AI Finds Hidden Self-Harm in 7.9% of Veteran Records

University of New Mexico researchers just found a way to make invisible mental health histories visible, and it could transform how healthcare systems support veterans in crisis.

Using a novel machine learning method, the team analyzed records for 1.3 million Veterans Health Administration patients. They discovered something startling: diagnosis codes captured only one-fourth of documented self-harm history.

The real number was four times higher than what showed up in standard tracking systems. About 7.9% of veterans had self-harm documented somewhere in their records, compared to just 1.85% visible through diagnosis codes alone.

"For research and planning, if we only count what is easy to see in diagnosis codes, we may substantially underestimate the need for mental health services," said Dr. Christophe Lambert, who led the study published in the Journal of Medical Internet Research.

The gap happens because crucial information gets buried in clinical notes. Some patient records contained more than 500,000 lines of notes, far too much for any clinician to review during a normal visit.

Past self-harm matters enormously because it's one of the strongest predictors of future suicide risk. It also shapes how doctors approach treating depression, PTSD, bipolar disorder, and other conditions that often occur alongside self-harm.

The team used a method called PULSNAR that works with messy real-world data. Unlike typical machine learning that needs clear yes/no examples, PULSNAR recognizes that missing information doesn't mean something never happened.

"Medical records can make self-harm hard to see in more than one way," said first author Dr. Praveen Kumar. "Sometimes the history is in a clinician's note but not in the diagnosis codes."

The study brought together experts from UNM, Veterans Affairs medical centers, Vanderbilt, and the VA Office of Mental Health. They combined expertise in informatics, psychiatry, computer science, and health services research.

The Ripple Effect

Better visibility means better care planning at every level. Health systems can allocate mental health resources more accurately. Researchers can study treatment outcomes more reliably. Clinical teams can identify patients who might need extra support before a crisis hits.

The VHA already uses specialized suicide monitoring tools and doesn't rely solely on diagnosis codes for immediate risk assessment. This research addresses a different challenge: making historical information accessible when planning long-term care at scale.

The breakthrough shows how artificial intelligence can help healthcare systems see what's been hiding in plain sight all along.