MIT tool identifies suicide-risk signals in crisis text conversations
Researchers at MIT’s McGovern Institute for Brain Research have developed a language-processing tool that accurately predicts suicide-risk severity from text conversations between people in crisis and Crisis Text Line counselors. The team analysed de-identified text from approximately 16,000 conversations and reported the results in the Journal of Psychopathology and Clinical Science.
The model is built to distinguish three groups used in Crisis Text Line assessments: non-suicidal conversations, suicidal ideation without imminent risk, and imminent risk. The latter category includes people with a suicide plan or an intent to die within the next 48 hours, making rapid identification especially important.
A lexicon linked to 49 risk factors
Daniel Low, formerly a graduate student in Satra Ghosh’s Senseable Intelligence Group and now a research scientist at the Child Mind Institute, developed the work with Ghosh and colleagues. They first used AI to produce a preliminary vocabulary connected to established suicide risk factors, then manually reviewed and curated it.
The final suicide-risk lexicon contains about 60 words or phrases for each of 49 factors, with relevance confirmed by expert clinicians. The machine-learning model searches conversations for those signals and assigns weights based on their contribution to predicted risk. This design lets researchers connect a risk estimate to identifiable categories rather than producing only an opaque score.
The analysis found patterns consistent with earlier research, but also highlighted distinctions relevant during an active crisis. Mentions of lethal means, such as “cut” or “pills”, and substance use were more often expressed by the highest-risk group than depressed mood or fatigue. Active suicidal ideation and self-injury were also strong predictors, while anxiety, post-traumatic stress disorder and emotional pain were intermediate predictors.
Interpretability, privacy and clinical limits
Unlike a large language model, the prediction model is described as lightweight and can run on a personal computer. MIT says that lowers computational cost and can reduce privacy concerns. It can also flag the words of concern that underpin an assessment, giving users visibility into why the model assigned a particular level of risk.
The researchers note important limitations. A lexicon does not interpret the context surrounding a term and may miss similar language that is not explicitly listed. The group has also developed approaches using large language models for suicide-risk detection, but uses the lexicon in parallel to ensure that specified terms are flagged and to help maintain data privacy.
Ghosh stresses that a human in the loop will remain critical in such a complex, high-stakes setting. The team says predictive models require thorough validation before clinical use and may need continuing refinement as language and target populations change. It is sharing the lexicon and its software package so other researchers can build comparable vocabularies for additional mental-health conditions.
What organisations should take from the work
For organisations evaluating text-based support and assessment systems, the practical implication is to prioritise tools that expose their signals, protect sensitive data and remain embedded in trained human decision-making rather than treating a model output as a standalone clinical judgment.