A 9 + 1 multi-agent collaborative teaching model for clinical reasoning: integrating large language models into medical education

Abstract

Clinical reasoning—the cognitive and metacognitive processes through which clinicians collect and interpret patient information, generate and test diagnostic hypotheses, and make clinical decisions—is a core competency in medical education. Since 1956, clinical educators have innovated methods to build this key skill. Artificial intelligence (AI) technologies, especially large language models (LLMs), have advanced rapidly. For example, OpenAI’s inaugural LLM scored 96.5% on the Medical Question Answering (MedQA) benchmark, rivaling top human performers. Integrating AI into teaching prompts key questions: Can it support students’ clinical reasoning development? This requires careful investigation. Guided by the Analysis, Design, Development, Implementation, and Evaluation (ADDIE) instructional design model and Design-Based Research (DBR) methodology, this study explored approaches for fostering clinical reasoning through LLM integration. We traced the educational evolution of clinical reasoning since 1956 and identified five core competencies: organizing knowledge structures, reasoning and judgment, cognitive regulation, communication and collaboration, and self-development. Building on these findings, we designed a “9 + 1” multi-agent collaborative teaching model covering an end-to-end closed-loop training process from instructional design to reflective feedback. We conducted a preliminary A/B controlled experiment comparing traditional methods (Group A, n = 30) with the 9 + 1 model (Group B, n = 30). Learning outcomes were assessed using a Mini-Objective Structured Clinical Examination (Mini-OSCE) tool. In the “chest pain case” experiment, Group B students showed marked gains in hypothesis completeness, cognitive error spotting, and reflection depth compared to Group A. These differences were statistically significant. Our “9 + 1” model enables collaborative engagement among teachers, students, and AI agents within Team-Based Learning (TBL) scenarios. Preliminary findings suggest it may help students develop knowledge structures and critical thinking in a more structured manner, while allowing teachers to identify challenging learning modules and offer targeted support. This approach shows potential for broader application in clinical education, though further validation with larger samples and diverse settings is needed.