Comparing Risk Prediction for Suicide Attempts and Deaths After Emergency Department Visits for Individuals With...

Objective: To identify and compare predictors of nonfatal and fatal suicidal events within 180 days of emergency department (ED) visits for mental health disorders.

Methods: This longitudinal cohort analysis assessed risk of nonfatal and fatal suicide events among a national sample of 511,559 patients presenting to an ED between 2015 and 2022 with 1 or more mental health disorder. Machine learning models were used to predict events incorporating demographic and clinical characteristics, incremental value of clinical feature sets, feature importance rankings across models, and accuracy with which a model trained to predict nonfatal outcomes predicted suicide deaths. Nonfatal attempts within 6 months of an ED visit were identified using electronic health records (n=4,525), while fatal events were identified from the National Death Index (n=434).

Results: The final sample of 872,627 ED episodes represented 511,559 individuals. The full model with 143 features achieved stronger performance for nonfatal events (area under the curve [AUC]=0.874) than for suicide deaths (AUC=0.787). The relative importance of features was highly correlated across models (r=0.82), with 73% of overlap between the top 15 predictors for nonfatal and fatal events. When the nonfatal model was applied to fatal outcomes on the same sample, performance declined (AUC=0.724), indicating limited generalizability despite similar predictors (sensitivity=78.1%, specificity=67.3%).

Conclusions: Although nonfatal and fatal suicidal events share predictors, their predictive strength and function differ. These differences underscore challenges in using models trained on nonfatal outcomes to identify risk of suicide death and contribute to the ongoing debate over whether suicidal thoughts and behaviors reflect a continuum of severity or distinct clinical pathways.