Evaluating natural language processing derived linguistic features associated with current suicidal ideation, past...

People with psychosis have a higher suicide risk than the general population. Natural language processing (NLP) has been used to understand communication in psychosis and suicide risk prediction, but not to predict future suicidal behavior in people with psychosis. We utilized NLP-derived linguistic features from a dyadic task among people with psychotic disorders to predict current suicide ideation, past suicide attempts, and future suicidal behavior.

N = 112 adults with psychotic disorders completed the Columbia-Suicide Severity Rating Scale at baseline and one-year follow-up to capture lifetime suicide attempts, current suicidal ideation, and suicidal behavior during the follow-up period. At baseline, participants completed a dyadic role-play task called the Social Skills Performance Assessment. Lexical features, lexical diversity, and sentiment features were extracted from task transcripts. Models for each outcome were generated using a 70 %-30 % train-test split with MLPRegressor. SHapley Additive exPlanations (SHAP) was utilized for feature analysis.

A total of 42.9 % of participants had baseline suicidal ideation, 67.9 % had at least one past suicide attempt, and 13.3 % had at least one reported new suicidal behavior at one-year follow-up. The models had strong predictive performance for past attempts (F1 = 0.75) and current ideation (F1 = 0.74-0.79), with future suicide behavior models showing the strongest predictive performance (F1 = 0.86-0.93). The top features varied across suicide ideation, past attempts, and future behavior.

NLP-derived features from a dyadic task were associated with high predictive accuracy for future suicidal behavior. Pending replication, these findings suggest that NLP-derived linguistic features from dyadic tasks could contribute to understanding suicide risk among people with psychosis.