns, generated from 264 ROIs, were significantly associated with HAMA scores. These ROIs, belonging to the SN encompassed the left anterior insula and SMA (4 ROIs), and those to the DMN encompassed the bilateral lateral temporal lobe and right angular gyrus (7 ROIs). Validation analysis based on brain nodes from alternative functional atlases also confirmed these findings. Previous studies have suggested that the connectivity within SN may be responsible for why stimuli more easily evoke anxiety in GAD patients (Li et al. 2016). Furthermore, evidence has shown that the strength of FC within the SN is closely related to the visual analog scores from the pre-scan anxiety assessment, indicating an intricate link between SN’s strength and the states of vigilance (Smith et al. 2019). With regard to DMN, it is involved in the integration and self-monitoring of autobiographical memories and emotion regulation (Greicius et al. 2003), however, when a task requires attention, the activation of this network is suppressed. Deficits in DMN suppression exist in a variety of mental illnesses, especially anxiety disorders (Anticevic et al. 2012). The involvement of DMN may be the basis in self-information processing and emotional processing bias in GAD (Geng and Li 2013), which is consistent with the cognitive models of GAD that hypothesize individuals with GAD show extreme hyper-vigilance for threatening ideas (Bar-Haim et al. 2007; Behar et al. 2009).
Despite the evidence of a large number of network alterations in GAD, it remains unclear whether somatic and psychological anxiety symptoms have distinct or common network mechanisms. Follow-up analyses showed that patients with higher somatic scores were associated with increased SN-based connectivity with the regions of SN involving anterior insula, supramarginal gyrus, anterior precuneus, and SMN, whereas patients with higher psychological scores and somatic scores were both associated with increased DMN-based connectivity with regions involving dorsomedial prefrontal, angular gyrus and posterior cingulate. Previous literature on SN has reported that anterior insula activity was positively correlated with body awareness in GAD (Cui et al. 2020). The SMA is mainly activated by motor tasks, and GAD patients show somatic symptoms, including excessive tension and inability to sit still (Li et al. 2013). While in the case of DMN, it has been shown that the activation of the dorsomedial prefrontal cortex is associated with participants’ fear evaluations (Raczka et al. 2010). In addition to the evidence of DMN’s involvements in psychological dimension, some other studies have showed that DMN are also associated with somatic symptoms. The regional activity and FC of DMN have been found to be involved in somatic symptoms of patients with somatization disorder (Su et al. 2014; Kim et al. 2019). Here, we extent the existing literature by demonstrating that SN-based connectivity was specific to the somatic symptoms of GAD, while DMN-based connectivity affected both dimensions of somatic and psychological symptoms. Although our findings need to be tested more thoroughly in future studies, however, they provide a potential possibility for clinical personalized precision treatment for GAD. Previous clinical trials have demonstrated the potential beneficial effects of repetitive transcranial magnetic stimulation (rTMS) on GAD patients (Cui et al. 2019), however, its effects on somatic and psychological anxiety symptoms are still unknown. According to our findings, the method of combining rTMS with fMRI to inhibit TMS-accessible nodes within the SN or DMN (Chen et al. 2013) may be applied to treat individuals who mainly exhibit somatic and/or psychological anxiety symptoms.
In addition to the abnormal within-network functional connectivity in GAD, we also observed that GAD patients had increased negative FC between SN and DMN compared to HCs. These results indicated that hypo-connectivity between the SN and the DMN is associated with both anxiety symptoms. Recent studies have also shown that the HAMA somatic subscale was negatively correlated with FC between the left anterior insula and left mPFC in GAD patients (Cui et al. 2020). The possible explanations for the association between anxiety symptoms and the negative FC between SN and DMN may be because of SN’s tendency to respond to external events that are behaviorally salient (Seeley et al. 2007), whereas DMN shows high activity when the subjects have an internal focus of attention, such as during internally directed thought (Gusnard et al. 2001; Buckner et al. 2008). The transition from an automatic behavior, where attention is usually internally focused, to a behavior guided by external events is accompanied by increased activation within the SN and deactivation of the DMN (Sharp et al. 2011). It is well established that inter-network interactions between SN and DMN is essential for the attention control. In healthy young subjects, a greater negative correlation between SN and DMN is associated with more effective cognitive control (Kelly et al. 2008). It is likely that anxiety leads to excessive processing of threatening stimuli at the expense of ongoing activities (Bar-Haim et al. 2007). Patients with anxiety disorder automatically focus their primary attention on threats in the initial response, thus increasing the possibility of negative information entering their working memory (Bar-Haim et al. 2007; Grant and White 2016). Our finding of increased negative FC between SN and DMN in GAD patients may suggest that GAD-mediated significantly enhanced cognitive control of negative information, such as errors and threats, thereby reducing the cognitive control resources required for effective working memory performance and correct behavior.
Despite delineating several critical observations, there were several limitations that should be considered. Firstly, this cross-sectional study could not examine the causal relationships between network abnormalities and behaviors of GAD. Secondly, although the antidepressant agents were adjusted in the analysis, the influences of different antipsychotics could not be eliminated entirely. Future studies on drug-naive patients with GAD should be conducted to validate our results. Thirdly, our results showed that the DMN network had a relatively stable predictive effect on the occurrence of psychological anxiety, whereas the prediction effect of the SN network on physical anxiety was not ideal. One possible reason is that patients’ psychological dimension is mainly dominated by the network connectivity within the DMN, while the somatic anxiety might be modulated by dysfunction within and between DMN and SN. Therefore, the features of the SN alone may not be sufficient to predict patients’ somatic symptoms. Further, the number of features of within SN connectivity was too small (only 6 edges) than the predictive sample (n = 28), which may lead to the issue of underfitting in the process of SVR model training (Zheng et al. 2006). Finally, we speculate that the activation of SN in the resting state in patients with somatic anxiety may be not prominent. In the future, task experiments can be designed to evaluate the predictive effect of SN activation/connectivity on somatic anxiety symptoms in patients with GAD.
Taken together, this study provides initial evidence for unique connectome-wide functional dysconnectivity of the anxiety symptoms in GAD patients, suggesting the dissociable neural network mechanisms may underlay the somatic and psychological symptoms. Further studies focusing on the specific functional network circuits and anxiety symptoms are needed to improve the personalized precision therapy for GAD patients.
Acknowledgements and disclosures
This work was supported by the National Natural Science Foundation of China (Nos. 82072006, 61906034, 62073058, U1808204), and the Medico-Engineering Cooperation Funds from University of Electronic Science and Technology of China (No. ZYGX2021YGLH201). We are grateful to the participants in the MRI scans used here.
Conflict of interest statement
The authors declare no competing financial interests.
Data availability
Data for the GAD and HCs are available from the corresponding author subject to anonymization to protect privacy of clinical data and implementation of a data sharing agreement as required by the local IRB.
Authors’ contribution
R.L., F.S., and H.C. designed the study; R.L., F.S., L.L., X.D., and Z.H. collected the data; R.L., F.S., X.S., T.Z., X.W., and C.D. performed the data analyses; R.L., F.S., and Z.L. drafted the manuscript. M.Y. and H.C. revised the manuscript. All authors commented on the manuscript. R.L., M.Y., and H.C. provided grant support.
