Applying Machine Learning to Investigate Future Life-Oriented Predictors of Adolescents’ Near-Term Suicidal Ideation

Background Adolescents contemplating suicide consider both the discontinuation and continuation of their lives. Yet prior research has mainly focused on what suicidal youth want to escape from (i.e., distress-oriented predictors) rather than what may keep them living into the future. This study applies machine learning (ML) to examine how adolescents’ momentary Future Life-Oriented (FLO) processes predict their near-term changes in suicidal ideation (SI). Methods Adolescents (N = 121; 15-17 years) completed a 14-day, 5x/day smartphone-based ecological momentary assessment protocol, reporting on a range of FLO predictors (e.g., vividness, importance, emotional valence of future life events), distress-oriented predictors (e.g., prior SI, stressors, affect), and SI. ML algorithms were trained on time-series data to predict future SI based on FLO and distress-oriented predictors. To identify optimal temporal prediction windows, we implemented a sliding-window approach and compared models across multiple temporal frames. Results Models including FLO variables significantly improved prediction of subsequent SI (AUC = .880) relative to models using prior SI (AUC = .783) alone or prior SI with stress/affect (AUC = .781). The best-performing model (Random Forest; AUC = .956) achieved optimal prediction using seven prior observations to forecast SI two timepoints ahead (avg. 11 hours). SHAP analyses identified prior SI and two FLO variables (desire to live, future thought uncontrollability) as the most influential predictors of SI. Conclusions Adolescents’ momentary FLO processes meaningfully enhance near-term prediction of SI, highlighting a potentially promising collection of modifiable targets for suicide prevention.