Machine learning-based prediction of suicidal ideation, plans, and attempts among school-going adolescents

Suicidal ideation, plans, and attempts represent significant mental health challenges among adolescents worldwide. Early identification of adolescents at risk is critical for preventing suicidal behaviors. Machine learning provides an advanced, data-driven approach to forecasting these risks and enabling prompt intervention accurately. This study utilized data from the Global School-based Student Health Survey (2019–2023), analyzing 16,704 records for each suicidal behavior (ideation, plan, attempt). SMOTE was applied for data balancing, and predictions were performed using both the balanced and imbalanced datasets. A variety of classifiers were employed, including GNB, SVC, DT, LR, RF, KNN, XGBoost, Bagging, AdaBoost, LightGBM, and CatBoost. The models were evaluated based on accuracy, recall, precision, F1 score, and AUC. Additionally, SHAP values were used to measure the importance of each feature in the predictions. All analyses were conducted using Python. The Random Forest classifier demonstrated strong predictive ability on the balanced datasets, achieving AUC scores of 84.25 % (attempts), 82.91 % (plans), and 80.74 % (ideation). The model achieved accuracies of 74.93 %, 73.47 %, and 76.38 %, with recall values of 74.56 %, 71.11 %, and 77.92 % for suicide plans, ideation, and attempts. Bagging and Decision Tree performed well, while XGBoost, CatBoost, and LightGBM showed strong results. In contrast, performance on imbalanced data was limited by low recall. Gender, loneliness, and parental attachment were key predictors of suicidal behavior. The Random Forest classifier displayed powerful predictive capabilities, effectively identifying suicide plans, ideation, and attempts. This model can be used to identify adolescents at risk for suicide, enabling early intervention through targeted mental health support, counseling, and timely preventive measures. Despite strong performance, the limitation of relying on self-reported data remains. Future research should focus on using longitudinal data to track the development of suicidal behavior over time. • Random Forest: Best AUC (Plans: 82.91 %, Ideation: 80.74 %, Attempts: 84.25 %). • Bagging AUC: Plans (82.20 %), Ideation (80.17 %), Attempts (83.62 %). • Decision Tree AUC: Plans (81.02 %), Ideation (78.94 %), Attempts (82.70 %). • XGBoost, CatBoost, and LightGBM also achieved impressive results. • Gender, loneliness, and parental attachment were key predictors of suicidal behavior.