How to Use a Nonfatal Suicide Model for Death Risk
How can clinicians use a model trained on nonfatal suicidal events to help identify suicide death risk after an ED mental health visit?
Suicide death is rare and harder to model directly, yet clinicians still need a practical way to identify high-risk patients after ED discharge. This guide applies when a health system has a risk model trained on nonfatal suicidal events and needs to understand how that model may still inform fatal-risk screening over the next 180 days.
-
Apply the nonfatal-event model to the post-ED population
Use the existing model that estimates 180-day risk of nonfatal suicidal behavior in patients discharged after an ED visit for a mental health disorder. In this study, the nonfatal model itself had strong discrimination for nonfatal outcomes with an AUC of 0.874.
-
Use the same predicted probabilities to screen for fatal outcomes
Interpret the model's predicted probabilities as a proxy screen for suicide death risk rather than as a death-specific estimate. When applied to fatal outcomes in this study, the nonfatal model achieved moderate discrimination with an AUC of 0.724 despite not being trained on fatalities.
-
Select the operating threshold with the Youden cutpoint
Use the Youden index to identify the probability cutpoint that maximizes sensitivity and specificity if a single threshold is needed. At the Youden cutpoint in this study, the nonfatal model achieved 78.1% sensitivity and 67.3% specificity for fatal outcomes, compared with 76.7% sensitivity and 82.4% specificity for nonfatal outcomes.
-
Recognize where the proxy model is most and least informative
Use the overlap between fatal and nonfatal risk architecture to justify the approach, but expect lower accuracy for suicide death than for nonfatal events. The correlation between feature-importance rankings across the nonfatal and fatal models was 0.82, yet some predictors diverged meaningfully, especially prior suicide-related events, male sex, sleep disorders, and opioid-related symptoms.
Clinical Considerations
- A model trained on nonfatal events predicts suicide death less accurately than it predicts nonfatal outcomes.
- Suicide death may require models tailored specifically to lethality and potentially additional contextual or social risk factors.
- This approach was evaluated only in patients with documented mental health disorders and EHR data sufficient for modeling.
Bottom Line
If a death-specific model is unavailable, a nonfatal suicidal-event model can still be used to screen for suicide death risk after ED discharge, but clinicians should expect only moderate discrimination and should give extra attention to lethality-associated features.
