An Artificial Intelligence-Based Framework for Predicting Emergency Department Overcrowding: Development and Evaluation Study.

Abstract

Background: Emergency department (ED) overcrowding remains a critical challenge, leading to delays in patient care and increased operational strain. Current hospital management strategies often rely on reactive decision-making, addressing congestion only after it occurs. However, effective patient flow management requires early identification of overcrowding risks to allow timely interventions. Machine learning (ML)-based predictive modeling offers a solution by forecasting key patient flow measures, such as waiting count, enabling proactive resource allocation and improved hospital efficiency.

Objective: The aim of this study is to develop ML models that predict ED waiting room occupancy (waiting count) at 2 temporal resolutions. The first approach is the hourly prediction model, which estimates the waiting count exactly 6 hours ahead at each prediction time (eg, a 1 PM prediction forecasts 7 PM). The second approach is the daily prediction model, which forecasts the average waiting count for the next 24-hour period (eg, a 5 PM prediction estimates the following day's average). These predictive tools support resource allocation and help mitigate overcrowding by enabling proactive interventions before congestion occurs.

Methods: Data from a partner hospital's ED in the southeastern United States were used, integrating internal and external sources. Eleven different ML algorithms, ranging from traditional approaches to deep learning architectures, were systematically trained and evaluated on both hourly and daily predictions to determine the models that achieved the lowest prediction error. Experiments optimized feature combinations, and the best models were tested under high patient volume and across different hours to assess temporal accuracy.

Results: The best hourly prediction performance was achieved by time series vision transformer plus (TSiTPlus) with a mean absolute error (MAE) of 4.19 and a mean squared error (MSE) of 29.36. The overall hourly waiting count had a mean of 18.11 and a SD (σ) of 9.77. Prediction accuracy varied by time of day, with the lowest MAE at 11 PM (2.45) and the highest at 8 PM (5.45). Extreme case analysis at (mean + 1σ), (mean + 2σ), and (mean + 3σ) resulted in MAEs of 6.16, 10.16, and 15.59, respectively. For daily predictions, an explainable convolutional neural network plus (XCMPlus) achieved the best results with an MAE of 2.00 and a MSE of 6.64. The daily waiting count had a mean of 18.11 and a SD of 4.51. Both models outperformed traditional forecasting approaches across multiple evaluation metrics.

Conclusions: The proposed prediction models effectively forecast ED waiting count at both hourly and daily intervals. The results demonstrate the value of integrating diverse data sources and applying advanced modeling techniques to support proactive resource allocation decisions. The implementation of these forecasting tools within hospital management systems has the potential to improve patient flow and reduce overcrowding in emergency care settings. The code is available in our GitHub repository.