<![CDATA[Emergency departments face increasing pressure from unpredictable patient surges that can intensify overcrowding, treatment delays, staff workload, and resource shortages. This study aimed to develop and evaluate machine learning algorithms for predicting emergency demand across multiple hospitals and forecast horizons. A quantitative multi-site predictive modeling design combined retrospective records from 1,427,860 emergency encounters with prospective validation data from 151,632 visits. Seasonal ARIMA, Poisson regression, support vector regression, random forest, long short-term memory, extreme gradient boosting, and hybrid ensemble models were compared using temporal and external validation. Results showed that the hybrid ensemble achieved the strongest twenty-four-hour performance, with a mean absolute error of 21.84 patients, mean absolute percentage error of 6.58%, and an R-squared value of 0.91. The model also produced high six-hour surge sensitivity, while recent arrivals, occupancy, ambulance activity, infectious disease indicators, bed availability, weather, air quality, and public events emerged as influential predictors. Predictive accuracy declined for longer horizons and less frequent clinical categories. The study concludes that machine learning can strengthen emergency preparedness when forecasts are explainable, continuously validated, and connected to predefined staffing, triage, bed-management, and ambulance-response protocols. ]]>
Copyrights © 2026
