USU Study Leverages Machine Learning to Improve Suicide Risk Detection in Soldiers
New research from Uniformed Services University indicates machine learning models significantly enhance the ability to predict suicide attempts among U.S. Army soldiers, a crucial step for military suicide prevention.
 |
| Graphic credit: Sofia Echelmeyer, USU |
May 28, 2025 by Vivian Mason
Suicidal behavior remains a significant concern within the U.S. Army, presenting a persistent challenge for military mental health professionals in accurately identifying soldiers at risk. A recent Uniformed Services University (USU) study, however, offers a promising new approach, utilizing machine learning to better predict suicide attempts among U.S. Army soldiers using data available from their Periodic Health Assessment (PHA). This vital research was conducted and published in the "Nature Mental Health" journal by Dr. James Naifeh, assistant scientific director at the Center for the Study of Traumatic Stress (CSTS) and a research associate professor in the Department of Psychiatry at USU’s School of Medicine, along with colleagues from USU and Harvard Medical School.
 |
Dr. James Naifeh, assistant scientific director at CSTS. (Courtesy photo) |
"What we found was that very few soldiers reported suicidal thoughts on their PHA," Naifeh reveals. "In fact, 99.8% of soldiers denied having suicidal thoughts when they completed the PHA." He further notes, "Those PHA screening questions weren't very good at predicting future suicidal behavior. Of the soldiers who in the next six months made a suicide attempt, 95% of them said that they were not having suicidal thoughts on the PHA."
To address these shortcomings in suicide risk detection, the study found that risk prediction was substantially improved by employing machine learning models, such as LASSO (Least Absolute Shrinkage and Selection Operator) regression. These advanced models incorporated administrative medical and personnel data available at the time of the PHA, including health care history, sociodemographic information, and service-related variables for the soldiers. This new model for predicting suicide attempts, developed by Naifeh and his colleagues, analyzed risk over the six months following a soldier's PHA and was initially designed to determine eligibility for a planned high-risk suicide attempt prevention intervention. Interestingly, the researchers discovered that adding more extensive administrative data beyond what was available at the PHA, such as Armed Forces Qualification Test scores or social determinants of health based on pre-service ZIP codes, did not significantly improve the machine learning prediction accuracy.
"Findings suggest that relying on soldiers’ self-report to predict their future risk of suicidal behavior does not seem to be a very fruitful way of approaching risk prediction,” Naifeh explains. “By bringing in other information available at the time of the PHA, we can do a better job predicting risk that doesn't rely on self-report screening responses."
This important military health research utilized data from the Army STARRS (Study to Assess Risk and Resilience in Servicemembers), the largest epidemiological and neurobiological study of suicidal behavior ever conducted in the U.S. Army. Naifeh has been involved with Army STARRS, a cornerstone of military suicide research, since shortly after its inception in 2009; the study has since transitioned to STARRS-LS, a longitudinal study that began in 2015.
The STARRS team, comprising investigators from USU, the University of California San Diego, Harvard Medical School, and the University of Michigan, integrated over 50 Army and Department of Defense administrative data systems with the records of all soldiers who served on active duty from 2004 to 2019. This comprehensive database has been instrumental for more than a decade in improving the prediction of suicidal behavior among soldiers using advanced statistical methods. Building upon this extensive dataset, Naifeh and his colleagues assessed the utility of the PHA’s suicide risk screening questions and developed a series of machine learning models, progressively adding more information to enhance the prediction of suicide attempts in the six months following the PHA.
The machine learning models for suicide risk assessment developed as part of this USU research assign each individual a risk score, representing their predicted probability of future suicidal behavior. "Based on our preferred model," Naifeh explains, "the 25% of soldiers with the highest predicted risk account for about 70% of the suicide attempts that occurred in the six months following the PHA." This phenomenon is known as a 'concentration of risk'. It is crucial for military suicide prevention because it allows for more intensive interventions to be targeted toward a smaller, high-risk group of soldiers who account for a large proportion of suicide attempts.
 |
Army STARRS study findings will help Army leaders address important mental and behavioral health issues in the military. (Photo courtesy of Army STARRS) |
The urgency for such targeted interventions for soldier mental health is highlighted by historical trends. The suicide rate in the military, particularly the U.S. Army, began to rise in 2004. While historically the military suicide rate was below that of the civilian population, by 2008, it surpassed the civilian rate when adjusted for sex and age. This concerning trend in military suicides was the impetus for the Army STARRS study. Suicide rates have unfortunately remained elevated in the years since. "I think it’s an ongoing concern and challenge to try and prevent suicidal behavior in service members in general and in the Army in particular, which is where our work has been primarily focused," says Naifeh.
Although suicide attempts are more common than suicide deaths, both are relatively rare events, which complicates identification efforts for at-risk soldiers. For instance, if the suicide death rate is 30 per 100,000 soldiers, this means 30 out of 100,000 soldiers will die by suicide. Identifying and assisting such a small group effectively presents a significant challenge for military suicide prevention programs. It is in this context that machine learning techniques have been increasingly used since 2015 to try and predict suicide among military personnel.
The work of the Army STARRS team, including Naifeh’s recent article, has already informed a new military suicide prevention initiative called SAFEGUARD (Suicide Avoidance Focused Enhanced Group Using Algorithm Risk Detection). This initiative uses machine learning to identify U.S. Army soldiers at the highest risk of suicidal behavior in different situations, such as after discharge from psychiatric hospitalization or upon arrival at their first duty station, with the goal of targeting scalable interventions to prevent suicidal behavior.
Recognizing that the period following discharge from psychiatric hospitalization is identified as a particularly high-risk time for suicide, Naifeh remarks, "In SAFEGUARD, we are trying to identify those with the highest risk of suicide over the next year. Then, we can do a remote intensive case management intervention to help soldiers as they transition out of the hospital and back into service. By providing that additional support and helping soldiers identify resources that can help them, we hope to reduce the rate of suicidal behavior in that group." This approach aims to improve soldier well-being and contribute to a healthier force.
