Objective: Using two-year follow-up data from the Shandong Adolescent Behavior and Health Cohort (SABHC), this study evaluated four algorithms in predicting future suicidal behavior among a school-based sample of adolescents.
Methods: Participants were 7072 adolescents who were surveyed at baseline and were followed up one year later in the SABHC. Demographic information, mental and physical health assessments collected at baseline were utilized to predict suicidal behavior at follow-up. After excluding observations with missing features, 3,169 participants remained for analysis. Logistic regression, XGBoost, decision tree, and random forest were employed predicting suicidal behavior over a one-year period. Synthetic minority over-sampling technique and 10-fold cross-validation were applied to address class imbalance and prevent over-fitting. The area under the receiver operating characteristic curve (AUC) served as the primary metric for assessing predictive performance.
Results & Discussion: The AUCs of the four predictive models, logistic regression, XGBoost, decision tree, and random forest, were 0.682, 0.710, 0.650, and 0.740, respectively. All algorithms identified history of suicidal behavior as the most important predictor of subsequent suicidal behavior. Other common predictors across all models were female gender, life events, history of non-suicidal self-injury, and hopelessness.
Conclusions: In the study, random forest outperformed the other algorithms in predicting one-year suicidal behaviour. The addition of more detailed data from various domains, along with elements from suicide theoretical models, might enhance the predictive power of machine learning algorithms.

suicidal behavior;adolescents;Machine Learning;longitudinal modeling