Objective:

To develop and evaluate machine learning models for predicting adolescent suicide attempts and to examine predictor contributions using explainable artificial intelligence.

Approach:

• Study Design: A repeated cross-sectional study using pooled data from the 2017–2024 Korea Youth Risk Behavior Web-Based Survey (n=448,798).
• Model Development: Models including logistic regression, random forest, and XGBoost were developed to classify suicide attempts.
• Performance Evaluation: Performance was evaluated using AUC, F1 score, and sensitivity-oriented screening thresholds.
• Predictor Contribution Analysis: SHapley Additive exPlanations (SHAP) quantified predictor contributions.

Key Findings:

• XGBoost achieved the highest F1 score.
• Random forest showed the highest sensitivity under screening conditions.
• Hopelessness was the strongest predictor, followed by school violence and perceived stress.
• Additional contributors included self-rated health, household economic status, sleep satisfaction, and behavioral indicators.
• Sensitivity improved under screening conditions, but positive predictive values remained low.

Interpretation:

Machine learning models demonstrated moderate performance in predicting adolescent suicide attempts, with psychological and social variables being the most significant predictors.

Limitations:

• The study may not capture all relevant predictors due to the complexity of adolescent suicide risk.
• Positive predictive values were low, indicating potential limitations in practical application.

Conclusion:

Machine learning models showed moderate predictive performance for adolescent suicide attempts, with SHAP enhancing interpretability.