MedKGEval: A Knowledge Graph-Based Multi-Turn Evaluation Framework for Open-Ended Patient Interactions with Clinical LLMs

TL;DR

MedKGEval introduces a knowledge-graph–driven, multi-turn evaluation framework for clinical LLMs that simulates doctor–patient dialogues with four specialized agents (Doctor, Patient, Judge, Director) and a MedKG knowledge graph. The framework enables turn-level assessment of clinical correctness, safety, comprehensiveness, and history-taking, addressing limitations of static benchmarks and retrospective analyses. It demonstrates its utility by benchmarking eight state-of-the-art LLMs on two tasks (Medication Consultation and Disease Diagnosis), revealing nuanced strengths and safety concerns that surface only in multi-turn, knowledge-grounded settings. The approach is scalable, language-agnostic via input KG switching, and extensible to additional medical tasks and RL-based improvements, offering a practical path toward more reliable and safety-conscious clinical LLMs.

Abstract

The reliable evaluation of large language models (LLMs) in medical applications remains an open challenge, particularly in capturing the complexity of multi-turn doctor-patient interactions that unfold in real clinical environments. Existing evaluation methods typically rely on post hoc review of full conversation transcripts, thereby neglecting the dynamic, context-sensitive nature of medical dialogues and the evolving informational needs of patients. In this work, we present MedKGEval, a novel multi-turn evaluation framework for clinical LLMs grounded in structured medical knowledge. Our approach introduces three key contributions: (1) a knowledge graph-driven patient simulation mechanism, where a dedicated control module retrieves relevant medical facts from a curated knowledge graph, thereby endowing the patient agent with human-like and realistic conversational behavior. This knowledge graph is constructed by integrating open-source resources with additional triples extracted from expert-annotated datasets; (2) an in-situ, turn-level evaluation framework, where each model response is assessed by a Judge Agent for clinical appropriateness, factual correctness, and safety as the dialogue progresses using a suite of fine-grained, task-specific metrics; (3) a comprehensive multi-turn benchmark of eight state-of-the-art LLMs, demonstrating MedKGEval's ability to identify subtle behavioral flaws and safety risks that are often overlooked by conventional evaluation pipelines. Although initially designed for Chinese and English medical applications, our framework can be readily extended to additional languages by switching the input knowledge graphs, ensuring seamless bilingual support and domain-specific applicability.

Figures (5)

• Figure 1: An example case of MedKGEval in the disease-diagnosis task.
• Figure 2: Illustration of the curated medical knowledge graph used for evaluation.
• Figure 3: Correctness (a) and History taking scores (b) of eight models in a Chinese disease-diagnosis scenario at different turn counts. Larger circles indicate more cases that ended the dialogue at that turn and received the corresponding score.
• Figure 4: Correctness (a) and History taking scores (b) of eight models in an English disease-diagnosis scenario at different turn counts. Larger circles indicate more cases that ended the dialogue at that turn and received the corresponding score.
• Figure 5: Prompt templates of the Director Agent for disease diagnosis and medication consultation in English.