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.

Background/aims: To compare the performance and cost-effectiveness of DeepSeek-R1 with OpenAI o1 in diagnosing and managing ophthalmology clinical cases. Methods: In this cross-sectional study, a total of 300 clinical cases spanning 10 ophthalmology subspecialties were collected from StatPearls, each with a multiple-choice question on diagnosis or management. DeepSeek-R1 was accessed through its public chat interface, while OpenAI o1 was queried via its Application Programming Interface with a standardised temperature of 0.3. Both models were prompted using plan-and-solve+. Performance was calculated as the proportion of correct answers. McNemar’s test was employed to compare the two models’ performance on paired data. Intermodel agreement for correct diagnoses was evaluated via Cohen’s kappa. Token-based cost analyses were performed to estimate the comparative expenditures of running each model at scale, including input prompts and model-generated output. Results: DeepSeek-R1 and OpenAI o1 achieved an identical overall performance of 82.0% (n=246/300; 95% CI: 77.3 to 85.9). Subspecialty-specific analysis revealed numerical variation in performance, though none of these comparisons reached statistical significance (p>0.05). Agreement in performance between the models was moderate overall (κ=0.503, p<0.001), with substantial agreement in refractive management/intervention (κ=0.698, p<0.001) and moderate agreement in retina/vitreous (κ=0.561, p<0.001) and ocular pathology/oncology (κ=0.495, p<0.01) cases. Cost analysis indicated an approximately 15-fold reduction in per-query, token-related expenses when using DeepSeek-R1 vs OpenAI o1 for the same workload. Conclusions: DeepSeek-R1 shows strong diagnostic and management performance comparable to OpenAI o1 across ophthalmic subspecialties, while significantly reducing costs. These results support its use as a cost-effective, open-weight alternative to proprietary models.

To evaluate the performance of large language models (LLMs), specifically Microsoft Copilot, GPT-4 (GPT-4o and GPT-4o mini), and Google Gemini (Gemini and Gemini Advanced), in answering ophthalmological questions and assessing the impact of prompting techniques on their accuracy. Prospective qualitative study. Microsoft Copilot, GPT-4 (GPT-4o and GPT-4o mini), and Google Gemini (Gemini and Gemini Advanced). A total of 300 ophthalmological questions from StatPearls were tested, covering a range of subspecialties and image-based tasks. Each question was evaluated using 2 prompting techniques: zero-shot forced prompting (prompt 1) and combined role-based and zero-shot plan-and-solve+ prompting (prompt 2). With zero-shot forced prompting, GPT-4o demonstrated significantly superior overall performance, correctly answering 72.3% of questions and outperforming all other models, including Copilot (53.7%), GPT-4o mini (62.0%), Gemini (54.3%), and Gemini Advanced (62.0%) (p < 0.0001). Both Copilot and GPT-4o showed notable improvements with Prompt 2 over Prompt 1, elevating Copilot's accuracy from the lowest (53.7%) to the second highest (72.3%) among the evaluated LLMs. While newer iterations of LLMs, such as GPT-4o and Gemini Advanced, outperformed their less advanced counterparts (GPT-4o mini and Gemini), this study emphasizes the need for caution in clinical applications of these models. The choice of prompting techniques significantly influences performance, highlighting the necessity for further research to refine LLMs capabilities, particularly in visual data interpretation, to ensure their safe integration into medical practice.

To assess the performance of Chat Generative Pre-Trained Transformer-4 in providing accurate diagnoses to retina teaching cases from OCTCases. Cross-sectional study. Retina teaching cases from OCTCases. We prompted a custom chatbot with 69 retina cases containing multimodal ophthalmic images, asking it to provide the most likely diagnosis. In a sensitivity analysis, we inputted increasing amounts of clinical information pertaining to each case until the chatbot achieved a correct diagnosis. We performed multivariable logistic regressions on Stata v17.0 (StataCorp LLC) to investigate associations between the amount of text-based information inputted per prompt and the odds of the chatbot achieving a correct diagnosis, adjusting for the laterality of cases, number of ophthalmic images inputted, and imaging modalities. Our primary outcome was the proportion of cases for which the chatbot was able to provide a correct diagnosis. Our secondary outcome was the chatbot's performance in relation to the amount of text-based information accompanying ophthalmic images. Across 69 retina cases collectively containing 139 ophthalmic images, the chatbot was able to provide a definitive, correct diagnosis for 35 (50.7%) cases. The chatbot needed variable amounts of clinical information to achieve a correct diagnosis, where the entire patient description as presented by OCTCases was required for a majority of correctly diagnosed cases (23 of 35 cases, 65.7%). Relative to when the chatbot was only prompted with a patient's age and sex, the chatbot achieved a higher odds of a correct diagnosis when prompted with an entire patient description (odds ratio = 10.1, 95% confidence interval = 3.3-30.3, P < 0.01). Despite providing an incorrect diagnosis for 34 (49.3%) cases, the chatbot listed the correct diagnosis within its differential diagnosis for 7 (20.6%) of these incorrectly answered cases. This custom chatbot was able to accurately diagnose approximately half of the retina cases requiring multimodal input, albeit relying heavily on text-based contextual information that accompanied ophthalmic images. The diagnostic ability of the chatbot in interpretation of multimodal imaging without text-based information is currently limited. The appropriate use of the chatbot in this setting is of utmost importance, given bioethical concerns. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Objectives: The integration of multimodal capabilities into GPT-4 represents a transformative leap for artificial intelligence in ophthalmology, yet its utility in scenarios requiring advanced reasoning remains underexplored. This study evaluates GPT-4’s multimodal performance on open-ended diagnostic and next-step reasoning tasks in complex ophthalmology cases, comparing it against human expertise. Methods: GPT-4 was assessed across three study arms: (1) text-based case details with figure descriptions, (2) cases with text and accompanying ophthalmic figures, and (3) cases with figures only (no figure descriptions). We compared GPT-4’s diagnostic and next-step accuracy across arms and benchmarked its performance against three board-certified ophthalmologists. Results: GPT-4 achieved 38.4% (95% CI [33.9%, 43.1%]) diagnostic accuracy and 57.8% (95% CI [52.8%, 62.2%]) next-step accuracy when prompted with figures without descriptions. Diagnostic accuracy declined significantly compared to text-only prompts (p = 0.007), though the next-step performance was similar (p = 0.140). Adding figure descriptions restored diagnostic accuracy (49.3%) to near parity with text-only prompts (p = 0.684). Using figures without descriptions, GPT-4’s diagnostic accuracy was comparable to two ophthalmologists (p = 0.30, p = 0.41) but fell short of the highest-performing ophthalmologist (p = 0.0004). For next-step accuracy, GPT-4 was similar to one ophthalmologist (p = 0.22) but underperformed relative to the other two (p = 0.0015, p = 0.0017). Conclusions: GPT-4’s diagnostic performance diminishes when relying solely on ophthalmic images without textual context, highlighting limitations in its current multimodal capabilities. Despite this, GPT-4 demonstrated comparable performance to at least one ophthalmologist on both diagnostic and next-step reasoning tasks, emphasizing its potential as an assistive tool. Future research should refine multimodal prompts and explore iterative or sequential prompting strategies to optimize AI-driven interpretation of complex ophthalmic datasets.

Code-free deep learning (CFDL) allows clinicians with no coding experience to build their own artificial intelligence models. This study assesses the performance of CFDL in glaucoma detection from fundus images in comparison to expert-designed models. Deep learning model development, testing, and validation. A total of 101 442 labeled fundus images from the Rotterdam EyePACS Artificial Intelligence for Robust Glaucoma Screening (AIROGS) dataset were included. Ophthalmology trainees without coding experience designed a CFDL binary model using the Rotterdam EyePACS AIROGS dataset of fundus images (101 442 labeled images) to differentiate glaucoma from normal optic nerves. We compared our results with bespoke models from the literature. We then proceeded to externally validate our model using 2 datasets, the Retinal Fundus Glaucoma Challenge (REFUGE) and the Glaucoma grading from Multi-Modality imAges (GAMMA) at 0.1, 0.3, and 0.5 confidence thresholds. Area under the precision-recall curve (AuPRC), sensitivity at 95% specificity (SE@95SP), accuracy, area under the receiver operating curve (AUC), and positive predictive value (PPV). The CFDL model showed high performance metrics that were comparable to the bespoke deep learning models. Our single-label classification model had an AuPRC of 0.988, an SE@95SP of 95%, and an accuracy of 91% (compared with 85% SE@95SP for the top bespoke models). Using the REFUGE dataset for external validation, our model had an SE@95SP, AUC, PPV, and accuracy of 83%, 0.960%, 73% to 94%, and 95% to 98%, respectively, at the 0.1, 0.3, and 0.5 confidence threshold cutoffs. Using the GAMMA dataset for external validation at the same confidence threshold cutoffs, our model had an SE@95SP, AUC, PPV, and accuracy of 98%, 0.994%, 94% to 96%, and 94% to 97%, respectively. The capacity of CFDL models to perform glaucoma screening using fundus images presents a compelling proof of concept, empowering clinicians to explore innovative model designs for broad glaucoma screening in the near future. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.