Predicting US Army Soldier Suicide-Related Behaviors at the Time of Periodic Health Assessments
Emily R. Edwards, Shelby Borowski, Matthew K. Nock, David M. Benedek, Kate H. Bentley, Joseph Geraci, Sarah M. Gildea, Marianne Goodman, Amit Gupta, Chris J. Kennedy, Andrew J. King, Evan M. Kleiman, Howard Liu, Alex Luedtke, Sarah Maggio, James A. Naifeh, Thomas H. Nassif, Chris Paine, Nancy A. Sampson, Nur Hani Zainal, James Wagner, Murray B. Stein, Vincent F. Capaldi, Robert J. Ursano, Ronald C. KesslerImportance
Suicide is the leading cause of death among active-duty US Army soldiers. Evidence-based preventive interventions exist but need to be targeted to be cost-effective.
Objectives
To develop machine learning models using administrative data available during periodic health assessments (PHAs) to determine eligibility for a remote group dialectical behavior therapy-based skills training intervention for soldiers with elevated risk of suicide or nonfatal suicide attempt.
Design, Setting, and Participants
From regular US Army soldier PHAs completed from 2015 to 2019, separate 70% training samples were created for suicides and SAs. Each training sample contained all cases and a stratified equal-probability sample of 10 times as many controls, with inverse-probability of selection weights applied to controls. Model performance was evaluated in the remaining 30% test sample. Data analysis was conducted September 2025 through January 2026.
Main Outcomes and Measures
Suicides were recorded in the National Death Index for soldiers either still in or out of active service. SAs were recorded in US Army records only for soldiers still in active service. A 24-month risk horizon was used because successive PHAs, though designed to occur annually, are sometimes separated by this much time. A loss to follow-up weight was used in the SA model to adjust for leaving service within 24 months.
Results
PHA data were available from 668 684 regular US Army soldiers. A total of 1 932 243 PHAs (including 85.5% by male soldiers, 46.7% by soldiers aged 28 years or older, and 58.1% by married soldiers) completed during the 2015 to 2019 period were included in analyses. Over 24 months, suicide prevalence was 66.7 per 100 000 (SE = 3.6) and SA prevalence was 704.1 per 100 000 (SE = 14.7) adjusted for loss to follow-up. Test sample area under the receiver operating characteristic curve (AUROC) for suicide was 0.72 (SE = 0.17), with Integrated Calibration Index (ICI) of 0.0003 and Brier score of 0.0007. Test sample AUROC for SA was 0.81 (SE = 0.02), with ICI of 0.0014 and Brier score of 0.0070. Suicide prevalence was meaningfully elevated (at least twice that expected by chance) only in the 5% of PHAs with highest predicted risk (sensitivity = 18.6%; SE = 2.0%). SA risk was elevated only in the 10% of PHAs with highest predicted risk (sensitivity = 46.5%; SE = 1.0%). Predicted probabilities of suicide and SA were correlated (Pearson
Conclusions and Relevance
Results of this prognostic study suggest that as suicides and SAs have distinct predictors, an attempt to optimize resource allocation will require thoughtful postintervention considerations of costs and benefits of interventions to optimize net benefit.