REACH VET

REACH VET is a suicide-prevention program the Veterans Health Administration (VHA) has run nationwide since April 2017. Each month its statistical model scores every patient who used VHA care in the prior two years, using their health records. It flags the 0.1% at highest estimated risk at each facility, so a clinician can re-evaluate them and reach out.

How the model works

REACH VET stands for Recovery Engagement and Coordination for Health, Veterans Enhanced Treatment. VHA is the health-care arm of the Department of Veterans Affairs (VA).

The model is a penalized regression, known as Lasso: a statistical method that shrinks the weight of variables adding little to a prediction. Variables are the pieces of patient information it weighs. It uses 61 of them in six groups. The first four are demographics (including age, gender, marital status, and race or ethnicity), diagnoses, medications, and use of health services. The last two are prior suicide attempt and combinations such as marital status by gender. It was cut down from a 2015 proof-of-concept with 381 variables, which was too unwieldy to compute. Ronald Kessler of Harvard refined it and added machine-learning methods.

Each month it scores about 6.28 million people and lists the top 0.1% at each VA facility. That is roughly 6,300 to 6,700 veterans a month, and more than 130,000 since 2017.

Who decides

A facility Suicide Prevention Coordinator receives each flagged name on an internal dashboard and notifies the veteran's clinician. The clinician re-evaluates the veteran's suicide risk and treatment. They then reach out without a script, offering enhanced care, safety planning, closer monitoring, and support with coping.

The flag is advisory. VA frames the model as a prompt for clinician review, not an automated decision, and states that AI will "never replace human intervention." Veterans may decline the outreach.

So this human review is real and staffed, in two stages: the coordinator, then the clinician.

What the flag catches and misses

The flagged group is at far higher risk. It dies by suicide at roughly 19 to 30 times the overall VHA rate.

But suicide is rare: about 179 a month among 6.28 million scored patients. An independent re-analysis of 2018 data found that about 0.05% of flagged veterans died by suicide. The flag missed about 98% of suicides, catching about 2%. For suicide attempts and deaths combined, about 5% of flagged veterans had the outcome.

Predictive value here means the share of flagged veterans who had the outcome. A separate retraining experiment, for veterans involved in the legal system, changed the predictive value for the combined outcome only slightly, and not significantly. That suggests the limit is the rarity of suicide itself, not a modeling flaw that could be fixed.

What the evaluations found

VA's own 2021 evaluation compared 173,313 veterans across 141 facilities. It used triple differences, a statistical method that compares changes over time across several groups. That is an observational design, not a randomized trial. REACH VET was associated with more completed outpatient appointments, more new safety plans, fewer mental-health admissions, and fewer documented suicide attempts. It was not associated with fewer deaths by suicide, or fewer deaths from any cause.

A 2025 follow-up of 266,246 observations found the same: no reliable difference in deaths. The program improved what it could measure quickly. The outcome it was built to change did not change in either study.

The 2024 investigation

In May 2024 The Markup and The Fuller Project, two news organizations, published an investigation. It reported that the model treated being a white man as a stronger risk signal than factors affecting women. It reported that the model gave weight to veterans who were "divorced and male" or "widowed and male," but to no female group.

The investigation reported that military sexual trauma and intimate-partner violence were excluded from the model. Both are linked to higher suicide risk in women veterans. It also reported that the model did not account for LGBTQ+ identity. VA research finds transgender veterans die by suicide at about twice the rate of veterans who are not transgender.

VA's suicide-prevention executive framed the exclusions as a judgment about predictive strength. The executive said military sexual trauma "was not among the most powerful for us to be able to predict suicide risk." So the finding is documented but contested, not a design intent VA has confirmed.

The female-veteran suicide rate rose about 24% from 2020 to 2021. That is roughly four times the rise among male veterans.

REACH VET 2.0

In October 2024 VA announced REACH VET 2.0. It adds military sexual trauma, intimate-partner violence, and medical factors specific to women. VA committed to evaluate it "for performance and bias before it is deployed." The announcement followed a bill from Senator Jon Tester requiring those changes.

By December 2025, trade press reported that 2.0 had launched in 2025 with the new factors. It also reported that race and ethnicity had been removed as variables. No peer-reviewed description of 2.0 or independent bias audit had been published, so its performance and fairness are not independently verified.

Oversight and support

The 2019 Hannon Act, a federal law, required a review by the Government Accountability Office (GAO). GAO's 2022 review described the program. It noted that VHA planned, but had not yet completed, studies of how the model's performance varied by age, sex, and race. It made no recommendations specific to REACH VET.

Congress has kept backing the approach. The appropriations law for military construction and VA for fiscal year 2026, signed in November 2025, funds suicide prevention and encourages expanded predictive modeling.

What this network is drawn from

This network follows the pattern the case file describes. It is not a reconstruction of the actual tool. It shows the model, the monthly flag list, the coordinator, the clinician, the health record, the mortality repository, and the program's evaluators. The veterans served, and whether they live or die, are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Here a mistake is, for example, a veteran at real risk ranked off the list. Closing a failure pathway means mistakes stop passing along it. The work along it goes on: the coordinator still notifies the clinician.

This case has a budget of 11 units. Each tool costs the same at every target level.

Before any tool is used, mistakes are copied about as fast as they are corrected. Six failure pathways are open. They are Monthly flag list to coordinator, Coordinator notifies clinician, and Clinician weighs the flag. They are also Care actions written to the health record, Health record supplies the model's variables, and Clinician reads the veteran's history.

Three checks do nothing at the start: Pre-deployment subgroup and bias check, Care measures checked against deaths, and Program review by subgroup.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets are not met before any tool is used. That level asks for mistakes to be caught faster than they are copied, and for REACH VET to be helping the work.

Mark AI-written records meets that level's targets on its own, for 2 units. It closes Health record supplies the model's variables and Clinician reads the veteran's history. Peer sharing rules does too at its stronger setting, for 3 units. Lingering effects is a Dynamics setting, on by default, in which damage outlasts its cause. With it off, the standard setting of Peer sharing rules also does, for 2 units. In all, more than 3,600 sets of tools within the budget meet them, counting each stronger setting as a set of its own.

Under Service and Safety Targets and All Governance Targets, the targets can be met, but only by spending the whole budget. Both levels ask you to close every failure pathway, among other targets. Closing all six takes five tools.

Escalate checks closes Monthly flag list to coordinator. Peer sharing rules, which sets rules for what people pass to each other, closes Coordinator notifies clinician. It also starts the check named Program review by subgroup.

Keep prompts neutral closes Clinician weighs the flag. In the Lab it keeps questions put to an AI neutral. REACH VET takes no questions, so here it stands for keeping the clinician's own re-evaluation independent of the flag.

Mark AI-written records closes the two pathways from the health record, as above. In the Lab it marks machine-written content in the records so readers can weigh it. Here that means marking the model's risk flag wherever it is logged in the health record.

Gate record entries, which requires sign-off before anything enters the records, closes Care actions written to the health record. Store less data closes it instead.

So exactly two sets of tools meet the targets at these levels, and each costs all 11 units. They differ only in Gate record entries or Store less data.

Neither set includes the tools that start the two other checks. Check with a second model starts Pre-deployment subgroup and bias check, but the work then falls behind demand. Check copied records starts Care measures checked against deaths, but closes no failure pathway.

Using every tool on offer, each at its strongest setting, costs 46 units, more than four times the budget. It closes every failure pathway and starts all three checks. But the benefit REACH VET adds would fall too far, so the targets would be met at no level.