Text-based suicide risk lexicon showed promise in crisis counseling chats

The study tested a simpler kind of artificial-intelligence tool for analyzing text in high-stakes mental health settings. Instead of relying only on a complex “black-box” model, which can be hard to inspect, the researchers built a lexicon: a structured list of words and phrases tied to specific concepts.

The new list, called the Suicide Risk Lexicon, covers 49 risk factors linked with suicidal thoughts and behaviors. These included types of suicidal thinking, self-injury, mood symptoms, and social or environmental experiences described in the source material. The researchers used GPT-4 Turbo to help generate the lexicon, then released it for other researchers to use.

The journal abstract says the team evaluated the lexicon in crisis-counseling conversations. News coverage of the work reported that the analysis involved approximately 16,000 de-identified conversations with Crisis Text Line volunteer counselors. In that coverage, conversations were grouped by Crisis Text Line assessments into nonsuicidal, suicidal thoughts without imminent risk, and imminent risk. “Imminent risk” referred to a plan for suicide or an intent to die within the next 48 hours.

The researchers also validated the lexicon with clinician ratings. That step matters because a list of relevant-sounding words is not necessarily a reliable way to measure what is happening in a real conversation. The abstract reports that the tool had high content validity, meaning it captured much of the material experts consider relevant to the risk factors it was designed to measure.

The lexicon accurately predicted risk in the crisis-counseling conversations, according to the abstract. It modestly outperformed the Linguistic Inquiry and Word Count lexicon, an established word-count approach that the authors said has limited content validity for mental illness. It also performed similarly to some deep-learning models.

The study offered one practical clue about which signals appeared most important in this setting. Active suicidal thoughts and direct self-injury were stronger indicators of imminent risk than passive suicidal thoughts and depressed mood. Active suicidal thoughts generally means thoughts involving wanting or planning to end one’s life, while passive thoughts can involve wishing not to be alive without an expressed plan or intent.

That finding describes patterns in these crisis chats; it does not mean any one phrase can determine a person’s safety or future. Suicide risk is complex, and the study does not establish that the tool can forecast later suicide attempts or deaths. Its reported outcome was the risk category assigned within the crisis-support context.

Important details are still missing from the supplied abstract and coverage. They do not provide numerical measures of the tool’s accuracy, details of how conversations were divided between development and testing, or information about participants’ ages, locations, languages, or backgrounds. They also do not report follow-up after the chats or testing in independent clinical settings.

For now, this is best viewed as a promising research tool rather than proof of a ready-to-use clinical system. Further validation would need to show how well it works with new populations and different kinds of support services, and whether it helps trained people make safer, more accurate decisions without missing people who need urgent help.