Enhancing forecasting of emergency department healthcare demand in western australia using hybrid

Abstract

Accurate forecasting of healthcare demand is essential for optimizing resource allocation and planning in Emergency Departments (EDs). This study refines the Relative Utilisation (RU) metric and introduces a novel ensemble forecasting framework that integrates machine learning and deep learning neural network algorithms to improve prediction accuracy in Western Australia (WA). Using population and demand data from the Australian Bureau of Statistics and WA health repositories for 2015–2023, we analyzed 762 monthly time series spanning age group, health region, and urgency disposition group. The refined RU metric excluded each health region’s data from overall state totals, yielding a more precise measure of local demand. To enhance forecast reliability, we aggregated infrequent events, applied a Kalman smoother to adjust for COVID-19–related disruptions, and implemented a hybrid modelling strategy combining machine learning models (XGBoost, LightGBM) with deep learning models. An ensemble method then integrated model outputs using SMAPE (Symmetric Mean Absolute Percentage Error)–based weighting for optimal performance. The ensemble achieved SMAPE errors of approximately 12% for ED presentation forecasts, outperforming individual models and traditional approaches such as Facebook Prophet and ARIMA, which recorded SMAPE errors of 17–18%. These findings demonstrate that a multi-model ensemble, combined with a refined RU metric, substantially enhances the accuracy and reliability of ED demand forecasts. The proposed framework supports more informed resource allocation and strategic planning at the health region level, enabling more efficient healthcare service delivery across WA’s health system.