Predictive Performance of Machine Learning for Suicide in Adolescents: Systematic Review and Meta-Analysis

BACKGROUND: In the context of escalating global mental health challenges, adolescent suicide has become a critical public health concern. In current clinical practices, considerable challenges are encountered in the early identification of suicide risk, as traditional assessment tools demonstrate limited predictive accuracy. Recent advancements in machine learning (ML) present promising solutions for risk prediction. However, comprehensive evaluations of their efficacy in adolescent populations remain insufficient. OBJECTIVE: This study systematically assessed the performance of ML-based prediction models across various suicide-related behaviors in adolescents, aiming to establish an evidence-based foundation for the development of clinically applicable risk assessment tools. METHODS: This review assessed ML for predicting adolescent suicide-related behaviors. PubMed, Embase, Cochrane, and Web of Science databases were rigorously searched until April 20, 2024, and a multivariate prediction model was employed to assess the risk of bias. The c-index was used as the primary outcome measure to conduct a meta-analysis on nonsuicidal self-injury (NSSI), suicidal ideation, suicide attempts, suicide attempts combined with suicidal ideation, and suicide attempts combined with NSSI, evaluating their accuracy in the validation set. RESULTS: A total of 42 studies published from 2018 to 2024 were included, encompassing 104 distinct ML models and 1,408,375 adolescents aged 11 to 20 years. The combined area under the receiver operating characteristic curve values for ML models in predicting NSSI, suicidal ideation, suicide attempts, suicide attempts combined with suicidal ideation, and suicide attempts combined with NSSI were 0.79 (95% CI 0.72-0.86), 0.77 (95% CI 0.71-0.83), 0.84 (95% CI 0.83-0.86), 0.82 (95% CI 0.79-0.84), and 0.75 (95% CI 0.73-0.76), respectively. The ML models demonstrated the highest combined sensitivity for suicide attempt prediction, with a value of 0.80 (95% CI 0.75-0.84), and the highest combined specificity for NSSI prediction, with a value of 0.96 (95% CI 0.94-0.99). CONCLUSIONS: Our findings suggest that ML techniques exhibit promising predictive performance for forecasting suicide risk in adolescents, particularly in predicting suicide attempts. Notably, ensemble methods, such as random forest and extreme gradient boosting, showed superior performance across multiple outcome types. However, this study has several limitations, including the predominance of internal validation methods employed in the included literature, with few studies employing external validation, which may limit the generalizability of the results. Future research should incorporate larger and more diverse datasets and conduct external validation to improve the prediction capability of these models, ultimately contributing to the development of ML-based adolescent suicide risk prediction tools.