Answers are grounded in this paper and cite their sources.
AIMS: Suicide is a leading cause of preventable death among U.S. veterans. Machine learning (ML) prediction is promising, yet most ML models for suicide rely on predominantly non-Hispanic White samples and rarely assess racial/ethnic status. METHODS: We analysed baseline data from the Behavioural Research and Analysis of Veterans' Outcomes (BRAVO) cohort, a national sample of post-9/11 veterans. A total of 2,083 veterans (37% female; mean age = 38.5 years) completed a survey assessing demographics, military history, behavioural health, discrimination, neighbourhood context and treatment access; 40.7% met the cut-score (≥7 on the Suicide Behaviours Questionnaire-Revised [SBQ-R]) for elevated suicide-related thoughts/behaviours. Using 90 variables assessed from self-report and census tract linking, we fit random-forest classifiers with multiple imputation and race-stratified models. Permutation importance and Shapley Additive Explanations (SHAP) identified influential predictors and their directional contributions; performance was summarised with area-under-the-curve (AUC). RESULTS: Overall model performance was adequate (AUC = 0.77) and similar across racial/ethnic groups (AUCs = 0.72-0.77). Across the full sample and race/ethnicity-stratified models, suicide risk was dominated by shared transdiagnostic and non-transdiagnostic phenotypes. The rank order of variable importance was higher neuroticism, emotion-regulation difficulties, irritability/anger, trauma-related shame and moral injury, internalizing symptoms, adverse childhood experiences and difficulty transitioning out of the military. Group-specific predictors (e.g. neighbourhood disadvantage, substance use patterns and minority-stress exposures) added meaningful signal. CONCLUSION: Interpretable ML in an intentionally diverse veteran cohort identifies a robust, shared suicide-risk profile with contextually distinct amplifiers across race/ethnicity.
No takes yet. Share an insight, caveat, or question.
Davis et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: