Advances in Large Language Models and the Case for GPT-4o in Virtual Patient Simulation for Health Professions Education

Abstract

Large language models (LLMs) have advanced rapidly, and the release of GPT-4o introduced a natively multimodal model capable of processing and generating text, audio, and images in near real time. These capabilities make LLMs newly practical for virtual patient simulation, a long-standing method in health professions education. This article argues that GPT-4o is a strong foundation for simulated patients by connecting three lines of evidence: the demonstrated medical knowledge of contemporary LLMs, the established educational value of simulation and standardized patients, and emerging studies of LLM-powered simulated patients. Conversational realism, the degree to which a simulated encounter reproduces the unscripted, responsive dialogue of a real clinical interview, is central to the educational value of these experiences, and it is precisely where earlier scripted virtual patients fell short and where generative models excel. Early studies show that GPT-based simulated patients produce largely plausible responses and are well received by students, and that pairing simulation with automated feedback improves clinical decision making. Realizing this potential requires attention to fabrication, bias, and the preservation of human interaction.