Bold claim: AI-driven eye screening is transforming care, yet clinicians still underutilize it. A recent systematic review and meta-analysis in the American Journal of Ophthalmology examined EyeArt, an autonomous AI system for detecting diabetic retinopathy (DR) from fundus photographs, and evaluated its accuracy across multiple studies. To set the stage: early detection of DR through regular dilated retinal exams lowers the risk of severe vision loss, but less than half of people with diabetes receive recommended screening due to workforce shortages and limited access. Autonomous AI tools like EyeArt, IDx-DR, and AEYE-DS offer point-of-care screening without direct clinician oversight, addressing these barriers.
How AI screening helps patients
The ACCESS randomized trial in young people showed dramatic improvements in screening uptake and follow-up when AI-generated results were provided immediately along with patient education. Additional health-economics research supports cost savings from autonomous AI DR screening, particularly in pediatric populations and primary care settings.
EyeArt in focus
EyeArt is Eyenuk’s flagship AI system. It first gained FDA clearance in 2020 and received expanded clearance in 2023 for use with several retinal cameras (Topcon NW400, Canon CR-2 AF, Canon CR-2 Plus AF). The system analyzes retinal images automatically and delivers a concise report in under 60 seconds, indicating overall eye- and patient-level DR status and whether the case is referable DR (rDR) or vision-threatening DR.
What prior data show
EyeArt has been validated in a pivotal prospective multicenter trial against the ETDRS grading standard and has undergone a large real-world clinical validation with over 100,000 patient visits, representing one of the largest datasets for DR screening technology in typical practice.
New study overview
This systematic review and meta-analysis followed PRISMA-DTA guidelines to assess EyeArt’s ability to detect referrable DR from color fundus photographs in adults. Researchers searched PubMed, Embase, and ClinicalTrials.gov through April 2025 and included eligible studies involving EyeArt screening.
Key findings
- Combined data from 17 studies totaling 162,695 examinations show EyeArt’s pooled sensitivity at 95% (95% CI: 92–97%) and specificity at 81% (95% CI: 74–87%).
- Subgroup analyses indicated consistent accuracy across study designs, economic settings, healthcare contexts, device types, external validation, and image gradability. Specificity varied slightly depending on vendor involvement.
Expert interpretation
Despite EyeArt’s strong diagnostic performance, real-world uptake of autonomous AI DR screening remains low. In the United States, CPT code 92229 (remote retinal imaging with automated analysis) was billed in only about 0.09% of adult diabetics with eye imaging encounters from 2021–2023, reflecting limited adoption. Reimbursement for DR screening is modest compared with other AI-enabled or imaging services, and there are upfront costs for cameras and IT integration. Authors conclude that these economic factors, plus integration barriers, deter primary-care adoption.
Limitations to consider
- Inconsistent reporting across studies, especially regarding ungradable images and how AI handles them.
- Variability in specificity across studies, likely tied to ungradable image management and hybrid workflows not consistently described.
- Only 11 studies offered true cross-country external validation, so applicability to regions with different retinal image characteristics may be overestimated.
- Most data come from high-income settings, with limited evidence from low-resource contexts where screening needs are greatest. This raises questions about generalizability and implementation challenges where the burden is highest.
Take-home message
EyeArt demonstrates high diagnostic accuracy for detecting referrable DR, with pooled sensitivity around 95% and specificity around 81%. The strong sensitivity supports autonomous screening in primary care, while the more variable specificity and inconsistent reporting of ungradable images call for standardized quality assurance and clearer workflows.
What comes next
To unlock the full public-health potential of AI DR screening, attention must extend beyond test accuracy to the post-diagnosis pathway. Real-world impact depends on robust electronic health record connectivity, well-defined referral processes, sustainable reimbursement, and targeted deployment in underserved populations. Prospective implementation studies and cost-effectiveness analyses are needed to guide policy and widespread adoption.
