Early‑risk from records
Researchers at UCLA showed that new methods for analyzing existing emergency and service records — including large language models — can flag evidence of suicide risk earlier than waiting for direct disclosure. The work isn’t a plug‑and‑play school tool, but it frames prevention as pattern recognition across attendance, nurse visits and behaviour data rather than single events. (newsroom.ucla.edu)
Most suicide tracking still works like a checklist: was there a diagnosis, was depression recorded, did someone explicitly say they were at risk. A UCLA-led study argues that approach misses what is often visible in the notes around a death, not just the boxes on the form. (jamanetwork.com) The records in this study did not come from therapy sessions. They came from the United States National Violent Death Reporting System, which combines death certificates with reports from law enforcement, coroners, and medical examiners into one anonymous database. (cdc.gov) That database is rich in narrative detail but poor at capturing distress in a standard way. The federal system tracks more than 600 data elements, yet the UCLA team says the usual mental-health fields still make suicide look less tied to acute emotional disruption than clinicians would expect. (cdc.gov) (newsroom.ucla.edu) The researchers used a large language model, which is a computer system trained to spot patterns in ordinary writing, on short death narratives written by investigators. They were asking whether the model could detect signs of emotional dysregulation, meaning severe trouble controlling feelings and behavior under stress. (jamanetwork.com) (newsroom.ucla.edu) They analyzed 72,585 suicide deaths from 2020 and 2021 in all 50 states, limiting the sample to people age 12 and older whose records included both a law-enforcement narrative and a coroner or medical examiner narrative of at least 20 words. The average age was 46.3 years, and 57,770 of the decedents, or 80.6%, were male. (jamanetwork.com) The striking result is that the model found clinically elevated distress across most domains, even though standard records often did not mark the person as depressed or diagnosed. UCLA said indicators of clinically relevant emotional distress may have been present in about 90% of suicides, far above what current coded measures suggest. (newsroom.ucla.edu) (jamanetwork.com) The age and sex pattern also mattered. In the journal paper, female decedents and younger decedents scored higher on these model-based symptom measures than male and older decedents. (jamanetwork.com) This is not a chatbot deciding who is in danger at school next Tuesday. The paper studied people who had already died, and it used retrospective narratives to show that hidden signals can sit inside existing records long before a system treats them as actionable risk. (jamanetwork.com) That shift is the real idea here: stop waiting for one dramatic event or one direct confession, and look for a pattern spread across ordinary documents. If future systems can read attendance changes, nurse visits, emergency calls, and behavior reports the way this model read death narratives, prevention starts to look more like weather forecasting than a single alarm bell. (newsroom.ucla.edu)
