MedRoute: RL-Based Dynamic Specialist Routing in Multi-Agent Medical Diagnosis

A research team from the University of Central Florida and other institutions has introduced MedRoute, a novel framework designed to revolutionize AI-assisted medical diagnosis. The system addresses a key limitation of current Large Multimodal Models (LMMs), which often act as overly generalists. Instead, MedRoute creates a collaborative team of AI agents, each modeled as a medical specialist (e.g., a radiologist, dermatologist). At its core is a reinforcement learning (RL)-trained router that dynamically decides which specialist agent is best suited for a given case, much like a real-world General Practitioner making a referral.

This dynamic routing is a significant advancement over previous multi-agent approaches that relied on static or predefined specialist selection, which couldn't adapt to complex, real-world scenarios. The framework's workflow includes the RL router for initial triage, the selected specialist agents for domain-specific analysis, and a final Moderator agent that synthesizes all inputs to produce a consolidated diagnosis. In extensive evaluations on standard medical datasets, MedRoute demonstrated superior diagnostic accuracy compared to existing state-of-the-art baselines, proving the effectiveness of its dynamic, team-based approach.

The researchers have made MedRoute's code and models publicly available, encouraging further development in this critical area. By closely emulating the collaborative, specialist-driven nature of real clinical practice, the work provides a strong, scalable foundation for building more reliable and trustworthy AI diagnostic tools. This moves the field beyond single, monolithic models toward flexible, multi-expert systems that can handle the vast and varied landscape of medical conditions.