Predicting adolescent suicide risk using integrated data from adolescents, parents and siblings: An analysis of...

Adolescent suicide is a critical public health issue, yet accurately predicting suicide risk remains challenging. Few studies integrate adolescents' self-reports with mental health, especially suicidality assessments from parents and siblings. This study employed machine learning (ML) models - Random Forest, XGBoost and LightGBM - to analyze 169 demographic, socio-relational, mental health, and functional variables from 7286 adolescents and one of their parents and siblings (N = 21858) to identify key predictors of suicide risk. There were 14.8 % of the adolescents at risk. Family-triad-models, which used all data from target adolescent, their sibling and parent, achieved consistent patterns in variable importance for predicting suicidality, with AUROC scores ranging from 0.875 to 0.877. All nine top predictors identified by all three ML methods originated from adolescents' self-reports, with emotional difficulties as the most important predictor, adjusted odds ratio (aOR) per ±1SD being 1.59 [95 % CI 1.46-1.73]. In family-proxy-models, which excluded the 56 self-reported variables from the target adolescent, predictive accuracy declined but remained sufficient (AUROC: 0.621 to 0.625). Four variables were identified by all three ML methods, with parents' reports of the target adolescent's emotional difficulties being the strongest, aOR ± 1SD being 1.47 [95 % CI 1.39-1.56]. Machine learning can effectively leverage adolescents' self-reports to predict suicide risk accurately. Even when adolescents' self-reports are unavailable due to their unwillingness to disclose information, family members' reports alone provide a sufficiently accurate basis for prediction. Moreover, emotional difficulties perceived by both adolescents and parents are crucial indicators of suicidality.