Autonomous AI brought diabetic eye screening to the point of care for youth with diabetes
Minority and underserved youth often miss diabetic retinopathy screening even though their risk of eye disease is higher. In the ACCESS2 study, autonomous artificial intelligence moved the exam into routine diabetes visits and produced nearly universal completion across groups. The practical effect was immediate: a screening process that usually depends on referral, follow-through, transportation, and family bandwidth was placed directly in the clinic workflow.
The study enrolled 397 youth with diabetes seen in diabetes care visits. Most had type 1 diabetes (72%), and the remaining 28% had type 2 diabetes. Each participant underwent autonomous AI diabetic eye examination and optical coherence tomography imaging. The central question was whether the technology could close the screening gap between White and non-White youth, a gap that was already visible before the intervention began.
Baseline data showed a clear divide. White participants were more likely to have completed a prior diabetic eye exam than non-White participants, 79.8% versus 61.8% with p < 0.0001. In routine care, differences like this often reflect more than simple awareness. They can emerge from competing demands, fewer specialty appointments, inconsistent referral completion, and the reality that preventive care is easier to lose when diabetes care already asks families to manage glucose checks, medications, school schedules, and follow-up visits.
Once the autonomous AI exam was embedded into the visit, the completion gap essentially disappeared. Screening completion reached 99.5% in the White cohort and 98% in the non-White cohort, meeting the study’s primary outcome of equivalence in screening completion after implementation. The result is notable because it shifts the burden from patient navigation to in-clinic execution. If the eye exam is performed while the patient is already present for diabetes care, one of the most common causes of missed screening falls away.
Diagnostic performance and the meaning of a high-sensitivity AI screen
Accuracy remained the second test of whether this model could function in real clinical practice. Compared with standard grading, the AI system showed 100% sensitivity with a 95% confidence interval of 67.6% to 100% for detecting more than mild diabetic retinopathy. Specificity was 76%, with a 95% confidence interval of 71.4% to 79.9%.
For a screening tool, the clinical message is straightforward. High sensitivity is what protects against missed disease, especially in a pediatric population where the point of screening is to identify children who need closer ophthalmic evaluation before vision-threatening progression occurs. A specificity in the mid-70s means some children without clinically important disease may still generate positive findings, but screening programs tolerate that tradeoff far more readily than they tolerate false reassurance. The question is not whether every image will be perfect; the question is whether the workflow reliably finds the youth who cannot afford to be missed.
The use of optical coherence tomography alongside autonomous AI also matters conceptually. Retinal imaging at the diabetes visit reduces reliance on a separate ophthalmology appointment and creates a more integrated pathway for care. In youth, where visit adherence can be uneven and disease duration accumulates over years, embedding eye screening into diabetes management may be one of the few ways to normalize surveillance rather than making it an optional add-on.
Why access and equity converge in a single clinic visit
Diabetic retinopathy disparities often persist because the screening system is built around multiple handoffs. Primary diabetes care identifies the need, a separate eye appointment must be scheduled, the family must get there, and the examination must be completed. Each handoff creates a failure point. Autonomous AI shifts the exam to the same place where diabetes is already being treated, which is exactly why the completion rates converged so sharply in this cohort.
There is also a broader equity implication. Non-White youth entered the study with lower prior screening completion, yet after implementation they reached the same high screening completion range as White youth. In a setting where access gaps are often explained away as differences in patient behavior, the data point to structure instead of blame. When the service is delivered where the patient already is, the disparity narrows without asking families to solve the logistics alone.
The study does not claim that autonomous AI replaces ophthalmologists. It does show that an AI-based screen can function as a reliable access tool at the point of care, with enough sensitivity to serve as a gatekeeper for further evaluation. In that role, the technology is not competing with specialty care; it is making specialty care more reachable by identifying who needs it and removing the first barrier to getting them into the system.
The limits of a screening solution that still needs follow-up pathways
Even strong screening completion does not by itself prove improved vision outcomes. This cohort demonstrates that autonomous AI can get the exam done and can detect more than mild diabetic retinopathy with high sensitivity, but the downstream pathway still matters. A positive screen has to lead to timely ophthalmic assessment, and the durability of this model depends on whether clinics can maintain the technology, interpret results consistently, and ensure referral completion when disease is found.
The diagnostic specificity also leaves room for false positives, which can add follow-up visits and anxiety if the care pathway is not carefully designed. In a pediatric setting, that operational detail is not trivial. Families will accept additional review when it is clearly tied to prevention, but repeated alarms without clear next steps can strain confidence in the process.
Still, the ACCESS2 study offers something concrete: autonomous AI diabetic eye exams can be delivered during diabetes care visits and can achieve equivalent screening completion among White and non-White youth. The study was registered as NCT05463289 and was supported by the National Eye Institute and Breakthrough T1D, with autonomous AI interpretations provided in kind by Digital Diagnostics. The next unanswered clinical question is not whether the exam can be placed at the point of care. It is how broadly this workflow can be scaled while preserving follow-up and durable access for the youth most likely to lose the chance for screening in the first place.