VA is updating its AI suicide risk model to reach more women

The Department of Veterans Affairs is in the process of adding additional risk factors to its artificial intelligence-powered tool for identifying veterans at high risk of suicide to better account for the experiences of women. 

The effort comes after a report released in March by nonprofit Disabled American Veterans warned that the department’s suicide prevention tool — the Recovery Engagement and Coordination for Health — Veterans Enhanced Treatment program — used retired male servicemembers as its baseline. 

The organization noted that the REACH VET model does not factor in military sexual trauma, or MST. VA data shows that one in three women and one in 50 men have confided in their clinical provider that they have experienced MST. The report recommended that VA revise its algorithm to include risk factors for MST, as well as intimate partner violence. 

A subsequent investigation conducted by The Fuller Project, in partnership with Military Times and Military.com, also found that the program’s algorithm considered being a white man more of an indicator of potential self-harm than other factors that fully or largely affect women. 

Naomi Mathis, Disabled American Veterans’ assistant national legislative director, told Nextgov/FCW that, as VA adopts modern capabilities like REACH VET, “you would think that these tools would enhance or enable the VA system to work better for the modern service or the modern veteran.”

REACH VET, which fully launched in 2017, uses a predictive model to analyze data from veterans’ electronic health records to identify those in the top 0.1% tier of suicide risk. REACH VET has identified approximately 6,700 veterans per month for additional healthcare assistance.

The program currently uses 61 indicators across six different categories to identify veterans at risk of self-harm, including across “demographics, diagnosis, medications, utilization and interaction terms, such as the interaction between marital statuses with gender.” 

A VA official told Nextgov/FCW that the department is “in the process of updating the REACH VET predictive algorithm to consider additional variables specific to women veterans.”

The new risk factors under consideration include intimate partner violence and MST, as well as medical conditions that affect women, such as pregnancy, fibroids, endometriosis and ovarian cysts.

“As we update the model, it will be evaluated for performance and bias before it is deployed,” the spokesperson said, adding that the goal is to launch the new algorithm in early 2025.

VA has worked in recent years to adopt novel approaches — such as REACH VET — to support veterans in need of mental healthcare services, even as the number of retired servicemembers who have died by suicide remains high. 

According to VA’s 2023 National Veteran Suicide Prevention Annual Report — which was based on data from 2021 — the rate of veteran suicide increased by 11.6% from 2020. The same review found a 24.1% increase in the age-adjusted suicide rate for women veterans from 2020 to 2021, compared to an increase of 6.3% among male veterans during the same period. 

Mathis said it was a positive step that VA was finally moving to add MST as a risk variable in its upgraded REACH VET algorithm but questioned why it didn’t factor it into its initial model, especially since it affects both men and women.

“You have underreporting of MST, and you're saying, ‘MST is not statistically significant,’” she added. “Well, then you're missing all of those people that do feel comfortable enough to report it.”

The VA representative pushed back on claims that the current REACH VET tool prioritizes men, saying that it only prioritizes individuals at the highest risk of self-harm. They noted, however, that “sex is included as a variable in the model and being male does have a positive value.”

The department has already submitted REACH VET’s predictive model and use case for a “safe AI review,” as outlined in President Joe Biden’s October 2023 executive order on the secure and trustworthy use of AI.

“We have tested the current REACH VET algorithm and are also currently testing all candidate models for the new REACH VET algorithm to ensure that they fairly estimate suicide risk across key demographic populations,” the spokesperson said.

Concerns about specific suicide risk factors not being included in REACH VET’s predictive algorithm have also attracted the attention of some top lawmakers. 

Senate Veterans’ Affairs Committee Chair Jon Tester, D-Mont., introduced legislation on Sept. 25 to improve mental healthcare services for veterans, with the proposal also seeking to address some of the concerns raised around the factors included in REACH VET’s algorithm. 

The legislation included a provision that would require VA to modify REACH VET to include “risk factors weighted for women,” including MST and intimate partner violence.

Tester said in a statement to Nextgov/FCW that VA needs to add the variables to its model to ensure that all veterans are receiving the mental health services they need.

“Women veterans are the fastest growing demographic of veterans, and my legislation will ensure VA takes the experiences of women veterans, specifically survivors of military sexual trauma and intimate partner violence, into account when treating veterans’ mental health to help make sure no veteran is falling through the cracks,” he added.