Prediction of Suicidal Ideation, Suicide Plans, and Suicide Attempts Among College Students: Developing a Theory-Driven Machine Learning Model

Abstract: Suicide among college students has emerged as an increasingly urgent public health issue. Traditional statistical methods can face challenges when handling numerous predictors and complex interactions among them, whereas existing machine learning studies frequently lack theoretical grounding and often focus on a single suicide-related outcome. To address these gaps, this study adopted the Narrative-Crisis Model of Suicide as a theoretical framework and developed separate predictive models for suicidal ideation, suicide plans, and suicide attempts, with the aim of comparing the predictive contributions of features across these three suicide-related outcomes. Following feature importance analysis, simplified models were constructed and further evaluated in a same-school temporal validation sample.
This study drew on two large-scale survey datasets collected from Chinese university students. In Dataset 1 (n = 5,296), baseline models were developed using 74 features primarily derived from the Narrative-Crisis Model of Suicide, supplemented by variables informed by related theoretical frameworks. Random Forest and Extreme Gradient Boosting were used for model training, with a linear support vector machine included as a supplementary comparator. Random Forest was used for the primary feature contribution analyses. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data. Model performance was evaluated across 10 repeated runs using AUC, AUPRC, accuracy, precision, recall, specificity, and F1 score. Based on SHapley Additive exPlanations (SHAP) values and feature importance rankings across the three outcome models, 16 variables were retained to construct simplified models, which were evaluated on both the internal hold-out test set and a same-school temporal validation sample collected at a later time point. All three suicide-related outcomes were assessed over the six months preceding each survey.
The results showed distinct predictive patterns across the three suicide-related outcomes. Tobacco and alcohol use showed relatively greater predictive contributions in the suicide plan and suicide attempt models than in the suicidal ideation model, whereas interpersonal trauma, including verbal and relational bullying, emotional abuse, and domestic violence, showed relatively greater contributions in the suicidal ideation model. Variables related to suicide crisis syndrome, suicidal narratives, and self-harm behavior also ranked highly across the models. Despite a substantial reduction in the number of predictors, the simplified models maintained good discrimination in both internal testing and same-school temporal validation. In the temporal validation sample, suicide plans and attempts were reported by only 25 and 23 participants, respectively, warranting caution when interpreting their high AUC estimates. Supplementary analyses further indicated that calibration remained imperfect and that decision-analytic performance varied across risk thresholds.
Overall, the findings suggest that combining theory-informed variable selection with data-driven machine learning may help identify outcome-specific predictive patterns across suicidal ideation, suicide plans, and suicide attempts. The simplified feature set may provide a useful basis for further validation and refinement of suicide risk assessment models in college populations.