ARVO 2026: Integrating LLMs for community-based triage of red eye diseases

Kaden Bunch, a fourth-year medical student at the Warren Alpert Medical School of Brown University, initiated a project aimed at improving the diagnosis and management of external red eye diseases by leveraging machine learning and large language models. The idea for this project emerged during his third-year ophthalmology rotation, where he noticed a significant volume of consults for red eye conditions such as allergic conjunctivitis, viral conjunctivitis, uveitis, and styes. Bunch observed that many of these consults could potentially be managed by primary care providers, but due to limited exposure to ophthalmic diseases, these providers often referred cases to specialists out of caution.

Recognizing the visual nature of ophthalmology and the subtle but distinct differences between various red eye conditions, Bunch utilized his background in data science to address this inefficiency. He sourced a comprehensive dataset containing images of external red eye diseases and trained several convolutional neural network (CNN) models, including MobileNet, ResNet, and EfficientNet, to classify these conditions. The models achieved a high accuracy rate of approximately 96% in distinguishing between different types of red eye diseases.

To further enhance the utility of his system, Bunch integrated the image classification models with large language models. This integration was designed to generate actionable, evidence-based recommendations for primary care and emergency medicine providers, rather than providing definitive diagnoses, in order to mitigate the risk of AI hallucinations. The system outputs include likely causes, common symptoms, red flag signs, and clear guidance on when to consult an ophthalmologist. Additionally, the system offers visual explanations by highlighting the image features that the model used to make its assessment, such as peri-limbal injection for uveitis or localized lesions for styes.

Overall, Bunch's project demonstrates that combining machine learning with large language models can not only achieve high diagnostic accuracy for external eye diseases but also empower non-specialist providers with practical recommendations. This approach aims to improve patient care, streamline consult workflows, and reduce unnecessary specialist referrals by supporting primary care providers in managing common ophthalmic conditions more confidently.