Predicting imminent suicide risk in a crisis hotline chat using machine learning

Identifying individuals at the highest risk of suicide in real time is one of the critical tasks in suicide prevention. However, understanding the mental processes of at-risk individuals is a formidable endeavor due to the rapid, impulsive, and temporally dynamic characteristics of severe suicide crises. Recent theories, such as the suicide crisis syndrome, have suggested sets of short-term predictors of imminent suicide risk (IMSR), defined as an acute state that requires urgent intervention to prevent suicide. However, to date, no study has examined the prediction value of these theories in real time. In this study, we used machine learning to investigate its potential for predicting IMSR during internet-based crisis hotline chat sessions. We analyzed 3309 anonymized chat sessions from an internet-based crisis hotline, 312 of which were classified as IMSR chats (i.e., requiring immediate intervention). We compiled a lexicon of psychological factors derived from the main theories of suicide crisis and extracted language patterns associated with these key theoretical constructs. A logistic regression model within a machine learning framework was used to determine the odds ratio of each predictor, while temporal analysis was used to examine the stability of predictors throughout the chat duration. Suicidal ideation with a specific plan and intent emerged as the strongest predictor of IMSR. Additionally, pain tolerance and deliberate self-harm, components of acquired capability, were significantly associated with IMSR, aligning with the interpersonal theory of suicide. Cognitive rigidity and impulsivity, markers of cognitive deficiencies, also played a key predictive role. Conversely, perceived burdensomeness, depressive symptoms, and emotional pain were negatively associated with IMSR. Temporal analysis revealed that most factors remained stable throughout the chats. IMSR is best understood through an integrated approach that combines cognitive, affective, and behavioral components from multiple theoretical frameworks. As many individuals do not explicitly disclose their suicidal intentions, it is crucial to identify indirect risk factors to improve real-time risk detection. This study advances the theoretical understanding of imminent suicide risk and the development of practical tools for real-time crisis intervention through machine learning-driven analysis of crisis chat interactions.