Toward precision psychological rehabilitation: predicting CBT efficacy in post-stroke depression using machine learning

TL;DR

This study explores how machine learning can predict which post-stroke depression patients will benefit most from cognitive behavioral therapy.

Contribution

The study introduces an interpretable machine learning model to predict individual CBT response in post-stroke depression patients.

Findings

CBT was associated with greater improvement in depressive symptoms compared to a control group.

Random forest classifier outperformed other models with an AUC of 0.897 and accuracy of 0.861.

Baseline depressive severity, anxiety, self-efficacy, and social support were key predictors of CBT response.

Abstract

This study retrospectively examined the potential benefits of cognitive behavioral therapy (CBT) for post-stroke depression (PSD) and developed an interpretable machine learning model to predict individual treatment response.

Clinical and psychological data from 120 PSD patients receiving CBT and 123 patients in a control group were analyzed. Changes in PHQ-9, GAD-7, and General Self-Efficacy Scale (GSE) scores were compared between groups. Within the CBT cohort, a random forest classifier was trained to predict treatment response and compared with logistic regression and gradient boosting models. SHAP values and ablation analyses were used to assess feature contributions and model interpretability.

Baseline characteristics were comparable between groups. The CBT group showed greater improvement in depressive symptoms than the control group. Among predictive models, the random forest classifier demonstrated the highest performance (AUC = 0.897; accuracy = 0.861). SHAP and ablation analyses consistently highlighted baseline depressive severity (PHQ-9), anxiety (GAD-7), self-efficacy (GSE), and social support (SSRS) as the most influential predictors of CBT response.

CBT was associated with greater improvement in depressive symptoms among patients with post-stroke depression; however, causal inferences should be made cautiously given the retrospective design. The proposed machine learning model shows preliminary promise for predicting treatment response, but further validation in prospective and multi-center studies is needed before clinical implementation.

Figures