Key Takeaways
Extended Takeaways
- In this cohort of 872,627 ED episodes, suicidal events within 180 days were uncommon but clinically significant, with 4,525 nonfatal and 434 fatal events, underscoring the challenge of identifying a rare outcome while making real-time disposition decisions.
- Demographics alone provided only moderate discrimination, with AUCs of 0.705 for nonfatal events and 0.680 for fatal events; incorporating ED mental health features increased performance to 0.838 and 0.752, showing the added value of diagnosis- and note-derived clinical information available during the visit.
- The full 143-feature model performed best for both outcomes, reaching AUC=0.874 for nonfatal events and 0.787 for fatal events, so the broadest risk stratification appears to come from combining current-visit findings with 180-day historical mental health, physical health, and treatment data.
- A recent suicide-related event in the prior 180 days ranked among the top 3 predictors for nonfatal events but only 63rd for fatal events, suggesting that recent suicidal care history may be especially useful for anticipating repeat nonfatal behavior rather than suicide death.
- Sleep disorders and opioid-related symptoms were considerably more predictive of fatal events than nonfatal ones, indicating that ED evaluations may miss important lethality signals if they focus primarily on overt suicidal presentation without probing sleep disruption and opioid-related clinical features.
- A model trained only on nonfatal suicidal behavior still identified suicide death with AUC=0.724; at the Youden cutpoint it achieved 78.1% sensitivity and 67.3% specificity for fatal outcomes, which suggests nonfatal-event models may support death-risk screening when fatal-event data are limited.
