Uncovering Language Disparity of ChatGPT on Retinal Vascular Disease Classification: Cross-Sectional Study
Open accessPublished version of a preprint
Abstract
Background: Benefiting from rich knowledge and the exceptional ability to understand text, large language models like ChatGPT have shown great potential in English clinical environments. However, the performance of ChatGPT in non-English clinical settings, as well as its reasoning, have not been explored in depth. Objective: This study aimed to evaluate ChatGPT’s diagnostic performance and inference abilities for retinal vascular diseases in a non-English clinical environment. Methods: In this cross-sectional study, we collected 1226 fundus fluorescein angiography reports and corresponding diagnoses written in Chinese and tested ChatGPT with 4 prompting strategies (direct diagnosis or diagnosis with a step-by-step reasoning process and in Chinese or English). Results: Compared with ChatGPT using Chinese prompts for direct diagnosis that achieved an F1-score of 70.47%, ChatGPT using English prompts for direct diagnosis achieved the best diagnostic performance (80.05%), which was inferior to ophthalmologists (89.35%) but close to ophthalmologist interns (82.69%). As for its inference abilities, although ChatGPT can derive a reasoning process with a low error rate (0.4 per report) for both Chinese and English prompts, ophthalmologists identified that the latter brought more reasoning steps with less incompleteness (44.31%), misinformation (1.96%), and hallucinations (0.59%) (all P<.001). Also, analysis of the robustness of ChatGPT with different language prompts indicated significant differences in the recall (P=.03) and F1-score (P=.04) between Chinese and English prompts. In short, when prompted in English, ChatGPT exhibited enhanced diagnostic and inference capabilities for retinal vascular disease classification based on Chinese fundus fluorescein angiography reports. Conclusions: ChatGPT can serve as a helpful medical assistant to provide diagnosis in non-English clinical environments, but there are still performance gaps, language disparities, and errors compared to professionals, which demonstrate the potential limitations and the need to continually explore more robust large language models in ophthalmology practice.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
A primary cause of low explicit probability is insufficient knowledge, which can drive LLMs to produce hallucinated information for predictions.
“A primary cause of low explicit probability is insufficient knowledge, which can drive LLMs to produce hallucinated information for predictions40,41.”
From Probabilistic medical predictions of large language models (2024) · cites this paper as [41]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 34 · Cited by 52
Research works printed but not linked: 3
- [23]Wei J Wang X Schuurmans D Bosma M Ichter B Xia F Chi EH Le QV Zhou D Chain-of-thought prompting elicits reasoning in large language models 2022 36th Conference on Neural Information Processing Systems (NeurIPS 2022) November 28-December 9, 2022 New Orleans, Louisiana
- [28]Ji Z Lee N Frieske R Yu T Su D Xu Y Ishii E Bang Y Dai W Madotto A Fung P Survey of hallucination in natural language generation arXiv 2022 2 8 2023-08-17 https://arxiv.org/abs/2202.03629 · arXiv
- [32]Wu J Wu X Qiu Z Li M Zheng Y Yang J Qualifying chinese medical licensing examination with knowledge enhanced generative pre-training model arXiv 2023 5 17 2023-08-17 http://arxiv.org/abs/2305.10163 · arXiv
- Authors
- Liu, X., Wu, J., Shao, A., Shen, W., Ye, P., Wang, Y., Ye, J., Jin, K., & Yang, J.
- Affiliations
- Second Affiliated Hospital of Zhejiang University (CN)
- Zhejiang University (CN)
- Corresponding author
- Jin, Kai
- Venue
- Journal of Medical Internet Research
- Publisher
- JMIR Publications Inc.
- Published
- 2024-01-22
- Volume
- 26
- Pages
- e51926
- ISSN
- 1438-8871 (ISSN-L)
- PMC
- PMC10845019
- OpenAlex
- W4389182159
- Cited by
- 52
- References
- 34
- Citations by year
- 2026: 9 · 2025: 16 · 2024: 27
- FWCI
- 2.07
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- Journal of Medical Internet Research (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- National Natural Science Foundation of China (82201195)
- Medical Scientific Research Foundation of Zhejiang Province, China
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Retinal Imaging and Analysis
- Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
- Keywords
- ChatGPT, large language models, language disparities, prompt language, hallucinations, diagnostic performance, ophthalmologist comparison
- MeSH terms
- Artificial Intelligence, Cross-Sectional Studies, Diagnostic Errors, Fluorescein Angiography, Humans, Language, Retinal Diseases, Vascular Diseases
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
