Federated learning for privacy-preserving ophthalmic artificial intelligence: clinical applications and translational challenges.

Federated learning (FL) is increasingly relevant to ophthalmology because retinal photographs, optical coherence tomography (OCT), OCT angiography, visual fields, and linked clinical records are clinically valuable but difficult to pool across institutions. In this narrative review, we synthesize ophthalmology-focused FL literature across diabetic retinopathy (DR), glaucoma, age-related macular degeneration (AMD), pediatric retinal disease, multi-disease retinal diagnostics, and emerging ophthalmic platforms. Current evidence suggests that FL can support collaborative AI development without centralizing raw patient data, and selected studies show performance close to centralized training under controlled retrospective or multicenter experimental conditions. For example, multicenter glaucoma detection from volumetric OCT achieved an AUC of 0.92 with FL compared with 0.94 for centralized training. However, FL is privacy-enhancing rather than privacy-complete, and most ophthalmic FL systems have not yet undergone prospective clinical validation. Model updates may remain vulnerable to gradient inversion, membership inference, poisoning, site-level bias, and latent identity or attribute leakage. For eye-care networks, the main value of FL is therefore not simply algorithmic performance but a governance model for privacy-conscious collaboration. Prospective validation, interoperability, explainability, workflow integration, privacy auditing, and clear responsibility for monitoring are needed before FL-enabled ophthalmic AI can be deployed routinely.