Point-of-care AI eye exams narrow a screening gap among young people with diabetes

Bringing an autonomous AI eye examination into a routine diabetes visit nearly eliminated a screening-completion gap in a study of 397 young people. Before enrolment, 79.8% of White participants and 61.8% of non-White participants reported having completed a diabetic eye examination. When the study offered the test at the point of care, completion reached 99.5% and 98%, respectively.[1]

The prospective ACCESS2 study was published in Nature Communications on 23 October. Its primary endpoint was equivalence in screening completion between the two racial cohorts, which the authors report achieving after the intervention. The baseline percentages describe whether participants had ever completed a prior examination; the later percentages describe completion of the study’s offered test. They are different service conditions and time windows, rather than matched annual screening rates.[2][1]

Putting the examination inside the existing visit

Diabetic retinopathy can develop without symptoms that alert a patient early enough. The purpose of routine eye screening is to identify changes before they progress to serious vision loss. Traditional pathways may ask a family to book a separate appointment, travel to another practice or wait for specialist interpretation. Each added step can reduce follow-through. An autonomous camera-and-AI examination at an existing care visit seeks to remove some of those steps by bringing image capture and a screening decision into the same location.[1]

Participants had type 1 diabetes in 72% of cases and type 2 in 28%. They received autonomous AI eye examinations alongside optical coherence tomography imaging during diabetes care. The study evaluated access and diagnostic performance separately: completing a test answers a different question from how accurately that test identifies disease.[1]

The accuracy result and its precision

For detecting more-than-mild diabetic retinopathy, sensitivity was 100%, with a 95% confidence interval of 67.6%–100%. Specificity was 76%, with a 95% confidence interval of 71.4%–79.9%. The wide sensitivity interval reflects uncertainty around the estimate; the result does not establish that the system will never miss disease. Specificity also matters for the referral workload because some participants without the target level of retinopathy were flagged for further assessment.[1]

A point-of-care result is most useful when it connects to appropriate follow-up. The camera and software can bring screening into the diabetes clinic, but patients with abnormal or unusable images still need a clear route to further assessment. That hand-off is part of the service being built around the algorithm.[1]

What changed for access

The findings support reducing separate appointments and referral steps. They do not isolate which combination of cost, travel, scheduling or other barriers created the prior screening gap; offering the test during an existing visit changes several conditions together. The measured achievement is high, statistically equivalent completion across the study’s two cohorts.[1]

The company Digital Diagnostics supplied AI interpretations in kind, and the authors disclosed investigator relationships with the company. Those disclosures accompany the result and are relevant when comparing it with future independent studies or assessing implementation in other clinics.[1]

ACCESS2 shows that the location of a test can be as consequential as the method used to interpret it. The study brought an examination into a visit young people were already attending and achieved similarly high uptake in both cohorts. The next service challenge is to make an abnormal result equally easy to act on, so the shorter screening pathway leads to accessible specialist care.[1]

Report an error or suggest a correction