What happened
Researchers from Moorfields Eye Hospital, three NHS trusts and St George's, University of London tested EyeArt v2.1.0 on 30,405 consecutive diabetic eye screening episodes. The episodes came from the North East London, South East London and Gloucestershire screening programs. EyeArt graded the images on servers at Homerton Hospital from January to April 2017, in parallel with usual care. The final human grades served as the reference. EyeArt flagged 95.7% of episodes with referable retinopathy and all 462 with moderate-to-severe non-proliferative or proliferative disease. It agreed with a human grade of no retinopathy 68% of the time. About half of the episodes, 15,091 of 30,405, would still have needed human grading. The authors conclude that EyeArt could halve the workload of human graders. Since the pathway did not change, no work was saved during the study. The NIHR funded part of the study, and the authors declare no competing interests.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Heydon P, Egan C, Bolter L, et al. Prospective evaluation of an artificial intelligence-enabled algorithm for automated diabetic retinopathy screening of 30 000 patients. British Journal of Ophthalmology 2021, 105(5):723-728. doi:10.1136/bjophthalmol-2020-316594. Funded in part by the NIHR. The authors declare no competing interests.
The algorithm demonstrated safe levels of sensitivity for high-risk retinopathy in a real-world screening service, with specificity that could halve the workload for human graders.
-
Assessment report, by authors with no tie to the system.
UK National Screening Committee (UK NSC) (2026). UK NSC consults on use of artificial intelligence in the diabetic eye screening programme (UK NSC blog, on its 2025 evidence summary), Paragraph on the 2024 resubmission; 'The evidence' section.
The new review concludes that some ARIAS are as good as the first human grader and could be tested in real world NHS services, enabling information to be collected about the effect of ARIAS on patients, staff, and costs.
Details
AI system
- Reported configuration
-
V2.1.0
- Workflow
-
EyeArt classified each screening episode as test-positive or test-negative, and episodes that it could not assess counted as positive. All images from each episode were processed without editing or selection.
Follow-up
- Conditions for replication
-
Human graders still review every episode that EyeArt flags as positive. Programs check local performance, since cameras, image quality and grading differ between sites.
Entry details
- Sector
-
Healthcare (diabetic eye screening).
- Purpose
-
Eyenuk built EyeArt to grade retinal photographs for diabetic retinopathy and flag the patients who need a specialist.
29 more questions not yet checked
model, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). EyeArt could halve retinal grading with no sight-threatening misses (AIBD-0931). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-0931.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)