Applications of Fairness Training in Artificial Intelligence to Ophthalmic Disease Diagnosis using External Eye Photographs

Diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma are leading causes of preventable blindness, yet whether external eye photographs carry phenotypic information about posterior segment disease remains unclear, and methods to mitigate demographic disparities in such models are largely unexplored. We developed and evaluated deep learning and vision-language models for binary classification of vision-threatening DR (VTDR), advanced AMD, and glaucoma using 40,687 external eye photographs from 9,878 patients at Massachusetts Eye and Ear. Convolutional networks (WResNet, EfficientNet), a vision transformer (ViT), and a fine-tuned vision-language model (LLaVA) were assessed by area under the receiver operating characteristic curve (AUC), with fairness across sex, race, and ethnicity quantified by the equity-scaled AUC (ES-AUC). Five fairness-training strategies were applied to the best models. LLaVA achieved the highest baseline AUC for VTDR (85.4%, comparable to ViT 82.6%); advanced AMD was similar across architectures (EfficientNet 86.3%; LLaVA 84.8%; ViT 84.7%), whereas glaucoma was substantially worse (67-74%). Fairness training improved VTDR and AMD AUCs to 86.3% and 85.9% while partially mitigating demographic disparities. In summary, although AI models that rely on external eye photography to detect posterior segment disease may not yet be suitable for clinical deployment, the fairness methods demonstrated in this proof-of-concept study could be applied to other AI systems to decrease performance disparities and potentially improve model performance.