Multivariate machine learning regression approaches to predict adolescent suicide risk.

BACKGROUND: Suicide is a leading cause of death among adolescents, yet most prediction models rely on binary classification and limited predictors, failing to capture the complexity of suicide risk. This study aimed to develop a more accurate and comprehensive prediction model using machine learning regression with a broad set of psychological and behavioral variables. METHOD: A total of 2241 adolescents were recruited from school settings and completed self-report questionnaires assessing 31 variables. Three machine learning regression algorithms: Lasso regression, support vector regression, and random forest regression were applied. Model performance was evaluated using the coefficient of determination (R²). To obtain robust estimates of predictive performance, all three models were evaluated using a 10-fold cross-validation procedure. RESULTS: (1) 13 features were selected by the Lasso regression from the original set of 31 suicide risk-related variables; (2) Among the three models tested, the random forest regression model demonstrated the best predictive performance, with an R² of 0.61 [95 % CI: 0.60, 0.62]; (3) Further analysis revealed that the five most important predictors of suicide risk were depression, emotion regulation, perceived burdensomeness, non-suicidal self-injury, and family function. CONCLUSION: A multivariate, integrative approach significantly improves the accuracy and precision of suicide risk prediction among school-based community sample. Beyond depression and emotion-related factors, perceived burdensomeness, non-suicidal self-injury, and family function also played important roles in predicting suicide risk, suggesting that these variables should be considered in future screening frameworks to improve early detection and targeted intervention.