Clinical Summary: Comparing Risk Prediction for Suicide Attempts and Deaths After Emergency Department Visits for Individuals With Mental Health Disorders
Emergency clinicians must make disposition decisions for patients with mental health disorders even though suicide risk after discharge is high and many patients who later die by suicide are not identified during the ED visit. This study clarifies which factors better predict nonfatal versus fatal suicidal events within 180 days and whether a model trained on nonfatal events can still identify patients at risk for suicide death.
Key Findings
- Among 872,627 ED episodes, there were 4,525 nonfatal and 434 fatal suicidal events in the 180 days post-ED discharge.
- The full model with all 143 features achieved the highest performance, with AUC=0.874 for nonfatal events and 0.787 for fatal events, compared with 0.705 and 0.680, respectively, for the demographic-only model.
- Adding the 54 mental health features from the ED visit represented an 18.9% gain over the demographic-only model and achieved an AUC of 0.838 for nonfatal events; and a 10.6% gain and an AUC of 0.752 for fatal events.
- The model predicting nonfatal events (AUC=0.874) showed moderate discrimination when applied to the fatal event outcome (AUC=0.724), with r = 0.82 and R2 = 0.683.
- At the Youden cutpoint, the nonfatal model achieved 76.7% sensitivity and 82.4% specificity for nonfatal outcomes and 78.1% sensitivity and 67.3% specificity for fatal outcomes.
For ED patients with mental health disorders, broad EHR-based models predict nonfatal suicidal events better than suicide deaths, but the overlap is strong enough that a nonfatal-event model still identifies many patients at risk for fatal outcomes. Fatal-risk assessment should give added weight to male sex, depressive disorders, white race, sleep disorders, and opioid-related symptoms rather than relying only on recent suicidal presentation.
Practice Implications
- Use more than demographics when assessing post-ED suicide risk: demographic-only discrimination was 0.705 for nonfatal events and 0.680 for fatal events, while the full 143-feature model improved performance to AUC=0.874 and 0.787.
- Do not assume the strongest predictors of repeat nonfatal behavior are the strongest predictors of suicide death: a suicide-related event in the past 180 days was one of the top 3 predictors for nonfatal events but ranked 63rd for fatal events, while male sex ranked first for fatal events but 48th for nonfatal events.
- In ED evaluations, actively assess sleep disorders and opioid-related symptoms, which were considerably more predictive of fatal events than nonfatal ones and may signal greater lethality risk.
- When fatal-event training data are limited, a model trained on nonfatal suicidal behavior can still support death-risk screening, with AUC=0.724, 78.1% sensitivity, and 67.3% specificity for fatal outcomes.
