Adaptive cognitive control circuit changes associated with problem-solving ability and depression symptom outcomes over 24 months

Abstract

Mechanistically targeted behavioral interventions are a much-needed strategy for improving outcomes in depression, especially for vulnerable populations with comorbidities such as obesity. Such interventions may change behavior and outcome by changing underlying neural circuit function. However, it is unknown how these circuit-level modifications unfold over intervention and how individual differences in early circuit-level modifications may explain the heterogeneity of treatment effects. We addressed this need within a clinical trial of problem-solving therapy for participants with depression symptoms and comorbid obesity, focusing on the cognitive control circuit as a putative neural mechanism of action. Functional magnetic resonance imaging was applied to measure the cognitive control circuit activity at five time points over 24 months. Compared with participants who received usual care, those receiving problem-solving therapy showed that attenuations in cognitive control circuit activity were associated with enhanced problem-solving ability, which suggests that this circuit plays a key role in the mechanisms of problem-solving therapy. Attenuations in circuit activity were also associated with improved depression symptoms. Changes in cognitive control circuit activity at 2 months better predicted changes in problem-solving ability and depression symptoms at 6, 12, and 24 months, with predictive improvements ranging from 17.8 to 104.0%, exceeding baseline demographic and symptom characteristics. Our findings suggest that targeting the circuit mechanism of action could enhance the prediction of treatment outcomes, warranting future model refinement and improvement to pave the way for its clinical application.

INTRODUCTION

Depression is one of the most prevalent, adverse, and chronic psychiatric disorders. It affects 18.3% of the adult population, a rate that has increased since the coronavirus disease 2019 pandemic (1), and it is the leading cause of disability worldwide (2). Poor outcomes for depression are exacerbated when other commonly co-occurring chronic conditions, such as obesity, are present (3). The rapidly emerging field of precision medicine in psychiatry highlights the need for personalized tailoring of treatments on the basis of individual neural circuit biotypes (4). We must address this need for individuals with comorbidities, because they have particularly poor outcomes with currently available approaches (3).

In a randomized clinical trial, Ma et al. (5, 6) demonstrated the efficacy of an integrated collaborative care intervention (I-CARE) for patients with depression symptoms and comorbid obesity. I-CARE combines in-person behavioral problem-solving therapy for depression and a video-based behavioral lifestyle intervention for obesity, and it has been demonstrated to produce better depression and weight loss outcomes than usual care (U-CARE). As is the case for all depression interventions, not all patients in I-CARE improved. The modest effect size of the intervention underscores our emphasis on a precision medicine strategy to understand and subsequently target the underlying mechanisms of action for optimizing treatment responses. For example, in cases in which individuals do not demonstrate early engagement with the target mechanism, clinicians can consider augmentation of treatments aimed at more effectively modulating these mechanisms.

Cognitive control plays a pivotal role in problem-solving and the regulation of symptoms that occur with depression, including negative emotion and poor motivation (7, 8); therefore, the neural circuit mechanisms associated with cognitive control are directly pertinent to the mechanisms of action targeted by problem-solving therapy. Cognitive control involves goal-directed processes that select responses to task-relevant material while inhibiting responses to task-irrelevant material in a cognitive neuroscience framework (9). Tasks such as the Go-NoGo are designed to engage these cognitive control processes. This operational definition is grounded in meta-analyses that support the involvement of a superordinate cognitive control circuit in such processes (10). Tasks that require control commonly activate the lateral prefrontal, dorsal anterior cingulate, and parietal cortices, whereas some specific task processes recruit subcortical regions such as the caudate. Prefrontal regions are thought to exert control by modulating how posterior brain regions process task-relevant and task-irrelevant material (11). The anterior insula (AI) is also involved in signaling the need for cognitive control (12).

Precision medicine trials of pharmacotherapy for depression report that pretreatment cognitive control circuit activity and connectivity distinguish acute responders from nonresponders (7, 13, 14) and differentially predict response to different types of medication (15). Patients with cognitive control circuit dysfunction respond poorly to antidepressants (16). Cognitive behavioral intervention is a promising alternative for inducing changes in cognitive control circuit function that in turn improve behavior, as illustrated in meta-analyses (17). Cognitive control circuit function also identifies depressed individuals at risk of recurrence over 1 year and who may benefit from different intervention strategies (7). These prior findings motivated our focus on cognitive control circuit mechanisms that underlie response to the I-CARE intervention, which is designed specifically for a vulnerable patient population with depression symptoms, comorbid obesity, and low responses to pharmacotherapy.

In this companion study of the parent I-CARE trial, we targeted modification of the cognitive control circuit as the a priori circuit mechanism of interest (18). Using a longitudinal design, we investigated the cognitive control circuit as the mechanism that underlies the response to I-CARE versus U-CARE, which is the response marker, and as the predictive marker of treatment outcomes. A unique design was the inclusion of five repeat assessments of the cognitive control circuit evoked by a Go-NoGo task and the concurrent evaluation of problem-solving ability and depression symptom outcomes over 24 months (18).

We first tested the cognitive control circuit as a potential response marker (19). Our first hypothesis was that, within our longitudinal design, changes in the cognitive control circuit activity would be directly associated with changes in problem-solving ability. Our second hypothesis was that these changes in circuit activity would also be directly associated with changes in depression symptoms. For each of these hypotheses, we anticipated that circuit activity changes associated with each of the outcome measures would be modified by I-CARE versus U-CARE. We then tested whether the cognitive control circuit might serve as a predictive marker (19). We focused on the first 2 months of I-CARE, in which intervention focused specifically on depression. This 2-month period is also considered the period in which we could expect to measure a meaningful change in brain function indicative of treatment-related neural plasticity (20). Our third hypothesis was that changes in circuit activity over the first 2 months of treatment predict subsequent outcomes for both problem-solving ability and depression symptoms at 6, 12, and 24 months. We explored the clinical translational value of circuit prediction by including the categorical symptom response as an additional outcome measure.

RESULTS

We analyzed data from 108 participants with depression symptoms and comorbid obesity from the ENGAGE neuroimaging trial (“Engaging self-regulation targets to understand the mechanisms of behavior change and improve mood and weight outcomes”) (see fig. S1 for CONSORT diagram) (18). ENGAGE was conducted in tandem with the parent “Research aimed at improving both mood and weight” (RAINBOW) trial (6). On average, participants had moderate depression and moderately severe obesity. Baseline demographic and symptom characteristics are summarized in Table 1. Participants were randomized to receive the behavioral intervention I-CARE (n = 59) or U-CARE (n = 49) (Fig. 1A). Our primary outcomes, problem-solving ability and depression symptoms (Fig. 1B), were assessed using the Social Problem-Solving Inventory—Revised Short Form (SPSI) and the 20-item Depression Symptom Checklist (SCL-20) at baseline, 6, 12, and 24 months. Compared with those in the U-CARE group, participants in the I-CARE group showed greater improvement in depression symptoms over the 24-month intervention period (F = 5.55, P = 0.02; fig. S2A), which is consistent with the parent RAINBOW study (5). Problem-solving ability also showed improvement over the 24-month period (F = 3.01, P = 0.05; fig. S2B).

Table 1. Baseline demographic and symptom characteristics by treatment group.

GED, General Educational Development.

Characteristics	I-CARE, N = 59, [N(%)]	U-CARE, N = 49, [N(%)]	Total, N = 108, [N(%)]
Sex
Male	17 (28.8%)	18 (36.7%)	35 (32.4%)
Female	42 (71.2%)	31 (63.3%)	73 (67.6%)
Race
Non-Hispanic white	46 (78.0%)	35 (71.4%)	81 (75.0%)
Black	1 (1.7%)	0 (0%)	1 (0.9%)
Asian/Pacific Islander	5 (8.5%)	3 (6.1%)	8 (7.4%)
Hispanic	4 (6.8%)	7 (14.3%)	11 (10.2%)
Other or not reported	3 (5.1%)	4 (8.2%)	7 (6.5%)
Education
High school graduate or GED	2 (3.4%)	4 (8.2%)	6 (5.6%)
Some college	10 (16.9%)	14 (28.6%)	24 (22.2%)
Undergraduate degree	28 (47.5%)	15 (30.6%)	43 (39.8%)
Graduate level work or degree	19 (32.2%)	16 (32.7%)	35 (32.4%)
Age (years)
Mean (SD)	52.4 (11.6)	51.6 (12.0)	52.0 (11.7)
Median (min, max)	52.6 (22.8), (76.0)	55.4 [24.9, 73.6]	53.1 [22.8, 76.0]
Body mass index (kg/m2)
Mean (SD)	34.9 (5.16)	36.3 (4.91)	35.5 (5.07)
Median (min, max)	33.3 (27.2), (50.2)	35.1 [27.1, 48.8]	33.9 [27.1, 50.2]
20-Item Depression Symptom Checklist score
Mean (SD)	1.49 (0.555)	1.59 (0.526)	1.53 (0.542)
Median (min, max)	1.60 (0.350), (2.45)	1.65 [0.350, 2.75]	1.60 [0.350, 2.75]
9-Item Patient Health Questionnaire score
Mean (SD)	14.0 (3.11)	13.4 (2.88)	13.7 (3.01)
Median (min, max)	14.0 (10.0), (23.0)	13.0 [10.0, 20.0]	13.0 [10.0, 23.0]
7-Item Generalized Anxiety Disorder scale score
Mean (SD)	7.80 (4.25)	8.02 (5.02)	7.90 (4.59)
Median (min, max)	8.00 (0), (17.0)	7.00 [0, 21.0]	7.00 [0, 21.0]
SPSI scale score
Mean (SD)	12.1 (2.77)	11.0 (2.66)	11.6 (2.76)
Median (min, max)	12.6 (5.00), (16.8)	11.0 [6.00, 15.8]	12.0 [5.00, 16.8]

Fig. 1. Study design and analysis strategy.

(A) Study design timeline. One hundred eight participants were randomized into I-CARE or U-CARE. Cognitive control circuit and treatment outcomes were measured at baseline (BL), 6, 12, and 24 months (MO), with an additional circuit measure at 2 months. I-CARE included problem-solving therapy. U-CARE participants continued to receive U-CARE from their primary care physician. *: No intervention between 12 and 24 months. (B) Treatment outcomes were problem-solving ability (self-administered SPSI) and depression symptoms (SCL-20). (C) The cognitive control circuit was evoked by a Go-NoGo functional magnetic resonance imaging task. Regions of the circuit include dLPFC, AI, CdN, IPL, SPL, amygdala (Amy), FFG, dACC, and PCUN. (D) Illustration of workflow to investigate potential circuit markers underlying treatment outcome improvement. We used voxel-wise whole-brain linear mixed models with change in treatment outcomes, specifically change in problem-solving ability or change in depression symptoms, at 6, 12, and 24 months (ΔOutcome) as the dependent variable. Fixed effects include change in cognitive control circuit activity from corresponding time points (ΔCognitive control circuit activity or ΔActivity), its interaction with treatment groups (I-CARE and U-CARE; Treatment × ΔActivity) and time (6, 12, and 24 months; Time × ΔActivity), and the interaction of all three factors (Treatment × Time × ΔActivity). Model-captured effects are illustrated in the four graphs on the right. The focal effects of interest were the treatment–by–circuit activity change interaction (Treatment × ΔActivity) and the main effect of circuit activity change (ΔActivity) across treatment groups. (E) Illustration of workflow to investigate potential predictive circuit markers of treatment outcome. Similar linear mixed models as in (D) were used, except that we used the circuit activity change at 2 months for the main and interaction terms (ΔActivity_2MO). Model-captured effects are illustrated in the four graphs on the right. Similarly, the effects of focal interest were the treatment–by–circuit activity change at 2 months interaction (Treatment × ΔActivity_2MO) and the main effect of cognitive control circuit activity change at 2 months (ΔActivity_2MO) across treatment groups. All changes are relative to the baseline.

The cognitive control circuit was assessed by functional magnetic resonance imaging (fMRI) during a Go-NoGo task at baseline, 2, 6, 12, and 24 months (Fig. 1C). Circuit activity was derived from the contrast of NoGo > Go. To identify response and predictive markers, we used voxel-wise whole-brain linear mixed models (LMMs) with change in treatment outcomes (ΔOutcome), specifically change in problem-solving ability (ΔSPSI; hypotheses 1 and 3) or change in depression symptoms (ΔSCL-20; hypotheses 2 and 3) at 6, 12, and 24 months relative to baseline, as the dependent variable. For testing response markers (hypotheses 1 and 2), the focal effects of interest were the treatment–by–circuit activity change interaction (Treatment × ΔActivity) and the main effect of circuit activity change (ΔActivity). For testing predictive markers (hypothesis 3), the effects of focal interest were the treatment–by–circuit activity change at 2 months interaction (Treatment × ΔActivity_2MO) and the main effect of cognitive control circuit activity change at 2 months (ΔAc-tivity_2MO) across treatment groups. Interaction effects of time were also included. The analysis strategy and example of model-captured effects are illustrated in Fig. 1 (D and E).

I-CARE modifies cognitive control circuit mechanisms that are associated with improved problem-solving ability

As hypothesized, the relationship between changes in cognitive control circuit activity and changes in problem-solving ability was modified by treatment across 6, 12, and 24 months, as reflected in a treatment–by–circuit activity change interaction (Treatment × ΔActivity). This significant interaction was localized within the left dorsal lateral prefrontal cortex (dLPFC; F = 24.79, P < 0.001), right inferior parietal lobule (IPL; F = 19.74, P < 0.001), precuneus (PCUN; F = 24.93, P < 0.001), and left fusiform gyrus (FFG; F = 25.31, P < 0.001) (Fig. 2A; see Table 2 for peak coordinates and summary statistics). Specifically, attenuation of the cognitive control circuit activity was associated with improvements in problem-solving ability in the I-CARE group, whereas the same circuit activity change was correlated with decrements in problem-solving in the U-CARE group (Fig. 2B). We verified the goodness of fit for our hypothesis-testing model using a likelihood ratio test and showed that our circuit model based on each of the above regions outperformed a baseline model in log-likelihood, Akaike information criterion (AIC), and Bayesian information criterion (BIC; table S1). These results indicate that the I-CARE intervention, designed to focus on problem-solving therapy, specifically modifies an a priori–determined cognitive control circuit as a neural mechanism of problem-solving outcomes.

Fig. 2. Distinct associations of change in cognitive control circuit activity and change in problem-solving ability for participants who received I-CARE versus U-CARE.

(A) Brain maps showing left dLPFC, left IPL, PCUN, and left FFG regions of the cognitive control circuit for which the association between change in cognitive control circuit activity and change in problem-solving ability was different for I-CARE versus U-CARE. Color shading from red to yellow indicates increasing significance levels, with P values ranging from 0.001 to 0.0002 and lower. (B) Scatter plots presenting the relationship between changes in cognitive control circuit activity at 6, 12, and 24 months relative to baseline (ΔCognitive control activity on the x axis) and changes in problem-solving ability at 6, 12, and 24 months relative to baseline (ΔProblem-solving ability on the y axis) for each of the regions shown in (A). Problem-solving ability was measured using the SPSI. Fitted lines in each plot represent trends for I-CARE (red) and U-CARE participants (gray). Beta estimates for interaction effects are shown in the figure, with estimates for each treatment group provided in table S2.

Table 2. Summary of response and predictive markers from linear mixed models.

Notes: Problem-solving ability was measured using the SPSI; depression symptoms were measured using the SCL-20. BA, Brodmann area; DF, degree of freedom; MNI, Montreal Neurological Institute.

Regions	BA	Cluster (mm3)	F stat	DF1	DF2	z Score	P value	Coordinates in MNI space
I-CARE modifies cognitive control circuit mechanisms that are associated with improved problem-solving ability (hypothesis 1)
Left dLPFC	9	5504	24.79	1	147.35	4.64	<0.001	−28, 34, 36
Right IPL	48	3800	19.74	1	153.93	4.15	<0.001	54, −30, 28
PCUN	7	2693	24.93	1	150.85	4.65	<0.001	6, −58, 50
Left FFG	37	4248	25.31	1	154.21	4.69	<0.001	−30, −36, −22
Longitudinal changes in the cognitive control circuit are associated with longitudinal changes in depression symptoms at 6, 12, and 24 months (hypothesis 2)
Left AI	47	368	11.95	1	175.33	3.20	<0.001	−24, 24, −10
Right IPL	40/48	1064	11.86	1	175.98	3.19	<0.001	62, −52, 24
Left CdN	25	760	12.59	1	168.28	3.29	<0.001	−6, 8, 0
Longitudinal changes in the cognitive control circuit are associated with longitudinal changes in depression symptoms, as a function of time (hypothesis 2)
Right dLPFC	45/47	416	8.34	2	126.89	3.36	<0.001	50, 36, −4
Early cognitive control circuit changes at 2 months are associated with improved problem-solving ability at 6, 12, and 24 months (hypothesis 3)
Right IPL	48/40	864	18.72	1	69.92	3.89	<0.001	46, −46, 30
Left IPL	40	2360	14.06	1	69.11	3.38	<0.001	−48, −38, 50
Early cognitive control circuit changes at 2 months are associated with improved depression symptoms at 6, 12, and 24 months (hypothesis 3)
Right IPL	40/48	464	11.85	1	69.04	3.10	<0.001	49, −42, 30
Left SPL	7/40	872	14.25	1	70.53	3.41	<0.001	−26, −50, 50

To better understand the potential mechanisms by which I-CARE may be modifying the cognitive control circuit to improve problem-solving ability, we examined behavioral performance changes that corresponded to circuit changes while participants were being scanned. We examined false alarms (commission errors) and response time during the Go-NoGo task that was used to engage the cognitive control circuit in the scanner. In the I-CARE group, attenuation of the right IPL and PCUN—cognitive control regions associated with improved problem-solving ability—was also associated with a reduction in commission errors to NoGo stimuli and slowed responses to Go stimuli (fig. S3 A to D). The observed relationship between changes in cognitive control circuit activity and changes in problem-solving ability remained when adding Go-NoGo performance measures as a confounding variable (table S4).

Because I-CARE was supplemented with pharmacotherapy as needed, we explored the confounding effect of medication prescription on the association between changes in cognitive control circuit activity and changes in problem-solving ability (see Supplementary Methods). We did not observe any difference in medication prescribed between the I-CARE and U-CARE groups at 2 (χ² = 1.99, P = 0.37), 6 (χ² = 1.83, P = 0.40), or 12 months (χ² = 1.76, P = 0.41). This indicates that participants in both arms had similar access to medication prescriptions. Furthermore, the association between changes in cognitive control circuit activity and changes in problem-solving ability was unaffected by the inclusion of medication prescription as a covariate (table S4).

Longitudinal changes in the cognitive control circuit are associated with longitudinal changes in depression symptoms at 6, 12, and 24 months

Contrary to our hypothesis, we did not observe a treatment–by–circuit activity change interaction indicating that I-CARE specifically modifies the cognitive control circuit to improve overall depression symptoms. However, we did observe significant main effects of circuit activity change across treatment groups localized in the IPL [beta estimate (b) = 0.23, 95% confidence interval (0.10, 0.35), P < 0.001], AI [b = 0.52, (0.23, 0.80), P < 0.001], and caudate nucleus [CdN; b = 0.41, (0.20, 0.63), P < 0.001] across 6, 12, and 24 months (Fig. 3A; see Table 2 for peak coordinates and summary statistics). Specifically, attenuation of the cognitive control circuit activity (ΔActivity) was associated with improved depression symptoms (Fig. 3B). In addition, another region within the cognitive control circuit, the right dLPFC, exhibited time-dependent associations with symptom changes (Time × ΔActivity: F = 8.34, P < 0.001; fig. S4A; see Table 2 for peak coordinates and summary statistics). Specifically, the attenuation of right dLPFC activity was similarly associated with symptom improvement at 6 and 12 months, but not at 24 months (fig. S4B). The observed relationships between changes in cognitive control circuit activity and changes in depression symptoms remained when adding Go-NoGo performance measures and medication prescription as confounding variables (table S4).

Fig. 3. Decreased cognitive control circuit activity is associated with improved depression symptoms at 6, 12, and 24 months across both treatment groups.

(A) Brain maps showing the left AI, the left CdN, and the right IPL regions of the cognitive control circuit for which an association between change in cognitive control circuit activity and change in depression symptom was observed across both treatment groups. Color shading from red to yellow indicates increasing significance levels, with P values ranging from 0.001 to 0.0002 and lower. (B) Scatter plots presenting the relationship between changes in cognitive control circuit activity at 6, 12, and 24 months relative to baseline (ΔCognitive control activity on the x axis) and changes in depression symptoms at 6, 12, and 24 months relative to baseline (ΔDepression symptoms on the y axis) for each of the regions shown in (A). Depression symptoms were measured using the SCL-20. Fitted lines in each plot represent trends for I-CARE (red) and U-CARE participants (gray). Beta estimates for interaction effects are shown in the figure, with estimates for each treatment group provided in table S2. Δ: change at a follow-up time point relative to baseline.

We further explored this main effect in relation to individual depression symptoms because the SCL-20 measure of depression symptoms comprises sub-measures of poor motivation relevant to cognitive control, as well as of dysphoria, as established in a prior analysis of the sample (see Supplementary Methods) (21). We observed that I-CARE (versus U-CARE) modified the relationship between changes in the cognitive control circuit activity (ΔActivity) and changes in “feeling everything is an effort,” as reflected in the significant treatment–by–circuit activity change interaction (Treatment × ΔActivity) for the right IPL (F = 4.53, P = 0.03) and FFG (F = 4.23, P = 0.04) and trends for the left dLPFC (F = 2.82, P = 0.10) and PCUN (F = 2.76, P = 0.10). We note that no multiple comparison correction was implemented for this analysis.

Early cognitive control circuit changes at 2 months are associated with improved problem-solving ability at 6, 12, and 24 months

To test whether changes in cognitive control circuit activity could be used as a predictive marker for treatment outcome, we focused on circuit activity change over the first 2 months of the intervention. We observed that changes in circuit activity at 2 months were associated with improvement of problem-solving ability at 6, 12, and 24 months across treatment groups, reflected in a significant main effect for the bilateral IPL [left IPL: b = −0.99, (−1.51, −0.47), P < 0.001; right IPL: b = −2.28, (−3.32, −1.24), P < 0.001] (Fig. 4A). Specifically, attenuation of cognitive control activity at 2 months relative to baseline predicted improvements in problem-solving ability that were sustained across 6, 12, and 24 months (Fig. 4B).

Fig. 4. Cognitive control circuit changes at 2 months predict improved problem-solving ability and improved depression symptoms at 6, 12, and 24 months.

(A) Brain maps showing bilateral IPL regions of the cognitive control circuit for which activity change at 2 months predicted improved problem-solving ability. Color shading from red to yellow indicates increasing significance levels, with P values ranging from 0.001 to 0.0002 and lower. (B) Scatter plots presenting the relationship between changes in cognitive control circuit activity at 2 months relative to baseline (ΔCognitive control activity at 2 months on the x axis) and changes in problem-solving ability at 6, 12, and 24 months relative to baseline (ΔProblem-solving ability on the y axis), for each of the regions shown in (A). Problem-solving ability was measured using the SPSI. (C) Brain maps showing right IPL and left SPL regions of the cognitive control circuit for which activity change at 2 months predicted improved depression symptoms. Color shading from red to yellow indicates increasing significance levels, with P values ranging from 0.001 to 0.0002 and lower. (D) Scatter plots presenting the relationship between changes in cognitive control circuit activity at 2 months relative to baseline (ΔCognitive control activity at 2 months on the x axis) and changes in depression symptoms at 6, 12, and 24 months relative to baseline (ΔDepression symptoms on the y axis) for each of the regions shown in (C). Depression symptoms were measured using the SCL-20. Fitted lines in each plot of (B) and (D) represent trends for I-CARE (red) and U-CARE participants (gray). Beta estimates for the overall effect are shown in (B) and (D), with estimates for each treatment group provided in table S2.

The circuit prediction model with IPL region predictors was verified compared with a baseline model with only baseline demographics and characteristics as predictors using the likelihood ratio test. The circuit predictor model exceeded the baseline model for each goodness-of-fit metric (table S1). We further showed that the circuit predictor model had robust generalizability using fivefold cross-validation with out-of-sample testing data (see fig. S5 for analysis strategy and fig. S6A for circuit predictors; see also Table 3 for overall model performance and table S3 for the performance of each fold). Using this approach, the full sample was divided into five subsets, and the model was trained and tested five times, each time with a different subset as the testing set, to ensure a robust validation of the model’s generalizability. Specifically, the average Pearson correlation coefficient between predicted and actual change in problem-solving was 0.53 for our circuit model, compared with 0.45 for the baseline model, resulting in a percentage improvement of 17.8%.

Table 3. Performance of circuit predictor models using 2-month cognitive control circuit changes to predict future improvements of treatment outcome for hypothesis 3.

Circuit model: using 2-month cognitive control circuit activity changes as well as baseline demographic and symptom characteristics as predictors to predict treatment outcomes at 6, 12, and 24 months. Baseline model: excluding 2-month cognitive control circuit measures and only including baseline demographic and symptom characteristics. Problem-solving ability was measured using the SPSI; depression symptoms were measured using the SCL-20. r_p, Pearson correlation coefficient; MAE, mean absolute error.

Measure	Problem-solving ability		Depression symptoms
	Baseline	Circuit	Baseline	Circuit
rp (SD)	0.45 (0.12)	0.53 (0.13)	0.25 (0.09)	0.51 (0.14)
MAE (SD)	1.76 (0.18)	1.62 (0.26)	0.60 (0.10)	0.49 (0.07)
AIC (SD)	756.62 (21.25)	579.54 (13.69)	439.24 (14.81)	260.64 (14.71)
BIC (SD)	775.81 (21.40)	605.93 (10.79)	459.79 (14.74)	289.53 (15.64)

Early cognitive control circuit changes at 2 months are associated with improved depression symptoms at 6, 12, and 24 months

Next, we applied our predictive model to the depression symptoms outcome. A significant main effect observed across treatment groups was specific to the right IPL [b = 0.62, (0.28, 0.97), P < 0.001] and left superior parietal cortex [SPL; b = 0.67, (0.36, 0.99), P < 0.001] (Fig. 4C). Mirroring the prediction for problem-solving outcomes, attenuation of cognitive control circuit activity at 2 months relative to baseline predicted improvements in depression symptoms sustained across 6, 12, and 24 months (Fig. 4D).

We also verified that this circuit predictor model for depression symptoms exceeded the performance of a baseline model on the goodness of fit (table S1) and showed generalizability using the same fivefold cross-validation with out-of-sample testing data (Fig. 5A; see fig. S6B for circuit predictors; see also Table 3 for overall model performance and table S3 for the performance of each fold). Specifically, the average Pearson correlation coefficient between predicted and actual change in depression symptoms was 0.51 for our circuit model as compared with 0.25 for the baseline model, resulting in a percentage improvement of 104%.

Fig. 5. Improved performance of a circuit-based model as compared with a baseline model for the prediction of depression symptom outcomes.

(A) Prediction of continuous depression symptom change at 6, 12, and 24 months with the 2-month circuit activity change of the cognitive control circuit as the predictor. The predicted SCL-20 change (Predicted Δdepression symptoms on the x axis) concatenated from the held-out data in fivefold cross-validation was plotted against the actual SCL-20 change (Actual Δdepression symptoms on the y axis). Depression symptoms were measured using the SCL-20. (B) Prediction of the categorical treatment response with the 2-month circuit change of the cognitive control circuit as the predictor in twofold cross-validation. Response was defined as ≥50% reduction of SCL-20 at 6 months, and nonresponse was defined as <50% reduction. The AUC for the precision-recall curve is shown for our circuit-based model and for a baseline model that only included the baseline demographic and symptom characteristics as predictors. The dotted gray line indicates the performance by chance.

As an exploratory step, we examined alternative behavioral and circuit predictors, including change in Go-NoGo performance at 2 months and baseline cognitive control circuit activity (see Supplementary Methods). However, no main effect or interaction effect was observed.

Considering the translational application of using early cognitive control circuit changes at 2 months to predict subsequent clinical outcomes

In precision medicine trials and clinical practice, the typical clinical outcome is expressed as a categorical value. For depression, the primary outcome is a common response as defined by a ≥50% reduction in symptom severity (22). Thus, we tested the potential translational application of the early change in circuit activity at 2 months as a predictor of response defined by ≥50% reduction on the SCL-20 at the subsequent 6-month time point. Using a logistic regression model combined with twofold cross-validation and out-of-sample testing data, we tested a model based on circuit predictors—the circuit model—compared with a baseline model and evaluated the model using precision-recall (PR) curve, receiver operating curve (ROC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) metrics. The model using 2-month cognitive control circuit predictors had a robust performance [PR–area under the curve (PR-AUC) = 46%; ROC-AUC = 0.65; sensitivity, 90%; specificity, 40%; PPV, 77%; NPV, 62%; Fig. 5B], exceeding that of the baseline model (PR-AUC = 31%; ROC-AUC = 0.52; sensitivity, 45%; specificity, 43%; PPV, 65%; NPV, 25%). The model localized predictors in the IPL, SPL, dLPFC, dorsal anterior cingulate cortex (dACC), AI, and putamen (fig. S7). The use of the circuit model increased PPV by 0.12, resulting in a percentage improvement of 39% and corresponding to a number needed to treat (NNT) of 8, or the number of patients needed to substantially change the outcome for one individual of eight, beyond an NNT of 60, which is meaningful for analogous cardiovascular disease (23).

Test-retest reliability and generalizability of the cognitive control circuit

Our findings suggest a strong overlap between the identified response and predictive markers and the cognitive control circuit (fig. S8). This overlap was also observed with models that excluded data at 24 months (see Supplementary Methods), as shown in figs. S9 to S11, corresponding to Figs. 2 to 4. The prerequisite of a marker is being reliable over time. Our reliability analysis showed that the test-retest reliability over 2 months (fig. S12A), quantified by intraclass correlation coefficient (ICC), met or exceeded the psychometric criteria (ICC > 0.6) in a majority of circuit regions of interest evoked by the Go-NoGo task, except the bilateral FFG (fig. S12B). In addition, the cognitive control circuit evoked by the NoGo > Go contrast from multiple datasets (fig. S13 A to D) showed that our Go-NoGo task consistently activated typical Go-NoGo regions in the cognitive control circuit, including the bilateral dLPFC, bilateral IPL, and the dACC.

DISCUSSION

Using fMRI data collected five times over 24 months within a randomized controlled trial, we have shown that the I-CARE behavioral intervention modifies the brain’s cognitive control circuit and improves outcomes in depression with obesity. Using LMMs and predictive models, we showed that circuit activity change after 2 months predicted subsequent outcomes at 6, 12, and 24 months. These findings suggest that imaging the cognitive control circuit might track responses to behavioral interventions and provide insight into future clinical outcomes in these populations.

Attenuation of the cognitive control circuit activity over 6, 12, and 24 months was specifically associated with improved problem-solving in the I-CARE group, whereas the opposing effect was observed in the U-CARE group. This specific effect of I-CARE was anatomically localized to the dLPFC, IPL, PCUN, and FFG, regions that define key nodes of the cognitive control circuit (10). Our finding suggests that with the problem-solving intervention, I-CARE may down-regulate cognitive control circuitry over time, shifting from cognitive overload to a more efficient strategy that manages irrelevant stimuli and achieves adaptive problem-solving behavior. In line with this view, attenuation in the cognitive control circuit correlated with fewer commission errors on the Go-NoGo task during scans and better problem-solving outcomes. This finding also supports our previous research (16, 24) showing that both excessive and insufficient activity of cognitive control circuitry contributes to different biotypes of depression.

Tempering of the cognitive control circuit activity was also linked with lower overall depression symptoms at 6, 12, and 24 months. A closer look at individual symptoms showed that I-CARE’s effects on this circuit specifically improved poor motivation compared with U-CARE, aligning with known connections between cognitive control and motivation in depression (25). The potential of I-CARE to specifically enhance motivation by modifying cognitive control activity deserves further detailed study.

This study demonstrates the potential value of early changes in the cognitive control circuit activity for predicting long-term treatment outcomes for depression at 6, 12, and 24 months. Specifically, changes in parietal regions’ activity after 2 months consistently predict improvements in problem-solving ability and depression symptoms sustained over 24 months. These early circuit changes also better distinguish treatment responders from nonresponders beyond baseline patient clinical demographic assessments, which suggests a critical 2-month window for neural adaptations under behavioral interventions. Our findings further validate the good test-retest reliability of these circuit changes as predictive markers.The strengths of our study include the longitudinal randomized controlled design with multiple assessment time points, a customized analysis strategy for testing circuit prediction within this design, and the integration of imaging with treatment outcomes in a vulnerable depression population.

We also recognize several limitations. Because of the natural attrition at the 12- and 24-month time points, we need to be cautious about the stability of the longer-term effects. Given our a priori focus on cognitive control and problem-solving, we used a Go-NoGo task to engage the cognitive control circuit. The use of this task limits our interpretation of processes of goal-directed inhibition of irrelevant information, so future studies should expand to other tasks, processes, and circuits. Depression is highly heterogeneous. Although our findings of intervention-induced circuit changes can improve specific symptoms like motivation, we did not comprehensively assess individual symptoms. Future investigations could expand to explore these individual symptoms in more detail. Demographically, our sample had a majority of college-educated non-Hispanic white women, which limits generalization to a more diverse population. Although a majority of regions that defined the cognitive control circuit as predictive of outcomes showed good test-retest reliability, the bilateral FFG showed more moderate reliability, which indicates the need for ongoing technical development to advance the clinical translation of circuit-based measures.

Despite these limitations, our findings indicate that changes in the brain’s cognitive control circuit during a behavioral intervention can promote more efficient processing of irrelevant stimuli, improving the patient’s problem-solving ability. These results suggest that changing behavior can induce neural plasticity, which improves functional ability even in vulnerable groups like the participants with depression symptoms and comorbid obesity studied here. Early circuit changes in the first 2 months of the intervention were critical predictors of better problem-solving and reduced depression symptoms over the 24-month timeline of the study. To evaluate whether activity changes in the cognitive control circuit could serve as responsive and predictive markers, future studies should adopt longitudinal, prospective trial designs and include a broader range of behavioral interventions. The ultimate goal is to integrate these circuit markers into clinical practice to guide treatment choices and enhance outcomes, which aligns with the move toward precision psychiatry. Specifically, early detection of circuit changes could provide actionable feedback for patients and providers to enhance treatment engagement, whereas a lack of change may prompt adjustments such as personalized problem-solving sessions, medication intensification, or alternative treatments.

MATERIALS AND METHODS

Study design

As outlined in the prespecified protocol for ENGAGE (18), participants were recruited within four medical centers of Sutter Health’s Palo Alto Medical Foundation (San Francisco, CA, USA) from 29 November 2015 through 2 November 2018 until the completion of the 24-month data collection. Inclusion criteria were (i) nine-item Patient Health Questionnaire score ≥ 10, indicating moderate or more severe depression, (ii) body mass indices (BMI) ≥ 30 (≥27 for Asian participants), (iii) age ≥ 18 years, and (iv) no alcohol or substance use disorder or any other exclusion criteria, including pregnancy, inability to communicate well in English, or plans to relocate. Participants were excluded if they had an active Axis I disorder other than major or minor depressive disorder or dysthymia, with the exception of comorbid anxiety disorders. Additional imaging exclusion criteria for ENGAGE included weight ≥ 350 pounds because of scanner constraints, magnetic resonance imaging (MRI) contraindications, traumatic brain injuries, presence of a tumor, or other known structural brain abnormality. The sample and design were powered to detect medium-sized effects (Cohen’s d 0.3 to 0.5). The study followed the CONSORT guidelines and is registered with ClinicalTrials.gov (NCT02246413). The study was approved by the Institutional Review Boards (IRB) at Stanford University (IRB 35,732 and 41,837) and University of Illinois at Chicago (2015–1324). All participants in the study provided written informed consent.

Participants were randomly assigned to receive either I-CARE or U-CARE (26). The I-CARE intervention integrated two evidence-based behavioral interventions, one for depression and one for obesity. The Program to Encourage Active, Rewarding Lives for Seniors (PEARLS) was an intervention for depression with problem-solving as its core component and as the first-line therapy (27, 28) as well as synergistic behavioral activation therapy, supplemented with stepwise increases in doses and number of pharmacotherapy as needed. The Group Lifestyle Balance (GLB) program was adapted from the Diabetes Prevention Program (29). It was grounded in social cognitive theory using a goal-based approach and consisted of videos for self-study, which were previously demonstrated to be effective for weight loss and cardiometabolic risk reduction in primary care. The intensive intervention phase for the first 6 months included nine individual face-to-face sessions lasting 60 min each (four weekly sessions, followed by two biweekly and then three monthly sessions) and the watching of 11 home-viewed GLB videos lasting 20 to 30 min each. This was followed by another 6 months of the maintenance phase, which included monthly telephonic problem-solving sessions lasting 15 to 30 min each. Participants in both the I-CARE and U-CARE groups continued to receive medical care from their personal physicians. U-CARE participants received a wireless activity tracker with batteries but not other self-care materials, such as intervention handouts, DVD sets, or an online access code for GLB videos. Investigators, the data and safety monitoring board, outcome assessors, and data analysts were blinded to the intervention assignment until after the completion of the primary data review through 24 months.

Treatment outcome measures

Problem-solving ability and depression symptom severity were the prespecified primary outcomes for depression and were collected at baseline, 6, 12, and 24 months (Fig. 1, A and B) (24). The self-administered SPSI (30) was used to assess social problem-solving. The SPSI total score ranges from 0 to 20, with higher scores being indicative of better problem-solving abilities. Depression symptoms were measured using the SCL-20 (31). Treatment response was defined as a ≥50% reduction of symptoms rated by the SCL-20 (22).

fMRI data acquisition

fMRI data were collected at baseline and at the end of initial intervention (2 months), intensive intervention (6 months), maintenance (12 months), and follow-up (24 months; Fig. 1A). As prespecified in the protocol paper, we focused on the cognitive control circuit as the brain circuit target (18). We used a Go-NoGo task that is well established for engaging this circuit (Fig. 1C) and a well-established imaging sequence (13, 15, 16, 32). In summary, participants were instructed to respond quickly to green stimuli (Go trials) and to inhibit responses to red stimuli (NoGo trials). Stimuli were presented in mini-blocks of two Go or two NoGo stimuli per block to allow for event-related analysis at the mini-block level. A total of 180 Go trials and 60 NoGo trials were presented for 0.5 s each, with an interstimulus interval of 0.75 s. Behavioral data for reaction time and false alarms (commission errors) were recorded. We focused on activation for the NoGo > Go contrast specific to our construct of cognitive control.

Statistical analysis

Our analysis strategy is summarized in Fig. 1 (D and E).

Testing hypothesis 1

Changes in the cognitive control circuit activity will be directly associated with changes in problem-solving ability, and this association will be modified by I-CARE. To test our first preregistered hypothesis (18), we constructed an LMM (Eq. 1; Fig. 1D) and ran it voxel-wise across cortical and subcortical gray matter using the following terms:

$$\begin{gathered} \text{ΔSPSI}=1+\text{ΔActivity}+\text{Treatment}\times \text{ΔActivity}+\text{Time}\times \text{ΔActivity} \\ +\text{Treatment}\times \text{Time}\times \text{ΔActivity}+{\text{SPSI}}_{\text{baseline}} \\ +\text{BM}{\text{I}}_{\text{baseline}}+\text{Age+Sex}+(1\mid \text{Participants}) \end{gathered}$$ (1)

where ΔSPSI is the change in problem-solving ability at 6, 12, and 24 months (primary outcome); ΔActivity is the change in NoGo > Go elicited cognitive control circuit activity at the same time points relative to baseline; Treatment is I-CARE versus U-CARE; Time is 6, 12, or 24 months; and SPSI_baseline, BMI_baseline, Age, and Sex are baseline demographic and symptom characteristic covariates of SPSI, BMI, age, and sex.

The dependent measure is the change in problem-solving ability (ΔSPSI) at 6, 12, and 24 months relative to baseline. The focal effects of interest were the treatment–by–circuit activity change interaction (Treatment × ΔActivity) and the main effect of circuit activity change (ΔActivity). Interaction effects of time were also included. Interaction and main effects in the model were fixed-effect terms. Baseline SPSI, BMI, age, and sex were included as regressors. A random intercept was added for each participant to account for repeated measures.

LMMs were implemented using the fitlme function of the Statistics and Machine Learning Toolbox in MATLAB (www.mathworks.com/products/matlab.html) and customized scripts. The significance of effects was tested under the type III hypotheses using the analysis of variance (ANOVA) function in MATLAB, with degrees of freedom estimated by the Satterthwaite method (33). We corrected for multiple comparisons using a voxel threshold of P < 0.001 and a Gaussian random field (GRF) theory family-wise error cluster-level threshold at P < 0.05 using DPABI V5.1 (http://rfmri.org/dpabi). Data smoothness for GRF correction was estimated on the statistical image following a procedure similar to Functional Magnetic Resonance Imaging of the Brain Software Library (FSL) easythresh. The corrected brain results were expressed in F scores and were also standardized into z scores, enabling easier comparison and interpretation of statistical significance. For regions that survived correction, the beta estimates of each model and the 95% confidence intervals were evaluated by emmeans packages (https://cran.r-project.org/web/packages/emmeans/index.html) in R version 4.0.5 (www.r-project.org/) and are shown in figures when applicable. Each of the models from the surviving regions was compared with a baseline model (Eq. 2) using a likelihood ratio test.

$$\text{ΔSPSI}=1+\text{SPS}{\text{I}}_{\text{baseline}}+\text{BM}{\text{I}}_{\text{baseline}}+\text{Age}+\text{Sex}+(1\mid \text{Participants})$$ (2)

Testing hypothesis 2

Changes in cognitive control circuit activity will be associated with changes in depression symptoms, and this association will be modified by I-CARE. Corresponding LMMs were used to test hypothesis 2 (Eq. 3; Fig. 1D) using the following terms:

$$\begin{gathered} \text{ΔSCL-20}=1+\text{ΔActivity}+\text{Treatment}\times \text{ΔActivity}+\text{Time} \\ \times \text{ΔActivity}+\text{Treatment}\times \text{Time}\times \text{ΔActivity}+\text{SCL-}{20}_{\text{baseline}} \\ +\text{BM}{\text{I}}_{\text{baseline}}+\text{Age+Sex}+(1\mid \text{Participants}) \end{gathered}$$ (3)

where ΔSCL-20 is the change in depression symptoms at 6, 12, and 24 months (primary outcome) and SCL-20_baseline is the symptom characteristic covariate of SCL-20.

Here, the dependent measure is the change in depression symptoms (ΔSCL-20) at 6, 12, and 24 months relative to baseline. We tested the significance of the LMM, correction for multiple comparisons, generated beta estimates, and verified goodness of fit compared with a baseline model (Eq. 4) using the procedure described under hypothesis 1.

$$\text{ΔSCL}-20=1+\text{SCL}-{20}_{\text{baseline}}+\text{BM}{\text{I}}_{\text{baseline}}+\text{Age+Sex}+(1\mid \text{Participants})$$ (4)

Testing hypothesis 3

Early changes in the cognitive control circuit activity over 2 months predict subsequent outcomes for both problem-solving ability and depression symptoms at 6, 12, and 24 months. LMMs were then conducted to test whether early change in the cognitive control circuit after 2 months predicts subsequent SPSI and SCL-20 outcomes as dependent measures at 6, 12, and 24 months relative to baseline (Eqs. 5 and 6; Fig. 1E). These models were also undertaken with a voxel-wise whole-brain analysis using the following terms:

$$\begin{gathered} \text{ΔSPSI}=1+{\text{ΔActivity}}_{2\text{MO}}+\text{Treatment}\times {\text{ΔActivity}}_{2\text{MO}}+\text{Time} \\ \times {\text{ΔActivity}}_{2\text{MO}}+\text{Treatment}\times \text{Time}\times {\text{ΔActivity}}_{2\text{MO}} \\ +{\text{SPSI}}_{\text{baseline}}+{\text{BMI}}_{\text{baseline}}+\text{Age}+\text{Sex}+(\text{1|Participants}) \end{gathered}$$ (5)

$$\begin{gathered} \text{ΔSCL-20}=1+{\text{ΔActivity}}_{2\text{MO}}+\text{Treatment}\times {\text{ΔActivity}}_{2\text{MO}}+\text{Time} \\ \times {\text{ΔActivity}}_{2\text{MO}}+\text{Treatment}\times \text{Time}\times {\text{ΔActivity}}_{2\text{MO}} \\ +\text{SCL-}{20}_{\text{baseline}}+BM{I}_{\text{baseline}}+\text{Age}+\text{Sex}+(1\mid \text{Participants}) \end{gathered}$$ (6)

where ΔActivity_2MO is the change in NoGo > Go elicited cognitive control circuit activity at 2 months relative to baseline.

Here, the effects of focal interest were the treatment–by–circuit activity change at 2 months interaction (Treatment × ΔActivity_2MO) and the main effect of cognitive control circuit activity change at 2 months (ΔActivity_2MO) across treatment groups. Fixed-effect activity terms were circuit activity change from baseline to 2 months instead of circuit activity change at the same time point as the treatment outcome. We also tested the significance of the LMM, corrected for multiple comparisons, generated beta estimates, and verified goodness of fit compared with a baseline model (Eq. 2 or 4) using the procedure described under hypothesis 1.

Validating the generalizability of circuit predictor models used to test hypothesis 3

We validated the generalizability of circuit predictor models with SPSI and SCL-20 outcomes using fivefold cross-validation (fig. S5). We reported the mean value of the correlation coefficient of the predicted versus actual values of held-out data, as well as mean absolute error, AIC, and BIC across each fold. A simplified version of Eq. 5 or 6 was used by removing all interaction effects of treatment or time (Eqs. 7 and 8).

$$\begin{gathered} \text{ΔSPSI}=1+{\text{ΔActivity}}_{2\text{MO}}+{\text{SPSI}}_{\text{baseline}} \\ +\text{BM}{\text{I}}_{\text{baseline}}+\text{Age}+\text{Sex}+(1\mid \text{Participants}) \end{gathered}$$ (7)

$$\begin{gathered} \text{ΔSCL-20}=1+{\text{ΔActivity}}_{2\text{MO}}+{\text{SCL-20}}_{\text{baseline}} \\ +\text{BM}{\text{I}}_{\text{baseline}}+\text{Age}+\text{Sex}+(\text{1|Participants}) \end{gathered}$$ (8)

In each of the fivefold iterations, we combined the feature selection and model training process using 80% of the data and tested the performance of the model using the held-out 20% of the data (fig. S5). Specifically, in each iteration, we first ran the LMM of Eq. 7 or 8 only on the training data (80%) for each voxel of the whole brain. We then applied a multiple comparison correction procedure with a voxel threshold of P < 0.01 and a GRF family-wise error cluster-level correction at P < 0.05 to retain survived regions as circuit regions of interest. Activity changes of these regions at 2 months were selected as circuit features. All circuit features were extracted from the training data and were collectively entered into the Eq. 7 or 8 model to train a final model using the same training data. This trained model would be able to take cognitive control circuit activity from new participants and output change in their outcome measures.

When testing the final model on the held-out testing data (20%), we similarly extracted from the testing data the circuit features initiated from the same circuit regions of interest derived above in the training process and entered them into the final model to generate the predicted outcomes (change in problem-solving ability or depression symptoms) for the testing samples. The predicted outcomes of the testing data from each fold were then combined to assess the overall performance of our model, in addition to the average performance of each fold. Observations from any participant were included only in the training or in the testing dataset to maintain the independence of the training and testing data. The random effect of participants was not removed to retain the partial pooling effect, which facilitates robust estimations in the training model (34). For held-out participants in the testing group, only fixed effects were used to predict future treatment outcomes. Circuit regions of interest from each fold were averaged to generate an average circuit region map to show which regions contributed to at least one fold of the cross-validated model. To confirm that our circuit model provided information in addition to the baseline symptom or problem-solving scores, we also compared the above metrics with a baseline model that only included the baseline demographic and symptom characteristics as predictors (Eqs. 2 and 4).

We also explored the clinical translational value of using the 2-month cognitive control circuit activity change in predicting categorical treatment response at the primary 6-month time point. We ran a voxel-wise whole-brain logistic regression with 2-month circuit activity change as the independent variable and the categorical response as the dependent variable (Eq. 9), embedded within a cross-validation, similar to the above continuous outcome prediction, illustrated in fig. S5.

$$\begin{gathered} \text{logit}\left({\text{Response}}_{6\text{MO}}\right)=1+{\text{ΔActivity}}_{2\text{MO}}+\text{SCL}-{20}_{\text{baseline}} \\ +\text{BM}{\text{I}}_{\text{baseline}}+\text{Age}+\text{Sex} \end{gathered}$$ (9)

where logit(Response_6MO) is the log odds of the categorical response at 6 months.

Given our response rate of 32%, two folds were implemented to ensure an adequate proportion of responders in each fold. The same training procedure used for circuit predictor models with SPSI and SCL-20 outcomes was applied here to select circuit regions of interest and was used to train a final model based on Eq. 9. The same testing procedure was conducted to assess the overall and average performance of our model.

For the categorical response prediction, we chose the optimal threshold by maximizing the F1 score, the first harmonic mean of true positive rate (also known as sensitivity or recall), and PPV (also known as precision), thus maximizing the identification of the number of actual responders being correctly identified (high sensitivity) and minimizing the number of actual nonresponders being identified as responders (high PPV). The performance of our model was primarily evaluated by visualizing the PR curve and the estimation of the PR-AUC (35). We also compared our circuit and baseline models (Eq. 10) on other common metrics, including sensitivity, specificity, PPV, NPV, and the receiver operating characteristic curve. Circuit regions of interest from each fold were averaged to generate an average circuit region map.

$$\text{logit}\left({\text{Response}}_{6\text{MO}}\right)=1+\text{SCL}-{20}_{\text{baseline}}+\text{BM}{\text{I}}_{\text{baseline}}+\text{Age+Sex}$$ (10)

Data and materials availability:

The full analysis codes are available at Zenodo (36) and https://github.com/WilliamsPanLab/ENGAGE-CognitiveControl. The ENGAGE study dataset is available upon request from Stanford BRAINnet at www.stanfordpmhw.com/datasets. The BRAINnet repository meets the requirements for being public but also aligns with the procedures of other official public and scientific repositories such as HCP, ABCD, and NDA. This choice aligns with the FAIRness guidelines and respects the original funding requirements, allowing for appropriate source contributions and citations. Our approach is specifically designed for scientific use, which includes limiting access to for-profit entities to comply with the original funding stipulations and participant consent. Therefore, total open access is not feasible. We intend to provide public access that is consistent with the consent agreements and the original funding intentions, similar to the data shared through NIH repositories. On Stanford BRAINnet, we established a data access request form that screens users, similar to other public repositories.