A Novel Approach to Modeling Likelihood of Emergency Department to Hospital Admission

Early identification of emergency department (ED) patients requiring admission could improve ED throughput and mitigate crowding. We used a novel prospective modeling approach to evaluate practical utility of the Epic ED Likelihood to Occupy an Inpatient Bed predictive model.

All adult ED visits (age >18 years) at an academic quaternary-care system from July 2023 – June 2025 were included. The primary outcome was inpatient admission. Consistent with prior studies, we first evaluated performance with a retrospective event horizon approach over a range of score thresholds at 1, 2, 4 and 8 hours. We also evaluated encounter-level performance with the maximum score prior to the outcome. We then used logistic regression to prospectively evaluate score performance. The area under the receiver operator curve (AUROC) was computed for each 1-hour interval from arrival to evaluate how the discriminative ability of the total score changed over time.

Of the 128,077 included encounters, there were 36,134 (28.2%) admissions. In the retrospective analyses, the AUROC was highest for the 1-hour event horizon 0.757 (95% confidence interval [CI] 0.755, 0.760) declining to 0.700 (95% CI 0.700, 0.701) at 8 hours. The encounter-level AUROC was 0.804 (95% CI 0.801-0.806). In the prospective analyses, the AUROC was highest at 1-2 hours 0.778 (95% CI 0.777-0.780) but declined to 0.635 (95% CI 0.630-0.639) at 8-9 hours. The AUROC was 0.811 (95% CI 0.809-0.814) for time to event 0-4 hours, but declined for later events, 0.736 (0.734-0.739) for 4-8 hours, and 0.633 (95% CI 0.630-0.635) for >8 hours.

Our prospective analyses revealed important temporal patterns not evident in the traditional, retrospective approach. The EPIC model has modest capabilities to predict admissions which are best early in the course but performance declines substantially in the encounter and with more distant admissions. The “look forward” approach more accurately reflects clinical decision-making and has implications for real-world deployment strategies.