Screening for diabetic retinopathy with AI, without a reading doctor

Why screen the eyes of people with diabetes?

Diabetes can damage the tiny blood vessels of the retina, the thin layer at the back of the eye that catches light. This damage is called diabetic retinopathy. Found early, it can be managed before eyesight drops.

The difficulty is that the check needs images of the back of the eye and a specialist to read them. The authors want to bring specialist-level diagnosis straight into primary care clinics, where people with diabetes are already followed.

How was the AI tested?

This was a pivotal trial: the final study used to ask for an authorisation. It enrolled 900 people with diabetes and no known retinopathy, at primary care clinics. Their median age was 59.

The AI's results were compared with a very demanding reference: a photograph reading centre in Wisconsin, which used wide-field images and an eye scan (OCT). The people operating the AI had received standard training before the trial began.

Each person was therefore examined twice, once by the AI in the clinic and once by the expert centre, so the two could be compared fairly. The study was registered on ClinicalTrials.gov (NCT02963441).

What were the results?

Among participants, 198 (23.8%) had retinopathy beyond the mild stage, or macular oedema (a swelling at the centre of the retina). The AI spotted these cases with a sensitivity of 87.2%, meaning it flagged about 87 out of 100 people who had the problem. Its specificity was 90.7%: it correctly cleared most people without it.

These figures beat every target set before the study (above 85% for sensitivity, above 82.5% for specificity). The images were also good enough to be used in 96.1% of cases.

The participants were varied: 28.6% were African American and 16.1% Hispanic. The share of usable images (96.1%) matters a great deal for a system meant to work without a specialist on site.

What does this change for patients?

Based on these results, the FDA authorised the system for use by healthcare providers. The authors stress that it is the first autonomous AI diagnostic system authorised in any field of medicine.

In practice, a person with diabetes can be screened during a routine appointment, without waiting for a specialist slot. The authors see this as a way to help prevent vision loss in thousands of people each year.

What are the limits?

Screening is not a full diagnosis: a positive result leads to a specialist, and a negative one does not remove the need for usual follow-up. The trial covered people with no known retinopathy, in specific US clinics.

The system was used by operators trained to a standard protocol. How it performs at larger scale, with very different teams and in other countries, remains to be seen.