Background: Suicide among adolescents is a major global health issue, and it remains challenging to identify those who are at risk for future suicidal behaviors. This study aimed to investigate the potential of machine learning (ML) based approaches to predict suicide attempts (SA) among at-risk adolescents.
Methods: We applied three ML-algorithms – elastic net (EN), random forest (RF) and extreme gradient boosting to longitudinal clinical data, and compared their performance to a traditional statistical approach – logistic regression (LR). Models were trained using patient data from a clinical cohort specialized on adolescents prone to self-harming and risk-taking behaviors (n=255). Forty-four predictor variables obtained at baseline were selected to predict future SAs within a two year follow-up period. Performance metrics included AUC, Brier score, sensitivity, specificity, PPV, and NPV. Moreover, individual predictor importance was explored.
Results: ML-algorithms outperformed LR in terms of overall predictive accuracy (AUC = 0.75 – 0.79 vs. 0.72) and model calibration (Brier scores = 0.18 – 0.20 vs. 0.24). A prior SA stood out as the most important predictor variable across all algorithms.
Discussion: Our findings demonstrate that SAs can be predicted with good overall accuracy, even among patients with high prevalences of suicidal behaviors. The clinical utility of ML-based SA-predictions is subject to several limitations. Alongside improvements of predictive performance, future research needs to address clinical and ethical implications of ML-based risk detection.