Inside VA’s yearslong AI effort to uncover veterans at high risk of suicide

Faced with a decades-long epidemic of troublingly high veteran suicide rates, the Department of Veterans Affairs has been exploring the adoption and use of artificial intelligence to help identify veterans at an increased risk of self-harm.

The statistics have been particularly alarming: More than 140,000 veterans have taken their lives since 2001, with an estimated 6,407 dying by suicide in 2022 alone, according to the most recent figures provided by the agency. Some mental health advocates have said these numbers likely represent an undercount, with one study finding that the veteran suicide rate was around 37% higher than VA reported from 2014 through 2018.

VA’s first use of advanced capabilities to identify veterans experiencing suicidal ideation predated the broader rollout of generative AI tools, like ChatGPT, that have come to dominate recent conversations about technological innovation. And continuing research efforts into the broader rollout of these tools across VA’s healthcare system — which serves an estimated 9 million veterans — portends a broader shift in the way mental healthcare providers identify those at a heightened risk of suicide. 

This article — the first in a series of pieces about VA’s embrace of AI to help prevent veteran suicides — is based on documents obtained through Freedom of Information Act requests and interviews with more than a dozen current and former VA officials and employees, researchers, veterans and veteran advocates over the past year. 

An approach ‘that helped us prioritize where we should focus’

VA’s earliest and most ambitious initiative to direct more targeted support to veterans at the highest risk of self harm is the Recovery Engagement and Coordination for Health-Veteran Enhanced Treatment — or REACH VET — program, a suicide prediction algorithm that was launched in April 2017. 

The model uses machine learning — which is a subset of AI that analyzes data to identify patterns and make decisions or predictions — to scan the department’s electronic health records and identify veterans in the top 0.1% tier of suicide risk. VA said the algorithm identified 30,000 veterans at high risk for suicide in its first year of use. 

The effort to implement the model across VA’s healthcare operations began under David Shulkin, who served as undersecretary of health and head of the Veterans Health Administration from 2015 until 2017 during the Obama administration, and then as VA secretary from 2017 to 2018 during the first Trump administration.

“We had not really had an approach that helped us prioritize where we should focus, that we were always sort of being reactive,” Shulkin said about VA’s pre-REACH VET suicide prevention efforts, adding that “what we really needed to do is try to understand why so many veterans were taking their lives and figure out a way to intervene before that happened.”

VA researchers published a proof of concept report in September 2015 about the use of predictive modeling to identify veterans at high risk of suicide. Sensing that the approach held promise, Shulkin asked VA personnel to put it into clinical practice as soon as possible. Following the model’s successful deployment, Shulkin said the feedback he received from providers convinced him that it was making a difference. 

“People would tell me stories about how the veterans would say, ‘I can't believe you're calling me. How did you know that this was the worst time of my life?’ And it was only years later that I saw the research that had been done showing that indeed this was making a real difference in terms of research, data, statistics, in saving veterans lives,” he said. 

How it works

The approach detailed in the 2015 report initially included 381 different variables in veterans’ electronic health records in its model. Dr. Ronald Kessler — the McNeil Family Professor of Health Care Policy at Harvard Medical School — was brought in to help refine the model and narrowed it down to a more concise grouping of 61 variables. VA credited Kessler with adding machine learning to the program.

Kessler said he put these factors “into what's called a Lasso — a penalized regression model — to show that, with a much smaller number of variables, [VA] could actually do a better job” with its model, adding that “as you add more and more variables, prediction accuracy will go up to a certain point, and then it will flatten off, and then at a certain point, it'll start going down.”

Some of the 61 variables in the deployed REACH VET model included recent suicide attempts, certain prescribed medications, the number of recent emergency room visits and certain medical diagnoses, such as depression and bipolar disorder. 

Dr. Matthew Miller, VA’s executive director of suicide prevention, said REACH VET provides “that pairing of the innovation and the technology with the human touch.” He stressed that the approach is designed to complement and enhance human-led interventions, rather than superseding the role of clinicians and other mental health personnel. 

Each month, REACH VET uses the 61 variables to scan the electronic health records of veterans engaged with the Veterans Health Administration to identify those who are in the top 0.1% of suicide risk. 

The list of high-risk veterans is then provided to VA facilities via a specialized dashboard, which is accessed by REACH VET coordinators. Each VA medical facility is directed to have these coordinators on their local suicide prevention teams, who then work with providers to engage the veterans and develop individualized safety plans. The coordinators and providers are also required to document their engagement with the identified veterans in their electronic health records.

Miller said these outreach conversations with veterans are not scripted, and that “it's going to go something like, ‘you've been identified as being at elevated risk for suicide at the present time; we'd like to offer you resources. We'd like to hear your input on that and we'd like to know what we can do together to assist you from there.’”

The REACH VET program and intervention approach have received plaudits from the medical community. An October 2021 report published in the JAMA medical journal concluded that the model “was associated with greater treatment engagement and safety plan documentation, fewer inpatient mental health admissions and emergency department visit days, and reduced prevalence of nonfatal suicide attempts” for veterans.

AI tools helping VA’s suicide prevention efforts

The predictive algorithm is just one of VA’s initiatives to use AI and ML to better identify veterans at an increased risk of suicide. Of the 227 total AI use cases that the agency included in its 2024 inventory, REACH VET and three other case studies were focused, in whole or large part, on identifying veterans at high risk of suicide. 

Other use cases include a predictive model to estimate the risk of a veteran having “a suicide or overdose-related health care event or death in the next year;” the use of natural language processing to parse the clinical notes of veterans in PTSD Specialty Care to identify signs of suicidal ideation; and a tool to identify veterans who have access to a gun and have an active opioid use disorder, which VA said are “two key risk factors for death by suicide.”

VA has also been updating REACH VET to a 2.0 model, which includes new risk factors such as military sexual trauma and intimate partner violence. The revised version comes after The Fuller Project reported last year that the algorithm considered being a white male a greater indicator of potential self-harm than other factors that primarily or fully affect women. 

Still, REACH VET only identifies a small portion of veterans at risk of self harm. Kessler said “this top 0.1% account for 2% to 3% of all suicides in VA,” and cited ongoing research he is conducting that has found a stronger correlation between retired servicemembers who were recently hospitalized and an increased suicide risk. 

“If they just went to the people who were hospitalized and applied this other model, they would get three times as many people who committed suicide with the same amount of money, same effort,” Kessler said. He added that he has not had any further engagement with VA about the REACH VET model since it was launched, but that he is working with the department on several other research projects. 

Suicide prevention coordinators who spoke with Nextgov/FCW on background also said their experiences have varied by VA medical facility, with some of their team supervisors offering them more support than at other sites where they worked. Others cited large caseloads and small teams as hindrances to engaging with non-responsive veterans. 

Since REACH VET coordinators are tasked with connecting providers with veterans identified by the algorithm, some coordinators also have much larger caseloads than others. A September 2022 Government Accountability Office report that reviewed the REACH VET program noted that “the number of veterans identified at each VHA facility varies based on the size of the veteran population that the facility serves.”

The model, however, was never meant to be a panacea for the veteran suicide crisis. REACH VET works alongside other VA initiatives designed to provide more streamlined care to veterans determined to be at high risk of self harm. Veterans can also contact the Veterans Crisis Line for support if they are experiencing a mental health crisis, and outside veteran service organizations also provide a variety of suicide prevention resources.

Despite efforts, veteran suicide rates remain high

Even as VA looks to expand its use of AI tools and other resources to better direct mental health services to veterans, the department is still struggling to bring down the overall suicide rate for retired servicemembers. 

This includes a heightened risk of self harm for veterans as they transition out of their military service. A September 2024 GAO report flagged a study that found that “the suicide rate within the first year of separation was about 2.5 times higher than for the active duty population.” 

Transitioning servicemembers are required to go through the Transition Assistance Program to facilitate their retirement from the military, although some veterans who spoke with Nextgov/FCW said it was often jarring and frustrating to go from active service back to civilian life. 

“I remember sitting in those classes and I'm just thinking, ’okay, you know, last week I was getting shot at. Now, I'm sitting in a classroom not really ingesting what they're saying,’” one veteran said, adding that it also took him months to receive his discharge papers. 

VA Secretary Doug Collins and Defense Secretary Pete Hegseth signed a memo on May 23 seeking to improve the transition process, including better continuity of healthcare and mental health services for retiring servicemembers.

Previous efforts to expand out resources for retired servicemembers at risk of self-harm, however, have not led to dramatic decreases in veteran suicides. And while REACH VET and many other VA initiatives are geared toward veterans who are already engaged with the agency, they often fail to account for those outside of its network. 

VA’s 2024 National Veteran Suicide Prevention Annual Report noted that 49.6% of veterans who took their own lives in 2022 did not receive Veterans Health Administration or Veterans Benefits Administration services. 

An agency official told lawmakers in April that the number of veterans who have died by suicide has remained “essentially unchanged” since 2008, despite the fact that “over that same period, VA spending on suicide prevention has increased by more than 11,000%, from $4.4 million per year in 2008 to $522 million per year in 2022.”

Collins reiterated those concerns during a June 24 congressional budget hearing and said he is looking at pursuing “a fresh, new approach” to suicide prevention. 

If you are a veterans in crisis or are having thoughts of suicide, or if you know a veteran in crisis, you should call the Veterans Crisis Line for confidential crisis support. Dial 988 then Press 1, chat online at VeteransCrisisLine.net/Chat or send a text message to 838255. The line is available 24 hours a day, 365 days a year.