Thinking Like a Doctor: Conversational Diagnosis through the Exploration of Diagnostic Knowledge Graphs

TL;DR

This work tackles the problem of conversational diagnosis under incomplete patient information by grounding reasoning in a diagnostic knowledge graph (KG). It introduces a two-step framework: Hypothesis Generation (HG) to propose a focused set of candidate diseases, and Hypothesis Verification (HV) to verify those hypotheses through clarifying questions, guided by a 3-hop subgraph expansion from anchor diseases. A realistic patient simulator, augmented to produce low-specificity symptom descriptions, and synthetic dialogue data generated with GPT-4o-mini train the HG and HV, enabling end-to-end performance gains that surpass parametric and retrieval-based baselines. Experimental results show improved Recall@k and fewer dialogue turns, with physician evaluations supporting the realism of the simulator and the clinical value of the generated questions, suggesting practical utility for automated triage and preliminary diagnosis in fast-paced clinical settings. The approach demonstrates the benefit of explicit diagnostic grounding for reliable, interpretable conversational diagnosis and provides a foundation for expanding disease coverage and multi-modal integration in the future.

Abstract

Conversational diagnosis requires multi-turn history-taking, where an agent asks clarifying questions to refine differential diagnoses under incomplete information. Existing approaches often rely on the parametric knowledge of a model or assume that patients provide rich and concrete information, which is unrealistic. To address these limitations, we propose a conversational diagnosis system that explores a diagnostic knowledge graph to reason in two steps: (i) generating diagnostic hypotheses from the dialogue context, and (ii) verifying hypotheses through clarifying questions, which are repeated until a final diagnosis is reached. Since evaluating the system requires a realistic patient simulator that responds to the system's questions, we adopt a well-established simulator along with patient profiles from MIMIC-IV. We further adapt it to describe symptoms vaguely to reflect real-world patients during early clinical encounters. Experiments show improved diagnostic accuracy and efficiency over strong baselines, and evaluations by physicians support the realism of our simulator and the clinical utility of the generated questions. Our code will be released upon publication.

Figures (4)

• Figure 1: Our system (i) generates diagnostic hypotheses from the dialogue context (§\ref{['subsubsec:HG']}), and (ii) verifies hypotheses through clarifying questions, which are repeated until the final diagnosis (§\ref{['subsubsec:HV']}).
• Figure 2: Comparison of disease hypothesis generation performance (Recall@$k$) across different methods.
• Figure 3: Robustness analysis under diverse patient personas (Recall@4).
• Figure 4: Schema of the diagnostic knowledge graph.