USU Study Uses Machine Learning Tools to Identify Those at Risk for Homelessness

Caption for Homeless Veteran.jpg: A new Army STARRS-LS study led by USU used  machine-learning tools to calculate risk of homelessness among veterans. ( photo)

By Vivian Mason

How does an Army veteran become homeless? 

Joseph T. fell through the cracks after returning home with a service-connected disability. He found himself penniless, homeless, relying on the kindness of friends with available sofas for a night’s sleep, eventually living in and out of shelters. 

Joseph’s story is not unique for veterans, with several thousands experiencing homelessness. But now, a new study aims to address that problem by identifying those who may be at risk, which will ultimately help offer targeted, personalized interventions to service members. 

The study, “Predicting Homelessness among U.S. Army Soldiers no longer on Active Duty,” is a collaboration between multiple universities and agencies led by researchers at USU’s Center for the Study of Traumatic Stress (CSTS).  It was recently published in the American Journal of Preventive Medicine. 

The study utilized surveys completed by 16,589 soldiers between 2011-2014 as part of Army STARRS-LS, also known as the Study to Assess Risk and Resilience in Servicemembers– Longitudinal Study, in conjunction with a novel machine learning approach to help address the estimated 1.4 million veterans at risk of homelessness.  The team created a prediction model to identify those U.S. Army soldiers who were at high risk of becoming homeless after transitioning to civilian life based on information available before the time of transition.

U.S. Army Soldiers - Most important predictors are Depression and PTSD, 60% of Soldiers in the highest 20% of risk can be predicted by machine learning models, From 2010 to 2020 the homeless population was reduced to nearly half.
Graphic by MC3 Brooks Smith, USU

Military veterans comprise a major group who are at high risk of homelessness and face unique challenges when it comes to finding safe and affordable housing. Often, they also encounter economic hardships and endure mental illness that affects their ability to find what they need, which in turn makes it difficult to provide them with the assistance they need. As a result, creating a proactive and accurate method of prediction for high risk soldiers is an essential step toward lessening the challenges faced by those who serve, and machine learning has proven to be a key step in the right direction. 

“The ability to predict and prevent homelessness has long been limited,” said Dr. Katherine Koh, lead author of the study and a psychiatrist at Massachusetts General Hospital and Harvard Medical School. “Developing a prediction model using machine learning methods that accurately predicts soldiers’ risk of becoming homeless opens up a novel way to target and help prevent soldiers from falling into homelessness.”

The machine learning approach itself is described as an equation to assess risk, similarly to how machine learning is used in many areas of medicine, such as when making decisions about whether to place a patient on a statin, a drug that lowers cholesterol. The responses from multiple predictor data sources were coded and used to establish a prediction model that would determine who might be most at risk for homelessness after returning to civilian life, a method that no prior prospective study had employed, but one that ultimately made all the difference.

One significant data source used in this study was the Army STARRS-LS, which found that two predictors were among the most important of the approximately 2,000 potential predictor variables: self-reported lifetime depression and posttraumatic stress disorder (PTSD). In addition, two self-reported indicators of suicidality (lifetime ideation and two or more lifetime suicide attempts) were also among the top predictors associated with risk of increased homelessness.

“This study is one of several in which the STARRS team, now linked with investigators from the VA [Veterans Administration], has shown the value of machine learning to predict those at risk of a number of behavioral health outcomes, including suicide and suicide attempt,” says Dr. Robert J. Ursano, a professor for the F. Edward Hébert School of Medicine’s Department of Psychiatry and director of the CSTS at USU, as well as co-author on the published study and one of its principal investigators. “The team was the first to use machine learning to predict suicide.”

Army Sgt. 1st Class Nicole Howell, 8th Theater Sustainment Command, talks with a homeless veteran in Honolulu, Aug. 5, 2015.  ( photo)
Army Sgt. 1st Class Nicole Howell, 8th Theater Sustainment Command, talks with a homeless veteran in Honolulu, Aug. 5, 2015.  ( photo)

With the information from the Army STARRS-LS surveys, the research team was able to use the machine learning approach to significantly predict homelessness, capturing about 60% of the soldiers who became homeless in the highest 20% of predicted risk. Though formerly found to be a strong factor in homelessness among veterans, substance use disorder was not identified as a strong predictor in this survey. “In this study, homelessness is the outcome examined,” Ursano adds. “Importantly, the study demonstrates the ability to predict this outcome, which means we can consider targeting those people identified as at risk for specific interventions to mitigate their chances of becoming homeless. This ‘personalized medicine’ approach to care can substantially help in sustaining our soldiers and veterans’ health.”

The results discovered in the study could find practical applications, including the potential of being added to DoD electronic health records (EHRs) to allow for specific interventions to be given to assist in preventing homelessness after separation. Additionally, the study could be implemented in the Department of the Army’s Career Engagement Survey, a self-report questionnaire wherein soldiers share their concerns/intentions about staying in the Army. All soldiers are required to complete the questionnaire before leaving the Army, and with access to an even greater sample size, the machine learning model results could possibly be even more significant, ultimately helping to reduce the prevalence of homelessness after separation/deactivation.