A severity-aware multi-agent soft actor-critic framework for medical resource coordination in plateau disaster rescue

Purpose Autonomous medical rescue on high-altitude plateaus faces formidable challenges, as extreme environmental conditions, including hypobaric hypoxia, severe cold, and inaccessible terrain, critically degrade human physiological resilience and logistical accessibility. The "platinum ten minutes" window becomes even more stringent, demanding rapid triage and precise resource allocation under severe infrastructure deficits. Existing multi-agent reinforcement learning approaches struggle to reconcile discrete tactical maneuvering with the continuous precision required for medical resource allocation, leading to sub-optimal logistics and coarse treatment decisions. Methods To address these limitations, we propose mSAC (injury severity-medical resource collaborative guided multi-agent soft actor-critic), a framework integrating a hybrid discrete-continuous action space with a severity-aware attention mechanism. By leveraging a maximum entropy objective, mSAC incentivizes diverse exploration strategies, mitigating premature convergence to local optima. The hybrid architecture facilitates a "micro-dosing" strategy, enabling agents to continuously modulate drug administration proportional to injury severity rather than relying on fixed dosages. Results Extensive evaluations demonstrate that mSAC achieves a survival rate of 84.2% and a drug efficiency ratio of 0.83, significantly outperforming baselines such as multi-agent deep deterministic policy gradient and independent proximal policy optimization. Qualitative analysis further reveals that agents balance operational agility with safety, adopting a high-speed rescue policy that trades off marginal collision risk for superior mission completion time and logistical sustainability. Conclusion The proposed mSAC framework effectively bridges the gap between discrete maneuver control and continuous medical resource allocation in extreme high-altitude rescue scenarios. Its severity-aware attention mechanism and hybrid action space enable fine-grained, adaptive treatment decisions, yielding substantial improvements in both casualty survival and drug utilization efficiency over existing multi-agent reinforcement learning methods.