MIT Researchers Develop Suicide-Risk Tool for Crisis Texts, but Clinical Use Needs Validation

MIT researchers have built a tool that estimates suicide risk from crisis text conversations and shows which words shaped its assessment. In an analysis of about 16,000 conversations, it identified patterns associated with Crisis Text Line’s highest-risk category. The finding concerns assessed risk levels, not whether the tool can predict a future suicide attempt.

A record of what people said during a crisis

The researchers wanted to identify which signs matter most while someone is in crisis, rather than rely on later recollections. Crisis Text Line gave them specialized training and controlled access to de-identified conversations with its counselors.

Crisis Text Line had classified the exchanges as non-suicidal, suicidal thoughts without imminent risk, or imminent risk. The last group included people with a suicide plan or an intent to die within 48 hours. The researchers used those assessments to test whether language could distinguish the groups.

About 16,000Conversations analyzed

The analysis used de-identified exchanges with Crisis Text Line counselors.

49Risk factors in the word list

The researchers linked words and phrases to 49 suicide risk factors.

From a word list to a risk estimate

The team used AI to draft a list of terms associated with established risk factors, then reviewed it by hand. Expert clinicians confirmed the terms’ relevance. The finished list has about 60 words or phrases per factor. A separate machine-learning model searches conversations for those terms and weighs their contribution to a risk estimate.

The prediction model is lightweight enough to run on a personal computer. It flags contributing words so a person can inspect an estimate’s basis. The researchers say using this simpler model rather than a large language model for prediction reduces cost and privacy concerns.

What the model picked up

Mentions of lethal means and substance use were more likely in the highest-risk group than mentions of depressed mood or fatigue. Active suicidal thoughts and self-injury were also strong indicators. In the model’s weighting, lethal-means terms contributed more to an estimate than terms associated with hopelessness.

On conversations it had not previously seen, the trained model predicted the assessed risk levels. The researchers describe its performance as accurate. Matching those categories does not establish that the tool can forecast an attempt or improve an intervention.

The boundary of a word-based alert

A word match is not a full reading of a message: the lexicon cannot account for context and may miss wording absent from its list. Researcher Satra Ghosh says human involvement in this complex work will remain critical.

The team says the model needs thorough validation before clinical use and may need updates as language or the people being assessed change. Researchers are sharing the suicide-risk lexicon and the software used to build it. The next challenge is establishing whether an inspectable alert is reliable enough to help people make decisions during a crisis.