To effectively perform differential diagnosis, an automated medical diagnostic system must possess the capability to conduct precise history-taking by asking clarifying questions to the patient. While recent Large Language Models in the medical domain...
To effectively perform differential diagnosis, an automated medical diagnostic system must possess the capability to conduct precise history-taking by asking clarifying questions to the patient. While recent Large Language Models in the medical domain exhibit impressive fluency, they face limitations regarding the scope of inquiry and a lack of factual grounding. To address these challenges, this study proposes a knowledge graph-based diagnostic system capable of generating context-appropriate clarifying questions.
First, we constructed a domain-specific knowledge graph based on a diagnostic schema, structured specifically for differential diagnosis rather than for simple biomedical information retrieval. Our proposed framework integrates a disease classifier and a dynamic subgraph extractor to selectively retrieve relevant clusters of symptoms and diseases based on the dialogue context. This enables the model to perform diagnostic reasoning grounded in explicit medical evidence, allowing it to generate appropriate clarifying questions or reach a final diagnosis. Furthermore, to evaluate the system's adaptability to real-world clinical noise, we employed a persona-based patient simulator that mimics diverse patient characteristics, including varying levels of language proficiency and ambiguity. Experimental results demonstrate that our fine-tuned model significantly outperforms the strong baseline, GPT-4.1-mini, in both diagnostic accuracy and conversational efficiency. Notably, by maintaining robust performance even against noisy patient inputs, we validate the effectiveness of structured knowledge grounding in medical dialogue systems.