LLM-based virtual patient versus traditional teaching in ophthalmology clerkship: a controlled trial
Region - Educational Research and Reviews · 24.07.2026 · 8 (4) · 32
DOI: 10.32629/rerr.v8i4.5344 ↗Kurz gefasst
Fachartikel in der Zeitschrift Region - Educational Research and Reviews, online erschienen am 24.07.2026. Verfasst von Dingqiao WANG und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Background
Clinical history-taking is central to medical education, yet practice opportunities remain scarce in short specialty rotations such as ophthalmology. No study has directly compared two commercial large language models (LLMs) operating within the same virtual patient framework.
Methods
We conducted a quasi-experimental sequential cohort study (TREND-reported). Forty-five medical students completing a five-day ophthalmology clerkship were allocated to one of three groups (n = 15 each): a control group receiving traditional teaching with real patients, and two intervention groups using a retrieval-augmented generation (RAG)-based virtual patient system powered by GPT-4o or Claude 3.5 Sonnet. Primary outcome was history-taking performance (100-point rubric, ICC = 0.89) on Day 2 and Day 5. Secondary outcomes were self-efficacy (SE-12 scale) and student satisfaction, assessed by paired t-test and Fisher's exact test.
Results
History-taking improved by 17.11 points (GPT-4o, p < 0.001) and 15.17 points (Claude, p = 0.001), versus a non-significant 5.87-point change in controls (p = 0.078). Self-efficacy gains were 25.77 points (GPT-4o, p < 0.001), 21.19 points (Claude, p = 0.001), and 2.39 points (control, p = 0.536). All intervention students rated satisfaction as basically satisfied or above, versus 80% in the control group (Fisher's exact p < 0.001). The two LLM groups did not differ significantly on any measure.
Conclusions
Both RAG-based virtual patient systems (GPT-4o and Claude 3.5 Sonnet) improved history-taking performance and clinical self-efficacy within five days; traditional teaching produced no significant change on either measure. The comparable outcomes across LLMs indicate that system architecture and case design matter more than model selection. These findings support the use of LLM-based virtual patients as a practical supplement to bedside teaching in compressed clerkships, subject to confirmation in randomized trials.
Zitieren
APA
WANG, D., MAO, J., & DU, W. (2026). LLM-based virtual patient versus traditional teaching in ophthalmology clerkship: a controlled trial. Region - Educational Research and Reviews, 8(4), 32. https://doi.org/10.32629/rerr.v8i4.5344
Vancouver
WANG D, MAO J, DU W. LLM-based virtual patient versus traditional teaching in ophthalmology clerkship: a controlled trial. Region - Educational Research and Reviews. 2026;8(4):32. doi:10.32629/rerr.v8i4.5344
BibTeX
@article{WANG2026,
author = {WANG, Dingqiao and MAO, Jun and DU, Wei},
title = {LLM-based virtual patient versus traditional teaching in ophthalmology clerkship: a controlled trial},
journal = {Region - Educational Research and Reviews},
year = {2026},
volume = {8},
number = {4},
pages = {32},
doi = {10.32629/rerr.v8i4.5344}
}
EUVIMED