MedCollab: Causal-Driven Multi-Agent Collaboration for Full-Cycle Clinical Diagnosis via IBIS-Structured Argumentation

A research team led by Yuqi Zhan has introduced MedCollab, a novel multi-agent AI framework designed to tackle the persistent challenges of diagnostic hallucinations and opaque reasoning in medical LLMs. The system fundamentally rethinks how AI approaches clinical diagnosis by emulating the hierarchical consultation workflow of modern hospitals, where different specialist agents collaborate throughout the full diagnostic cycle. Unlike single-model approaches, MedCollab dynamically assembles teams of clinical and examination agents based on patient-specific symptoms and test results, creating a more realistic and adaptive diagnostic process. This architecture represents a significant shift from treating AI as a monolithic oracle to viewing it as a collaborative team of experts.

The technical innovation lies in its structured reasoning protocols. MedCollab implements an Issue-Based Information System (IBIS) argumentation framework that forces each agent to declare a "Position" supported by traceable evidence from medical knowledge bases and clinical data, making the AI's reasoning auditable. Furthermore, it constructs Hierarchical Disease Causal Chains that transform flat predictions into structured models of pathological progression using explicit logical operators. A multi-round consensus mechanism with logic auditing and weighted voting iteratively filters out low-quality reasoning. Evaluated on real-world clinical datasets, MedCollab demonstrated superior performance in Accuracy and RaTEScore metrics compared to both pure LLMs and existing medical multi-agent systems, showing a marked reduction in dangerous medical hallucinations. This suggests a path toward more extensible, transparent, and clinically compliant AI decision support tools.