Deep Spatiotemporal Forecasting and Reinforcement Optimization for Ambulance Allocation

Emergency Medical Services (EMS) require timely and equitable ambulance allocation supported by accurate demand estimation. In our prior work, we developed a statistical forecasting module based on Overall Smoothed Average Demand (OSAD) and Average Maximum (AMX) to estimate proportional EMS demand across spatial zones. Although this approach was interpretable and computationally efficient, it was limited in modeling nonlinear spatiotemporal dependencies and adapting to dynamic demand variations. This paper presents a unified deep learning-based EMS planning framework that integrates spatiotemporal demand forecasting with adaptive ambulance allocation. Specifically, the statistical OSAD/AMX estimators are replaced by graph-based spatiotemporal forecasting models capable of capturing spatial interactions and temporal dynamics. The predicted demand is then incorporated into a reinforcement learning-based allocator that dynamically optimizes ambulance placement under fairness, coverage, and operational constraints. Experiments conducted on real-world EMS datasets demonstrate that the proposed end-to-end framework not only improves demand forecasting accuracy but also translates these improvements into tangible operational benefits, including enhanced equity in resource distribution and reduced response distance. Compared with traditional statistical and heuristic-based baselines, the proposed approach provides a more adaptive and decision-aware solution for EMS planning.