FDA approves first AI software that can identify disease, no specialists needed

April 11, 2018

Imaging of the retina is used to diagnose diabetic retinopathy. Adobe

ASHINGTON — The Food and Drug Administration has approved the first artificial intelligence software that can decide, without a clinician’s involvement, whether a patient might have a certain disease, the agency announced Wednesday.

The software, called IDx-DR, looks for diabetic retinopathy, an eye disease that afflicts individuals with diabetes. With minimal training, health care providers can use a special camera to take a picture of the back of the patient’s retina, which an algorithm then analyzes to look for the disease. If the software finds evidence of the disease, it recommends that a patient see an eye specialist.

A computer program that can analyze medical images could save time and money, cutting down on unnecessary, expensive trips to specialists. This particular software could be used by primary care physicians or other medical professionals who are not eye specialists.

“If you have an autonomous device like ours, and you can put it in a place where a diagnostic capability didn’t exist before, you’re increasing access, you’re reducing cost, you’re making things more productive because then only the patients who really need to see the ophthalmologist go there,” said Laura Shoemaker, director of marketing and communications for IDx, the company that sells the device.

The device is able to detect the disease once it has progressed past the “mild” stage, at which point the American Diabetes Association recommends an eye specialist take a look within three to six months. If the disease progresses untreated, it can lead to blindness, according to the American Optometric Association.

According to an FDA press release, the algorithm fails to detect about 13 percent of cases of diabetic retinopathy. Shoemaker said that this data is not yet published, but that it is under review at a “leading scientific journal.”

“There’s a margin of error even with human ophthalmologists,” Shoemaker said, citing a study that found an error rate of 20 to 30 percent for ophthalmologists who look at digital images, like this algorithm does.

Shoemaker said that the company had a threshold of success it had to meet in order for the software to be approved, which came out of “discussions with the FDA.” The algorithm needed to correctly identify at least 85 percent of patients who did have the disease, and needed to correctly identify at least 82.5 percent of patients who did not have the disease. This device passed the bar, with rates of 87.4 and 89.5 percent, respectively.

Shoemaker said she did not know if those thresholds would apply to other companies submitting artificial intelligence image recognition software to the FDA. The FDA did not immediately respond to that question.

The software is allowed to be used in the clinic today — but how much it costs is still up in the air.

“We’re currently in conversations about reimbursement with the important powers that be,” Shoemaker said. She said that IDx is talking with professional organizations such as the American Academy of Ophthalmology that influence coverage decisions by the Centers for Medicare and Medicaid Services, and that the cost should be less than the current standard of care.

Shoemaker said the plan is for IDx to charge per patient examination, the payment model the company is presently using in Europe, where the device is approved. Medical offices would also need to purchase a specific retina camera, which could increase the cost; IDx has an agreement with the camera manufacturer to distribute the two products bundled.