JMIR Publications

ABSTRACT

Background:

Recently, large language models (LLMs) have experienced rapid development, and virtual patients based on LLM (LLM-VPs) have garnered significant interest from researchers.

Objective:

This study aims to systematically analyze the current applications, research trends, and challenges of LLM-VPs in medical education, while exploring potential future directions.

Methods:

This study adhered to the PRISMA-ScR guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) for a scoping review. A comprehensive search was conducted across five databases (Web of Science, PubMed, IEEE Xplore, Embase, and Scopus) to identify studies related to the application of LLM-VPs in medical education.

Results:

This scoping review includes 22 studies. The analysis reveals that most of the literature has emerged in the past year, indicating that research on LLM-VPs is still in its early stages but is highly innovative. The research primarily focuses on medical training, with the ChatGPT series being the most commonly used model. Integration of other technologies and tools, such as social robots and mixed reality (MR), during the use of LLMs can enhance the realism of LLM-VPs and improve user interaction experiences. The studies span a broad range of medical fields, with the most common training scenario being medical history taking. The interaction process typically occurs in the first-person perspective, using either text or speech, with text being the dominant mode. Evaluations of LLM-VPs predominantly assess user experience; however, the methods for evaluation lack standardization, with only 5(23.8%) studies objectively measuring the learning outcomes facilitated by LLM-VPs. All the studies included in this review express a positive attitude towards the application of LLM-VPs.

Conclusions:

The application of LLM-VPs in medical training is still in its early stages, although there has been a growing interest in the recent two years. While LLM-VPs show great potential for communication skills training, they cannot replace real-world interactions. Furthermore, challenges such as heterogeneous research designs, lack of non-verbal cues in interactions, and privacy and security concerns hinder their broader implementation. Future research should focus on enhancing the reliability, authenticity, safety, and scientific validity of LLM-VPs.