Game-Theoretic Reinforcement Learning for Zonal Prevention and Early Warning in Psychiatric Hospitals

Psychiatric hospitals faced high-risk and highly uncertain management conditions during public health emergencies. To address these issues, this study proposed a zoned prevention and control early-warning management model based on game-theoretic reinforcement learning. A reinforcement learning mechanism was embedded to improve decision-making under dynamic and uncertain environments. Through continuous state–action–reward feedback, the model iteratively optimized strategies and generated adaptive zonal prevention and early-warning plans. For the experimental design, data were obtained from the COVID-19 Reported Patient Impact and Hospital Capacity dataset. Comparative models included Deep Q-Network (DQN), Proximal Policy Optimization (PPO), Graph Attention Network (GAT), Long Short-Term Memory Autoencoder (LSTM-AE), and Multi- Agent Reinforcement Learning (MARL). The proposed optimized model served as the core framework for performance evaluation. Results showed that the proposed model outperformed all benchmark models in prevention and control effectiveness. In terms of infection rate reduction, the model achieved scores of 0.714, 0.703, and 0.718 on data subsets A, B, and C, respectively, which were substantially higher than GAT’s 0.588 on subset A. Regarding epidemic duration, the optimized model reduced the duration to 16.431 days on subset A, compared with 21.587 days for LSTM-AE. For early warning performance, the model reached a warning precision of 0.912 on subset A, exceeding PPO’s 0.867. It also achieved a faster average response time of 8.973 minutes, compared with 12.384 minutes for DQN. Overall, the proposed model demonstrated superior integrated performance in early warning accuracy, response efficiency, resource utilization, and epidemic control effectiveness.