Improved understanding of factors affecting prolonged emergency department (ED) length of stay is crucial to improving patient outcomes. Our investigation builds on prior work by considering ED length of stay in operationally distinct time periods and using benchmark and novel machine learning techniques applied only to data that would be available to ED operators in real time.

This study was a retrospective review of patient visits over 1 year at 2 urban EDs, including 1 academic and 1 academically affiliated ED, and 2 suburban, community EDs. ED length of stay was partitioned into 3 components: arrival-to-room, room-to-disposition, and admit disposition to departure. Prolonged length of stay for each component was considered beyond 1, 3, and 2 hours, respectively. Classification models (logistic regression, random forest, and XGBoost) were applied, and important features were evaluated.

In total, 135,044 unique patient encounters were evaluated for the arrival-to-room, room-to-disposition, and admit disposition-to-departure models, which had accuracy ranges of 84% to 96%, 66% to 77%, and 62% to 72%, respectively. Waiting room and ED volumes were important features in all arrival-to-room models. Room-to-disposition results identified patient characteristics and ED volume as the most important features for prediction. Boarder volume was an important feature of the admit disposition-to-departure models for all sites. Academic site models noted nurse staffing ratios as important, whereas community site models noted hospital capacity and surgical volume as important for admit disposition-to-departure prediction.

This study identified granular capacity, flow, and nurse staffing predictors of ED length of stay not previously reported in the literature. Our novel methodology allowed for more accurate and operationally meaningful findings compared to prior modeling methods.