A machine-learning arrival-risk model for second-look risk stratification among intermediate- and lower-acuity emergency department patients

Introduction Emergency department triage must identify patients at risk for deterioration using limited arrival information. We evaluated whether a gradient-boosted arrival-risk model could provide second-look risk stratification among patients assigned intermediate or lower triage acuity. Methods This retrospective cohort study used MIMIC-IV-ED linked with hospital and intensive care unit data. Adult emergency department stays with non-missing triage acuity were included. The primary outcome was intensive care unit admission within 24 hours of arrival or death during the index hospitalization. Predictors included demographics, arrival transport, triage acuity, vital signs, pain score, and term frequency–inverse document frequency features from chief complaint text. Performance was evaluated using a patient-level training and held-out split. Results The cohort included 418,025 stays from 201,214 patients, with 25,205 primary outcome events. The held-out set included 83,502 stays and 5,020 events. Overall, receiver operating characteristic area under the curve was 0.907 (95% confidence interval, 0.903–0.912) and precision-recall area under the curve was 0.515 (0.500–0.530). Among 50,989 acuity 3–5 stays with 617 events, corresponding values were 0.833 (0.818–0.849) and 0.074 (0.062–0.091). The highest-risk 10% included 5,099 stays and captured 302 events, with 48.9% sensitivity, 5.9% positive predictive value, and a number needed to alert of 16.9. Results were similar after excluding race/ethnicity and in an additional deidentified-timestamp-based patient-level partition. Conclusions The model identified substantial residual risk heterogeneity among intermediate- and lower-acuity emergency department patients. Prospective external validation, workflow evaluation, and fairness assessment are required before implementation.