Researchers build suicide-risk text tool that explains its assessments

Daniel Low, Satra Ghosh and colleagues analysed approximately 16,000 de-identified Crisis Text Line conversations to study language associated with suicide risk. Jennifer Michalowski described the work for MIT’s McGovern Institute for Brain Research.

Their study, published in the Journal of Psychopathology and Clinical Science, used the service’s assessments to group conversations into three risk levels. The researchers report that their model predicted risk severity in conversations it had not seen during training.

Those assessments concern risk categories in crisis conversations, not proof that the system can predict who will later attempt suicide.

An assessment users can inspect

The team used AI to draft a vocabulary linked to established risk factors, then manually curated it and had expert clinicians confirm its relevance. The resulting lexicon covers 49 factors.

A simpler machine-learning model uses that vocabulary to estimate risk and flag the language behind its assessment. It can run on a personal computer, reducing computing demands and allowing local processing of sensitive text.

The trade-off is context: a lexicon can miss unfamiliar wording or misread a term’s significance. The researchers stress the need for thorough validation before clinical use and continued human involvement. They are sharing the lexicon and software used to build it, allowing researchers to develop vocabularies for other mental-health conditions.