AI Detects Suicide Risk Missed by Standard Assessments
New Study Finds That Large Language Models Can Detect Nuances of Language That Regular Assessments Miss
Researchers at Touro University have found that an AI tool identified suicide risk that standard diagnostic tools missed. The study, published in the Journal of Personality Assessment, provides evidence that large language models, a type of AI that processes and analyzes conversational language, show promise for detecting who is at risk of attempting suicide.
“It’s very difficult to predict who might attempt suicide, which makes it hard to know who to help and when to intervene,” says Yosef Sokol, PhD, clinical assistant professor at Touro's School of Health Sciences and the lead author of the study. Current methods for assessment are not very accurate, typically relying on people answering multiple choice questions about themselves. Those questions lack nuance and often don’t reflect respondents' unique experiences; respondents may also be worried about directly answering questions about suicide.
“I’m trying to create a set of methods that enable us to assess for suicide risk without asking direct questions,” says Sokol. “We can do that by using the characteristics of language to predict suicide.”
In the study, the researchers used machine-learning AI programs that comprehend and interpret human language, called large language models (LLM), to assess a concept called future self-continuity, which is how a person sees the connection between their present self and future self. “Having a sense of oneself that continues into the future is a core aspect of healthy identity and its lack is strongly related to suicide,” says Sokol. Because future self-continuity is very nuanced, standard tests are not good at detecting it.
The study used an LLM (Claude 3.5 Sonnet) to process audio responses to 15 interview prompts of people talking about themselves and their futures. It included 164 participants, 93 of whom had past-year suicide ideation. The researchers first asked the participants to rate their risk on a 1-7 scale of how likely they were to attempt suicide in the future. They then compared the results of the LLM and standard assessment tools to that self-reported perceived risk.
The LLM found patterns in natural speech that standard methods missed, such as how coherently someone described their future, their emotional tone, and the level of detail they provided. “We got a better measure of future self-continuity and were better at predicting if people thought they would attempt suicide,” says Sokol.
This study compared LLM insights to the perceived risk of suicide, not actual attempts at suicide, but perceived risk is important information for healthcare providers. Research shows a person’s perceived risk predicts actual suicidal behavior; even without a future attempt, providers want to identify and support people who feel that they are at risk of attempting suicide.
According to Sokol, LLM predictions could be used in any setting, such as a hospital, crisis hotline, or even with a therapist. “We want to create a system where a therapist could ask you a set of 2-3 questions and record it, and the LLM could produce a risk score,” says Sokol.
LLM could also potentially be used to help detect many mental health disorders such as depression or anxiety, and may be even easier to detect than risk of suicide, he says. “Suicide may be one of the hardest to detect, and, to save lives, the most important.”
