Human-Governed Multi-Agent Clinical Reasoning Framework

Abstract

Large language models (LLMs) show promise in medical reasoning, but their direct deployment in clinical settings faces challenges related to opacity, output instability, and lack of explicit uncertainty control. We present a Human-Governed Multi-Agent Clinical Reasoning Framework that combines LLM-based diagnosis and question generation with deterministic probability updates implemented in Python. The framework separates diagnosis generation, question selection, and belief updating into distinct, auditable components. The diagnosis agent generates candidate diagnoses, while the deterministic reasoning layer initializes probabilities uniformly and updates them based on structured evidence. A Python prototype demonstrates the framework for chest-pain diagnosis. Results show that combining LLM reasoning with deterministic probabilistic updates in a human-governed architecture provides a practical approach for transparent, safe, and clinically interpretable decision support.