Interpretable machine learning predicts risk of aggressive behavior in schizophrenia

Objective: To develop an online calculator for predicting aggressive behavior in schizophrenic patients based on shapley additive explanations (SHAP). Methods: A total of 238 schizophrenic patients were selected for the study.The important characteristic variables of aggressive behavior were screened by Boruta algorithm.The 238 patients were randomly divided into a training set (n=145) and a test set (n=93) in a 3∶ 2 ratio to train nine machine learning (ML) models and perform tenfold cross-validation.ML models were evaluated using receiver operating characteristic (ROC) curves to screen the best predictive performance model, and decision curve analysis was further used to assess the clinical benefit of ML models.An online calculator for predicting aggressive behavior in schizophrenic patients was constructed using SHAP add-on to interpret and visualize the ML models and via R package. Results: A total of 238 patients with schizophrenia were included and 76 patients (31.9%) developed aggressive behavior.Boruta algorithm screened for duration of disease, social support rating scale (SSRS), brief psychiatric rating scale (BPRS), high-density lipoprotein (HDL), neutrophil/lymphocyte ratio (NLR), and history of violence as important characteristic variables of aggressive behavior.Among the nine ML algorithms, the ROC of the training and test sets confirmed that the extreme gradient boost (XGBoost) model had the highest performance in predicting the risk of aggressive behavior.The results of the decision curve analysis indicated that the model had high predictive accuracy and clinical application value.The SHAP visualization showed that the rank order of risk for contributing to aggressive behavior in order is disease duration, SSRS, BPRS, HDL, NLR, history of violence.SHAP summary plots showing the global impact and distribution of each feature variable on model predictions.Online calculator based XGBoost effectively predicted risk of aggressive behavior in schizophrenic patients. Conclusions: XGBoost model online calculator for constructing SHAP additional explanations based on disease course, SSRS, BPRS, HDL, NLR, and history of violence accurately predicts patients′ risk of aggressive behavior and improves early prevention and intervention of aggressive behavior risk.