Identifying exudative recurrence in neovascular age-related macular degeneration using large language models

• ChatGPT-4o detects OCT changes from standard Heyex™ scrolling video • AI shows strong potential for nAMD recurrence screening • Segmentation + LLM may improve specificity and workflow To evaluate the ability of a large language model (LLM) to identify exudative recurrence in neovascular age-related macular degeneration (nAMD) from optical coherence tomography (OCT) scans and compare its performance with retina specialists. Retrospective, single-center, observational study. Two consecutive OCT scans of patients with nAMD under pro re nata (PRN) regimen were collected. Screen recordings of the complete OCT scan comparisons (previous vs. current) were analyzed by two retina specialists versus ChatGPT-4o after providing a standardized prompt. Main outcome measure was agreement between the LLM and the ophthalmologists in detecting exudative activity. Statistical measures included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Cohen’s kappa for agreement. A total of 102 eyes were included. Among 71 confirmed recurrences, ChatGPT-4o correctly identified 63 and missed 8; of 31 non-recurrences, it correctly identified 10 and misclassified 21 as positive. Overall accuracy was 71.57%, with sensitivity of 88.73% and specificity of 32.26%. Agreement with ophthalmologists was fair (Cohen’s k = 0.238). ChatGPT-4o demonstrated strong sensitivity in detecting exudative recurrences in nAMD, but limited specificity. Given the small, single-center cohort, the lack of external or same-dataset validation, these results should be considered preliminary, but highlights as LLM-assisted analysis, while not yet capable of replacing clinical expertise, may serve as a valuable adjunct for screening exudative changes in nAMD. Further refinements in the LLM could improve specificity and clinical utility.