BACKGROUND: Adolescent suicide remains a significant public health concern, yet existing suicide screening instruments primarily focus on already manifested suicidal phenomena, underscoring the need for reliable and practical tools to enable early identification and intervention. METHODS: Based on a large-scale school-based cohort study conducted in Southern China in 2022, this study aimed to develop and preliminarily validate two machine learning-based tools (a 51-item full version and an 11-item abbreviated version) designed to help identify adolescents at risk of developing suicide risk. The dataset was divided into two samples for tool development and longitudinal interview validation. During the tool development phase, LASSO regression was employed to select items with optimal contributions for recent suicide attempts from a multidimensional set of risk factors, followed by model training with multiple machine learning algorithms. The developed models were subsequently evaluated for their ability to predict suicide risk as assessed by the follow-up interview in the longitudinal validation phase. RESULTS: Both versions of the screening tool demonstrated adequate discriminative ability, with the CatBoost algorithm outperforming others (AUROC ≥ 0.87). The abbreviated tool showed a slight trade-off between model precision and practicality, with a 0.02 reduction in AUROC, while still maintaining appropriate discrimination. Longitudinal validation using follow-up interview outcomes supported the predictive validity of both tools. These findings provide preliminary evidence for the utility of machine learning-based suicide risk screening tools among adolescents. CONCLUSIONS: This study provides evidence supporting the machine learning-based screening tools for early suicide risk detection in adolescents that integrates multidimensional vulnerabilities. The tools show promise in facilitating early identification and targeted interventions in school settings, addressing a critical need in adolescent mental health care. Nonetheless, further research is warranted to confirm their efficacy and support broader implementation.
Li et al. (Sun,) studied this question.