A Novel Framework for Evaluating the Clinical Reasoning Process of Large Language Models: A Comparative Study in Nephrology

Although interest in the application of large language models (LLMs) in medicine is growing, accuracy evaluations have largely relied on static knowledge tests. However, discussions on clinical reasoning, the process most critical to real-world practice, remain limited. In this study, we propose a novel framework to evaluate not the final diagnosis generated by AI, but the reasoning process itself. This study proposes a novel framework that systematically evaluates the capabilities of LLMs (OpenAI GPT-o3, Gemini 2.5 Pro, DeepSeek-R1, Llama4-Marveric) by deconstructing the clinical reasoning process into discrete cognitive steps. We focused on nephrology cases, which often involve multiple organ systems and diverse pathologies, thus requiring a high level of reasoning. The four nephrologists independently evaluated the outputs. Our evaluation of four leading LLMs revealed that while Gemini 2.5 Pro demonstrated the best overall performance, all models exhibited common weaknesses in advanced, synthetic tasks such as "formulating differential diagnoses with rationale" and "treatment planning," particularly in dynamically changing clinical scenarios. Furthermore, a notable finding of our research is that the highest-performing model was not the most computationally intensive, demonstrating that reasoning quality and computational efficiency are not in a simple trade-off. In conclusion, our step-by-step evaluation method is an effective approach for identifying the specific strengths and weaknesses in an LLM's clinical reasoning. The weaknesses identified, particularly in formulating a differential diagnosis with a clear rationale and developing comprehensive treatment plans for dynamic scenarios, should become a primary target for future model development and for the creation of support system.