The Eye Exam That Happens Before You Leave

A child with diabetes can leave an endocrinology appointment with a list of things still to do. One of them may be an eye exam.

The recommendation is straightforward. The logistics are not. A family has to understand why the exam matters, find an eye-care professional, schedule another visit, arrange transportation, miss school or work, and make sure the result travels back to the diabetes clinic. Each step is small. Together, they can become a wall.

Researchers at Johns Hopkins tested a different path: bring the eye exam into the diabetes visit itself. A clinic worker takes retinal photographs. An autonomous AI system analyzes them on the spot. Within about a minute, it reports whether signs of referable diabetic eye disease are present, absent, or whether the images are not good enough to interpret.

The technology matters. But the deeper innovation is the redesign of the journey.

A randomized test of fewer steps

The ACCESS trial enrolled 164 young people with type 1 or type 2 diabetes at two pediatric diabetes sites in Baltimore. Participants were 8 to 21 years old, met screening criteria, had no known diabetic eye disease, and had not completed a diabetic eye exam in the previous six months.

Researchers randomly assigned 81 participants to receive the AI-based eye exam during the study visit. Eighty-three were assigned to an enhanced version of usual care: a referral to an optometrist or ophthalmologist, plus scripted education and a handout explaining how to obtain the exam. One control participant was excluded from the primary outcome after also enrolling in another eye-screening study, leaving 82 in that analysis.

The primary endpoint was documented completion of a diabetic eye exam within six months.

Every participant in the AI group completed the screening: 81 of 81, or 100%. In the control group, 18 of 82 completed an exam, or 22%. The absolute difference was 78 percentage points, with a 95% confidence interval of 69 to 87 points.

That is an unusually large effect. It is also important to understand what the experiment actually tested. The AI group received the exam immediately during the visit. The control group still had to complete a separate referral, even after receiving extra education. The result therefore supports an AI-enabled point-of-care workflow—not the claim that an algorithm alone caused people to become more adherent.

Research references: [1] [2]

The moment of knowing changed the next step

The study also looked at follow-through after screening.

The AI system flagged 25 of the 81 participants for referral. Sixteen of those 25, or 64%, completed an eye-care visit within six months. The control arm’s eye-care completion rate was 22%.

The comparison requires care because the groups reached eye care through different pathways. In the intervention group, the denominator included people with an immediate abnormal AI result. In the control group, completion meant following the original screening referral. The trial authors treated this as a prespecified secondary outcome, but it is not a simple like-for-like comparison of two groups with the same confirmed disease.

Still, the behavioral logic is plausible: “Please schedule a screening” is different from “This test found something that needs a specialist.” Immediate information can turn a preventive recommendation into a concrete next action.

The remaining gap matters too. Nine of the 25 young people with an abnormal AI result did not complete specialist follow-up within six months. AI removed one barrier, but it did not remove transportation, scheduling, cost, trust, or every other barrier between detection and care.

Research references: [1]

Access was the intervention

The trial population was deliberately relevant to the access problem. Thirty-five percent of participants were Black, 47% were insured through Medicaid, and 34% reported household income below $50,000. Before randomization, 35 of the 164 participants said they had never had a diabetic eye exam. In unadjusted comparisons, that group was more likely to be Black or Hispanic, to have Medicaid, and to report lower household income.

The randomized result did not show statistically significant differences in completion by race, ethnicity, socioeconomic status, or parental education within the study. But the trial was small and conducted at two sites in one academic health system, so it cannot settle whether the workflow will reduce disparities at population scale.

What it offers is a concrete design principle: when care depends on a second appointment, access is not merely a patient education problem. It is a systems problem. A useful AI deployment can compress the system—camera, interpretation, result, and referral decision—into the visit a patient is already attending.

Research references: [1]

What the study does not prove

The study did not measure visual acuity, prevented blindness, long-term disease progression, or health outcomes after treatment. It measured screening completion and follow-through over six months.

The device used in the trial was authorized for adults at the time of the study, not for patients 21 and younger. For safety, a board-certified retina specialist later reviewed all images. The study reported no adverse events during the visits or photography, but a single trial of 164 participants is not enough to establish performance across every pediatric population or clinic.

The researchers also noted that some participants had encountered autonomous AI screening in an earlier study, which may have increased acceptance. All intervention images were diagnosable in ACCESS; other settings, patient groups, cameras, and operators may produce more ungradable images. The study compared point-of-care AI with referral-based eye care, not with a well-run tele-retinal program.

There are relevant commercial ties. Digital Diagnostics supplied AI interpretations in kind. One author reported being an investor, director, and consultant to the company and holding relevant patents. The trial also received support from the National Eye Institute and the Diabetes Research Connection. These disclosures do not erase the randomized findings, but they strengthen the case for independent replication.

Research references: [1] [2]

The human measure of useful AI

Healthcare AI is often discussed as a contest of diagnostic accuracy: machine versus clinician, percentage point versus percentage point. ACCESS points toward another measure.

Did the patient receive the recommended care?

For 81 young people, the eye exam happened because it was placed where they already were. The camera was close. The answer was immediate. The referral, when needed, had a reason attached to it.

That is a humbler vision of AI than replacing a doctor. It may also be a more consequential one: use computation to shorten the distance between a person and care.