Large Language Model-Based Virtual Patient Systems for History-Taking in Medical Education: Comprehensive Systematic Review.

BACKGROUND

Large language models (LLMs), such as GPT-3.5 and GPT-4 (OpenAI), have been transforming virtual patient systems in medical education by providing scalable and cost-effective alternatives to standardized patients. However, systematic evaluations of their performance, particularly for multimorbidity scenarios involving multiple coexisting diseases, are still limited.

OBJECTIVE

This systematic review aimed to evaluate LLM-based virtual patient systems for medical history-taking, addressing four research questions: (1) simulated patient types and disease scope, (2) performance-enhancing techniques, (3) experimental designs and evaluation metrics, and (4) dataset characteristics and availability.

METHODS

Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020, 9 databases were searched (January 1, 2020, to August 18, 2025). Nontransformer LLMs and non-history-taking tasks were excluded. Multidimensional quality and bias assessments were conducted.

RESULTS

A total of 39 studies were included, screened by one computer science researcher under supervision. LLM-based virtual patient systems mainly simulated internal medicine and mental health disorders, with many addressing distinct single disease types but few covering multimorbidity or rare conditions. Techniques like role-based prompts, few-shot learning, multiagent frameworks, knowledge graph (KG) integration (top-k accuracy 16.02%), and fine-tuning enhanced dialogue and diagnostic accuracy. Multimodal inputs (eg, speech and imaging) improved immersion and realism. Evaluations, typically involving 10-50 students and 3-10 experts, demonstrated strong performance (top-k accuracy: 0.45-0.98, hallucination rate: 0.31%-5%, System Usability Scale [SUS] ≥80). However, small samples, inconsistent metrics, and limited controls restricted generalizability. Common datasets such as MIMIC-III (Medical Information Mart for Intensive Care-III) exhibited intensive care unit (ICU) bias and lacked diversity, affecting reproducibility and external validity.

CONCLUSIONS

Included studies showed moderate risk of bias, inconsistent metrics, small cohorts, and limited dataset transparency. LLM-based virtual patient systems excel in simulating multiple disease types but lack multimorbidity patient representation. KGs improve top-k accuracy and support structured disease representation and reasoning. Future research should prioritize hybrid KG-chain-of-thought architectures integrated with open-source KGs (eg, UMLS [Unified Medical Language System] and SNOMED-CT [Systematized Nomenclature of Medicine - Clinical Terms]), parameter-efficient fine-tuning, dialogue compression, multimodal LLMs, standardized metrics, larger cohorts, and open-access multimodal datasets to further enhance realism, diagnostic accuracy, fairness, and educational utility.