Clinical Evidence for AI Retinal Screening

Clinical evidence is the foundation for using artificial intelligence in retinal screening. EyeCheckup is designed to support clinical and screening workflows that use retinal fundus photographs to identify people who may need further professional assessment.

This page explains the evidence principles EyeCheckup follows, the validation questions healthcare organizations should ask, and how AI-assisted retinal screening should fit into a responsible care pathway.

Why clinical evidence matters

Retinal screening is not only a software task. It affects patient follow-up, referral decisions, workload for clinicians, and the confidence of hospitals, clinics, and screening programs. For that reason, an AI system should be evaluated for more than model accuracy alone.

A strong clinical evidence package should address:

  • Intended use and target population.
  • Image acquisition conditions and camera compatibility.
  • Sensitivity and specificity for the relevant screening tasks.
  • Performance across age, sex, ethnicity, image quality, and disease severity.
  • Handling of ungradable or low-quality images.
  • Human review and escalation pathways.
  • Clinical workflow impact and turnaround time.
  • Data protection, auditability, and post-market monitoring.

EyeCheckup positions AI as a clinical decision support and screening workflow tool. It should not replace a clinician's judgment, diagnosis, or treatment plan.

Published and independent clinical evidence

EyeCheckup's clinical evidence should be presented from primary sources wherever possible. Two important evidence signals are already available and should be visible from this page.

Nature Eye clinical study

The peer-reviewed article Head to head comparison of diagnostic performance of three non-mydriatic cameras for diabetic retinopathy screening with artificial intelligence was published in Eye, a Nature Portfolio ophthalmology journal.

The study evaluated EyeCheckup AI software in 900 patients with diabetes at Akdeniz University. Fundus photographs were captured with Canon CR2 AF, Topcon TRC-NW400, and Optomed Aurora non-mydriatic cameras. The AI outputs were compared with a clinical reference standard established by retina specialists.

The study reported high diagnostic performance across the tested camera workflows. For more-than-mild diabetic retinopathy, sensitivity was reported at 91% and above and specificity at 96% and above. For vision-threatening diabetic retinopathy, sensitivity was reported at 95% and above and specificity at 96% and above. For suspected clinically significant diabetic macular oedema, sensitivity and specificity were also reported at 95% and above across the tested cameras.

This publication is an important evidence asset because it directly names EyeCheckup, describes the patient population, explains the camera setup, and reports clinically relevant screening thresholds.

The Lancet Digital Health and ARIAS evaluation

EyeCheckup also reports that it was included in the peer-reviewed Lancet Digital Health article Automated retinal image analysis systems to triage for grading of diabetic retinopathy: a large-scale, open-label, national screening programme in England.

This article describes an independent real-world evaluation framework for automated retinal image analysis systems used in diabetic eye disease screening. The evaluation used data from the North East London Diabetic Eye Screening Programme and assessed commercial AI systems in a large screening-programme context.

Because some ARIAS reporting is designed around blinded or coded system comparisons, EyeCheckup should reference this evidence carefully. The current safest wording is that EyeCheckup has been reported by EyeCheckup and listed by the Royal College of Ophthalmologists as evaluated using North East London Diabetic Eye Screening Programme data in the ARIAS study. Per-system performance claims should only be shown when the source identifies EyeCheckup directly.

Clinical evidence directory listing

The Royal College of Ophthalmologists AI directory lists EyeCheckup AI as a CE Class IIa system for healthcare-professional use in detecting more-than-mild diabetic retinopathy and vision-threatening diabetic retinopathy, including severe non-proliferative or proliferative diabetic retinopathy and/or diabetic macular oedema in adults with diabetes who have not previously been diagnosed with diabetic retinopathy. The directory also links EyeCheckup to the Nature Eye publication and the ARIAS evaluation.

Evidence areas we prioritize

Retinal image quality

AI performance depends on the quality of the fundus image. A responsible screening workflow needs a clear process for image capture, image quality checks, repeated acquisition when needed, and referral or manual review when a result is not reliable.

EyeCheckup workflows should make image quality visible to operators and administrators. The goal is not only to produce a result, but to make sure that result is based on an image suitable for interpretation.

Disease-specific model performance

Clinical validation should be evaluated separately for each disease or risk signal. A model that performs well for one task should not automatically be assumed to perform equally well for another.

For diabetic retinopathy screening, relevant performance measures include sensitivity, specificity, referral threshold behavior, false positive and false negative handling, and comparison against a validated reference standard.

For broader retinal and systemic risk signals, evidence should clearly state whether the output is screening support, risk stratification, triage assistance, or research use.

Real workflow performance

AI screening only creates value when it works in the setting where it will be deployed. Hospitals, clinics, occupational health programs, mobile screening teams, and national screening projects may have different requirements.

Implementation should therefore evaluate:

  • Staff training needs.
  • Average capture and analysis time.
  • Integration with existing patient flow.
  • Referral and reporting rules.
  • Monitoring of rejected or ungradable images.
  • Local governance requirements.

Intended-use clarity

EyeCheckup content and deployment materials should clearly separate screening support from medical diagnosis. The software can help identify cases that may need professional evaluation, but final clinical responsibility remains with qualified healthcare professionals according to local rules and the configured workflow.

For procurement teams, this distinction matters. It helps define who can operate the system, where the system can be deployed, what patient communication should say, and when an ophthalmologist or other specialist must review the case.

What buyers should ask before deploying AI retinal screening

Healthcare organizations evaluating an AI retinal screening platform should ask:

  • What exact conditions and outputs are supported?
  • Which cameras, image types, and image fields are accepted?
  • What happens when image quality is insufficient?
  • What reference standard was used in validation?
  • How are false negatives and false positives monitored?
  • Does the system support audit logs and quality review?
  • How is health data protected?
  • What clinical oversight is required?
  • How are updates, model changes, and performance drift managed?

These questions help move AI procurement from a feature comparison to a patient-safety and workflow-readiness review.

Related EyeCheckup pages

Dedicated evidence pages

The existing Nature and Lancet posts are useful news posts, but procurement and clinical governance teams also need durable evidence pages that explain the studies in a structured way:

  • Nature Eye Diabetic Retinopathy Study summarizes the 900-patient clinical study, camera workflows, endpoints, sensitivity/specificity values, limitations, DOI, PubMed link, and practical deployment implications.
  • Lancet Digital Health ARIAS Evaluation summarizes the ARIAS evaluation framework, screening-programme setting, EyeCheckup participation evidence, careful wording about coded systems, and the Royal College of Ophthalmologists directory listing.

These pages would be stronger than short news posts because they can explain the evidence in procurement-ready language and give search engines a durable clinical evidence structure.

Sources and standards to consider

Medical disclaimer

EyeCheckup content is for clinical, operational, and educational discussion. It is not personal medical advice. Screening results, diagnosis, referral, and treatment decisions should be made by qualified healthcare professionals according to local clinical practice, regulatory requirements, and the configured intended use of the product.

Scroll to Top