How-To Guides

The study's best-performing model predicted nonfatal suicidal events better than suicide deaths, but it showed useful discrimination for both outcomes. The full 143-feature model achieved an AUC of 0.874 for nonfatal suicidal events and 0.787 for fatal suicidal events within 180 days after emergency department discharge. By comparison, a demographic-only model achieved AUCs of 0.705 for nonfatal events and 0.680 for fatal events.

The study examined suicidal events occurring within 180 days after discharge from an emergency department visit for a mental health disorder. Among 872,627 emergency department episodes representing 511,559 individuals, there were 4,525 nonfatal suicidal events and 434 fatal suicidal events during that 180-day follow-up period.

Yes. Adding clinical features improved prediction for both nonfatal and fatal suicidal events beyond demographics alone. When the 54 mental health features from the emergency department visit were added, the model reached AUC=0.838 for nonfatal events and AUC=0.752 for fatal events, representing gains of 18.9% and 10.6%, respectively, over the demographic-only model. The full model that combined all 143 features performed best, with AUCs of 0.874 for nonfatal events and 0.787 for fatal events.

The highest-ranking predictors of nonfatal suicidal events were suicidal ideation at the emergency department visit, age older than 65 years as a negative association, and a suicide-related event in the 180 days before the emergency department episode as a positive association. The authors also found that younger age, suicidal ideation, and Medicaid coverage were more predictive of nonfatal events than of fatal events.

The top predictors of suicide death were male sex, depressive disorders, and white race, all with positive associations in the fatal-event model. The study also found that sleep disorders and opioid-related symptoms ranked considerably higher for predicting fatal events than nonfatal events.

The study found substantial overlap, but not complete overlap, between predictors of nonfatal and fatal suicidal events. Eleven of the top 15 features (73.3%) were the same in both models, and the Shapley-based feature importance rankings were moderately to strongly correlated (r=0.82). However, the relative importance differed for several predictors, including prior suicide-related events, male sex, sleep disorders, and opioid-related symptoms.

Yes, but less accurately than it predicts nonfatal events. A model trained on nonfatal suicidal behavior had an AUC of 0.724 when applied to suicide deaths, even though it was not trained on fatal outcomes. At the Youden cutpoint, it achieved 78.1% sensitivity and 67.3% specificity for fatal outcomes, compared with 76.7% sensitivity and 82.4% specificity for nonfatal outcomes.

The findings support the idea that nonfatal and fatal suicidal events share a substantial risk architecture, while still showing meaningful differences in some predictors. The overlap in top features, the correlation in feature-importance rankings (r=0.82), and the ability of the nonfatal model to predict fatal events with AUC=0.724 were interpreted by the authors as evidence consistent with a continuum between nonfatal and fatal suicidal outcomes.

This study used electronic health record data and National Death Index linkage to develop machine learning models predicting nonfatal and fatal suicidal events within 180 days after emergency department discharge for patients with mental health disorders. The sample began with emergency department visits from October 1, 2015, to September 30, 2022, and the final analytic sample included 872,627 emergency department episodes. The models used gradient tree boosting, a two-thirds training and one-third test split, and predictors drawn from demographics, emergency department diagnoses and note-derived data, prior 180-day clinical history, physical health conditions, and prior treatment features.

The authors noted several limitations that affect interpretation and generalizability. The models may not generalize to other patient populations, health systems, or countries without comparable electronic health record data. The sample included only individuals with documented mental disorders, so people who attempt suicide without a recorded mental health diagnosis may have been missed, and the models were not compared with clinical suicide risk assessments.