Machine learning models were built using administrative data from 1.93 million US Army periodic health assessments (PHAs, 2015–2019) across 668,684 soldiers to predict 24-month risk of suicide death and nonfatal suicide attempt (SA), aiming to target a remote group DBT-based skills training intervention.
Suicide model AUROC was 0.72; SA model AUROC was 0.81. Suicide prevalence was meaningfully elevated only in the top 5% of PHAs by predicted risk (sensitivity 18.6%); SA risk was elevated only in the top 10% (sensitivity 46.5%). Overall, 11.9% of PHAs flagged soldiers at elevated risk of either outcome, and 3.1% were elevated for both.
- Suicide and SA predictors were distinct, limiting a single unified screening approach. - SA data only captured soldiers still in active service, requiring loss-to-follow-up weighting that may not fully correct for bias. - Suicide model AUROC had a very wide SE (0.17), reflecting rarity of the outcome and uncertainty in performance estimates.
Routine PHAs can be used with ML models to flag the roughly 12% of soldiers at elevated suicide-related risk for targeted DBT-skills interventions — but suicide and SA require separate risk scores, and the low sensitivity for suicide (18.6%) means most future suicides will not be flagged.
Explore related topics
