An AI eye exam made retinal screening part of a routine diabetes visit

A recommendation to book an eye examination still leaves a family with a booking, another journey and time away from work or school. The ACCESS trial moved diabetic eye screening into an appointment already taking place. All 81 young people offered the AI examination completed it at their diabetes visit. In the separate-referral group, 22% completed an eye examination within six months.[2]

Four photographs and a result before leaving

The Johns Hopkins study, published in January 2024, enrolled 164 participants aged 8–21 with type 1 or type 2 diabetes. The intervention added a five-to-ten-minute examination, without pupil-dilating drops, to a visit to their diabetes specialist. The comparison group received a referral and deliberate education about the importance of eye screening.[2]

IDx-DR, now called LumineticsCore, combines a retinal camera, software used by clinic staff and an image-analysis service. Its original FDA review describes two photographs of each eye: one centred on the optic disc and another on the macula. Images are transmitted for analysis and the software returns a disease assessment or feedback that the image quality is inadequate.[3]

That quality check is part of the clinical function. A camera operator can retake an image before the patient leaves, rather than discovering later that the screening could not be interpreted. The FDA’s pivotal study trained staff without previous ocular-imaging experience to capture and submit the images. The specialist task being automated is interpretation of the retinal images; acquiring usable photographs remains a staff responsibility.[3]

An abnormal screen starts another part of care

Of the 81 AI examinations, 25 produced an abnormal result. Sixteen of those 25 participants subsequently attended an eye-care provider within six months: 64% follow-through among those referred after an abnormal screen. Nine had not completed that next step. The initial screening and the later specialist appointment are separate outcomes, with different denominators.[1]

Because this was paediatric research, a retina specialist also overread every image set. The manufacturer’s US indication, checked in October 2026, covers people aged 22 and over with diabetes who have not previously been diagnosed with diabetic retinopathy. The youth trial therefore tested an investigational use with additional clinical review.[1][5]

The product is designed to identify more than mild diabetic retinopathy. It does not offer a comprehensive check for other eye diseases, such as glaucoma, and it does not treat the disease it identifies. A positive result connects the patient to further eye care; an unreadable examination requires retesting or referral.[5]

What happened when primary-care clinics adopted it

A separate Johns Hopkins evaluation, published in July 2024, compared adult screening adherence before and after some primary-care sites introduced autonomous AI. The researchers analysed more than 17,000 patients with diabetes in each of 2019 and 2021, using propensity weighting to compare sites that adopted AI with those that did not.[4]

The increase in annual testing was 7.6 percentage points greater at the adopting sites. Among Black patients, adherence rose by 12.2 percentage points at those sites, while it fell by 0.6 points at non-adopting sites. These are changes in screening attendance within an observational study, rather than proof that the software alone caused the improvement or prevented blindness.[4]

The trials illustrate why accuracy and access have to be evaluated together. ACCESS changed where and when the examination happened, as well as who interpreted it. Its result supports that combined service design; it does not isolate the effect of AI from the benefit of having any suitable screening service in the same clinic.

The strongest practical achievement is a completed first step while the patient is already present. The follow-through figures then reveal the next job for the service: getting the people with abnormal results into the appointments where further assessment and treatment can happen.

Report an error or suggest a correction