Human-Likeness is Multi-Dimensional: Evaluating LLM Generated Virtual Patients Against Simulated Patient Dialogue

Abstract

Large language models (LLMs) are increasingly used to generate virtual patients (VPs) for medical training. As these systems move toward broader deployment in medical education, reliable evaluation is needed to identify potential failure modes and ensure that generated patient responses remain consistent with human simulated patient (SP) behavior. Inaccurate, unrealistic, or inconsistent responses may introduce educational risks by exposing learners to misleading patient interactions. However, current evaluation approaches rely heavily on human judgment, making large-scale evaluation difficult. In this work, we present a controlled generation pipeline for systematically comparing VP responses generated by 10 LLMs across multiple decoding temperatures with human respiratory SP responses. To support scalable evaluation, we introduce a multi-dimensional framework for measuring similarity between LLM-generated and human SP dialogue across semantic, conversational grounding, linguistic, and emotional dimensions. Across semantic, conversational, linguistic, and emotional dimensions, no single model consistently produced the most human-like responses. From a practical perspective, Gemini 2.5 Flash demonstrated the most consistently human-like performance across semantic, linguistic, grammatical, and emotional dimensions, while Claude Sonnet 4.6 achieved the strongest semantic fidelity. These findings highlight the importance of multi-dimensional evaluation and demonstrate how automated metrics can provide a scalable approach for assessing virtual patient systems before broader deployment in medical education.