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

Comprehensive Summary

This study by Lin investigates whether interpretable machine-learning approaches can predict individual response to repetitive transcranial magnetic stimulation (rTMS) in patients with schizophrenia using baseline clinical characteristics. The authors conducted a retrospective analysis of 156 patients and trained several machine-learning models (including Random Forest, XGBoost, support vector machine, and logistic regression) using demographic variables, illness duration, and clinical severity measures such as PANSS and GAF scores, with nested cross-validation and temporal hold-out testing to evaluate performance. Among the models, the Random Forest classifier achieved the highest predictive accuracy, with an AUC of approximately 0.84 in cross-validation and 0.70 in temporal validation, indicating moderate discriminative capability. SHAP-based interpretability analyses identified baseline functional status, symptom severity, and illness duration as the most influential predictors, and further demonstrated a nonlinear relationship in which patients with moderate functional impairment showed the highest probability of response. In the discussion, the authors emphasize that despite the absence of statistically significant group-level treatment effects, predictive modeling may still hold clinical utility by identifying responsive subgroups, though they stress that the findings remain exploratory and require external validation.

Outcomes and Implications

This research is important because variability in rTMS response currently limits its effective deployment in schizophrenia treatment, and there are no established tools to guide patient selection. By demonstrating that routinely available clinical data can predict response with moderate accuracy, the findings support the feasibility of applying precision-medicine frameworks within psychiatric neuromodulation. Clinically, such predictive models could help prioritize patients with intermediate functional impairment and higher symptom burden, thereby optimizing resource allocation and minimizing exposure to ineffective interventions. However, the authors note that immediate clinical translation is constrained by the study’s retrospective design, modest sample size, and lack of external validation, as well as the need for prospective trials and integration into decision-support systems. Consequently, while the work provides a proof-of-concept for personalized rTMS stratification, widespread clinical implementation will likely require further multimodal validation and development over the coming years.