Accuracy Evaluation of Machine Learning-Based Hospital Resource Demand Forecasting During Infectious Disease Surges: A Comparative Analysis

The COVID-19 pandemic exposed critical vulnerabilities in hospital resource allocation during infectious disease surges, necessitating accurate demand forecasting capabilities. This study evaluates machine learning-based prediction algorithms for hospital resource demand (e.g., ICU occupancy) through comparative analysis. We assess time series methods, ensemble learning techniques, and deep learning architectures using historical utilization data from multiple healthcare facilities. Performance metrics, including MAE, RMSE, and MAPE, were computed for short-term and medium-term prediction horizons. Results demonstrate that ensemble approaches achieve higher accuracy than traditional methods. Across 7-21-day horizons, the ensemble model (XGBoost + Random Forest + LSTM) achieved the lowest prediction errors, with a 7-day MAPE of 7.64% and sustained advantages over ARIMA/SARIMA baselines. These findings provide evidence-based guidelines for healthcare coordinators aligned with AHRQ emergency preparedness priorities.