Prediction of generalized anxiety disorder treatment outcomes with neurobehavioral responses to approach-avoidance conflict: a randomized clinical trial

Abstract

Treatments for generalized anxiety disorder (GAD) often aim to address maladaptive approach-avoidance behavior patterns. Approach-avoidance conflict (AAC) offers a potential framework for identifying treatment outcome predictors and informing optimization of GAD treatment. The current study examined whether pre-treatment neurobehavioral AAC indices predict symptom improvement in behavioral activation (BA) and exposure therapy (EXP) for GAD. Treatment-seeking adults meeting criteria for GAD completed a randomized clinical trial with pre-treatment blinding, conducted from 2016–2021. Participants were randomized to complete 10 manualized sessions of BA or EXP. Participants completed an AAC task during functional magnetic resonance imaging pre-treatment. Computational parameters of task behavior were derived, and neural activity was assessed during decision-making and positive and negative outcomes of decisions. Outcome measures were GAD symptoms and depressive symptoms. Of 121 participants recruited, 56 (29 BA, 27 EXP; mean age 33.0 years; 12.5% male) treatment completers were included in analyses. Greater AAC task avoidance (d = −0.28) and greater left dorsolateral prefrontal cortex activation during negative outcomes (d = −0.32), predicted greater symptom reduction across treatments. Blunted left amygdala activation to positive outcomes was associated at a trend level with favorable symptom reduction for BA but not EXP (d = −0.20). The dorsolateral prefrontal cortex may be a target for enhancing behavior therapy outcomes generally, while left amygdala activation to positive affect may be a target for enhancing outcomes for BA. These findings may inform the optimization of behavioral therapies for GAD and hold potential for transdiagnostic applications, warranting larger, longitudinal studies in clinical settings. Clinical Trials Registration: ClinicalTrials.gov NCT02807480.

Introduction

Generalized anxiety disorder (GAD) affects 4% of the population [1] and confers significant individual and socioeconomic burden [2]. First-line treatments include medication (e.g., selective serotonin reuptake inhibitors) and psychotherapy (e.g., cognitive behavioral therapy [CBT]). Unfortunately, less than half of patients obtain long-lasting improvements from treatment [3]. Symptom severity, illness duration, and comorbidity have been associated with poorer treatment outcomes, but these findings do not provide specific strategies for improving treatment effectiveness. Thus, identifying actionable predictors of treatment success is needed to substantially improve clinical outcomes.

Task-based neuroimaging is a promising avenue for identifying such predictors, affording the ability to examine brain activity during clinically-relevant experimental tasks. A recent meta-analysis found that greater activity in prefrontal cortex (PFC) regions (e.g., dorsomedial PFC, dorsal anterior cingulate cortex [dACC], inferior frontal gyrus [IFG]), and anterior insula during various affective and attentional tasks predicted positive outcomes in CBT for anxiety disorders [4]. None of the studies included in this meta-analysis focused on GAD treatment, although other work has implicated the dACC in predicting response to pharmacotherapy for GAD [5, 6]. These prior studies also did not examine decision-making tasks, despite the central relevance of maladaptive decision-making to the etiology and maintenance of anxiety disorders [7].

GAD, like other anxiety disorders, is characterized by maladaptive behavior patterns that maintain and exacerbate negative affect. Such maladaptive behavior often occurs under conditions of simultaneous conflicting motivations wherein approaching a valued reward also confers risk; i.e., approach-avoidance conflict [8]. The relevance of approach-avoidance conflict to clinical anxiety is supported by evidence from animal models that approach-avoidance conflict responses are sensitive to anxiolytics [9]. Human translational studies have reported that individuals with anxiety disorders demonstrate either more or less avoidance during approach-avoidance conflict tasks; these mixed findings may reflect greater decision uncertainty during approach-avoidance conflict among anxious individuals [10–13]. Regarding neural responses during approach-avoidance conflict, anxiety disorders have been associated with reduced striatal response to rewards and punishment, blunted amygdala activity during decision-making, and greater striatal engagement during affective outcomes, with striatal engagement during affective outcomes tracking avoidance behavior [13, 14].

Gold-standard cognitive-behavioral psychotherapies for anxiety disorders target maladaptive avoidance (i.e., avoidance under approach-avoidance conflict) in different ways. For example, exposure-based therapies (EXP) use behavioral strategies to test negative expectancies concerning avoided situations [15]. Behavioral activation (BA), originally developed for the treatment of depression, primarily aims to increase engagement in rewarding and meaningful activities [16]. These techniques, which are not typically used within the same course of treatment, represent distinct frameworks for addressing imbalanced patterns of approach-avoidance conflict responding. To the extent that BA and EXP preferentially target avoidance versus approach-related drives, patients’ response patterns during approach-avoidance conflict may provide insight concerning who is most likely to benefit from one intervention or the other. The relatively small effect sizes observed in the literature for relationships between neurobiological measures and self-report measures may limit the ability to use fMRI to assign individuals to specific treatments [17]. However, identifying brain or behavioral responses that relate to treatment outcome for a specific intervention or across interventions could inform future experimental studies testing the effectiveness of combining, augmenting, or personalizing treatment approaches to enhance efficacy. Previous research has identified a number of neural predictors of treatment outcomes for anxiety [4] and depression [18], employing resting-state and various executive functioning and emotional processing paradigms. However, these studies did not employ paradigms eliciting approach-avoidance conflict decision-making, limiting their potential to inform our understanding of this crucial behavioral aspect of symptom maintenance and treatment.

Here, we present a randomized clinical fMRI investigation of behavioral and neural responses to approach-avoidance conflict as predictors of psychotherapy response for adults with GAD. Treatment-arms were EXP and BA, two commonly-employed behavioral treatment strategies. We focused on behavioral therapeutic techniques alone, without cognitive techniques, to allow more precise interpretations of treatment predictors. We have previously reported on the clinical outcomes of BA and EXP in a GAD sample overlapping with that of the current study. Findings indicated that compared to EXP, BA yielded more rapid improvement in anxiety and depression scores during therapy and lower scores post-treatment, whereas treatment effects were comparable at 3- and 6-month follow-timepoints [19]. We focused the present analyses on a priori hypotheses that pre-treatment behavioral and neural responses to approach-avoidance conflict would predict treatment response above and beyond pre-treatment symptom severity. Using a previously employed computational modeling approach to behaviorally disentangle decision uncertainty and emotional conflict, we hypothesized that (1) greater decision uncertainty would interfere with engagement in therapeutic behavioral change, and would therefore be associated with worse outcomes across treatments and (2) greater emotional conflict and avoidance would relate to better outcomes in EXP than in BA, given that EXP directly addresses avoidance of anxiety-provoking situations. In regard to neural responses, we focused on a priori region of interest analyses to test additional hypotheses that (3) individuals with weaker pre-treatment striatum and amygdala engagement during reward and positive affective outcomes would benefit more from BA than EXP, given BA’s focus on increasing reinforcement and reward; (4) individuals with greater pre-treatment amygdala and prefrontal engagement during negative outcomes would benefit more from EXP than BA, given EXP’s focus on dampening threat reactivity; and (5) prefrontal activation during decision-making would either be associated with better or worse treatment outcomes across EXP and BA, given contrasting theories about GAD relating to either insufficient top-down regulation versus hyper-active top-down control by PFC regions [20] and thus, the potential for PFC engagement to either facilitate or disrupt one’s ability to respond to behavioral therapy.

Materials and methods

Procedure

The study was a randomized clinical trial conducted in Tulsa, Oklahoma examining neural predictors of behavioral therapy outcomes in adults with GAD, preregistered at ClinicalTrials.gov NCT02807480. Before treatment, participants completed the approach-avoidance conflict task during fMRI, followed by 10 weeks of manualized treatment.

Ethics approval and consent to participate

Participants provided informed consent after receiving a complete description of the study and received monetary compensation for study procedures. The study protocol followed the guidelines of the Western (WCG) Institutional Review Board, which approved the study protocol (#20151232). Research was conducted in accordance with the World Medical Association Declaration of Helsinki.

Participants

Participants were recruited from local mental health clinics and the community via electronic and print advertisements. Participants were aged 18–55, were fluent in English, and met criteria for GAD per the Mini International Neuropsychiatric Interview (MINI) [21]. Exclusion criteria included severe depression and active suicidal ideation; many psychiatric medications were also exclusionary (e.g., anxiolytics, mood stabilizers), but stable SSRI use was allowed (see Supplement 1). Only participants who completed at least 7 sessions of treatment were included in the present analyses, to ensure an adequate dose of treatment.

Interventions

Therapies are described in detail in our published protocol paper [22] and in Supplement 1. Briefly, EXP and BA were both manualized, ten-session interventions focused on identifying maladaptive behavior patterns, assigning weekly behavior homework, and planning for long-term behavior change; cognitive techniques (e.g., cognitive restructuring) were not included in either intervention in order to maintain homogeneity and support interpretability concerning predictors of response. The aim of BA was engaging in behaviors that increase opportunities for reward, especially those in line with the patient’s values. The aim of EXP was approaching rather than avoiding anxiogenic stimuli in order to disconfirm threat expectancies. As reported previously, both treatment-arms yielded substantial improvements in GAD symptoms, anxiety, and depression, as well as increased behavioral engagement and decreased functional impairment [19]. It is also noteworthy that BA and EXP share some similarities. Both interventions focus on identifying maladaptive behavior patterns, assigning weekly behavioral homework, and planning for long-term behavior change. Patients with similar presentations may have overlapping behavioral change goals in both treatments, whereas the focus and framing of the specific therapy activities would differ between treatments. The interventions were delivered in group- and individual-based sessions (see Supplement 1 for details).

Self-report measures

Participants completed self-report symptom measures at pre-treatment, at each treatment session, and at post-treatment, via RedCap [23]. Given that the primary inclusion criterion was a diagnosis of GAD, the primary outcome measure was the GAD-7 [24]. To enable additional assessment of depressive symptoms, in light of the well-documented comorbidity and overlap between GAD and depression [25], the secondary outcome measure was the NIH Patient-Reported Outcomes Measurement Information System (PROMIS) Depression scale [26]. We included PROMIS Anxiety [26] and Behavioral Inhibition and Approach Systems (BIS/BAS) Scales [27] in exploratory analyses.

Experimental paradigm

Participants completed the approach-avoidance conflict task, as previously described [28] (Fig. 1), then rated aspects of their experience of the task (Table S1 in Supplement 1). Briefly, each trial of the task consisted of 3 phases: decision-making, affective outcome, and reward feedback. During decision-making, participants were shown a runway with images on each side that represented 2 possible outcomes; each outcome was a positive or negative affective stimulus (pictures and sounds drawn from the International Affective Picture System [29], International Affective Digitized Sounds [30], and other public-domain audio files) paired with a monetary reward (in US cents). On approach-reward trials, positive affective outcomes are offered on both sides of the runway, with monetary reward on one side (thus, these trials are designed to only elicit motivation to approach the monetary reward). On avoid-threat trials, a positive affective outcome is offered on one side of the runway and a negative affective outcome is offered on the other side, with no monetary reward possible (thus, these trials are designed to only elicit motivation to avoid the negative affective outcome). On conflict trials, a positive affective outcome is offered on one side of the runway and a negative affective outcome is offered on the other; 2, 4, or 6 cents were offered for negative affective outcomes, and 0 cents were offered for positive affective outcomes. Conflict trials are thus designed to elicit approach-avoidance conflict, as the motivation to avoid the negative affective outcome is in conflict with motivation to approach the reward. Participants used a joystick to move along the runway and indicate their preference for the probability of each of the 2 affective outcomes. We used model parameter estimates from a computational model previously derived from this task [11]. Derivation of the computational parameters decision uncertainty (DU) and emotional conflict (EC) is described in detail elsewhere [11]. Briefly, the parameter decision uncertainty (DU) reflects inconsistency of choices across trials with similar contingencies (e.g., choosing to approach the negative stimulus sometimes but not others when the same number of points could be won), and relates to self-reported difficulty making decisions and slower response times. The parameter emotional conflict (EC) reflects avoidance tendencies towards negative affective outcomes relative to reward value (e.g., higher EC could reflect only choosing to approach the negative stimuli when 6 points could be won, whereas lower EC could reflect choosing to approach when both 4 or 6 points could be won). Therefore, EC reflects the magnitude of reward required to prevent avoidance, after accounting for choice inconsistencies. To derive these parameters, the model included observable task stimuli (o; runway cues, trial type cues, affective stimuli, points), task states (s; runway position, trial type), and action policies (π; choice of runway position). Different matrices specified the probability that runway positions would lead to observing different task stimuli, P(o_t|s_t), depending on actions, P(s^t+1|s^t,π), and the subjective preference for (or aversion to) different task stimuli (see Supplement 1).

Fig. 1. The approach-avoidance conflict task.

During the task, participants used a joystick to indicate their preference for the probability of each of the 2 affective outcomes shown (each runway position equated to 10–90% probability on each side). The avatar’s starting position and the side on which the offered outcome or reward appeared were counterbalanced across trials. An image of a sun or cloud indicated a stimulus outcome with a positive or negative valence, respectively. The filling (in red) of an adjacent rectangular meter indicated the level of reward. Approach–reward (APP) trials involved no threat (i.e., positive affective outcomes on both sides) and 2 cents were offered for 1 side. Avoid–threat (AV) trials had no possibility of reward: 0 cents were offered for both positive and negative affective outcomes. In conflict (CONF) trials, 2, 4 or 6 cents were offered for negative affective outcomes (CONF2, CONF4 or CONF6), and 0 cents were offered for positive affective outcomes. Neural activity was examined for decision-making, affective outcome, and points outcome phases across trials.

Neuroimaging

Acquisition parameters for fMRI are described in Supplement 1. Analysis focused on a priori regions of interest (ROIs) from the Brainnetome atlas [31] that we have previously found to be robustly and reliably active during the approach-avoidance conflict task [32]. Neural activity during decision-making and affective outcome phases was assessed in left and right amygdala, bilateral dACC, bilateral striatum (combined caudate and nucleus accumbens), and left and right dlPFC. Neural activity during monetary reward outcomes was assessed in the striatum, given its established role in reward processing [33].

Data analysis

To assess whether the approach-avoidance conflict task elicited the expected behavior patterns as observed in prior work [28, 32], task behavior was analyzed with a linear mixed-effects regression (LME), with avoidance behavior as the outcome variable, and trial-type, treatment-arm, and their interaction as predictors, with random effects for participant ID. F-tests were used to compare variances in AAC indices between treatment-arms.

To identify predictors of symptom trajectory, LMEs were conducted with symptom score as the outcome variable, and effects of time (Week 1 of treatment through post-treatment), treatment-arm, and pre-treatment approach-avoidance conflict response (behavioral or neural), and the interactions among these, as predictors. Pre-treatment symptom score was a covariate (and thus, the pre-treatment timepoint was not included in the time variable), as was pre-treatment fMRI motion for neural predictors; a random effect for participant ID was included. Additional covariates of age, sex, and education were considered, as were random effects for treatment cohort. However, based on model comparison using the Aikake information criterion, none of these were included in final models (Fig. S1). Effects of interest included the time task measure interactions and the time × treatment-arm × task measure interactions. Exploratory whole-brain analyses were conducted for each task condition of interest, using 3dLMEr in AFNI [34], with models mirroring the ROI analyses, but conducted in each voxel individually (see Supplement 1 for details).

To aid in interpretation of findings, LMEs were repeated with self-reported ratings of approach-avoidance conflict task experiences, and the Behavioral Approach System / Behavioral Inhibition System (BIS/BAS) Scales as predictors. Post-hoc correlations were conducted between a priori task indices with significant prediction effects and self-reported task experiences and BIS/BAS Scales.

For analyses of approach-avoidance conflict indices, Bonferroni-corrected alpha levels were 0.025 for the two behavioral variables, 0.007 for the seven ROIs in decision-making and affective outcome phases respectively, and 0.05 for the one ROI in the points outcome phase. Analyses were repeated with datapoints that were both influential (Cook’s d > 4/n) and an outlier on the approach-avoidance conflict variable of interest (>±3 standard deviations from the mean) excluded. Results did not substantially differ with these datapoints excluded, therefore the results from the full dataset are presented.

Results

Pre-treatment characteristics and clinical outcomes

The dropout rate from treatment was 21% in BA and 33% in EXP, which is similar to typical CBT dropout rates [35]. The 56 (27 EXP, 29 BA) participants who completed treatment and had useable fMRI data were included in the present analyses (Fig. 2). The EXP group had fewer male participants and lower pre-treatment PROMIS Depression scores than the BA group; treatment groups were comparable on other pre-treatment clinical characteristics and medication status (Table 1, Table S2, Table S3). Both treatments significantly reduced symptoms, as reported previously [19]. Variance in AAC behavioral and neural indices was not significantly different between treatment-arms (ps > 0.157).

Fig. 2. CONSORT diagram.

Participants who discontinued treatment were asked to remotely complete a survey assessing reasons for discontinuation, which included ratings of various potential reasons as well as a space to provide free text. Participants were allowed to indicate as many reasons for discontinuation as applicable. Responses were summarized using the categories shown above.

Table 1.

Demographic and clinical characteristics.

	BA (N = 29)		EXP (N = 27)		p
	Mean	SD	Mean	SD
Age	33.10	9.98	33.00	10.71	0.970
	N	%	N	%
Sex = Male	7	24.1	0	0.0	0.011†
Sex = Female	22	75.9	27	100
Income					0.126†
Not provided	0	0.0	2	7.4
Less than 50,000	16	55.2	7	25.9
50,000–100,000	5	17.2	8	29.6
100,000–150,000	7	24.1	7	25.9
over 150,000	1	3.4	3	11.1
Race					0.076†
American Indian	9	32.1	3	12.0
Asian	0	0.0	2	8.0
Black	2	7.1	0	0.0
Hispanic	1	3.6	1	4.0
Multi	0	0.0	0	0.0
White	16	57.1	19	76.0
Race = White	16	55.2	19	70.4	0.369
Education					0.358†
Less than high school	1	3.4	0	0.0
High school or GED	0	0.0	2	7.4
Some college, no degree	9	31.0	5	18.5
Bachelor’s degree	11	37.9	7	25.9
Associate’s degree	2	6.9	3	11.1
Master’s degree	6	20.7	9	33.3
Professional/doctoral degree	0	0.0	1	3.7
GAD7	11.46	3.34	12.85	5.33	0.251
PROMIS Anxiety	64.94	4.30	65.17	4.83	0.852
PROMIS Depression	59.23	5.79	55.70	6.97	0.044
SDS	12.69	5.02	11.44	5.54	0.382
	N	%	N	%
MDD Current Episode	13	44.8	6	22.2	0.133
MDD Lifetime	27	93.1	21	77.8	0.137†
Social Anxiety Disorder	8	27.6	10	37.0	0.638
Agoraphobia	0	0.0	3	11.1	0.106†
Panic Disorder	0	0.0	2	7.4	0.228†
Suicidality	11	37.9	7	25.9	0.500
PTSD	0	0.0	0	0.0	>0.99†
Binge Eating Disorder	1	3.4	1	3.7	>0.99†
Concurrent Medication	9	31.0	8	29.6	>0.99
Sessions Attended	8.93	1.00	8.89	1.05	0.878

Age, sex, race, income, education, and symptom scores were identified via self-report. Diagnoses were identified via the Mini International Neuropsychiatric Interview (MINI).

GAD7 generalized anxiety disorder scale–7, PROMIS patient-reported outcomes measurement information system, SDS sheehan disability scale, MDD major depressive disorder, PTSD post-traumatic stress disorder.

†Fisher's exact test.

Approach-avoidance conflict task behavior

In the approach-avoidance conflict task, as expected, participants exhibited the least avoidance behavior during approach-reward trials, the most avoidance behavior during avoid-threat trials, and moderate avoidance behavior during conflict trials, with a pattern of increasing approach behavior with increasing reward levels during conflict. This was reflected in a significant effect of trial-type on pre-treatment avoidance behavior (p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.67, Fig. S2), with similar patterns of pre-treatment behavior across treatment-arms (treatment-arm effect and treatment-arm by trial-type interaction effect ps > 0.915).

Prediction of treatment outcomes

Neural activity in a priori regions of interest during conflict decision-making (ps > 0.08) and points outcomes (ps > 0.88) did not predict common or differential treatment outcomes between treatment-arms and are not shown in tables. Results for task behavior and neural activation to negative and positive affective outcomes are described here (Tables S4–S7).

Common predictors of treatment outcome

Greater pre-treatment emotional conflict (β = −0.32[−0.52– −0.13], d = −0.28, p = 0.001), and greater left dlPFC activation during negative affective outcomes (β = −0.18[−0.27–−0.08], d = −0.32, p < 0.001), predicted greater PROMIS Depression improvements across treatment-arms (Fig. 3). Whole-brain analyses revealed a similar finding in a cluster sharing 55% of its voxels with the a priori left dlPFC ROI and extending posteriorly (see Supplement 1 for whole-brain results: Table S9, Fig. S3). Other findings met a p < 0.05 threshold but not the Bonferroni-corrected threshold, specifically: lower decision uncertainty (β = 0.23[0.01–0.44], d = 0.18, p = 0.036) and more activation to negative affective outcomes in the a priori right dlPFC (β = −0.14[−0.25–−0.04], d = −0.23, p = 0.009) and striatum (β = −0.09[−0.18–−0.01], d = −0.20, p = 0.023) predicted better PROMIS Depression improvement across treatment arms. Other a priori indices did not predict depression outcomes (ps > 0.08). Task measures did not predict GAD-7 symptom trajectory across treatment-arms (ps > 0.09, |ds| < 0.15).

Fig. 3. Common predictors of treatment outcomes.

Effects are shown for A. emotional conflict, a computational measure reflecting avoidance drive relative to reward value; and neural activation to negative affective outcomes in B. left dlPFC, C. right dlPFC, and D. bilateral striatum. Points are actual data, with each panel representing a tertile of the respective task measure. Slopes are estimated effects from models with the form symptom score ~ time × treatment × task measure + pre-treatment symptom score, with random effects for participant, modeled within each panel at the mean +/− 1 standard deviation of the respective task measure. BA behavioral activation, EXP exposure, Dep depression, Wk week of treatment, Post post-treatment timepoint, EC emotional conflict, R right, L left, dlPFC left dorsolateral prefrontal cortex, Str striatum, Neg activation to negative affective outcomes, M mean, SD standard deviation.

As an exploratory analysis, to investigate whether PROMIS Depression effects were specific to depression, we assessed the extent to which these pre-treatment approach-avoidance conflict measures would also predict trajectory of PROMIS Anxiety scores. Echoing PROMIS Depression findings, greater pre-treatment emotional conflict (β = −0.27[−0.47–−0.07], d = −0.23, p = 0.008) and greater left dlPFC activation during negative outcomes (β = −0.18[−0.28–−0.09], d = −0.33, p < 0.001) predicted greater PROMIS Anxiety improvements across treatment-arms. This suggests that these findings are not specific to depression, but rather to an anxiety- and depression-related construct captured by the PROMIS measures.

Differential prediction of treatment outcomes

Lower levels of pre-treatment left amygdala activation to positive affective outcomes predicted favorable GAD symptom improvements in BA and limited GAD symptom improvements in EXP, though this did not meet Bonferroni-corrected p-value thresholds (time × left-amygdala_pos × treatment-arm interaction: β = −0.12[95% CI: −0.22–−0.02], d = −0.20, p = 0.019; Fig. 4). Post-hoc analyses indicated that in EXP, greater symptom improvement was associated with greater pre-treatment left amygdala activity to positive affective outcomes (time × left-amygdala_pos interaction for EXP: β = −0.12, d = −0.49, p < 0.001), whereas in BA, this neural activity was not predictive of symptom improvement (β = −0.003, d = −0.08, p = 0.933).

Fig. 4. Differential predictors of treatment outcomes.

Effects are shown for neural activation to positive affective outcomes in A. left amygdala and B. right dlPFC; as well as for C. decision uncertainty, a computational measure reflecting the consistency of choices. Points are actual data, with each panel representing a tertile of the respective task measure. Slopes are estimated effects from models with the form symptom score ~ time × treatment × task measure + pre-treatment symptom score, with random effects for participant, modeled within each panel at the mean +/− 1 standard deviation of the respective task measure. BA behavioral activation, EXP exposure, Dep depression, Wk week of treatment, Post post-treatment timepoint. R. Amyg right amygdala, L. dlPFC left dorsolateral prefrontal cortex, Pos activation to positive affective outcomes, DU decision uncertainty, M mean, SD standard deviation.

Regarding PROMIS Depression trajectories, there were also effects (though not meeting Bonferroni-corrected p-value thresholds) for lower decision uncertainty (i.e., more consistent choices) (β = −0.34 [−0.64 – −0.04], d = −0.19, p = 0.027) and lower levels of right dlPFC activation to positive outcomes (β = −0.21[−0.41–−0.01], d = −0.18, p = 0.036) predicting favorable PROMIS Depression improvements in BA and limited improvements in EXP. Upon exploratory follow-up analysis, these findings were less robust for PROMIS Anxiety scores (DU: d = −0.16, p = 0.069; right dlPFC: d = −0.09, p = 0.321).

Post-hoc correlations and prediction analyses

Self-reported task experiences

Greater avoidance (EC) and greater left and right dlPFC and striatal activation to negative outcomes were associated with self-reports of greater motivation to avoid negative outcomes and less motivation to approach positive outcomes, more anxiety toward negative outcomes, and less looking at the negative pictures. Greater avoidance (EC) and striatal activation to negative outcomes were also associated with more self-reported distraction during negative outcomes; these indices and left dlPFC activation to negative outcomes were associated with higher ratings of the task overall as upsetting. Higher DU was associated with self-reports of understanding the task less, having more difficulty deciding during the task, not having enough time to decide during the task; higher right dlPFC activation to positive outcomes was also associated with self-reports of understanding the task less. (ps < 0.05, Table S11).

Greater self-reported anxiety and distraction attempts during negative outcomes predicted greater improvements in both GAD-7 and PROMIS Depression across treatment-arms. (ps < 0.03, ds < −0.19). Greater improvements in PROMIS Depression scores were also seen in those who rated the task as more interesting and enjoyable, reported approaching positive outcomes less and avoiding negative outcomes more, and reported less looking at the picture during negative outcomes (ps < 0.043, |ds| > 0.17). Task experience measures did not differentially predict treatment outcomes (Tables S12–S14).

Self-reported trait measures

Greater pre-treatment EC and left dlPFC activation to negative outcomes were associated with greater BAS Drive, Fun-Seeking, and Reward scores; greater right dlPFC and striatum activation to negative outcomes were also associated with higher scores on some BAS subscales. Self-report trait measures did not predict improvement in GAD-7 (|ds| < 0.14, ps > 0.09) symptoms. Greater pre-treatment BAS Drive predicted greater improvements in PROMIS Depression symptoms (d = −0.25, p = 0.004); other trait measures did not predict improvement in PROMIS Depression symptoms (|ds| < 0.07, ps > 0.07, Table S12–S14).

Discussion

This study examined whether pre-treatment approach-avoidance conflict indices could predict treatment outcomes in BA and EXP for patients with GAD. Three key findings emerged. First, greater pre-treatment behavioral avoidance and greater dlPFC activity during negative outcomes predicted greater improvements in depressive and anxiety symptoms across treatment modalities. Second, small effects were observed for relationships between pre-treatment left amygdala activation to positive decision outcomes and greater GAD symptom improvements in BA than EXP; as well as similar relationships between behavioral decision uncertainty and right dlPFC activity during positive outcomes and depressive symptom improvements. Third, neural activity related to decision outcomes, rather than the decision-making process itself, was most consistent in predicting treatment response, a finding that aligns with our previous observations that neural activation during outcomes relates most robustly to approach-avoidance conflict behavior [14]. Taken together, these results not only provide a first step towards using neuroscience-based approaches to individualize treatment plans for GAD based on neurobiological predictors but also underscore the potential for these markers to have transdiagnostic utility in optimizing treatment outcomes across various anxiety and mood disorders.

The finding that greater dlPFC response to negative affective outcomes predicted greater symptom improvements across treatments is contrary to our original hypothesis that greater prefrontal-amygdala reactivity to negative affective outcomes would predict superior response to EXP. These findings are consistent with prior meta-analytic findings in which greater prefrontal activation to a variety of tasks predicted favorable outcomes in CBT for anxiety, including greater dlPFC activation predicting superior outcomes for social anxiety disorder [4]. Interpretation of our result is aided by our finding that bilateral dlPFC activation during negative affective outcomes was positively correlated with self-reported anxiety during negative affective outcomes, as well as with activations in salience-network regions (anterior insula, dACC, and striatum) [36] during negative affective outcomes. One possibility is that the dlPFC plays a role in down-regulation of negative affect in the present data, in line with prior findings associating the dlPFC with emotion regulation, particularly among anxious individuals [37]. Both EXP and BA involve monitoring emotions and modifying behaviors; patients who tend to engage in more dlPFC-mediated regulation during negative decisional outcomes may have a greater ability to benefit from these therapeutic techniques. A second possibility is that heightened dlPFC activity during negative outcomes reflects a maladaptive “over-processing” of negative emotion, which is corrected by EXP and BA. Future research examining the effects of therapeutic interventions on dlPFC activity will shed light on whether these therapies serve to capitalize on or correct heightened dlPFC activity. Both interpretations suggest that individuals with less propensity to engage the dlPFC during affective processing may require more time in treatment or different interventions to support optimal symptom improvement.

Greater emotional conflict (i.e., greater behavioral avoidance) also predicted greater improvements in anxiety and depression scores across treatments, contrary to our hypotheses that these measures would relate to better outcomes in EXP only. It is noteworthy that, in the approach-avoidance conflict task, avoidance involves moving away from the negative affective outcome (paired with points), and toward the positive affective outcome. The positive correlation of emotional conflict with reward-seeking, fun-seeking, and drive scales suggests that this task measure may reflect the individual’s motivation to pursue positive emotional experiences, even when costly, rather than endure a distressing experience. Thus, low emotional conflict may reflect a propensity to endure anxiogenic situations with intense distress rather than avoiding them. The present findings are striking when considered alongside previous findings using this task [11, 12] indicating lower emotional conflict among those with depression and anxiety compared to healthy controls. Taken together, these findings suggest that those with low emotional conflict (a) represent a substantial proportion of the GAD population, and (b) may be inadequately treated with existing behavioral intervention techniques focused on reducing maladaptive avoidance.

Consistent with our reward-related hypotheses, left amygdala activity to positive decision outcomes differentially predicted treatment outcomes at a trend level, with attenuated amygdala activity predicting GAD symptom improvements in BA but not EXP. Relevant to interpreting this finding is that BA yielded greater symptom reduction at post-treatment than EXP overall in the current study (as reported elsewhere [19]). The present findings shed further light on this difference, indicating specific characteristics of participants for whom EXP was less effective. This finding did not meet our corrected threshold for significance, though it should be noted that due to sample size the present study had less power to detect differential predictors compared to common predictors. Larger clinical trials would be useful for corroborating or refuting this finding. While much of the literature in anxiety disorders has focused on negative affective processing, current results suggest that reactivity to positive stimuli may be useful to consider in treatment selection. Given the amygdala’s role in responding to motivationally salient stimuli, with the left amygdala implicated in positive valence processing [38], our finding suggests that patients with attenuated reactivity to the salience of positive outcomes may benefit most from an intervention that enhances this reactivity, such as BA, which is designed to increase response-contingent positive reinforcement. These results may relate to prior findings indicating superior CBT outcomes for adults with anxiety and depression who show attenuated reward reactivity ([39–41], but see opposite findings involving CBT for youth with anxiety and depression [42–44]). Similar patterns were seen among those with lower right dlPFC reactivity to positive images and lower decision uncertainty (i.e., more consistent choices), contrary to our hypothesis that greater decision uncertainty would hinder treatment. Taken together, these trend-level findings provide some support for the hypothesis that BA, more than EXP, is better suited for patients with blunted reactivity to positive decision outcomes and/or rigid behavioral patterns.

It is interesting to note that some findings were evident for measures of depression and anxiety more generally (PROMIS measures) but not for the GAD-7 (e.g., dlPFC and emotional conflict predictors) or vice-versa (e.g., left amygdala differential outcome prediction). This may indicate that engagement of dlPFC when processing affective outcomes is more relevant to the targeting of general negative affect in treatment, while left amygdala or positive affect processing may be more relevant to the reduction of GAD-specific symptoms or worry. However, future larger studies examining depression, anxiety, and worry-related symptoms using multiple measures is needed to further clarify the symptom-specificity of the presented findings.

Limitations

Interpretations are limited by a small sample size, though the current sample size is larger than any other previously-published fMRI treatment prediction study of GAD. In addition, while the inclusion of two active interventions is a strength, the absence of a non-behavioral treatment control condition (e.g., supportive therapy or attentional tasks) limits our ability to determine the extent to which observed symptom changes are caused by the treatments themselves. Moreover, the conceptual overlap between EXP and BA may limit our power to detect differential prediction of outcomes. In addition, because we only analyzed pre-treatment neural activity, these analyses cannot address the question of whether these treatments act via distinct mechanisms. Of note, these findings do not provide individual-level prediction that could potentially be attained through pursuit of cross-validated, machine-learning approaches (which would require a larger sample). While the machine-learning approach has its merits, our goal in the present analysis was to advance theories about specific brain networks and psychological constructs implicated in treatment prediction. We therefore focused on prespecified regions of interest based on a priori hypotheses, with the goal of informing future research aimed at treatment optimization and enhancement. Lastly, the effect sizes identified here and in prior fMRI-based treatment prediction studies would preclude their use in prescribing one intervention over another for individual patients in the clinic. However, such findings could be important for informing future experimental studies testing the effectiveness of augmentations targeting these behavioral or neural predictors to potentially enhance treatment efficacy.

Conclusions

This fMRI study elucidates the neurobiological predictors of treatment response in patients with GAD undergoing BA or EXP. Our findings demonstrate that greater prefrontal activation, particularly in the dlPFC, predicts improved outcomes across both treatments, consistent with the dlPFC’s role in emotion regulation. Greater avoidance during approach-avoidance conflict also predicted superior treatment outcomes. Avoidance findings support a compensatory approach to reducing avoidance, whereas dlPFC findings could support either capitalization on existing executive and emotion-regulation skills, or compensatory reduction of excessive processing of negative emotion [45]. In addition, there was some evidence that individuals with blunted left amygdala reactivity to positive affective outcomes experienced greater improvements to BA but not EXP. These neurobehavioral markers were predictive of treatment success even when accounting for initial symptom severity, lending credence to their potential utility in translational research. Our findings represent an important step toward a neuroscience-based approach to GAD treatment, and potentially other mood and anxiety disorders, using prognostic markers. Further research is warranted to extend these findings and develop targeted interventions for subgroups of patients with differing neurobehavioral profiles.

Author contributions

Concept and design: RLA, MPP, MGC, JLA, WT. Acquisition, analysis, or interpretation of data: HB, TJM, RK, HY, WKT, RS, EA, NK, AC, MC, EW, RLA. Drafting of the manuscript: HB. Critical review of the manuscript for important intellectual content: All authors. Statistical analysis: HB, TJM, RK, HY, WKT, RS, RLA. Obtained funding: RLA. Administrative, technical, or material support: RK, HY, WKT, RS, CRM, KBW, MGC. Supervision: CRM, KBW, MGC, RLA.

Data availability

Analysis scripts, data, and treatment manuals are publicly available in a data repository at https://osf.io/8cfwk/?view_only=2f7fb33971204f2b8201e1e43d1faf01. Dr. Berg and Dr. Aupperle had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Competing interests

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Robin Aupperle reports financial support was provided by the National Institute of Mental Health. Christopher Martell reports receiving royalties on four books on the topic of behavioral activation. Martin Paulus reports receiving royalties for an article about methamphetamine published in UpToDate. The other authors report no biomedical financial interests or potential conflicts of interest.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-025-03460-x.