Neural and behavioral markers of inhibitory control predict symptom improvement during internet-delivered cognitive behavioral therapy for depression

The study investigates whether baseline inhibitory control and intrinsic brain connectivity within overlapping inhibitory control–emotion regulation networks predict response to internet-delivered CBT (iCBT) for major depressive disorder. Executive functioning deficits, including inhibitory control, are common in MDD, persist across episodes, and relate to emotion regulation difficulties central to CBT. Prior work suggests poorer executive function predicts poorer treatment response, but few studies have tested EF as a predictor specifically for psychotherapy, particularly iCBT, and often lacked control conditions. Given the overlap between neural systems subserving inhibitory control and emotion regulation (notably dACC, anterior insula, TPJ, and prefrontal regions), the authors hypothesized that better baseline inhibitory control performance and stronger rsFC within this network would predict greater symptom improvement, with stronger effects in iCBT versus a monitored attention control (MAC).

- Behavioral inhibitory control in MDD is mixed: classic Flanker interference differences vs. healthy controls are often absent, but nuanced measures (post-error adjustments, sequential dependencies/Gratton effect) can reveal abnormalities. MDD often shows atypical post-error adjustments and weaker Gratton effects. - Neuroimaging shows more consistent alterations: MDD exhibits hypoactivity in inhibitory control regions and compensatory hyperactivity in prefrontal areas during tasks. dACC activation differences during error processing have been reported. - Overlap exists between neural substrates of inhibitory control and emotion regulation (dACC, anterior insula, TPJ, inferior/middle frontal gyri). These regions form an intrinsic resting-state network; rsFC within this network relates to inhibitory control performance. - Prior predictors of treatment response in MDD include ACC activity/connectivity, but fewer studies target the broader inhibitory control–emotion regulation network; one study linked right insula–middle temporal gyrus rsFC to psychotherapy response (behavioral activation) without a control condition. - iCBT is scalable and effective; predictors of response have been explored, but baseline inhibitory control as a predictor for iCBT had not been examined with a randomized control. - The present work addresses gaps by testing EF and rsFC predictors in a randomized iCBT vs MAC design, focusing on prescriptive (treatment-specific) and prognostic effects.

Design: Randomized controlled trial comparing internet-based CBT (iCBT) versus a monitored attention control (MAC) over 10 weeks. Analyses focus on secondary outcomes predicting post-treatment PHQ-9 scores. Participants: 266 consented; inclusion required current MDD diagnosis and baseline PHQ-9 between 10–23; no psychotropic medications. After exclusions for non-completion (iCBT n=7; MAC n=10) and one inconsistent psychiatric history, 60 participants remained (iCBT n=30; MAC n=30). IRB-approved; registered (NCT01598922). Interventions and measures: - iCBT: Six online CBT lessons with weekly homework over 10 weeks. MAC: Completed online symptom questionnaires only (including PHQ-9) at each of six sessions; no CBT content. - Primary predictor measures at baseline: • Flanker task (modified Eriksen arrow task) outside scanner: 30 practice, five blocks x 70 trials (46 congruent, 24 incongruent). Derived indices: Flanker interference (accuracy, RT), Gratton effect (sequential dependency; accuracy, RT), post-error adjustments (accuracy, RT). Higher Flanker interference indicates worse inhibitory control; higher Gratton and post-error adjustment scores indicate better inhibitory control. • Resting-state fMRI: Siemens Tim Trio 3T, 32-channel coil. Preprocessing SPM8; denoising/analysis in CONN toolbox v15.d. ROIs from Langner et al.'s conjunction of cognitive emotion regulation and cognitive action regulation meta-analyses, segmented into: right anterior insula (AI), left AI, frontal medial (used as dACC ROI per text description for overlapping self-regulation network), and right temporoparietal junction (TPJ). Pairwise ROI time-series correlations computed and Fisher z-transformed: dACC–left AI, dACC–right AI, dACC–TPJ, left AI–right AI, left AI–TPJ, right AI–TPJ. • Clinical outcome: PHQ-9 at baseline and post-treatment (self-report). HAMD was also collected and examined in supplemental analyses. Data handling and statistics: - Missing baseline data imputed via missForest (random forest-based). Missingness: 0% for demographics and PHQ-9; up to 15% for rsFC. QC failures for some Flanker variables were also imputed (7 Gratton, 12 post-error). - Group comparisons (iCBT vs MAC): t-tests and chi-square on demographics, PHQ-9, Flanker metrics, rsFC. - Predictive modeling: Elastic net regression (glmnet) with 10-fold CV for hyperparameter tuning (alpha and lambda). Three-stage modeling: Stage 1: Demographics/treatment (age, gender, treatment group, baseline PHQ-9) + Flanker variables (accuracy, RT interference; Gratton accuracy/RT; post-error accuracy/RT). Stage 2: Demographics/treatment + rsFC variables among dACC, left AI, right AI, TPJ. Stage 3: Variables retained from Stages 1–2 plus baseline PHQ-9 entered jointly (Flanker and rsFC key candidates). - Stability: Each elastic net analysis repeated 10,000 times; variables retained if non-zero coefficients in ≥75% of replicates. Essential predictors defined as those with absolute coefficients exceeding the mean absolute coefficient of all retained variables in that stage (full list in Supplementary Table S2). - Model performance: Nested 10-fold CV within each of 10,000 runs to compute mean squared error (MSE). Compared Stage models to a null model via corrected t-tests. - Post hoc interpretability: Ordinary least squares regression including essential Stage 3 predictors; partial regression plots produced (effects package in R).

- Sample and baseline differences: • N=60 (iCBT n=30; MAC n=30), unmedicated MDD. • MAC showed lower baseline right AI–left AI rsFC than iCBT (t(58)=2.50, p=0.015). • As previously reported, MAC had higher post-treatment PHQ-9 than iCBT (t(58)=2.77, p=0.007). No other significant baseline group differences. - Flanker task effects across the combined sample: • Significant Flanker interference: accuracy t(59)=12.84, p<0.001, d=1.66; RT t(59)=39.08, p<0.001, d=5.04. • Gratton effect significant for accuracy (t(59)=5.80, p<0.001, d=0.75), not for RT (t(59)=-0.57, p=0.57). • Post-error effect significant for RT (t(59)=2.82, p=0.006, d=0.36), not for accuracy (t(59)=-1.16, p=0.25). - Elastic net retained predictors (essential): Stage 1 (behavioral): treatment group; gender; interactions of treatment group with gender, Flanker RT, Gratton accuracy, post-error accuracy. Stage 2 (rsFC): treatment group; right AI–TPJ rsFC; interactions of treatment group with age, gender, and right AI–TPJ rsFC. Stage 3 (combined): treatment group; gender; Flanker RT; right AI–TPJ rsFC; left AI–right AI rsFC; interactions of treatment group with baseline PHQ-9, age, gender, Flanker RT, Gratton accuracy, post-error accuracy, right AI–TPJ rsFC, and left AI–right AI rsFC. - Model performance (10,000 repetitions with nested CV): • Stage 3 MSE=0.89 significantly lower than null MSE=0.99 (Mdiff=-0.098, SDdiff=0.31, t=3.14, p=0.0017), Stage 1 MSE=0.94 (Mdiff=-0.042, SDdiff=0.21, t=2.73, p=0.0062), and Stage 2 MSE=0.99 (Mdiff=-0.092, SDdiff=0.28, t=3.43, p=0.00059). • Stage 1 (p=0.071) and Stage 2 (p=0.91) did not differ from the null. - Linear model with essential Stage 3 predictors: • Explained variance R^2=0.46; F(17,42)=2.07, p=0.03. • Higher Flanker RT interference (worse inhibitory control) predicted higher post-treatment PHQ-9; effect stronger in iCBT than MAC. • Stronger right AI–TPJ rsFC and weaker right AI–left AI rsFC predicted higher post-treatment PHQ-9 (poorer outcome); the beneficial pattern was lower right AI–TPJ and higher right AI–left AI rsFC; effects stronger in iCBT than MAC. • In iCBT only: better Gratton accuracy predicted lower post-treatment PHQ-9; unexpectedly, higher post-error accuracy predicted higher post-treatment PHQ-9 (worse outcome). - Overall: Baseline inhibitory control (behavioral) and rsFC within inhibitory control–emotion regulation regions predicted lower depression severity after iCBT, with stronger prescriptive effects for iCBT compared to MAC.

Findings support the hypothesis that baseline inhibitory control and intrinsic connectivity within overlapping inhibitory control–emotion regulation networks are meaningful predictors of response to iCBT in MDD. Better behavioral inhibitory control (lower Flanker RT interference; stronger sequential dependency/Gratton accuracy) was associated with greater symptom improvement, particularly within the iCBT arm, indicating treatment-specific (prescriptive) utility beyond nonspecific factors present in the control condition. The unexpected negative association between post-error accuracy and iCBT improvement may reflect that individuals with hypersensitive error processing benefit more from CBT’s cognitive-emotional strategies, although this requires direct testing. Neurally, connectivity patterns centered on the anterior insula emerged as predictive, while dACC-related rsFC did not survive as essential predictors. Higher inter-insular (right–left AI) connectivity and lower right AI–TPJ connectivity were associated with better outcomes, consistent with the AI’s role in salience detection, network switching, and error monitoring, and the TPJ’s role in reorienting that can compete with sustained control. Lower AI–TPJ coupling may favor inhibitory control over stimulus-driven reorienting, aligning with better treatment response. Together, these behavioral and neural markers suggest that preserved or more efficiently organized inhibitory control systems facilitate engagement and benefit from CBT techniques (e.g., reappraisal), making them promising candidates for guiding personalized treatment selection.

This randomized study identifies baseline inhibitory control performance and anterior insula-centered rsFC as predictors of symptom improvement during iCBT for MDD, with stronger effects in iCBT than a monitored attention control. Specifically, lower Flanker RT interference, stronger inter-insular rsFC, and weaker AI–TPJ rsFC predicted better outcomes, alongside interactions indicating prescriptive utility for iCBT. These results add to evidence that preserved executive function supports better psychotherapy response and highlight insula–TPJ network dynamics as a neural substrate of CBT benefit. Future research should replicate in larger samples, examine long-term outcomes, incorporate broader cognitive batteries and multimodal biomarkers, and directly compare multiple treatment modalities to advance treatment selection algorithms.

- Baseline group difference: the iCBT group had higher left AI–right AI rsFC at baseline than MAC despite randomization. - Causality and unmeasured factors: Associations may be influenced by related constructs (e.g., emotion regulation) not fully captured; directionality cannot be confirmed. - Sample size: Modest N limits power, especially for interaction tests; replication in larger cohorts is needed. - Follow-up: Outcomes beyond the 10-week window (e.g., 6–12 months) were not assessed. - Outcome measurement: Primary outcome emphasized self-report PHQ-9; generalizability to clinician-rated measures may differ (supplemental HAMD analyses provided). - Generalizability across treatments: Findings may not extend to other modalities (e.g., pharmacotherapy, ECT); treatment-specific effects require direct comparisons. - Cognitive task scope: Only the Flanker task was used; inhibitory control is multifaceted and results may not generalize to other EF measures.