MATH & TECHNOLOGY

Computers can diagnose eye problems Artificial intelligence may boost health care in underserved areas

BY MARIA TEMMING

The computer will see you now. Artificial intelligence algorithms may

soon bring the diagnostic know-how of an eye doctor to primary care offices and walk-in clinics, speeding up the detec- tion of health problems and the start of treatment, especially in areas where specialized doctors are scarce. The first such program pending approval by the U.S. Food and Drug Administration is trained to spot symptoms of diabetes- related vision loss in eye images.

While other already approved AI pro- grams help doctors examine medical images, there’s “not a specialist looking over the shoulder of [this] algorithm,” says Michael Abràmoff. He founded the company, IDx, that developed the system under FDA review, dubbed IDx-DR. “It makes the clinical decision on its own.”

IDx-DR and similar AI programs — which are learning to predict everything from age-related sight loss to heart prob- lems just by looking at eye images — don’t follow preprogrammed guidelines for how to diagnose a disease. Instead, these programs are machine-learning algo- rithms that researchers teach to recog- nize symptoms of a particular condition by using example images labeled with whether a patient had that condition.

An algorithm still in testing uses eye images to predict cardiovascular health. A green heat map (left) overlaid on a retinal fundus image (right) highlights the areas — most notably blood ves- sels — that factor most heavily into the program’s prediction about a patient’s blood pressure.

IDx-DR studied over 1 million eye images to learn how to identify symptoms of diabetic retinopathy, which develops when high blood sugar damages retinal blood vessels. As many as 24,000 people in the United States lose their vision to diabetic retinopathy yearly, but the con- dition can be treated if caught early.

Abràmoff, a retinal specialist at the University of Iowa in Iowa City, and colleagues compared how well IDx-DR detected diabetic retinopathy in more than 800 U.S. patients with diagnoses made by three human specialists. Of the patients identified by IDx-DR as hav- ing at least moderate diabetic retinopa- thy, more than 85 percent actually did. And IDx-DR was correct in more than 82.5 percent of the cases where it con- cluded a patient had mild or no diabetic retinopathy, the researchers reported February 22 in Beverly Hills, Calif., at the annual meeting of the Macula Society.

The FDA’s decision about IDx-DR is expected within a few months, Abràmoff says. If approved, it would become the first autonomous AI to be used in pri- mary care offices and clinics.

AI algorithms to diagnose other eye diseases are in the works, too. An algorithm described in the Feb. 22 Cell studied over 100,000 eye images to learn

the signs of different eye conditions. These included age-related macular degeneration, or AMD — a leading cause of vision loss in adults over age 50 — and diabetic macular edema, a condition that develops from diabetic retinopathy.

This AI was designed to flag advanced AMD or diabetic macular edema for urgent treatment and to refer less severe cases for routine checkups. In tests, the algorithm was 96.6 percent accurate in diagnosing eye conditions from 1,000 pic- tures. Six ophthalmologists made similar referrals based on the same eye images.

Researchers still need to test how this algorithm fares in the real world where the quality of images may vary from clinic to clinic, says Aaron Lee, an ophthalmologist at the University of Washington in Seattle. But this kind of AI could be especially useful in rural and developing regions where medical resources and specialists are scarce and people otherwise wouldn’t have easy access to in-person eye exams.

AI might also be able to use eye pic- tures to identify other kinds of health problems. One algorithm that studied ret- inal images from over 284,000 patients could predict cardiovascular health risk factors such as high blood pressure.

When researchers tested the algo- rithm’s powers of prediction, the program was 71 percent accurate in distinguish- ing between eye images from smoking and nonsmoking patients, according to a report published February 19 in Nature Biomedical Engineering. In 70 percent of cases, the program correctly predicted which patients went on to have a major cardiovascular event, such as a heart attack, within five years of when the pho- tos were taken.

With AI getting more adept at screen- ing for a growing list of conditions, “some people might be concerned that this is machines taking over” health care, says Caroline Baumal, an ophthalmologist at Tufts Medical Center in Boston. But diag- nostic AI can’t replace the human touch. “Doctors will still need to be there to see patients and treat patients and talk to patients,” Baumal says. AI may just help people who need treatment get it faster. s

Copyright of Science News is the property of Society for Science & the Public and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder’s express written permission. However, users may print, download, or email articles for individual use.