Multi-Agent Reinforcement Learning for Emergency Medical Supply Request Regulation in Healthcare Supply Chains

Public health emergencies place acute pressure on medical supply systems. This study develops Med-iDR-MARL, a hierarchical multi-agent reinforcement learning (MARL) framework for regulating hospital requests before allocation. Regional incentives influence hospital responses for deferrable (DD) and reducible (RD) demand, while a local scheduler retains non-deferrable (ND) demand and applies hospital-level service bounds. Across 20 simulated 30-day windows, the framework reduced DD/RD requests by 37.71% with mean incentive expenditure of CNY 274,844.23, compared with 31.03% and CNY 350,038.43 for Rule-Adaptive. Independent hospital-agent deep Q-network learning achieved greater reduction, lower expenditure, and better pooled reduction per unit expenditure. The proposed framework also had the highest peak-to-average ratio despite lower absolute peaks than No-DR. In a separate five-seed ablation, removing the minimum-service guard increased the hospital-day violation rate from zero to 58.95%. Local scheduler diagnostics found feasible solutions in 132,000 tested states but material optimality gaps and no consistent runtime advantage over vectorized exact enumeration. The results identify conditional efficiency and service trade-offs. The Shanghai COVID-19 series supplies an external risk trajectory; hospital behavior, clinical outcomes, and operational logistics remain unvalidated.