Large language models approach expert-level clinical knowledge and reasoning in ophthalmology: A head-to-head cross-sectional study
Open accessThe DOI matches its Crossref record; 14 of the 14 references it cites were found in a registry or an index.
Abstract
Large language models (LLMs) underlie remarkable recent advanced in natural language processing, and they are beginning to be applied in clinical contexts. We aimed to evaluate the clinical potential of state-of-the-art LLMs in ophthalmology using a more robust benchmark than raw examination scores. We trialled GPT-3.5 and GPT-4 on 347 ophthalmology questions before GPT-3.5, GPT-4, PaLM 2, LLaMA, expert ophthalmologists, and doctors in training were trialled on a mock examination of 87 questions. Performance was analysed with respect to question subject and type (first order recall and higher order reasoning). Masked ophthalmologists graded the accuracy, relevance, and overall preference of GPT-3.5 and GPT-4 responses to the same questions. The performance of GPT-4 (69%) was superior to GPT-3.5 (48%), LLaMA (32%), and PaLM 2 (56%). GPT-4 compared favourably with expert ophthalmologists (median 76%, range 64–90%), ophthalmology trainees (median 59%, range 57–63%), and unspecialised junior doctors (median 43%, range 41–44%). Low agreement between LLMs and doctors reflected idiosyncratic differences in knowledge and reasoning with overall consistency across subjects and types ( p >0.05). All ophthalmologists preferred GPT-4 responses over GPT-3.5 and rated the accuracy and relevance of GPT-4 as higher ( p <0.05). LLMs are approaching expert-level knowledge and reasoning skills in ophthalmology. In view of the comparable or superior performance to trainee-grade ophthalmologists and unspecialised junior doctors, state-of-the-art LLMs such as GPT-4 may provide useful medical advice and assistance where access to expert ophthalmologists is limited. Clinical benchmarks provide useful assays of LLM capabilities in healthcare before clinical trials can be designed and conducted.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Kung, T. H., Cheatham, M., Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., Maningo, J., et al. (2023) · PLOS Digital Health · doi:10.1371/journal.pdig.0000198 · Cited by 4,115 · cited 2×
“Our null hypothesis was that LLMs and doctors would exhibit similar performance, supported by results in a wide range of medical examinations [3, 6].”
BackgroundResultKung, T. H., Cheatham, M., ChatGPT, Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., et al. (2022) · doi:10.1101/2022.12.19.22283643 · Cited by 4,115
Large Language Models Encode Clinical Knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2212.13138 · Cited by 4,083
Large language models encode clinical knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2023) · Nature · doi:10.1038/s41586-023-06291-2 · Cited by 4,083 · cited 1×
“…of LLM chatbots have suggested that LLM responses may be equivalent or even superior to human doctors in terms of accuracy and empathy, and experiments using complicated case studies suggest that LLMs operate well even outside typical presentations and more common medical conditions [4,25,26].”
BackgroundAnalyzing and Interpreting Data From Likert-Type Scales
Sullivan, G. M., & Artino, A. R. (2013) · Journal of Graduate Medical Education · doi:10.4300/jgme-5-4-18 · Cited by 3,044
Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., et al. (2023) · JAMA Internal Medicine · doi:10.1001/jamainternmed.2023.1838 · Cited by 2,613
Gilson, A., Safranek, C. W., Huang, T., Socrates, V., Chi, L., Taylor, R. A., & Chartash, D. (2023) · JMIR Medical Education · doi:10.2196/45312 · Cited by 2,078
The imperative for regulatory oversight of large language models (or generative AI) in healthcare
Meskó, B., & Topol, E. J. (2023) · npj Digital Medicine · doi:10.1038/s41746-023-00873-0 · Cited by 1,047
Bloom’s taxonomy of cognitive learning objectives
Adams, N. E. (2015) · Journal of the Medical Library Association : JMLA · doi:10.3163/1536-5050.103.3.010 · Cited by 1,030
Multimodal machine learning in precision health: A scoping review
Kline, A., Wang, H., Li, Y., Dennis, S., Hutch, M., Xu, Z., Wang, F., Cheng, F., & Luo, Y. (2022) · npj Digital Medicine · doi:10.1038/s41746-022-00712-8 · Cited by 528
Capabilities of GPT-4 on Medical Challenge Problems
Nori, H., King, N., McKinney, S. M., Carignan, D., & Horvitz, E. (2023) · arXiv · doi:10.48550/arxiv.2303.13375 · Cited by 509
Thirunavukarasu, A. J., Hassan, R., Mahmood, S., Sanghera, R., Barzangi, K., El Mukashfi, M., & Shah, S. (2023) · JMIR Medical Education · doi:10.2196/46599 · Cited by 187
Self-citationPythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Biderman, S., Schoelkopf, H., Anthony, Q., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. (2023) · arXiv · doi:10.48550/arxiv.2304.01373 · Cited by 160 · cited 1×
“As LLM performance often correlates with the frequency of query terms’ representation in the model’s training dataset, a simple indicator of ‘familiarity’ could be engineered by calculating the relative frequency of query term representation in the training data [7,33].”
BackgroundLarge language models will not replace healthcare professionals: curbing popular fears and hype.
Thirunavukarasu, A. J. (2023) · Journal of the Royal Society of Medicine · doi:10.1177/01410768231173123 · Cited by 63
Self-citation
Cited by
Generative artificial intelligence in medicine
Teo, Z. L., Thirunavukarasu, A. J., Elangovan, K., Cheng, H., Moova, P., Soetikno, B., Nielsen, C., Pollreisz, A., Ting, D. S. J., Morris, R. J. T., et al. (2025) · Nature Medicine · doi:10.1038/s41591-025-03983-2 · Cited by 119
Self-citationPrompt Engineering in Clinical Practice: Tutorial for Clinicians
Liu, J., Liu, F., Wang, C., & Liu, S. (2025) · Journal of Medical Internet Research · doi:10.2196/72644 · Cited by 76
Ethics of artificial intelligence in medicine
Savulescu, J., Giubilini, A., Vandersluis, R., & Mishra, A. (2024) · Singapore Medical Journal · doi:10.4103/singaporemedj.smj-2023-279 · Cited by 68
Wang, L., Li, J., Zhuang, B., Huang, S., Fang, M., Wang, C., Li, W., Zhang, M., & Gong, S. (2025) · Journal of Medical Internet Research · doi:10.2196/64486 · Cited by 51
Sezgin, E., Jackson, D. I., Kocaballi, A. B., Bibart, M., Zupanec, S., Landier, W., Audino, A., Ranalli, M., & Skeens, M. (2025) · Cancer Medicine · doi:10.1002/cam4.70554 · Cited by 30
Metadata
- Authors
- Thirunavukarasu, A. J., Mahmood, S., Malem, A., Foster, W. P., Sanghera, R., Hassan, R., Zhou, S., Wong, S. W., Wong, Y. L., Chong, Y. J., Shakeel, A., Chang, Y.-H., Tan, B. K. J., Jain, N., Tan, T. F., Rauz, S., Ting, D. S. W., & Ting, D. S. J.
- Venue
- PLOS Digital Health
- Publisher
- Public Library of Science (PLoS)
- Published
- 2024-04-17
- Volume
- 3
- Issue
- 4
- Pages
- e0000341
- ISSN
- 2767-3170 (ISSN-L)
- OpenAlex
- W4394892516
- Cited by
- 88
- References
- 14
- FWCI
- 6.97
- Type
- journal article
- Language
- en
- Open access
- gold
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
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.