MEDCO: Medical Education Copilots Based on A Multi-Agent Framework

TL;DR

MEDCO introduces a multi-agent copilot for medical education that simulates a realistic training environment with an agentic patient, a radiologist, and a medical expert to train a student through interactive diagnosis, feedback, and memory-enhanced learning. The framework demonstrates substantial performance gains for agentic students and human-like learning behaviors, across single- and multi-modal (image+text) scenarios, using a hierarchical ICD-10 based evaluation. It contributes a novel agent-based pedagogical paradigm, a memory-augmented learning approach, and a multi-modal assessment protocol that can inform AI-enabled medical education and potentially other domains. The work highlights the practical potential of coordinating multiple specialized LLM roles to emulate real-world clinical training and collaboration, with clear directions for future expansion and human-user validation.

Abstract

Large language models (LLMs) have had a significant impact on diverse research domains, including medicine and healthcare. However, the potential of LLMs as copilots in medical education remains underexplored. Current AI-assisted educational tools are limited by their solitary learning approach and inability to simulate the multi-disciplinary and interactive nature of actual medical training. To address these limitations, we propose MEDCO (Medical EDucation COpilots), a novel multi-agent-based copilot system specially developed to emulate real-world medical training environments. MEDCO incorporates three primary agents: an agentic patient, an expert doctor, and a radiologist, facilitating a multi-modal and interactive learning environment. Our framework emphasizes the learning of proficient question-asking skills, multi-disciplinary collaboration, and peer discussions between students. Our experiments show that simulated virtual students who underwent training with MEDCO not only achieved substantial performance enhancements comparable to those of advanced models, but also demonstrated human-like learning behaviors and improvements, coupled with an increase in the number of learning samples. This work contributes to medical education by introducing a copilot that implements an interactive and collaborative learning approach. It also provides valuable insights into the effectiveness of AI-integrated training paradigms.

Figures (11)

• Figure 1: The illustration of the MEDCO framework, which consists of three steps: 1) Initiating different roles and tools; 2) The medical expert evaluates the student's diagnosis and provides feedback, and the student digests and stores the feedback in their learning memory; 3) The student applies the knowledge within memory to improve diagnosis in future practicing scenarios.
• Figure 2: The pipeline of our MEDCO framework from learning to practicing scenario.
• Figure 3: The learning curve of an agentic student using rethinking strategies (knowledge and suggestion) or peer discussion at various retrieval ranges (as percentages) in the practicing scenario, where the performance of Claude3.5-Sonnet and 2 Agents serve as reference benchmarks.
• Figure 4: The influence of multi-modalities on the performance of the agentic student when initialized by different LLMs, where the red dashed line denotes the results of 2 Agentsfan2024ai, serving as the reference.
• Figure 5: The examples to show our collected visual data: (a) The Chest CT image from patient 1170; (b) The neck ultrasound report photo (Chinese text) from patient 1310.