Researchers led by Chris J. Kennedy at the Center for Precision Psychiatry at Massachusetts General Hospital have demonstrated that an AI model can predict suicide risk among U.S. Army veterans with moderate to good accuracy up to ten years out, using only data available before service members even leave active duty, a finding that could reshape how the Department of Veterans Affairs identifies at-risk veterans long before existing tools do.
REACH VET’s Track Record So Far
The VA’s existing suicide risk model, REACH VET, launched in 2017 and scans veterans’ electronic health records to identify those in the top 0.1 percent of predicted suicide risk, using variables like prior attempts, certain medications, diagnoses such as depression or bipolar disorder, and emergency room visits. In its first year, the VA said the algorithm identified roughly 30,000 veterans at high risk, and a 2021 study published in JAMA found REACH VET was associated with better treatment engagement, more documented safety plans, fewer psychiatric hospitalizations and fewer nonfatal suicide attempts, though it did not measurably reduce deaths from suicide within a six-month window.
Why a Ten-Year Model Changes the Calculation
REACH VET operates on relatively recent health record data and predicts near-term risk, but the Mass General Brigham-led research pushes prediction much further out, modeling risk using information available while service members are still in uniform, well before most of the health record signals REACH VET relies on would even exist. That distinction matters because it could allow the VA and Department of Defense to identify elevated long-term risk during active service or the transition to civilian life, a period researchers have long identified as particularly dangerous for veteran suicide risk.
REACH VET Is Also Being Rebuilt in Real Time
Separately, the VA has continued updating its existing REACH VET model, now on a version described as REACH VET 2.0, which incorporates new variables including military sexual trauma and intimate partner violence, while removing race and ethnicity as model inputs following concerns about algorithmic fairness and potential discriminatory impact. Researchers have also been testing REACH VET’s adaptability for specific subpopulations, including veterans involved in the legal system, a group that carries distinct risk factors not well captured by the original model.
Lawmakers See Political Momentum, and Money
The research has attracted attention on Capitol Hill: one lawmaker has proposed grant funding specifically to support development of veteran suicide prevention AI models, and congressional committees working on the fiscal year 2026 VA funding bill have signaled support for expanding AI-based suicide prevention tools, reflecting bipartisan appetite for tools that could address veteran suicide rates that have remained stubbornly elevated compared to the general population for years.
Clinicians Warn Prediction Alone Doesn’t Save Lives
Mental health researchers caution that identifying at-risk veterans, even years in advance, only helps if the VA has adequate staffing and resources to act on those predictions, echoing REACH VET’s own mixed track record of improving engagement and safety planning without necessarily reducing suicide deaths. Separately, generative AI large language models have also been tested against human mental health providers on veteran suicide risk stratification tasks, with mixed results that researchers say underscore AI should augment, not replace, clinical judgment in such high-stakes assessments.
What Comes Next
The Mass General Brigham team’s decade-long prediction model remains a research finding rather than a deployed clinical tool, and researchers say substantial validation work remains before the VA or Department of Defense could integrate it into an operational screening pipeline alongside REACH VET 2.0. If it holds up under further study, veteran advocates say it could shift suicide prevention efforts earlier into a service member’s career, rather than concentrating resources only after veterans have already left the military and accumulated the health record red flags current tools rely on.
Photo: sasint / PIXABAY via Pixabay
