Large Language Models (LLMs) demonstrate tremendous potential in healthcare, yet current applications predominantly rely on static pre-trained knowledge without mechanisms for continuous optimization through feedback. This paper proposes a reinforcement learning-based multi-source feedback enhancement framework (MS-RL-CoT), innovatively employing LLM as the core policy network of a reinforcement learning agent while designing a mechanism that integrates expert feedback, user feedback, and environment feedback. We introduce Chain-of-Thought (CoT), explicitly incorporating reasoning processes into reinforcement learning optimization objectives, and design an attention-based feedback fusion mechanism to dynamically balance the importance of different feedback sources. Experiments on three medical tasks—interactive diagnosis, treatment planning, and case analysis Q&A—demonstrate that MS-RL-CoT achieves significant improvements over existing methods in accuracy, adherence to medical protocols, and information acquisition efficiency. Notably, diagnostic accuracy improves by 19.2% points over baseline LLMs and 8.4% points over standard RLHF. Results confirm the effectiveness of multi-source feedback and chain-of-thought reasoning in enhancing LLM performance in medical applications.