While artificial intelligence suddenly seems to be everywhere, physician-scientists at Wilmer have been developing and evaluating AI tools for years. What started with one project by Neil Bressler, M.D., in partnership with the Johns Hopkins Applied Physics Lab (APL), has multiplied in number and complexity. Today, AI-related projects at Wilmer encompass everything from precision medicine to surgical training to drug discovery, with many patients and faculty members benefiting from the unique, collaborative environment at Johns Hopkins.

“What sets Wilmer apart from other academic eye centers is the breadth of projects using artificial intelligence,” says T.Y. Alvin Liu, M.D., director of Wilmer’s Precision Ophthalmology Center of Excellence. “We also have a deep bench of investigators involved in the entire life cycle of AI projects, from fundamental model design to translating data into useful AI clinical decision tools, and from implementation of FDA-approved AI tools on a health system level to the ethical and societal considerations of AI technology.”

Monitoring From A Distance

Fundus photographs can reveal subtle changes in the eye and are used to diagnose and monitor patients with conditions such as age-related macular degeneration, diabetic retinopathy and diabetic edema. To determine if treatment is working or if a disease is progressing, best corrected visual acuity (BCVA) is also closely monitored.

Considered the gold standard for measuring someone’s vision, BCVA is assessed in the clinic following refraction. But for patients undergoing treatments for common retinal diseases, measuring BCVA might be of value as often as monthly — time-consuming for ophthalmologists in short supply, and inconvenient or logistically challenging for many patients.

When Bressler, the James P. Gills Professor of Ophthalmology, learned that certain AI algorithms were able to estimate the age of a patient based on reading fundus photographs of their retinas, he had an idea. Could a fundus photograph provide the BCVA of a patient?

He and his team constructed an AI algorithm to read fundus images of patients and estimate their BCVA. They validated these estimates from known BCVA measurements matched to fundus photographs. “We discovered AI could estimate BCVA from fundus photographs without refracting human beings or having them read an eye chart — usually within 10 letters [on a standardized eye chart] of the actual BCVA. Recent work is getting close to the goal of within 5 letters across many retinal diseases,” Bressler says. That means an AI algorithm evaluating retinal images could be as accurate as BCVA determined by a clinical exam, saving time and money.

Bressler envisions a day when patients could use smartphones to acquire fundus images of their retinas that would estimate BCVA. Patients could check this BCVA as often as needed from their home. Their physicians could receive more frequent, accurate information — allowing ophthalmologists to monitor more people and at greater distances from their clinics.

Improving Outcomes + Health Equity

A disease with wide-reaching and potentially devastating consequences is diabetic retinopathy, which is caused by damage to blood vessels in the retina. Because early detection can be sight-saving, patients with diabetes should be screened annually by an ophthalmologist. The screening, though vision-saving for many patients, can be burdensome because it requires a separate trip to the eye doctor. As a result, many patients simply do not get their annual screenings.

Since 2020, Johns Hopkins Medicine has deployed an FDA-approved autonomous AI screening device for diabetic retinopathy in primary care physicians’ offices, with Wilmer playing a pivotal role in the implementation process.

“The doctor checks your A1C levels, and right then and there, you get your retinal photographs taken. In real time, those photos are analyzed by an AI model that can tell right away whether you have diabetic involvement in the back of the eye,” says Liu. “It’s a one-stop shop.”

T.Y. Alvin Liu

A research team at Wilmer led by Liu has been studying the effectiveness of the devices. “It significantly improved the adherence rate when it comes to annual diabetic retinopathy screening. And specifically, it improved the adherence rate for at-risk populations,” says Liu. “We showed that we improved outcomes and health equity on a population level.”

Predictive Medicine, Customized Care

                                                             Jithin Yohannan

If one impetus for using AI is its wide reach, another is its precision. Several years ago, using a database of visual field tests, Jithin Yohannan, M.D., M.P.H., constructed an AI algorithm to predict which patients were at risk for rapidly progressing glaucoma based on their very first field test, which measures how wide of an area your eye can see.

The algorithm achieved 90% accuracy. Yohannan is now inputting additional factors — optical coherence tomography images of the optic nerve, clinical information such as intraocular pressure, and demographic information — into training the algorithm to see if it can predict with even greater performance which patients will progress faster.

“The next step is developing a model that can predict what course of treatment would be best for the patient sitting in front of you. It could essentially give us customized recommendations and estimated impacts of treatment decisions on very specific patients,” says Yohannan.

AI For The Surgeon

Shameema Sikder, right, oversees surgical training at the Center of Excellence for Ophthalmic Surgical Education and Training.

A global challenge is the shortage of highly trained ophthalmologists who can perform sight-saving surgeries. Shameema Sikder, M.D., the L. Douglas Lee and Barbara Levinson-Lee Professor of Ophthalmology, leads an NIH-funded research team in developing AI to analyze surgeons’ performance using surgical videos, and provide feedback to surgeons. The AI is expected to shorten the time surgeons take to acquire skill and reduce variation in skill across surgeons who are in training and in practice.

As the director of the Center of Excellence for Ophthalmic Surgical Education and Training (OphSET) at Wilmer, Sikder aims to implement AI to influence how surgeons are trained in the future. With additional support from a Microsoft for Startups grant, Sikder’s team has developed a cloud-based platform, Circlage, for surgeons to use the AI analytics to acquire skill. Sikder’s team, including collaborators from the Malone Center for Engineering in Healthcare and the Whiting School of Engineering, are now working on AI to provide surgeons personalized feedback to optimize skill, with the goal of elevating standards of surgical care globally. Recently, the team received a Maryland Innovation Initiative Award to commercialize Circlage.

Targeted Therapies 

Laura Ensign

Laura Ensign, Ph.D., the Marcella E. Woll Professor of Ophthalmology, and her team at Wilmer's Center for Nanomedicine, focus on developing sustained-release drugs that can be targeted to a specific area of the body and designed to release therapeutic molecules over time. The goal is to reduce or eliminate the frequent reapplication of drugs, which in the case of eye diseases often involves multiple eyedrops or regular injections into the eye.

For this project involving AI, Ensign’s team partnered with Michael Cummings, Ph.D., director of the Center for Bioinformatics and Computational Biology at the University of Maryland. They used a drug already known to reduce intraocular pressure (IOP), which can cause vision loss in glaucoma. Their aim: to combine the drug with other molecules that cause the drug to stay in the body longer and release more slowly.

The researchers decided to use a peptide — a sequence of amino acids — as the helper molecule. They constructed machine-learning algorithms to predict which peptide out of thousands would have the characteristics needed to penetrate into cells with the IOP- lowering drug in tow and bind to melanin (the compound that provides color to the eye and protects eye  cells from UV exposure). When the  AI model revealed the best peptide candidate, the researchers tested it in animal models and found that the peptide and IOP-lowering drug did have a high concentration in melanin-containing cells — and that sustained therapeutic effect was achieved.

These represent just a sampling of dozens of projects underway at Wilmer that are harnessing the power of AI to improve clinical care in ophthalmology. “AI holds the promise of transforming much of ophthalmology, and Wilmer is at the forefront of it,” says Liu.