The Pulse
MIT Tests Suicide-Risk Model on 16,000 Crisis Chats
MIT researchers tested a lightweight text model on about 16,000 crisis-counseling conversations, using a lexicon linked to 49 suicide-risk factors. The tool can flag patterns in text, but the researchers say it needs thorough validation and
AI.info Team ·
A text tool faces a human judgment test
MIT researchers say a lightweight text model can help counselors sort crisis conversations by suicide-risk level. But the tool’s word-based approach can miss context, and its developers say it needs further validation and human oversight before clinical use. The study tests whether a model can identify patterns in what people say—not whether it can determine who will later attempt suicide.
The team analyzed about 16,000 de-identified conversations with Crisis Text Line volunteer counselors. The nonprofit provided specialized training and controlled access to the restricted dataset. Counselors’ assessments placed the conversations in three categories: non-suicidal, suicidal ideation without imminent risk, and imminent risk, defined as having a plan or intent to die within the next 48 hours.
49 factors become a searchable lexicon
To build the tool, the researchers used artificial intelligence to generate a preliminary list of words and phrases associated with 49 suicide-risk factors. They manually reviewed and curated the list, which contains about 60 terms or phrases for each factor; expert clinicians confirmed their relevance. A machine-learning model then searches text for those terms and uses their associated factors to estimate risk.
The model’s predictions also let the researchers examine which factors appeared more often in the highest-risk group. Mentions of lethal means and substance use were more associated with that group than depressed mood or fatigue, MIT reported. Active suicidal ideation and self-injury were also strong predictors, while anxiety, post-traumatic stress disorder and emotional pain fell in an intermediate range.
Readable scores, but limited context
The model assigns weights to risk factors, making it possible to see which factors contributed to a score. That differs from a system that returns a risk estimate without showing how it reached it. The large language model helped create the lexicon; the risk-prediction model itself is a simpler system that, MIT says, can run on a personal computer.
A lexicon can also treat a phrase differently from the way a person would. It does not understand context, and it may miss language that conveys a similar meaning but is absent from its list. The researchers say they sometimes use the lexicon alongside larger language models, both to ensure certain terms are flagged and to support data privacy.
Ghosh says human review will remain necessary
“This is such a complex space that having a human in the loop is, I think, going to be critical for a long, long time,”
Satra Ghosh, a senior research scientist and director of the Open Data in Neuroscience Initiative at MIT’s McGovern Institute, tells MIT News. The researchers say models need thorough validation before clinical use and may require updates as language and target populations change.
The team is sharing the lexicon and software for building similar tools, while Daniel Low says researchers are using the lexicon to study text from sources including social media and electronic health records. For now, the result is a research aid tested against crisis counselors’ assessments—not a substitute for their judgment.