UCLA: records + LLMs could flag risk earlier
Researchers at UCLA report that analyzing existing emergency‑response and administrative records with large language models can surface signals of suicide risk before an event occurs, suggesting earlier windows for intervention. The team cautions this approach is augmentation — not a replacement for human judgment — and raises obvious privacy and false‑positive issues for schools to consider. (newsroom.ucla.edu) (newsroom.ucla.edu)
Most suicide warning signs never arrive as a neat medical diagnosis. They often show up as scattered details in police reports, coroner notes, and other narrative records that were written for documentation, not prediction. (jamanetwork.com) A large language model is a pattern-finder for text. Feed it thousands of short case summaries, and it can spot recurring language the way a smoke alarm picks up particles before anyone sees flames. (jamanetwork.com) The UCLA-led team tested that idea on the National Violent Death Reporting System, the Centers for Disease Control and Prevention database that combines death certificates with law-enforcement and coroner or medical-examiner records. Those records are anonymous, and they are built from the source documents investigators already collect after violent deaths. (cdc.gov) The researchers looked at 72,585 suicide deaths from 2020 and 2021 for people age 12 and older whose files included both law-enforcement and coroner narratives of at least 20 words. The average age was 46.3 years, and 80.6% of the people in the sample were male. (jamanetwork.com) Instead of asking only whether a person had a documented psychiatric diagnosis, the team looked for “emotional dysregulation,” which means distress severe enough that feelings and impulses are not staying under control. That is closer to asking whether the person sounded overwhelmed in the final stretch, not whether they had the right billing code in a chart. (newsroom.ucla.edu) Current national coding often misses that picture. UCLA says the main federal tracking system suggests fewer than half of suicide decedents had a mental health disorder at death, and fewer than a third were known to have been depressed before death. (newsroom.ucla.edu) The language-model approach found something much larger. In the paper, clinically relevant dysfunction in the “negative valence” domain, which covers responses to threat, loss, and distress, appeared in 90.3% of the deaths. (jamanetwork.com) That gap matters because the raw material was already sitting in the files. The model was not discovering a new lab test; it was reading existing narratives from emergency-response and death-investigation records that standard coding had flattened into simpler categories. (newsroom.ucla.edu) The study also found differences by age and sex. Younger people and women showed higher levels of this distress signal, which suggests one-size-fits-all screening can miss how risk shows up in different groups. (newsroom.ucla.edu) This was not a live school monitoring system, and it did not prove a model can predict an individual crisis in real time. It was a cross-sectional study of death records, so the immediate result is better measurement of hidden warning signs, not a ready-made product for campuses. (jamanetwork.com) The pitch from the researchers is narrower than “let artificial intelligence decide who is at risk.” They describe large language models as a way to augment human review of records that schools, hospitals, or public agencies may already have, with earlier intervention as the goal. (ph.ucla.edu) The hard part comes next. Any attempt to use school or administrative records this way would have to deal with privacy rules, false positives, and the cost of responding to alerts, because a system that flags distress without trusted humans and real support behind it can create a new problem instead of solving the old one. (jmir.org)
