AI alone now making the diagnosis

For the first time, a computer can diagnose patients without a doctor's interpretation. In April 2018, the Food and Drug Administration approved IDx-DR, a tool driven by artificial intelligence that detects diabetic retinopathy by studying images of the back of the eye.

Within one minute, the software produces a diagnostic result, detecting either the presence or absence of the condition, caused when high blood sugar damages blood vessels in the retina, which can lead to vision loss.

To use the tool, a provider captures images of a patient's retina and then uploads them to the software. After determining that the images are of sufficient quality, the software applies deep learning and other detection techniques to identify the presence or absence of diabetic retinopathy. If the software finds the condition, a provider would then refer the patient to an ophthalmologist.

Because IDx-DR works on its own, providers who don't specialize in eyes can rely on the so-called “autonomous AI” system, which could make a big difference in advancing early treatment of retinopathy, since about half of people with diabetes do not see an eye doctor annually.

Although the FDA approval process only took 85 days, the work was eight years in the making.

“The biggest challenge we had is that no one had ever approved a system that makes a diagnosis without a human presence,” said Dr. Michael Abramoff, president and founder of Coralville, Iowa-based IDx. Because the FDA designated IDx-DR as a “breakthrough device,” the agency provided feedback and guidance throughout clinical trials.

“It's really important that this happens safely,” Abramoff said. After the autonomous AI is proven to be safe, it needs to be reproduced safely, he said. “There can be an enormous pushback if we do this too rapidly and are poorly prepared.”

In the clinical study, which involved images from 900 patients at 10 primary-care sites, IDx-DR correctly detected diabetic retinopathy that was more than mild 87.4% of the time and correctly detected the lack of the condition 89.5% of the time.

Since receiving approval, IDx has been ramping up implementations, drawing on a recent $33 million round of financing, led by venture-capital firm 8VC, to speed up market adoption. The company is working with academic medical centers, retail clinics and primary-care clinics nationwide to bring the system online. The company operates on a pay-per-click model: Providers bill insurers for the service, get reimbursed, and then share part of the reimbursement with IDx.

Though company officials didn't offer other details about reimbursement, Abramoff said he and his team have recommended specific models to their clients.

The system is live at one site already, the University of Iowa Hospitals and Clinics in Iowa City.

“It's nice for the patient to have that screening, because a lot of patients will put it off,” said Bianca Carlson, a physician assistant at the University of Iowa Hospitals and Clinics who uses the tool. “Our ultimate goal is to make sure we get the patient screened and that we can catch it ahead of time.”

Among the challenges of bringing the platform online is integrating it with the electronic health record.

“No one has ever integrated a diagnostic system where there's no human involved,” Abramoff said. “We're ramping up slowly because we want to make sure we work out all the kinks with the EHR and the workflow.”

To connect the system to the EHR at the University of Iowa, the company worked with the hospital, relying on HL7 interfacing. The company can also use application programming interface integration.

Now, when a person with diabetes comes in and gets imaged, the images go directly into the IDx-DR system, which nearly immediately produces a result that gets put into the EHR.

Tools like IDx's will help physicians be more productive and efficient, said Susan Etlinger, an industry analyst with the Altimeter Group. “AI tools can help physicians handle a lot more data a lot more quickly and help them prioritize,” she said. “Theoretically, that could give a family physician a lot more tools in their toolbox to be able to run an initial diagnostic on somebody and then refer that person for additional treatment.”