Can You Spot the Virtual Patient (VP)? Expert Evaluation, Turing Test, Linguistic Analysis, and Semantic Similarity Analysis

Abstract

Communication is a core clinical competency and essential for safe, effective patient care. Virtual patients (VPs) powered by large language models (LLMs) offer a promising alternative to traditional formats such as standardized patients (SPs). However, few studies have quantitatively as well as qualitatively evaluated their realism across diverse clinical scenarios using behavioral studies and linguistic dialog analyses. This study evaluated the realism of GPT-4o-generated VPs through expert review, Turing testing, linguistic analysis, and semantic analysis. We generated 44 VPs using structured illness scripts derived from doctor-simulated patient dialogues across 17 clinical categories. Experts annotated responses for hallucinations, omissions, and repetitions, with high interrater reliability (ICC > 0.77). A Turing test revealed that participants struggled to distinguish VP responses from those of real patients, with average classification accuracy for VP responses falling below chance level. Lexical and syntactic analyses of over 2,000 conversation turns showed that VP responses were generally realistic, more formal and lexically consistent, while human responses displayed greater emotional variation. Semantic similarity using BioClinicalBERT and cosine similarity averaged 0.871 at the response level and 0.842 at the transcript level. Our findings support the integration of LLM-based VPs in communication training while highlighting areas for trust, transparency, and refinement, contributing to both GenAI deployment in healthcare and its safety evaluation.