AI in Ophthalmology: The Key to the Present Success Includes a CPT Code. The Future? Agentic AI That Doesn’t Wait To Be Told What To Do | AAO 2025

Successful artificial intelligence (AI) depends on many things — a sound foundation model, data-dense datasets for those models to work on, an abundance of computer processing power and an equally large supply of electricity to keep that processing humming. But in the recent history of AI in ophthalmology, the number 92229 might be one of the most important ingredients in the recipe.

That is the number for the Current Procedural Terminology (CPT) code for using autonomous AI to diagnose diabetic retinopathy, a widespread application of AI in ophthalmology and one of the most successful in U.S. healthcare so far.

“Reimbursement and everything financial — the business model has to do with it — is very, very important,” said Michael Abramoff, M.D., Ph.D., at a session on AI yesterday at the American Academy of Ophthalmology’s annual meeting, which is being held in Orlando, Florida.

Abramoff is a professor of ophthalmology and visual science at the Carver College of Medicine at the University of Iowa. But far more relevant to AI in ophthalmology is that Abramoff founded Digital Diagnostics, the company that pioneered the development of retinal scanners that use autonomous AI to diagnose retinal disease caused by diabetes. The scanners are often used by primary care physicians to decide which patients need to be referred to a specialist. In rural Iowa, a patient may need to drive three to six hours to see an ophthalmologist, he said. “Now AI can make a high-quality diagnosis wherever the patient is, wherever there’s a primary care clinic, usually a federally qualified health center.

Abramoff also noted total reimbursement for autonomous AI for diabetic retinal exams has been climbing the past two years in contrast to other evaluation and management codes. He said strong supporting evidence translates into favorable conditions for reimbursement.

“The more you can show that you have better outcomes for patients in various ways, lower costs, better vision, better clinician productivity, better clinician job satisfaction, the easier it is for CMS to pay more rather than less,” said Abramoff, who also credited AAO President Michael X. Repka, M.D., MBA, for playing a role in favorable reimbursement.

The other panelists were J. Peter Campbell, M.D., M.P.H., a professor of ophthalmology at the Oregon Health & Science University and chair of the AAO’s committee on AI, and Pearse Keane, M.B.B.C.H., a professor of artificial medical intelligence at University College London and an ophthalmologist at Moorfields Eye Hospital NHS Foundation Trust.

Campbell helped set the stage for the discussion stage by delineating the difference between assistive and autonomous AI, which are what the words suggest. “If the computer is saying this is a macular hole and referring to the retina specialist without a physician in the loop, that would be an autonomous application. If it is simply providing the optometrist or ophthalmologist ‘I think this is what it is; you use your clinical judgment [about] what’s going on,’ that is an assistive technology.” Campbell said an algorithm could be used in any number of ways and that FDA regulation of AI hinges on the uses.

Keane, whose observations about AI had a more futuristic bent, mentioned a third type of AI: agentic. Referencing an article written by Eric Topol, M.D., a major proponent of AI, that was published in The Lancet earlier this year, Keane said agentic AI moves AI from being a narrow tool to working as an agent on our behalf.

“The key conceptual shift is, instead of using them as, effectively, like sophisticated calculators, to embrace them as teammates. So agentic AI is the potential to take initiative, rather than waiting for queries and data to proactively monitor and pull data and identify issues and propose solutions with long-term memory and context tracking complex patient histories and interactions over time,” Keane said.

Campbell also spoke about the “hype cycle” that many new technologies follow: “Everyone gets super excited about it, then you find all the problems with it, and then eventually reach a status quo.” He also thumbnailed the “implementation gap”: that it takes 17 years from when something is known to work for it to be commonly used in practice. Campbell said there are lots of reasons for that gap. Abramoff spoke about the fraught issues of liability and AI replacing human workers and professionals. He said he is a co-author of a paper that is about to be published about state medical boards playing a role in overseeing the use of AI by physicians.

“You can use AI all you want as a clinician,” Abramoff said. “But is this still safe? You know there will be a lawsuit where a patient says, ‘Hey, you clinician, you blinded me, or you killed my brother because you used ChatGPT or whatever.” Abramoff said that when the AI technology is autonomous, the medical liability falls on the creator of the AI. He said AI also creates a major change in medical ethics from the doctor-patient relationship because doctors may not be part of the decision.

Campbell shared a relatively short list of 10 AI ophthalmology products that the FDA has cleared that started with the 2018 clearance of Abramoff’s autonomous AI retinal scanning technology IDx-DR, which has been renamed LumineticsCore. He noted the yawning gap between the soaring number of papers published on AI and the application of AI, but he also said as the accuracy of imaging improves, so will the use of AI and, consequently, automation in ophthalmology. “I think we’re getting to the point where much of the eye exam can be automated,” he said. Campbell also noted that FDA clearances paint an incomplete picture of the use of AI and that “many, many tools might be built into our devices that don’t require FDA clearance if they’re only providing segmentation or assessment.”

Keane referenced findings reported by researchers at Google in Nature in April 2025 that showed an AI model navigating through differential diagnosis, a core responsibility and skill of clinicians. The model is equipped with a feature that allows a back-and-forth conversation about symptoms, how long they have been occurring, and other information. He also mentioned new AI models that will allow clinicians to upload images and then get answers to questions about the image from the AI model. “You can easily see a situation where you have a retinal photo and you say, ‘What’s that yellow dot beside the optic disc?’ And they say it is a drusen or an exudate,” Keane said.

Keane also explained his title as a professor of artificial medical intelligence was an intentional reference to developments in AI outside of healthcare that come under the heading of artificial general intelligence, which Keane said means "when will AI systems be able to do most cognitive tasks better than humans?" In ophthalmology, Keane said, artificial medical intelligence has the potential to eliminate the current cascades of diagnosis and care from generalists to specialists to subspecialists. Keane noted that he is an example of the super subspecialist who is ”very specialized in a tiny sliver of the eye and a tiny sliver of the medical retina.” He compared the AI-powered future of ophthalmology to the handheld device that the Dr. Leonard “Bones” McCoy character used in “Star Trek” that delivered a diagnosis on the spot. Ophthalmologists will not be as siloed as they are now, Keane said.

“I think the stuff that I do probably can be automated more easily than the stuff that the comprehensive ophthalmologist will do,” he said. “The comprehensive ophthalmologist would have the ability to pick up some weird type of keratoconus on one patient and then diagnose some esoteric type of drusen on another patient, and then the next patient would look at some better assessment of neuro-ophthalmic conditions.”