Predicting Self-Harm Events After Emergency Department Visits Associated With Mental or Substance Use Disorder Diagnoses.

OBJECTIVE: Using a machine learning model, the authors aimed to predict self-harm events within 90 days of an emergency department (ED) visit associated with a mental or substance use disorder diagnosis. METHODS: Electronic health record and claims data of patients ages ≥11 years from seven integrated health systems were used to identify ED visits and diagnoses. Data on 1,254 predictors were extracted for the 60 months before an ED visit, and self-harm was measured in the 90 days after this visit. Logistic regression models were used to estimate self-harm events, with performance estimated by using out-of-sample cross-validation predictions. RESULTS: In a sample of 1,554,828 ED visits for 803,413 patients, self-harm occurred for 19,603 (1.3%) visits. The best-fitting model yielded an area under the curve of 0.85, indicating good overall performance that was consistent across various subgroups. Patients with ED visits estimated to be at the 95th percentile in predicted risk for self-harm had a 10.8% rate of self-harm events, compared with 2.4% for visits estimated at the 50th percentile. Patients with visits estimated to be in the highest risk stratum had an observed risk for self-harm of 27.2%, approximately 136 times greater than the observed risk in the lowest risk stratum. CONCLUSIONS: The model performed well in predicting self-harm events in the 90 days after an ED visit for individuals with behavioral health diagnoses. Such machine learning models could be useful in identifying ED patients who may need safety planning and close follow-up after discharge.