For years, the specter of artificial intelligence (AI) has loomed over medicine, fueled by predictions of robotic surgical dominance and sweeping technological transformations. While much of the rhetoric has been overstated, I have noticed several changes even from my limited perspective as a medical student. As AI creeps into day-to-day clinical practice, the conversation has evolved from whether it will replace clinicians to how it is already changing healthcare workflows and redefining clinical authority. This is especially relevant in the field of ophthalmology, particularly in regards to screening for diabetic retinopathy.
In April 2018, LumineticsCore became the first AI diagnostic system to receive de novo FDA approval. LumineticsCore uses fundus photographs to screen for diabetic retinopathy (DR), with real-world studies demonstrating 87% sensitivity and 91% specificity. Since then, two other AI-based DR screening systems have received FDA approval. Across the country, testing and adoption of such technology by several academic institutions and federally qualified health centers is ongoing. The American Diabetes Association (ADA) recognizes AI systems as an alternative to traditional screening approaches. Furthermore, CMS established a specific billing code for AI diagnostic screening, which now allows reimbursement for retinal exams performed by a software algorithm. These developments represent the formal integration of AI in clinical care and therefore position such technology as a new source of diagnostic authority.
AI screening is poised to solve serious gaps in eye care. For individuals with Type 2 diabetes, ADA recommends a dilated eye exam at the time of diagnosis. However, in practice, this process often means months-long wait times with significant loss to follow-up. Studies have shown that only about half of Medicare beneficiaries with diabetes receive an eye exam. AI-based algorithms can mitigate these disparities, offering a way to integrate screening within routine primary care and minimize the time between diabetes diagnosis and diabetic eye exams. Preliminary data from institutions piloting AI-based screening systems in their clinics have shown promising results, with double-digit percentage point increases in screening rates.
Until recently, ophthalmologists held exclusive authority over the screening, diagnosis, and management of diabetic retinopathy. AI systems are breaking down this boundary by moving screening from ophthalmology clinics to primary care offices, community health centers, and even patient homes. Portable imaging devices and AI algorithms can theoretically extend screening to nearly anyone with diabetes without directly involving ophthalmologists. This expansion can address existing access issues, particularly in rural areas where the shortage of ophthalmologists is most dire. However, AI’s potential to expand access to eye exams does not immediately guarantee equity. Rather, the implementation of AI technology is creating new gatekeepers within an already imperfect and fragmented health system.
As screening moves from ophthalmology clinics into primary care settings, the authority and responsibility for deciding who is screened also shifts. Screening technology may be autonomous, but its implementation is not. In primary care settings, which are already strained by high patient volumes, the successful integration of AI screening requires staff to undergo dedicated training to operate screening cameras, follow up on abnormal results, and manage referrals to ophthalmologists. In addition, there are significant investments in infrastructure. This includes high-quality fundus cameras, reliable electricity and internet access, and EHR integration. Clinics also need to navigate new reimbursement pathways, justify upfront costs, and balance the return on investment to ensure long-term sustainability. Screening is therefore far from autonomous. Considerable staff and financial investment are needed, which presents additional barriers that can perpetuate existing disparities, particularly in rural and low-resource communities.
Algorithmic bias within AI technology is another challenge. Sociodemographic reporting in prior AI studies of retinal diseases is astonishingly lacking. In a recent systematic review of 360 published AI studies on retinal diseases, only 31% of studies reported gender or sex, and just 9% reported race and ethnicity. Limited diversity in training datasets and inadequate reporting of participant characteristics limit the generalizability of these algorithms. Head-to-head validation studies have demonstrated substantial differences in performance. In one multicenter study evaluating the real-world performance of seven automated AI-based DR screening algorithms across VA sites, the authors found considerable variations in sensitivity, ranging from 51% to 86%. For AI systems to truly address inequities in care, we need to use representative datasets, validate algorithms across diverse patient populations, and mandate transparency around the reporting of sociodemographic data.
The question is no longer whether AI should be used in diabetic retinopathy screening, and by extension, other contexts in healthcare. Its potential to expand access is evident. Rather, the more urgent questions are who benefits from its use and who controls its implementation and regulation. AI-based systems are not simply technical innovations; they also fundamentally change clinical authority and practice. These systems do more than detect disease: they shape how disease is defined, determine who is screened, and ultimately decide who receives care.
Bonnie is a third-year medical student at the Perelman School of Medicine at the University of Pennsylvania. She enjoys painting outside the lines, cultivating her ever-growing plant collection, and tossing together hearty salads. She is a 2025–2026 Doximity Op-Med Fellow.
Collage by Jennifer Bogartz / Shutterstock
