Optimizing Sepsis Treatment Through Reinforcement Learning: A Revisitation of Reward and Loss Functions in Dueling Double Deep Q-Network

With the development of artificial intelligence., an more and more studies are exploring the application of deep reinforcement learning to assist in the clinical treatment of sepsis. Most RL algorithms have shown promising results in formulating treatment decisions on datasets. Previous algorithms based on Dueling Double Deep Q-Network (DDDQN) employed relatively simple reward settings without considering additional patient indicators. Objective: In order to obtain better treatment strategies. Methods: In this study., we enhanced the DDDQN algorithm by introducing a novel reward function based on clinical treatment recommendations and selecting a more optimal loss function. We evaluated our algorithm using state-of-the-art value function methods in expected return., survival rate., and action distribution. Results: The results indicate that DDDQN., with the introduction of the new reward function., achieved higher survival rates in treatment decisions on the MIMIC-III dataset. Conclusion: Reaching up to 93.4%., a notable improvement of approximately 6.3% compared to clinical strategies. Significance: Our approach provides a more effective treatment strategy.