AbstractObjective To develop a machine learning algorithm to predict emergency department waiting room surge status (Green/Yellow/Red categories) 4, 8 and 12 hours in advance. Patients and Methods We designed a retrospective cohort with prospective validation conducted between July 1, 2023 and April 30, 2025. A total of 154,956 encounters were analyzed, features included ED operational metrics and timestamps with 72-hour lagging data. Deep neural network and gradient boosted decision tree (XGBOOST) models were trained to predict three pre-defined surge categories: Green (= 31 waiting patients; 12.3% of the hours). Results The XGBOOST model demonstrated strong predictive performance across all forecast horizons. AUC curve showed excellent discrimination between Green and Yellow levels ranging from 0.87 to 0.91 AUC across all time horizons, demonstrating reliable differentiation between normal and moderate surge conditions. The model showed acceptable discrimination for Red Levels with AUC of 0.76 and 0.77, meeting commonly accepted thresholds for clinical forecasting tools, particularly given the difficulty of predicting rare, high-volume situations. The operational accuracy of 68% to 70% demonstrated strong real-world multi-class operational forecasting over prolonged windows up to 12 hours in advance. Conclusion Using operational metrics with timestamps, XGBOOST models can differentiate between different levels of emergency department surge states with meaningful accuracy. This ability to forecast risk of high volumes provides a window of opportunity to proactively modify operations.
Jones et al. (Wed,) studied this question.