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

Area of Science:

• Neuroscience
• Psychiatry
• Artificial Intelligence

Background:

• Individual responses to repetitive transcranial magnetic stimulation (rTMS) in schizophrenia are highly variable.
• There is a lack of predictive clinical tools to guide rTMS treatment decisions.

Purpose of the Study:

• To develop and interpret machine learning models for predicting individual rTMS treatment response in schizophrenia patients.
• To identify baseline clinical features that predict response to rTMS therapy.

Main Methods:

• Retrospective analysis of 156 schizophrenia patients' data, including Positive and Negative Syndrome Scale (PANSS) and Global Assessment of Functioning (GAF) scores.
• Training and evaluating multiple machine learning models (Random Forest, XGBoost, SVM, logistic regression) using demographic and clinical features.
• Utilizing cross-validation, a temporal hold-out set, and Shapley Additive Explanations (SHAP) for model interpretation.

Main Results:

• The Random Forest model demonstrated the highest predictive performance with a cross-validated AUC of 0.84 and a temporal hold-out AUC of 0.70.
• Moderate baseline GAF scores and higher PANSS scores were identified as significant predictors of rTMS response.
• Model performance plateaued around 100 cases, suggesting sufficient data for reliable predictions.

Conclusions:

• Interpretable machine learning models can identify baseline features associated with individual rTMS response in schizophrenia.
• These findings support the potential for personalized interventions to optimize rTMS therapy.
• External validation is necessary to confirm the generalizability of these predictive models.