Evaluating the Performance of ChatGPT on Board-Style Examination Questions in Ophthalmology: A Meta-Analysis

To review empirical research on ChatGPT's accuracy in answering ophthalmology board-style examination questions up to March 2025 and to analyze the effects of GPT versions, question types, language differences, and ophthalmology topics on accuracy. A search was conducted in PubMed, Web of Science, Embase, Scopus, and the Cochrane Library in March 2025. Two authors extracted data and independently assessed study quality. Accuracy rates were calculated with Stata 17.0. GPT-4 had an integrated accuracy of 73%, higher than GPT-3.5's 54%. It scored 77% in text and 55% in image tasks. GPT-4's accuracy was 73% in English-speaking countries and 71% in non-English ones. In ophthalmology, General Medicine achieved the highest accuracy (80%), while Clinical Optics had the lowest performance (55%). GPT-4 outperforms GPT-3.5, but its image processing capability needs further validation. Performance varies by language and topic, suggesting the need for more research on cross-linguistic efficacy and error analysis.