MedCouncil: A Debate-Enabled Multi-Agent Framework for Second-Opinion Clinical Decision Support in General Internal Medicine

Abstract

Large Language Models (LLMs) show promise in healthcare, but single-model systems often suffer from hallucinations, lack transparency, and have limited clinical safety. We present MedCouncil, a multi-agent framework designed to serve as a second-opinion tool for general internal medicine. Inspired by multidisciplinary consultations, MedCouncil orchestrates nine specialized clinical agents through a centralized supervisor agent. The system features a structured, multi-round Debate Mode that mimics peer consultations and integrates dynamic toolcalling and persistent memory. We evaluated MedCouncil on 700 cases, comprising 500 synthetic clinical vignettes and 200 retrospectively de-identified patient cases from a tertiary hospital in Bangkok. MedCouncil achieved a Top-1 diagnostic accuracy of 89.4%, reduced the hallucination rate to 3.1%, and demonstrated a 98.9% safety compliance rate, outperforming single-model baselines. A blind user study with 30 practicing internists confirmed high explainability (4.6/5) and clinical utility. By embedding layered safety guardrails, MedCouncil provides a safe, explainable, and regulatory-compliant blueprint for deploying clinical AI in resource-constrained settings.