AI Speeds Diabetic Retinopathy Diagnosis Without Specialist

Marcia Frellick

August 28, 2018

An artificial intelligence (AI) system safely and effectively detects diabetic retinopathy (DR), a major cause of blindness, without the need for a physician to interpret the image or results, according to an article published online August 28 in npj Digital Medicine.

The trial results will change practice, one expert not involved with the study told Medscape Medical News. He predicts the system will be available in places such as major drugstore chains within 3 years.

Results from the prospective study evaluating the platform, called IDx-DR (IDx LLC), were convincing enough for the US Food and Drug Administration (FDA) to fast-track its approval in April.

It was "the first FDA authorized autonomous AI diagnostic system in any field of medicine with the potential to help prevent vision loss in thousands of people with diabetes annually," write lead author Michael D. Abràmoff, MD, PhD, founder and president of IDx. Abràmoff is also a professor of ophthalmology and visual sciences with University of Iowa Health Care, Iowa City.

Target Sensitivity and Specificity Endpoints Met

The results show that when used to detect DR in a primary care setting, the AI system exceeded primary endpoints with a sensitivity of 87.2% (95% confidence interval [CI], 81.8 - 91.2%; targeted endpoint, > 85%), a specificity of 90.7% (95% CI, 88.3 - 92.7%; targeted endpoint, > 82.5%), and an imageability rate of 96.1% (95% CI, 94.6 - 97.3%; no prespecified target reported).

High sensitivity is key because the AI system's main role is to identify people who likely have DR and need to see an eye care provider. The AI system is not intended to replace eye exams and does not suggest treatment. Rather, it flags people who need to see an eye specialist right away.

By comparison, the authors note that three previous studies showed board-certified ophthalmologists who perform indirect ophthalmoscopy have an average sensitivity of 33% to 73% compared with the same Early Treatment Diabetic Retinopathy Study Severity Scale standard.

In total, 900 patients with no history of DR were enrolled at 10 primary care sites across the United States. Participants were an average age of 59 years old, 47.5% were men, 16.1% were Hispanic, 28.6% were African American, 92.9% had type 2 diabetes, and 7.1% had type 1 diabetes. Ultimately, 23.8% of participants were found to have more-than-mild DR (mtmDR).

In a press release, the authors write that IDx-DR will enable millions of people in the United States to be diagnosed with DR in primary care and retail clinics.

Increased access is essential as fewer than half of people with diabetes meet the recommended eye exam schedule, despite intensive efforts to change those numbers.

"The Buck Stops With the Computer"

"It's very exciting and it's definitely paradigm shifting," T.Y. Alvin Liu, MD, retina specialist and assistant professor of ophthalmology at Wilmer Eye Institute at Johns Hopkins Medicine in Baltimore, Maryland, told Medscape Medical News. "The FDA thinks the results are so impressive, so reliable that whatever the machine says, it does not need a human for verification. The buck stops with the computer."

"It changes the way we think about AI, what it means to be a doctor, and what the future holds," he added.

The sensitivity and specificity are impressive, Liu said. "This machine is at least as good as humans if not better at picking up referable diabetic retinopathy."

Because the machine can be operated by someone who isn't a physician and it only takes nonmedical personnel a few hours to learn to take a picture, IDx-DR opens up access to screening for patients at risk, a huge plus for public health, he said.

Liu said he would expect to see the screening offered at drugstore chains such as Walgreens and CVS in addition to primary care offices within 3 years.

Though the system only detects DR, IDx is developing more AI diagnostic algorithms for detecting macular degeneration, glaucoma, Alzheimer's disease, cardiovascular disease, and stroke risk, according to the press release.

Study Details

AI system operators had no prior ocular imaging experience and were provided with 4 hours of standardized training. 

Participants were first imaged using a nonmydriatic retinal camera. Three attempts were made to obtain a usable image without dilation followed by three more attempts after dilation. Once the image was obtained for the AI system, or after six attempts, eye drops were administered for further assessment according to the gold standard Wisconsin Fundus Photograph Reading Center (FPRC) protocol.

Among the 900 patients screened by the system, 40 had images that were ungradable by the FPRC reading center and another 33 had insufficient images for AI, leaving 819 that could be fully evaluated.

Of those, 621 patients were DR-negative by FPRC; the AI system reported that mtmDR was present in 65.

Among the 198 participants diagnosed with mtmDR by the FPRC, the AI system correctly diagnosed 173. False negatives were reported by the AI system in 25.

The authors report that 29 participants had clinically significant diabetic macular edema according to fundus photography, 19 had center-involved diabetic macular edema by optical coherence tomography, and 42 had clinically significant and/or center-involved diabetic macular edema.

Abràmoff is director and a shareholder of IDx LLC and has relevant patents and patent applications assigned to the University of Iowa. The coauthors received fees from IDx LLC for statistical consultancy and one reported holding shares in IDx LLC. Liu has reported no relevant financial relationships.

npj Digital Medicine. Published online August 28, 2018. Full text

For more news, join us on Facebook and Twitter


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.