Foundation models are a novel type of artificial intelligence algorithms, in which models are pretrained at scale on unannotated data and fine-tuned for a myriad of downstream tasks, such as generating text. This study assessed the accuracy of ChatGPT, a large language model (LLM), in the ophthalmology question-answering space. Evaluation of diagnostic test or technology. ChatGPT is a publicly available LLM. We tested 2 versions of ChatGPT (January 9 "legacy" and ChatGPT Plus) on 2 popular multiple choice question banks commonly used to prepare for the high-stakes Ophthalmic Knowledge Assessment Program (OKAP) examination. We generated two 260-question simulated exams from the Basic and Clinical Science Course (BCSC) Self-Assessment Program and the OphthoQuestions online question bank. We carried out logistic regression to determine the effect of the examination section, cognitive level, and difficulty index on answer accuracy. We also performed a post hoc analysis using Tukey's test to decide if there were meaningful differences between the tested subspecialties. We reported the accuracy of ChatGPT for each examination section in percentage correct by comparing ChatGPT's outputs with the answer key provided by the question banks. We presented logistic regression results with a likelihood ratio (LR) chi-square. We considered differences between examination sections statistically significant at a P value of < 0.05. The legacy model achieved 55.8% accuracy on the BCSC set and 42.7% on the OphthoQuestions set. With ChatGPT Plus, accuracy increased to 59.4% ± 0.6% and 49.2% ± 1.0%, respectively. Accuracy improved with easier questions when controlling for the examination section and cognitive level. Logistic regression analysis of the legacy model showed that the examination section (LR, 27.57; P = 0.006) followed by question difficulty (LR, 24.05; P < 0.001) were most predictive of ChatGPT's answer accuracy. Although the legacy model performed best in general medicine and worst in neuro-ophthalmology (P < 0.001) and ocular pathology (P = 0.029), similar post hoc findings were not seen with ChatGPT Plus, suggesting more consistent results across examination sections. ChatGPT has encouraging performance on a simulated OKAP examination. Specializing LLMs through domain-specific pretraining may be necessary to improve their performance in ophthalmic subspecialties. Proprietary or commercial disclosure may be found after the references.

Background Evidence on the performance of Generative Pre-trained Transformer 4 (GPT-4), a large language model (LLM), in the ophthalmology question-answering domain is needed. Methods We tested GPT-4 on two 260-question multiple choice question sets from the Basic and Clinical Science Course (BCSC) Self-Assessment Program and the OphthoQuestions question banks. We compared the accuracy of GPT-4 models with varying temperatures (creativity setting) and evaluated their responses in a subset of questions. We also compared the best-performing GPT-4 model to GPT-3.5 and to historical human performance. Results GPT-4–0.3 (GPT-4 with a temperature of 0.3) achieved the highest accuracy among GPT-4 models, with 75.8% on the BCSC set and 70.0% on the OphthoQuestions set. The combined accuracy was 72.9%, which represents an 18.3% raw improvement in accuracy compared with GPT-3.5 (p<0.001). Human graders preferred responses from models with a temperature higher than 0 (more creative). Exam section, question difficulty and cognitive level were all predictive of GPT-4-0.3 answer accuracy. GPT-4-0.3’s performance was numerically superior to human performance on the BCSC (75.8% vs 73.3%) and OphthoQuestions (70.0% vs 63.0%), but the difference was not statistically significant (p=0.55 and p=0.09). Conclusion GPT-4, an LLM trained on non-ophthalmology-specific data, performs significantly better than its predecessor on simulated ophthalmology board-style exams. Remarkably, its performance tended to be superior to historical human performance, but that difference was not statistically significant in our study.

Background/aims: This study assesses the proficiency of Generative Pre-trained Transformer (GPT)-4 in answering questions about complex clinical ophthalmology cases.Methods: We tested GPT-4 on 422Journal of the American Medical AssociationOphthalmology Clinical Challenges, and prompted the model to determine the diagnosis (open-ended question) and identify the next-step (multiple-choice question). We generated responses using two zero-shot prompting strategies, including zero-shot plan-and-solve+ (PS+), to improve the reasoning of the model. We compared the best-performing model to human graders in a benchmarking effort.Results: Using PS+ prompting, GPT-4 achieved mean accuracies of 48.0% (95% CI (43.1% to 52.9%)) and 63.0% (95% CI (58.2% to 67.6%)) in diagnosis and next step, respectively. Next-step accuracy did not significantly differ by subspecialty (p=0.44). However, diagnostic accuracy in pathology and tumours was significantly higher than in uveitis (p=0.027). When the diagnosis was accurate, 75.2% (95% CI (68.6% to 80.9%)) of the next steps were correct. Conversely, when the diagnosis was incorrect, 50.2% (95% CI (43.8% to 56.6%)) of the next steps were accurate. The next step was three times more likely to be accurate when the initial diagnosis was correct (p<0.001). No significant differences were observed in diagnostic accuracy and decision-making between board-certified ophthalmologists and GPT-4. Among trainees, senior residents outperformed GPT-4 in diagnostic accuracy (p≤0.001 and 0.049) and in accuracy of next step (p=0.002 and 0.020).Conclusion: Improved prompting enhances GPT-4’s performance in complex clinical situations, although it does not surpass ophthalmology trainees in our context. Specialised large language models hold promise for future assistance in medical decision-making and diagnosis.

Aims Automated machine learning (AutoML) is a novel tool in artificial intelligence (AI). This study assessed the discriminative performance of AutoML in differentiating retinal vein occlusion (RVO), retinitis pigmentosa (RP) and retinal detachment (RD) from normal fundi using ultra-widefield (UWF) pseudocolour fundus images. Methods Two ophthalmologists without coding experience carried out AutoML model design using a publicly available image data set (2137 labelled images). The data set was reviewed for low-quality and mislabeled images and then uploaded to the Google Cloud AutoML Vision platform for training and testing. We designed multiple binary models to differentiate RVO, RP and RD from normal fundi and compared them to bespoke models obtained from the literature. We then devised a multiclass model to detect RVO, RP and RD. Saliency maps were generated to assess the interpretability of the model. Results The AutoML models demonstrated high diagnostic properties in the binary classification tasks that were generally comparable to bespoke deep-learning models (area under the precision-recall curve (AUPRC) 0.921–1, sensitivity 84.91%–89.77%, specificity 78.72%–100%). The multiclass AutoML model had an AUPRC of 0.876, a sensitivity of 77.93% and a positive predictive value of 82.59%. The per-label sensitivity and specificity, respectively, were normal fundi (91.49%, 86.75%), RVO (83.02%, 92.50%), RP (72.00%, 100%) and RD (79.55%,96.80%). Conclusion AutoML models created by ophthalmologists without coding experience can detect RVO, RP and RD in UWF images with very good diagnostic accuracy. The performance was comparable to bespoke deep-learning models derived by AI experts for RVO and RP but not for RD.

We tested the accuracy of ChatGPT, a large language model (LLM), in the ophthalmology question-answering space using two popular multiple choice question banks used for the high-stakes Ophthalmic Knowledge Assessment Program (OKAP) exam. The testing sets were of easy-to-moderate difficulty and were diversified, including recall, interpretation, practical and clinical decision-making problems. ChatGPT achieved 55.8% and 42.7% accuracy in the two 260-question simulated exams. Its performance varied across subspecialties, with the best results in general medicine and the worst in neuro-ophthalmology and ophthalmic pathology and intraocular tumors. These results are encouraging but suggest that specialising LLMs through domain-specific pre-training may be necessary to improve their performance in ophthalmic subspecialties.

Foundation models represent a paradigm shift in artificial intelligence (AI), evolving from narrow models designed for specific tasks to versatile, generalisable models adaptable to a myriad of diverse applications. Ophthalmology as a specialty has the potential to act as an exemplar for other medical specialties, offering a blueprint for integrating foundation models broadly into clinical practice. This review hopes to serve as a roadmap for eyecare professionals seeking to better understand foundation models, while equipping readers with the tools to explore the use of foundation models in their own research and practice. We begin by outlining the key concepts and technological advances which have enabled the development of these models, providing an overview of novel training approaches and modern AI architectures. Next, we summarise existing literature on the topic of foundation models in ophthalmology, encompassing progress in vision foundation models, large language models and large multimodal models. Finally, we outline major challenges relating to privacy, bias and clinical validation, and propose key steps forward to maximise the benefit of this powerful technology.