When individuals reach out during a mental‑health crisis, counselors must quickly identify those at high risk of suicide. Researchers at MIT’s McGovern Institute have built a language‑processing tool that extracts suicide‑risk signals from text conversations. The system was created by former graduate student Daniel Low in Satra Ghosh’s Senseable Intelligence Group; Low now leads the AI, Risk, and Contemplative Science Lab at the Child Mind Institute and is a visiting scholar at Harvard.
The core of the tool is a custom lexicon covering 49 suicide‑risk factors, each linked to roughly 60 curated words or phrases validated by clinicians. The team first used AI to generate a preliminary list, then manually refined it. A lightweight machine‑learning model scans crisis‑text messages for these terms and combines their weighted occurrences to estimate risk.
Data came from Crisis Text Line, a 24/7 free text‑based support nonprofit. Approximately 16,000 de‑identified conversations were analyzed and labeled into three categories: non‑suicidal, suicidal ideation without imminent risk, and imminent risk (a plan or intent to act within 48 hours). The model accurately predicted risk levels on unseen conversations.
Key findings align with prior research but are not always intuitive: mentions of lethal means (e.g., “cut”, “pills”) and substance use were stronger indicators of the highest‑risk group than depressed mood or fatigue. Active suicidal ideation and self‑injury also ranked high, while anxiety, PTSD, and emotional pain were intermediate predictors. Each factor receives a weight; lethal‑means terms carry the heaviest weight, whereas expressions of hopelessness contribute less.
Lexicon‑based approaches have limits—they ignore context and may miss synonyms not explicitly listed. The researchers are experimenting with large language models (LLMs) in parallel to improve coverage while preserving privacy. Unlike resource‑intensive LLMs, the current lightweight model runs on a personal computer, lowering cost and enhancing interpretability: flagged words reveal why a risk score was assigned, enabling counselors to act on concrete evidence.
The lexicon and the accompanying software package have been released publicly, allowing others to build similar resources for different mental‑health conditions. Ongoing work explores applying the method to social‑media streams and electronic health records to broaden and refine suicide‑risk estimation.
Review