Prediction Models for Suicidal Behavior Among Adolescent Students Based on Longitudinal Study Data: A Systematic Review and Meta-Analysis

Background: Suicidal behavior is common among adolescents, yet most affected individuals do not receive medical care. Predicting suicidal behavior in school settings may support early prevention. Methods: We systematically reviewed longitudinal studies of machine-learning models predicting future suicidal behavior in adolescent students. Eight databases were searched through September 2026; risk of bias and applicability were assessed using PROBAST + AI. A bivariate random-effects model jointly pooled sensitivity and specificity and generated a summary receiver operating characteristic curve; study-reported AUCs were pooled separately as a supplementary analysis. Results: Seven models were included. Under the bivariate random-effects model, pooled sensitivity was 0.70 (95% CI: 0.52–0.83) and pooled specificity was 0.80 (95% CI: 0.71–0.87); the summary AUC was 0.83. The supplementary univariate analysis yielded a pooled AUC of 0.79 (95% CI: 0.75–0.83; 95% prediction interval: 0.63–0.89). Heterogeneity was substantial (I2 = 95.5% for sensitivity and 99.1% for specificity). Exploratory subgroup and meta-regression analyses identified no statistically significant moderators, but the small number of models limited statistical power. All models were evaluated using internal validation and were judged to be at high overall risk of bias. Conclusions: Current models show moderate discrimination, but their sensitivity remains insufficient for use as standalone early-warning tools. Because the synthesis was based on the final models selected by the original study authors, the pooled performance estimates may be optimistic.