Multi-Agent LLM Reasoning for Clinical Procedure Sequencing from High-Granularity EHR Data

Real-world clinical decisions emerge from collaboration, yet most medical AI functions as a single expert. To better reflect clinical practice, we introduce a multi-agent framework where large language model (LLM) agents, each embodying a distinct specialty (e.g., cardiology, endocrinology), collaboratively recommend a sequence of patient procedures. Each agent reasons using a private knowledge base of domain-specific PubMed literature, a shared memory containing the patient's electronic health records (EHRs), and the history of the ongoing inter-agent discussion. We simulate two organizational models: one requiring unanimous consensus and another where a designated team leader makes the final call. Using the MIMIC-III dataset to predict procedural sequences, we show that the leader-based model consistently outperforms both the consensus and single-agent configurations, achieving higher accuracy on metrics including Mean Reciprocal Rank (MRR). Our work presents a robust and interpretable paradigm for multi-agent clinical decision support, more closely aligning AI with the collaborative nature of clinical practice.