Researchers at the Wilmer Eye Institute at Johns Hopkins Medicine published findings this year showing that an AI-assisted diagnostic tool used during routine primary care visits meaningfully narrowed a longstanding racial gap in diabetic eye care. The peer-reviewed results, published April 13 in the journal npj Digital Medicine, found that African American patients with diabetes were substantially more likely to get referred for a diabetic eye exam when an AI screening tool was used during their visit than when it was not.
The Core Numbers
The study found that 64.9% of African American patients received a referral for a diabetic eye exam when the AI-assisted tool was used, compared with only 44.4% when it was not, a gap of more than 20 percentage points. Among patients who both opted into AI-assisted screening and went on to attend their diabetic retinopathy evaluation, the study found they were 15% more likely to be African American than patients who skipped AI screening, suggesting the tool did not just flag more referrals overall but specifically reached a population that has historically been underserved by traditional eye-care referral pathways. The researchers also found elevated referral rates for patients with hypertension, 89.6% versus 82.6%, and chronic kidney disease, 26.2% versus 20.9%, when the AI tool was part of the visit.
How the Screening Actually Works
The tool works by capturing a retinal image with a specialized camera during a routine primary care appointment and analyzing it in real time for signs of diabetic retinopathy, the diabetes complication that damages blood vessels in the retina and remains a leading cause of preventable blindness in adults. Because the image is captured and interpreted on the spot, patients can be told the same day whether they need a follow-up eye exam, removing the extra visit, extra appointment, and extra transportation burden that a separate ophthalmology referral typically requires. The study evaluated this workflow specifically at community-based primary care sites that serve underserved populations, rather than at specialty eye clinics that patients already have to seek out.
Why This Gap Existed in the First Place
Diabetic retinopathy screening has long suffered from a structural problem: even when primary care doctors recommend a referral, many patients, especially those facing transportation, cost, or scheduling barriers, never make it to a separate ophthalmology appointment to actually get screened. National data has repeatedly shown African American patients with diabetes are less likely than white patients to receive recommended annual eye exams, contributing to higher rates of preventable vision loss. By moving the screening into the primary care visit itself and returning same-day, AI-generated results, the Hopkins-affiliated research team effectively removed one of the biggest points where patients had been falling out of the referral pipeline.
How Researchers Are Interpreting the Result
The study’s authors describe their findings as exploratory but describe the disparities reduction as one of the clearer examples yet of AI-assisted tools directly narrowing, rather than simply not worsening, a documented healthcare disparity. That framing matters because AI diagnostic tools have more often been criticized for the opposite effect: algorithms trained on datasets that underrepresent certain populations have, in other clinical contexts, been shown to perform less accurately for Black patients, reinforcing rather than closing gaps in care. The Hopkins team’s results suggest that when an AI tool changes the point-of-care workflow itself, in this case eliminating the need for a second appointment, the benefit can outweigh the risk of algorithmic bias in the underlying model, at least for this particular screening task.
What Skeptics and Cautious Voices Note
Health-equity researchers not involved in the study caution against treating a single, exploratory result as proof that AI screening alone can close disparities gaps more broadly. The observed benefit is specific to a workflow change, same-day results at the point of primary care, that could plausibly be replicated with non-AI interventions, such as training primary care staff to operate retinal cameras themselves and interpret basic findings without an AI tool at all. Some clinicians argue the more important variable may simply be removing the second appointment, with the AI serving mainly to make same-day interpretation logistically feasible rather than being the active ingredient that closed the gap.
Where This Goes From Here
The Wilmer Eye Institute team has framed the results as justification for expanding AI-assisted retinal screening to more community-based primary care sites serving underserved populations, and similar autonomous AI screening tools for diabetic retinopathy have already received FDA clearance from other companies in recent years. Broader adoption will likely hinge on reimbursement policy, since primary care practices need a sustainable payment pathway to buy the cameras and integrate screening into already-packed appointment slots. If insurers and Medicaid programs move to cover point-of-care AI retinal screening more broadly, the Hopkins findings give health systems a concrete, published data point to justify making that investment.
