Machine-learning identified suicide risk and emergency department inpatient admission

Accurate identification of patients at high risk for suicide following emergency department (ED) visits remains a critical clinical challenge. Although machine learning models using electronic health record (EHR) data can predict suicide risk, it remains unclear how predictions align with ED disposition decisions.

We conducted a retrospective cohort study using deidentified EHR data from the Optum Labs Data Warehouse linked to the National Death Index. The cohort included 249,310 ED encounters for mental health disorders among adults ≥18 years (2015-2022). A validated gradient boosting model estimated 180-day risk of suicide death and nonfatal attempts. The primary outcome was inpatient admission (psychiatric or medical) at the index ED visit. Agreement with algorithmically identified high-risk visits (top 15.8%) was assessed using Cohen's kappa, and patient characteristics of the groups were compared using standardized differences.

Inpatient admission occurred in 15.8% (n = 39,311) of visits. Agreement between admission and high-risk classification (top 15.8%) was low for suicide death (κ = 0.12) and nonfatal attempts (κ = 0.10). Among high-risk patients, 25.8% (fatal) and 24.3% (nonfatal) were admitted. Admitted patients were more likely to be aged ≥65 years and female, and less likely to be aged 18-34 years, Medicaid-insured, or diagnosed with suicidal ideation, substance use, or bipolar disorder.

ED inpatient admission decisions demonstrated limited concordance with machine learning-predicted suicide risk. Integrating predictive models into ED workflows may enhance identification of patients at elevated longer-term suicide risk and support more targeted care.