An In-Clinic AI Eye Exam Raised Screening Completion From 22% to 100%

The result was about access, not just accuracy

Clinical guidelines recommend regular eye screening for people with diabetes because retinal damage can advance before vision changes are obvious. Yet completing a separate visit with an eye-care professional can be difficult, especially for families already managing school, work, transportation and frequent diabetes appointments.

A randomized trial at a U.S. pediatric diabetes center tested a different workflow. Instead of referring every participant elsewhere for screening, the intervention brought a retinal camera and an autonomous AI system into the diabetes visit itself. Among 164 participants ages 8 to 21, all 81 assigned to the AI arm completed the study-defined eye exam. In the control arm, 18 of 82 evaluable participants completed a documented exam with an eye-care professional within six months: 100% versus 22%, a 78-percentage-point difference (95% confidence interval, 69 to 87; P<0.001).

That is a large randomized effect on screening completion in this setting. It does not show that AI improved vision, prevented blindness or outperformed an ophthalmologist at comprehensive eye care. It shows what happened when one narrow screening step became available at the point of care.

Research references: [1]

What the trial actually compared

Participants had type 1 or type 2 diabetes, were at least 8 years old and had not completed a diabetic eye exam in the previous six months. The researchers randomly assigned 81 people to the intervention and 83 to usual referral. One control participant was excluded from the primary analysis after a protocol violation, leaving 82 evaluable controls.

The intervention and control pathways used different completion definitions because they were different care models:

• In the AI arm, completion meant receiving the autonomous eye exam during the diabetes visit.
• In the control arm, completion meant having a documented dilated eye exam by an eye-care professional within six months.

Research references: [2]

Registration and workflow details

The trial was registered before recruitment began, and the public record reports the same arm sizes and outcome counts as the paper.

The AI workflow used a nonmydriatic retinal camera. A trained operator captured images, and the system returned one of three study-relevant outcomes: referable diabetic eye disease detected, not detected, or insufficient image quality. A positive or insufficient-quality result triggered referral to an eye-care professional. For the trial, a retina specialist also reviewed all images as a safety and reference procedure; that review was part of the research protocol, not evidence that the autonomous output itself required a specialist to produce a result.

Research references: [2] [1]

The second handoff was still hard

The most important limitation appears after the first screening step.

The AI system returned a positive result for 25 participants. Sixteen of those 25—64%—completed a follow-up visit with an eye-care professional within six months. In the control arm, 18 of 82 participants—22%—completed the referred eye exam. Those percentages should not be compared as if they describe the same denominator or the same event. The AI figure is follow-through after a positive point-of-care result; the control figure is initial screening completion after referral.

This distinction matters. Bringing the first exam into an existing visit removed one layer of friction, but it did not guarantee that everyone with a positive result reached specialist care. A screening system can close one gap while leaving the next handoff exposed.

The study team reported no adverse events during study visits or retinal photography. In the trial sample, comparison with the specialist reference yielded estimated sensitivity of 100% and specificity of 78.9%. The sample was too limited to treat those figures as definitive pediatric performance estimates, and some positive AI outputs would be expected not to match the specialist reference standard.

Research references: [1]

The device was used outside its labeled age range

The trial used IDx-DR, now marketed as LumineticsCore. The FDA-cleared indication is for adults with diabetes who have not previously been diagnosed with diabetic retinopathy. The trial population included people younger than 22, so the paper explicitly describes this as off-label use and added safeguards, including specialist overread of all images.

That boundary should travel with the result. A successful research workflow in one pediatric diabetes center is not, by itself, a regulatory authorization or a general recommendation for pediatric deployment.

The study also has other limits. It was conducted at a single center. Participants and clinicians could not be blinded to the workflow. The primary endpoint was completion, not a patient health outcome. And although the randomized design supports a causal conclusion about this bundled intervention in this setting, it cannot isolate how much of the effect came from the algorithm, the camera being on site, staff support, immediate results or the absence of a second appointment.

Research references: [3]

The human system around the model did the work

This trial is best read as a lesson in service design.

The AI mattered because it could produce a referral recommendation without requiring an eye specialist to be present at the diabetes visit. But the result depended on more than software: a camera, a trained operator, clinic time, image-quality checks, a referral pathway, family communication and a specialist safety process all surrounded the model.

The useful question for another clinic is therefore not simply whether the algorithm is accurate. It is whether the entire pathway can reliably do five things:

1. Reach people who are missing screening. 2. Capture usable images during a visit they already attend. 3. Communicate a narrow result without overstating what was examined. 4. Route positive or ungradable cases to qualified eye care. 5. Track whether that second handoff is completed.

What the result supports

The trial suggests that point-of-care autonomous screening can substantially increase completion of a narrow eye-exam step for young people with diabetes in a setting like the one studied. Its unfinished business is equally clear: access to an initial result is not the same as access to the care that may need to follow.

The study was funded by the National Eye Institute and Diabetes Research Connection. Digital Diagnostics supplied AI interpretations in kind. One coauthor disclosed investment, consulting, directorship and patents related to the company and technology; another disclosed research support from pharmaceutical companies outside the submitted work.

Research references: [1]