The American Diabetes Association estimates about 30 million Americans had diabetes in 2015, and a third of them have signs of diabetic retinopathy.1 Getting all them into ophthalmologists’ or optometrists’ offices for a diabetic retinal disease screening is impossible, let alone having certified readers review all those scans.

 

Michael Abramoff, MD, PhD, delivers results of the IDx-DR pivotal trial at the Macula Society meeting  in February.  

That drove University of Iowa retina specialist Michael Abramoff, MD, PhD, to develop an artificial intelligence (AI) system that moves screening into the primary-care practice and utilizes an algorithm to evaluate the scans for signs of diabetic retinopathy.

Dr. Abramoff is founder and president of IDx, a private company developing an autonomous AI platform for evaluating retinal images for diabetic retinopathy. The company recently took a couple of steps toward commercializing its technology. 

In February, the Iowa City-based venture submitted a de novo application to the Food and Drug Administration for the IDx-DR AI system. Two weeks later, Dr. Abramoff reported that a pivotal FDA trial of IDx-DR met its key endpoints for identifying DR in a diabetic population.2

Seed Planted In Residency

Dr. Abramoff, a professor at the University of Iowa Carver College of Medicine department of ophthalmology and visual science, also in Iowa City, tells Innovation Insight that the idea grew out of an observation he made as a resident evaluating retinas of patients with diabetes. “I realized what I was doing—looking at people who had nothing wrong with their retinas—maybe could be done more efficiently by a computer,” he says. It took a couple of decades for AI to emerge and make that possible.

The concept behind IDx-DR is to make early detection of diabetic eye disease more accessible and affordable for patients by pushing highly reproducible technology into primary-care offices. The IDx-DR system requires assistants to operate a robotic camera that captures retinal images with a high degree of reliability and reproducibility. 

Dr. Abramoff explains that if the AI algorithm determines any captured image is insufficient for analysis, the robotic camera instantaneously tells the operator, who can then capture new scans before the patient leaves the chair. The images are then sent to the AI database, which uses an algorithm to analyze them for signs of DR. 

The AI algorithm incorporates deep-learning-based lesion detectors to identify DR, including diabetic macular edema. The trial evaluated IDx-DR by comparing it to standardized imaging and grading protocols by the Wisconsin Fundus Photograph Reading Center (FPRC). FPRC grading includes Early Treatment Diabetic Retinopathy Study Severity Scale (ETDRS) and DME determinations from widefield stereoscopic photographs and macular optical coherence tomography.  

Pivotal Trial Findings

The FDA pivotal trial involved 10 different primary-care sites and prospectively enrolled 900 people with diabetes, 819 of whom could be fully evaluated by both AI and FRPC. The trial defined more-than-mild DR
(mtmDR) as ETDRS level 35 or higher and DME in at least one eye. System operators in the primary-care practices had four hours of training while FPRC-certified photographers conducted the FPRC imaging. 

Dr. Abramoff reported the results at the 41st annual Macula Society meeting in Beverly Hills, Calif.
According to the presentation, the AI system detected mtmDR at a sensitivity of 87.2 percent and specificity of 90.7 percent. In 96 percent of subjects, it obtained high-quality images. Overall, 23.8 percent of participants had mtmDR. 

Dr. Abramoff says AI-based systems have certain advantages over telemedicine, which he’s also investigated. “With AI, you don’t have the wait times. With AI, it takes minutes,” he says. “The patients get the results immediately, and they get the referral immediately.”

Another problem he points out with telemedicine is that the image quality is often insufficient, “but the patients have often left the primary-care office by the time a specialist reads it.” He adds, “Telemedicine never had to do clinical trials.”

The de novo FDA application means IDx-DR has no existing predicate in medicine. “There’s no AI system that is autonomous like ours,” Dr. Abramoff says. “There’s no supervision by a physician. A lot of these AI systems involve assisting specialists; this is different.”  RS

REFERENCES

1. Lee R, Wong TY, Sabanayagam C. Epidemiology of diabetic retinopathy, diabetic macular edema and related vision loss. Eye Vis (London). 2015;2:17. doi:10.1186/s40662-015-0026-2.

2. Abramoff M. Artificial intelligence for automated detection of diabetic retinopathy in primary care. Paper presented at: Macular Society; February 22, 2018; Beverly Hills, Calif. Available at: http://webeye.ophth.uiowa.edu/abramoff/MDA-MacSocAbst-2018-02-22.pdf