Predicting rTMS treatment response in schizophrenia using interpretable machine learning: a SHAP-based analysis

Individual responses to repetitive transcranial magnetic stimulation (rTMS) in schizophrenia vary, and predictive clinical tools are lacking.

To develop and interpret machine learning models predicting individual rTMS treatment response using baseline clinical features.

Exploratory, hypothesis-generating study using retrospective patient data with internal validation and interpretability analysis.

We analyzed 156 patients with schizophrenia, assessing baseline Positive and Negative Syndrome Scale (PANSS) and global assessment of functioning (GAF) scores. Machine learning models (Random Forest, XGBoost, support vector machine, logistic regression) were trained on demographic and clinical features. Performance was evaluated via cross-validation and a temporal hold-out test set. Shapley additive explanations (SHAP) were used for model interpretation.

Baseline characteristics were comparable between groups (all p > 0.1). Although both groups improved clinically, between-group differences were not statistically significant. The Random Forest model achieved the highest performance: cross-validated area under the receiver operating characteristic curve (AUC) = 0.84, temporal hold-out AUC = 0.70. Learning curve analysis indicated performance plateaued around 100 cases. SHAP and decision tree analysis highlighted moderate baseline GAF and higher PANSS as key predictors for response.

Despite modest group-level efficacy, interpretable machine learning models identified baseline features associated with individual response to rTMS. These findings can inform personalized interventions, though future external validation is needed.

Not applicable.