AI-based hourly prediction of emergency department demand integrating social determinants of health

Emergency department (ED) crowding is associated with delayed treatment and increased mortality, and proactive staffing requires reliable short-term demand forecasts. Most operational models rely on temporal or weather patterns but may insufficiently capture underlying population vulnerability. We evaluated whether community-level social determinants of health (SDoH) improve the prediction of near-term ED demand compared with weather factors using XGBOOST, and developed an interpretable model to inform resource allocation. The study included 213,138 encounters aggregated into 22,883 hourly observations at a tertiary teaching hospital in Wuhan, China. Neighbourhoods differed markedly in age structure and diagnostic profiles, indicating heterogeneous healthcare needs across the catchment area. Models incorporating dynamic SDoH features consistently outperformed those using weather variables alone. Overall, the composite model achieved the best performance with an RMSE of 3.48 visits per hour, an MAE of 2.62 visits per hour, and an R ² of 0.68. Mean absolute prediction errors were approximately 2 to 3 patients per hour across scenarios, suggesting practical utility for short-term staffing and capacity planning. The study indicates that community-level social vulnerability may provide more informative signals than weather conditions for anticipating short-term ED demand. Integrating SDoH-informed forecasts into hospital operations may enable earlier staffing adjustments and help mitigate crowding-related risks.