Predicting suicide attempts among US Army soldiers using information available at the time of periodic health assessments

Abstract

The value of population screening for suicide risk remains unclear. The US Army’s annual medical examination, the Periodic Health Assessment (PHA), screens for suicidality and other mental and physical health problems. Here in our 2014–2019 cohort study we used PHA and Army administrative data (n = 1,042,796 PHAs from 452,473 soldiers) to develop a model to predict 6-month nonfatal and fatal suicide attempts (SAs). The model was designed to establish eligibility for a planned high-risk SA prevention intervention. The PHA suicide risk screening questions had limited value, as 95% of SAs occurred among soldiers who denied suicidality. However, a simple least absolute shrinkage and selection operator (LASSO) penalized regression model that included a wide range of administrative predictors had good test sample discrimination (0.794 (standard error 0.009) area under the receiver operating characteristic curve) and calibration (integrated calibration index 0.0001). The 25% of soldiers at highest predicted risk accounted for 69.5% of 6-month SAs, supporting use of the model to target preventive interventions. A machine learning model incorporating a wide range of administrative medical and demographic data from the US Army outperformed suicide risk screening questions in predicting suicide attempts over the 6 month period following soldiers’ annual medical examinations.