AI Eye Exams Close Youth Diabetes Screening Gap

Nearly every young diabetes patient offered an autonomous artificial-intelligence eye exam at a routine clinic visit went on to complete the screening, and a longstanding gap between White and minority youth all but closed, according to a prospective study published on 23 October in Nature Communications. The ACCESS2 study, led by researchers at Johns Hopkins, enrolled 397 White and non-White young people with type 1 (72%) and type 2 (28%) diabetes and offered them an AI-read retinal scan during a diabetes care appointment rather than a separate trip to an eye clinic.

At the start the two groups were unequal. White participants were more likely to have had a prior diabetic eye exam, 79.8% against 61.8% for non-White youth (p < 0.0001). After the AI exam was offered at the point of care, 99.5% of the White cohort and 98% of the non-White cohort completed it — what the authors call “equivalent screening completion after implementation”. The trial is registered as NCT05463289, with Johns Hopkins University as lead sponsor, and is still recruiting toward a target of 500 participants.

In brief

An AI-read eye exam given during routine diabetes visits lifted screening completion to 98-99.5% across White and non-White youth in a Johns Hopkins-led trial.
Against expert grading the system caught every more-than-mild case (sensitivity 100%, 95% CI 67.6%-100%) but flagged false positives at 76% specificity.
The study ran at two Baltimore sites, used AI readings supplied in kind by the maker, and the authors say it “suggests the potential” to cut disparities, not that sight outcomes improved.

What the exam measured, and where the numbers stay uncertain

The screening tool is an autonomous system: a non-mydriatic fundus camera captures retinal images and AI software returns a result on the spot, without an eye specialist reading the scan first. Measured against standard expert grading for more-than-mild diabetic retinopathy, it reported a sensitivity of 100% — catching every case the graders flagged — but the 95% confidence interval runs from 67.6% to 100%, a wide band that reflects how few positive cases the young cohort contained. Specificity was 76% (95% CI 71.4%-79.9%), meaning roughly a quarter of screened eyes without more-than-mild disease would still be referred on, generating false alarms and follow-up appointments. The work has real limits that the paper states plainly: it ran only at two Baltimore sites, the Johns Hopkins Pediatric Diabetes Center and Mount Washington Pediatric Hospital, so it is a single-city result rather than a national one. The AI interpretations were provided in kind by the manufacturer, Digital Diagnostics, and one author is an investor in and director of that company, a conflict the authors declare. And the headline finding is about who got screened, not whether anyone’s sight was saved: the authors are careful to write that the approach “suggests the potential” to reduce disparities, stopping short of a clinical-outcomes claim.

Why point-of-care screening matters for diabetic eyes

Diabetic retinopathy is a leading cause of preventable sight loss, and it is largely silent until it is advanced, which is why guidelines push regular retinal checks for anyone living with diabetes. The problem screening programmes keep running into is attendance: a young patient already juggling a chronic condition is less likely to book, and then keep, a separate ophthalmology appointment, and those drop-offs fall hardest on underserved groups. Putting the exam inside the diabetes visit removes that extra trip, which is the mechanism the Hopkins team is testing. The study sits in a widening body of work using AI to read medical images where specialists are scarce or overstretched — from a King’s College London model that spots a DNA-repair defect in ovarian cancer from routine slides to Sophont’s pathology model predicting tumour gene activity, and Stanford’s use of immune “digital twins” to flag preterm-labour risk. The common thread is bringing a diagnostic judgement closer to the patient. In the UK the parallel infrastructure already exists: the NHS runs a national diabetic eye screening programme that checks people with diabetes for the very condition this AI is designed to catch, which is why a tool that lifts attendance, if it holds up elsewhere, would be of direct interest beyond Baltimore.

A proof of access, not yet of better sight

What this trial demonstrates well is reach: offered conveniently, an AI eye exam got almost every young patient screened and narrowed a gap that programmes have struggled with for years. What it does not yet show is that earlier detection translated into treatment and preserved vision, or that the same completion rates survive outside two Baltimore clinics with motivated staff and vendor-supplied AI. The next things to watch are concrete: whether the full ACCESS2 cohort, due to finish recruiting by the end of 2026, holds the accuracy seen so far, and whether a false-positive rate near a quarter proves tolerable once the referrals it generates land on real eye clinics.