Barriers to artificial intelligence applications in ophthalmology: beyond diagnostic accuracy toward patient...

Artificial intelligence (AI) has become a major area of ophthalmic innovation, with a rapidly expanding literature, yet routine clinical adoption remains limited. Diagnostic accuracy alone is an insufficient marker of clinical value because even technically strong systems may fail to improve referral, follow-up, treatment, clinician workload, or health-system value. This focused narrative review examines the barriers that separate ophthalmic AI performance from real-world implementation. We reviewed ophthalmic AI literature indexed in PubMed, Web of Science, and Google Scholar, with emphasis on studies published from January 2018 to September 2026 and on prospective, randomized, real-world, economic, governance, and patient-centred evidence. We organize translation across four linked domains: algorithmic validity, clinical utility, workflow integration, and health-system value, with ethical governance and trust as cross-cutting requirements. Diabetic retinopathy screening now provides a concrete translational example spanning large-scale retrospective validation, prospective deployment, workflow implementation, and economic evaluation. We also consider how foundation models and large language models alter evidence requirements because outputs can depend on prompts, retrieved context, tools, and model versions. For glaucoma, retinal disease management, retinopathy of prematurity, anterior segment applications, and automated reporting, evidence remains less mature. The central challenge for ophthalmic AI is therefore no longer model development alone, but disciplined clinical translation, including prospective evaluation, operational workflow measurement, equity assessment, lifecycle monitoring, transparent governance, and trustworthy implementation.