Estimating suicide risk from text

Scientists at MIT's McGovern Institute for Brain Research have developed a language-processing tool that estimates a person's suicide risk from text, the institute reports. The tool, developed by Daniel Low and colleagues, uses a custom-built list of words and phrases linked to 49 suicide risk factors to search text and estimate risk.

Low is a former graduate student in Senior Research Scientist Satra Ghosh's Senseable Intelligence Group, now a research scientist at the Child Mind Institute and a visiting scholar at Harvard University. Ghosh, Low, and colleagues report in the Journal of Psychopathology and Clinical Science that the tool accurately predicts suicide risk from text conversations with crisis counselors.

The researchers collaborated with Crisis Text Line, which provided specialized training and controlled access to a restricted dataset. They analyzed de-identified texts from approximately 16,000 conversations. Crisis Text Line assessments grouped these into three risk levels: non-suicidal, suicidal ideation without imminent risk, and imminent risk, the last defined as having a plan for suicide or intent to die within 48 hours.

To build the tool, the team used artificial intelligence to generate a preliminary list of words and phrases tied to established suicide risk factors, then manually reviewed and curated it. The final lexicon includes about 60 words or phrases for each of 49 risk factors, with relevance confirmed by expert clinicians. A machine learning model was then trained to search conversations for these terms and predict suicide risk.

The model found mentions of lethal means and substance use were more likely expressed by the highest-risk group than depressed mood or fatigue. Active suicidal ideation and self-injury were also strong predictors; anxiety, post-traumatic stress disorder, and emotional pain were intermediate predictors. The model assigns each risk factor a weight based on its contribution. Lethal-means terms such as "cut" or "pills" are weighed heavily, while hopelessness-related terms contribute less.

One limitation of lexicons, the researchers note, is that they do not consider term context and can miss similar terms not explicitly included. Low stresses the prediction model is a simpler, lightweight model that runs on a personal computer and is interpretable, telling users how it reached an assessment. Ghosh says having a human in the loop will remain critical for a long time. The researchers add that any predictive model must be thoroughly validated before clinical use and may need continual refinement. They are sharing the suicide risk lexicon and the software package used to build it.