Seminar Summary for HBHI Workgroup on AI and Healthcare 

 

At the final HBHI-AI seminar of the academic year, an engaging session brought together experts from across Johns Hopkins to reflect on one of the country’s most consequential medical AI deployments: point-of-care diabetic retinopathy screening using autonomous artificial intelligence. The HBHI Workgroup on AI and Healthcare—co-led by Dr. Tinglong Dai of the Carey Business School and Dr. Risa Wolf of the School of Medicine—has convened monthly seminars since 2023 to explore critical intersections of artificial intelligence and clinical care. This season-closing session featured Dr. T. Y. Alvin Liu, the inaugural James P. Gills Jr. and Heather Gills Rising Professor of AI in Ophthalmology, who walked the audience through a five-year journey of implementation, adoption, and lessons learned.

 

Speaker Bio: Dr. T. Y. Alvin Liu, MD, is the inaugural James P. Gills Jr. M.D. and Heather Gills Rising Professor of Artificial Intelligence in Ophthalmology at Johns Hopkins University. He holds joint appointments in the School of Medicine and the Whiting School of Engineering, and serves as the founding director of the James P. Gills Jr. and Heather Gills Artificial Intelligence Innovation Center—the first endowed AI center at the Johns Hopkins School of Medicine. Dr. Liu leads translational AI projects in both clinical and operational domains, with a focus on responsible deployment, real-world outcome measurement, and cross-disciplinary implementation. He also co-chairs the AI and Data Trust Council at Johns Hopkins Medicine and is a national thought leader in AI governance, with leadership roles in the American Academy of Ophthalmology and the AMA's AI Specialty Society Collaborative.

 

The Promise of Autonomous AI in Eye Care

Dr. Liu began by explaining why diabetic retinopathy (DR) is such a compelling target for AI: it's a leading cause of irreversible blindness in working-age adults, yet early detection and timely treatment can dramatically reduce vision loss. Traditionally, DR screening has been confined to specialty eye clinics. But thanks to autonomous AI tools—first approved by the FDA in 2018—it’s now possible to perform retinal screening in primary care offices, using a cloud-connected desktop fundus camera and an AI model that returns a binary refer/no-refer result in real time.

This shift is more than technological—it’s structural. A dedicated CPT code (92229), national coverage from Medicare and private payers, and endorsements from bodies like the American Diabetes Association and NCQA’s HEDIS program have smoothed the regulatory and financial path. “It’s a rare case where the incentives aligned,” Liu said, “and the technology was able to punch above its weight.”

 

What Happened When Hopkins Rolled It Out

Johns Hopkins began deploying the system in 2020 at primary care sites that collectively serve about 40% of the health system’s adult diabetic population. The results were striking. Screening adherence jumped 7.6 percentage points relative to matched control sites—and among African American patients, adherence increased from 45% to 57%. Contrary to fears that adoption would drop post-COVID, exam volume has remained stable at around 800 per year, with per-site usage actually increasing.

Yet success brought new questions. One surprise: a 90% “leakage rate,” meaning that very few patients screened in one year returned for screening the next. The team is now conducting a longitudinal analysis to map patient-level trajectories and understand this retention gap.

Another challenge: in real-world conditions, half of patients required pharmacologic dilation to obtain usable images—far more than in clinical trials. To address this, Hopkins instituted a standing dilation protocol and employed dedicated imaging staff, which Liu emphasized was essential. “I can’t imagine an undedicated team outperforming a dedicated one,” he said.

 

From Science to Systems: Lessons on Scaling Medical AI

In the discussion that followed, Dr. Wolf highlighted Hopkins’ national leadership not just in using the technology, but in building the IT and clinical infrastructure that made meaningful study possible. “It’s one of the few places,” she noted, “where we’ve been able to move beyond diagnostic accuracy and start asking questions about real-world impact.”

“Hopkins is ahead of the curve,” Wolf said. “Now the challenge is to help the rest of the country catch up.”

Among those questions: How do patients and providers actually perceive autonomous AI? And what does successful implementation look like in operational terms?

Dr. Liu offered a practical framework: “You don’t need 95% certainty for every implementation decision. Sometimes 70% and a top-down push are enough to move forward.” That mindset, he suggested, is crucial for health systems experimenting with AI-enabled workflows. In the U.S., where fee-for-service payment models still dominate, he warned that only a few vertically integrated systems like Kaiser have the incentive structure to make early screening pay off. Until value-based care models mature, widespread adoption of AI screening will remain an uphill climb.

 

Audience Engagement: Reimbursement, Reinforcement Learning, and Institutional Variation

Dr. David Newman-Toker asked a foundational question: how did this all get started? Who actually pushed to engage payers like CMS or Blue Cross—and who paid for the early effort? Dr. Liu acknowledged that this success story was, in many ways, a best-case scenario. “This is as good as it gets,” he said, emphasizing how rare it is for clinical, financial, and regulatory incentives to align so perfectly. Much of the early legwork was done by pioneers like Dr. Michael Abramoff, who spent nearly a decade lobbying for a new CPT code. “It took years of sweat equity,” Liu noted, “and it’s not something most medical AI products will be able to replicate under current incentives.”

During the same Q&A session, Dr. Antonio Trujillo posed one of the most incisive questions of the hour: If reinforcement learning is central to AI’s evolution, why can’t these screening models improve continuously with new real-world data? Dr. Liu confirmed that while such updates are technically straightforward, the regulatory environment remains rigid. Once an AI device is approved by the FDA, it’s effectively frozen—unable to evolve without re-submission. However, the agency now permits vendors to propose a Predetermined Change Control Plan (PCCP) at the time of submission, which would allow for future, structured updates. “The mechanism now exists in theory,” Liu noted, “but no company has done it in practice.” Tinglong Dai added that the FDA’s PCCP database is searchable and may soon offer a window into who takes the first leap.

 

Dr. Gordon Gao followed up with a question on site-level variation: Had some Johns Hopkins primary care clinics implemented the screening more effectively than others? Liu replied that while the initial deployment was not randomized—the sites were chosen based on diabetic population size—his team had used propensity-score weighting to mitigate bias. Even so, he emphasized a broader tension: “There’s always a tradeoff between scientific purity and real-world impact. If your goal is to maximize adherence, you don’t randomize—you bet on the sites where it’ll work best.”

Dr. Tinglong Dai closed the Q&A with a forward-looking question: could AI eventually play a second-opinion role within ophthalmology itself—not just primary care? Liu explained that while this kind of model is viable in countries with severe specialist shortages, it’s not financially feasible in the U.S. “Subspecialists here are already seeing 60 patients a day at five minutes each,” he said. “There’s no time—and no reimbursement—for AI as a second reader.” Still, he acknowledged that international models using AI as a co-pilot for human experts could offer compelling examples of future possibilities.

 

Looking Ahead: A Renewed Commitment

In closing, Dr. Dai and Dr. Wolf thanked the community for its engagement over the year and announced that the HBHI-AI Workgroup will continue in 2025–26. They invited attendees to share speaker ideas and promised an even richer set of conversations next season.