Neural functional architecture and modulation during decision making under uncertainty in individuals with generalized anxiety disorder

Abstract Background Recent evidence suggests that repetitive transcranial magnetic stimulation (rTMS) might be effective in treating generalized anxiety disorder (GAD). Cognitive models of GAD highlight the role of intolerance of uncertainty (IU) in precipitating and maintaining worry, and it has been hypothesized that patients with GAD exhibit decision‐making deficits under uncertain conditions. Improving understanding of the neural mechanisms underlying cognitive deficits associated with IU may lead to the identification of novel rTMS treatment targets and optimization of treatment parameters. The current report describes two interrelated studies designed to identify and verify a potential neural target for rTMS treatment of GAD. Methods Study I explored the integrity of prefrontal cortex (PFC) and amygdala neural networks, which underlie decision making under conditions of uncertainty, in GAD. Individuals diagnosed with GAD (n = 31) and healthy controls (n = 20) completed a functional magnetic resonance imaging (fMRI) gambling task that manipulated uncertainty using high versus low error rates. In a subsequent randomized‐controlled trial (Study II), a subset of the GAD sample (n = 16) completed the fMRI gambling task again after 30 sessions of active versus sham rTMS (1 Hz, right dorsolateral prefrontal cortex) to investigate the modulation of functional networks and symptoms. Results In Study I, participants with GAD demonstrated impairments in PFC‐PFC and PFC‐amygdala functional connectivity (FC) mostly during the high uncertainty condition. In Study II, one region of interest pair, dorsal anterior cingulate (ACC) – subgenual ACC, showed “normalization” of FC following active, but not sham, rTMS, and neural changes were associated with improvement in worry symptoms. Conclusions These results outline a possible treatment mechanism of rTMS in GAD, and pave the way for future studies of treatment optimization.

It is important to consider these neuronal abnormalities in the context of cognitive and emotional processes. A leading theory of GAD suggests a central role of "intolerance of uncertainty" (IU), a cognitive bias which interferes with information processing, including decision-making (DM) (Ladouceur, Talbot, & Dugas, 1997). The IU model proposes that excessive emotional response in uncertain situations contributes to the development and maintenance of worry (Dugas, Gagnon, Ladouceur, & Freeston, 1998).
Many of the neural areas believed to underlie these DM processes have also been found to be impaired in patients with GAD.
Integrity of these networks in GAD has mostly been explored in the context of emotion dysregulation; however, the neural mechanism of IU per se has not been established. Previous research has found that, unlike healthy control (HC) adults, those diagnosed with GAD experience decreased amygdala activation during a high versus low certainty gambling task (Yassa, Hazlett, Stark, & Hoehn-Saric, 2012). In addition, AI activations during an ambiguous affective DM task are significantly associated with self-reported IU in an unselected sample of young adults (Simmons, Matthews, Paulus, & Stein, 2008). Importantly, no studies have reported FC analysis of DM under uncertainty in GAD. This report describes two interrelated studies. Study I aimed to characterize the neural circuit FC underlying the cognitive processes related to DM under uncertainty, focusing on fronto-limbic FC, in patients with GAD versus HCs. We predicted that individuals with GAD would demonstrate weaker PFC-amygdala FC, evidencing less inhibition of emotional responses, and stronger reactivity of cognitive-emotional error monitoring and salience PFC circuit (i.e., ACC and AI) during high uncertainty trials (i.e., trial blocks involving high rates of error feedback or "lose" trials). Further we predicted that FC during high uncertainty trials would correlate with trait measures of GAD symptoms (i.e., worry) and IU.
Study II aimed to demonstrate modulation of fronto-limbic circuit FC following rTMS treatment. In a randomized control trial (RCT) we previously showed that, in GAD, right DLPFC-targeted lowfrequency rTMS, but not sham, improved anxiety, worry and depressive symptoms and altered local DLPFC activation during a gambling DM task under conditions of uncertainty (Diefenbach et al., 2016).
Since DLPFC, which has been implicated in GAD (e.g., Hilbert et al., 2014) is part of the DM network (Krain et al., 2006) and rTMS is believed to alter neural networks architecture (i.e., FC; Rossini et al., 2015;Wagner et al., 2009), we test the hypothesis that FC patterns during high uncertainty trials would normalize following treatment with active versus sham rTMS to this region, and that changes in FC would correlate with improvements in symptoms and IU trait in GAD participants receiving active rTMS.

| Participants
Fifty-one adults (≥18 years old) completed the fMRI gambling task during participation in either a single session neuroimaging study or during the baseline assessment of a randomized-controlled trial (Clinical Trials ID: NCT01607710). Participants in the GAD group (n = 31) were diagnosed with either principal or coprincipal GAD of at least moderate severity (Clinical Global Impression-Severity (Guy, 1976) Williams, 1988) ≤17. Psychiatric exclusions for the GAD group included post-traumatic stress disorder (current), substance use disorder (past 6 months); or lifetime bipolar, psychotic, developmental, or obsessive-compulsive disorder. Participants taking psychiatric medications were enrolled so long as pharmacotherapy was stabilized for 3 months prior to study entry, with the exception of benzodiazepines taken as needed, which were stabilized based upon medication half-life. Participants enrolled in the HC group (n = 20) reported no current psychiatric diagnoses or lifetime psychiatric treatment. Participants in both groups were excluded for medical disorders which could confound imaging (e.g., brain trauma) or situations that were unsafe (e.g., metal in body).
While there was no a priori IQ exclusion, all participants were assessed to have an estimated IQ >80 (measured by NeuroTrax ™ Comprehensive Testing Suite global cognitive score; NeuroTrax Corp., Bellaire, TX).

| Functional MRI task
During a computerized gambling task, adapted from Bystritsky et al. (2008) and described in our previous report (Diefenbach et al., 2016), participants were shown two cards (red and blue) and asked to predict which card would be drawn next. Participants were instructed to "look for a pattern." Unknown to them, trials were presented in Win and Lose Blocks, in which 75% of the trials showed participants correct or error feedback respectively. Thus, lose blocks constitutes a 'high uncertainty' condition, given that significantly more error feedback is presented. Each condition (Win/Lose) included six blocks with eight trials/block, with win/lose trials presented randomly. Trials were presented for 2.3 s each with feedback (correct or error) presented for 1.2 s (task block length = 28 s). Rest blocks showing a white cross over a black background for 18 s in length interleaved task blocks (total run length = 381 s, including 13 s of instructions). Before the task, participants were given 50 points (with no monetary value) and told that they could win or lose two points per trial based upon correct or incorrect predictions respectively. By design, all participants ended with a loss of 16 points.

| Image acquisition
MRI scans were conducted on a Siemens 3T Allegra MRI scanner.

| Data analysis
Regional activation analysis To assess brain regions functionally involved with the task (i.e., defining regions of interest; ROIs), individual statistical maps were entered into a mixed-effect repeated measures analysis of variance (ANOVA) with Task Condition (Win/Lose) as the within-subject effect and Group (GAD/HC) as the between-subject effect. While Task Condition effects were the primary focus for ROI definition (see below), Group main effect, as well as Group by Task Condition interaction were also explored.

ROI-to-ROI functional connectivity analysis
Functional connectivity analysis was performed using Functional Connectivity (CONN) toolbox version 14.n (http://web.mit. edu/swg/software). Preprocessing was redone using CONN's standard pipeline, including: realignment, coregistration with a high-resolution anatomic scan, slice time correction, structural segmentation, normalization to MNI template, and smoothing (FWHM 8 mm 3 ). White matter and cerebrospinal fluid were computed per subject, and entered as potential confound regressors along with realignment effects and scrubbing parameters (set according to CONN defaults: global-signal scan-to-scan Z-value = 9; motion threshold = 2 mm). Task Conditions (win/lose for Study I and II) and Time (pre/post rTMS for Study II) were entered as within-subjects regressors of interest while Group (GAD vs. HC for Study I) and Treatment Condition (active vs. sham rTMS for Study II) were entered as between-subjects regressors of interest, using CompCor (Behzadi, Restom, Liau, & Liu, 2007). Band-pass filter (0.008-0.09 Hz) was applied, followed by detrending (removal of linear trends within each functional session), to reduce noise influence.

ROIs definition
As mentioned above, brain regions functionally involved in the task were defined as having a significant Task Condition effect in the group activation analysis of Group by Task Condition ANOVA.
Spheres, 5 mm in diameter around point of maximal group activation, were defined and entered into CONN as ROIs. For ROIs not identified by GLM analysis binary masks were created based on the FSL Harvard-Oxford atlas (Desikan et al., 2006).

Functional connectivity analysis
Individual (first-level) ROI-to-ROI FC analysis was performed by calculating the time courses temporal weighted-correlations for all pair-wise ROI combinations. Next, these measures were entered into repeated-measures ANOVA across subjects (second-level analyses) using a standard mixed within-(Task Condition) and between-subjects (Group) GLM, as described above. Significant results were considered at FDR corrected p value (q FDR ) < 0.05.
To assess the relationship between GAD psychopathology and ROI-to-ROI FC patterns, correlation analyses were performed for each of the Task Conditions separately. Correlation of FC with PSWQ and IUS were first calculated in the entire sample and interpreted as significant at p < 0.0125, applying correction for each ROI pair for four comparisons (two measures for each of the two conditions). Follow-up exploratory correlation analyses within the GAD group were conducted for significant results at the entiresample level.

| Participants/Image acquisition
In Study II we present data from a GAD subgroup who completed the fMRI task a second time after a treatment course of active (n = 9) or sham (n = 7) rTMS (M = 6.06 ± 3.3, range = 1-12 days between final rTMS session and second fMRI). In addition to the exclusion criteria outlined in Study I, participants were also excluded from Study II for concurrent psychotherapy. Therefore, no participants in Study II were undergoing psychotherapy over the course of rTMS treatment. In addition, for those participants taking psychiatric medication, type and dose remained stable over the course of rTMS treatment.
Image acquisition parameters were identical to Study I.

| rTMS protocol
Participants completed 30 sessions (5 days/week for 6 weeks) of low-frequency (1 Hz; 90% of the resting motor threshold) rTMS for 900 pulses/session. These stimulation parameters were chosen to be the same as those used in a previous open trial of rTMS for GAD (Bystritsky et al., 2008), although a longer treatment course (i.e., 30 sessions) was administered in the current study to protect against inadequate dosing. rTMS was administered using the FDA-Cleared Neurostar TMS Therapy System (note that neither the use in GAD nor the protocol used here are FDA approved), and sham rTMS was administered using a sham coil (Neuronetics XPLOR) that delivers <10% of an active pulse. rTMS was administered to the right DLPFC (MNI coordinates: x = 42, y = 36, z = 32) using stereotactic neuronavigation system (Visor2, ANT Neuro, Enschede, Netherlands; http:// www.ant-neuro.com), as described previously (Diefenbach et al., 2016).

| Data analysis
Since no group effects were found between GAD and HC with GLM activation analysis in Study I (see Results), only FC analysis was performed for Study II, to assess the effects of active (vs. sham) rTMS on the ROIs pairs showing abnormal FC in GAD compared to HC in Study I.
Individual ROI-to-ROI FC analysis was calculated as described above for pre-and post-treatment scans. Next, Treatment Condition (Active/Sham) by Time (Pre/Post-treatment) repeated measures ANOVAs were calculated for either the Lose or Win task conditions separately, based on a-priori hypotheses and results from Study I, as described below. Due to relatively small sample size, threshold was set at uncorrected p < 0.05, and effect sizes are also presented to aid interpretation.
To assess the relationship between pre-to-post-treatment changes in brain functional architecture and GAD psychopathology, correlations between FC changes and IUS and PSWQ changes over time (post-pre) were calculated for the active rTMS group.  Table S1).  (Table 2). It is important to highlight that no region showed an activation main effect of Group or Group by Task Condition interaction for analyses of activation. These task-related ROIs were therefore used for subsequent functional connectivity analyses exploring a-priori hypothesized group effects, as the absence of a Group or Group by Task Interaction in the initial activation analysis minimizes introducing an ROI selection bias in the FC analyses.

| Functional connectivity analysis
We focused our FC analysis on PFC regions and amygdala given their documented role in IU (Krain et al., 2006). PFC ROIs were defined as spheres at regions showing a Task Condition main effect in GLM analysis (Table 2). Since no activation effects were found in the amygdala,   Figure S2 provides full FC maps for each group and task condition). Results indicated significant positive FC for dACC2-sgACC, dACC2-right AI, dACC2-left AI, and sgACC-right AI during Lose in the GAD group only, with significant group differences. In addition, dACC1-right Amygdala FC was significantly anti-correlated in HCs, but not GAD during Lose blocks, with significant group difference. dACC1-right Amygdala FC also differed significantly between HC and GAD groups during Win; however, FC did not differ significantly from zero (i.e., no significant correlation between these two regions) for either group.
dACC2-sgACC and sgACC-right AI also showed significant group differences during Win; again, FC did not differ significantly from zero. Finally, for dACC1-sgACC, only the GAD group showed significant anticorrelation during Win with significant group TA B L E 1 Study I: Sample demographic and clinical characteristics

| Participants
Characteristics of the participants in Study II are described in Table 4. The groups were matched on age, gender, race, estimated IQ and symptom severity. As we reported previously (Diefenbach et al., 2016), participants in the active rTMS group showed significantly more pre-to-post-treatment effect on symptoms with 7 (77.8%) compared with 2 (28.6%) meeting treatment responder status (defined as ≥50% HARS improvement) (χ 2 [1, N = 16] = 3.87, p = 0.049).

| Functional connectivity analysis
Analyses for Study II examined treatment effects on ROI pairs that were significantly related to GAD status (i.e., showed significant FC Group effects) in Study I (  Figure 1 and Table 3 and exploratory tested for Win condition for dACC1-sgACC only. See Supporting Information Figure S3 and Table S2 for bar graphs and statistical results). Pre-to post-treatment changes in dACC2-sgACC FC were moderately, though nonsignificantly, associated with changes in trait worry (PSWQ r = 0.53) and minimally associated with changes in IU (IUS r = 0.22) in the active rTMS group.
Study II's results indicated that right DLPFC-targeted rTMS modifies dACC-sgACC FC in patients with GAD in the direction of "normalization" (from positive to negative FC). Further, post-rTMS changes in dACC-sgACC FC were moderately associated with improvements in worry symptoms, although this result should be carefully interepreted as it was not significant (this can be attributed to low power but previous work also had not shown correlation between FC and symptom changes following rTMS; Liston et al., 2014).
These results are consistent with literature implicating the sgACC as a key structure in neuromodulation therapies for emotional disorders. The sgACC is a common deep brain stimulation target for major depressive disorder (MDD; Mayberg et al., 2005) as well as a proposed downstream mechanism through which cortical stimulation from rTMS improves symptoms (Pathak, Salami, Baillet, Li, & Butson, 2016). Previous research has also indicated that DLPFC-targeted rTMS decreases sgACC resting-state activity (Baeken et al., 2015;Fox, Halko, Eldaief, & Pascual-Leone, 2012;Noda et al., 2015) and FC with several brain areas in patients with MDD (Baeken et al., 2014;Liston et al., 2014;Taylor et al., 2018) and post-traumatic stress disorder (PTSD; Philip et al., 2018). Although preliminary, current findings suggest that sgACC, and specifically its FC with dACC, is a potential target for rTMS treatment for GAD as well.
As hypothesized, dACC-sgACC FC changes were associated with improvements in worry; however, contrary to hypothesis, not with changes in IU. This is surprising given previous research indicating that changes in IU may mediate clinical symptom improvements in GAD following psychological therapies (Bomyea et al., 2015). It is possible that neurostimulation such as rTMS may exert GAD treatment effects through a different mechanism. This interpretation,  (Manohar & Husain, 2016), including evaluation of rewards value (Levy & Glimcher, 2012). Thus, abnormally increased negative FC between these two emotion-cognitive control regions might indicate downregulation of (or low reactivity to) positively-valenced stimuli. This, in conjunction with hyper-reactivity to negative stimuli, may contribute to information processing biases in GAD (Hayes & Hirsch, 2007).

| Study limitations
We note several limitations of our studies. First, while GAD was the primary diagnosis in all patients, over half (61%) met criteria for other anxiety or depressive disorders, and there was significant correlation between depression and anxiety symptoms, limiting the speci-

| Study summary
To summarize, we demonstrated functional neural networks architecture abnormalities, focusing on PFC and amygdala, during a DM under uncertainty task in GAD versus HC and their relationship to trait worry and IU. Results suggest increased emotional reactivity combined with decreased emotional and cognitive regulation during the task characterized by high error feedback is associated with a core symptom of GAD, i.e., excessive worry. Furthermore, a follow-up RCT in a GAD subsample indicated that these abnormalities can be modulated by right DLPFC rTMS, leading to normalization of FC between key emotion regulation areas, sgACC and dACC, along with symptom improvement. These results outline a possible treatment mechanism, providing a target for future studies examining treatment optimization for GAD, preferably on an individualized level.

ACK N OWLED G M ENTS
This study was funded by a grant (#129522) from the Hartford HealthCare Research Funding initiative to Dr. Diefenbach. The funding source had no role in the study design; collection, analysis, interpretation of data; writing the report, or in making the decision to submit the article for publication. Material support was provided by Neuronetics. Neuronetics reviewed a draft of this report prior to submission and otherwise had no role in the study design; collection, analysis, interpretation of data; writing the report, or in making the decision to submit the article for publication. Data from this study were presented at the annual meeting of the Society for Biological Psychiatry, San Diego, CA, May 2017.

CO N FLI C T O F I NTE R E S T
Drs. Diefenbach and Goethe report that they receive material support from Neuronetics.
Dr. Goethe reports that he has received speaker fees to dis-

S U PP O RTI N G I N FO R M ATI O N
Additional supporting information may be found online in the Supporting Information section at the end of the article.
How to cite this article: Assaf M, Rabany L, Zertuche L, et al.