The artificial intelligence system, EyeArt, was able to detect diabetic retinopathy with a sensitivity of 95.5% and specificity of 86%.
Globally, the number of people with diabetes is growing, and outpacing the ability of health care providers to offer screening for diabetic retinopathy. To address this gap in care, investigators have worked to develop and test an artificial intelligence system to interpret images and determine whether they require a referral to an ophthalmologist for care.
Jennifer I. Lim, MD, Marion H. Schenk Esq. Chair, Professor of Ophthalmology, and Director of Retina Service at the University of Illinois at Chicago presented the study data at the 2019 Association for Research in Vision and Ophthalmology (ARVO) Imagine in the Eye Conference in Vancouver, BC.
Lim described the strong clinical trial results, stating that the EyeArt system achieved sensitivity of 95.5% and specificity of 86%.
In part 1 of the interview, Lim shared about the growing need for diabetic retinopathy screening as the population of patients with diabetes continues to grow worldwide. Part 2 covered how the EyeArt system was trained and then tested in a clinical trial.
Overall there were 942 patients that were enrolled into the study. Of these, a total of 1674 eyes had images that were gradable both by the EyeArt system and by the fundus photograph reading center of these 1674 eyes, 1364 eyes were negative for referable diabetic retinopathy [DR]. So, now we specifically looked and saw how many of these did the EyeArt system correctly identify, using the reading center as the gold standard. And of these, that was 1180 eyes that were correctly identified as not having referable DR there were 184 that were incorrectly identified as having referable DR. So, when we looked at those eyes, we found out that the vast majority did have diabetic retinopathy, but at a very mild level. So, didn't quite hit that level and the others had other eye disease that was not diabetic retinopathy. So, even though they were falsely positive, if you will, they still actually did have eye disease.
Okay so, let's look at the other set. So, of the 1674 eyes in the study, 310 eyes were identified by the reading center as being positive for referable diabetic retinopathy. So, looking at these 310 the EyeArt system correctly identified 296 of these eyes as having referable DR. There were 14, however, that it said did not have referable DR, so we looked closely at these eyes, and these were the false negative eyes, and they had very early stages of moderate and NPDR level 35 or lower. None of them had any higher levels.So, the sensitivity of the system was actually 95.5%—the EyeArt system correctly identified images as being referable DR if they truly had referable DR. And the specificity—that is correctly identifying them as not having referable DR when they did not have referable DR per the reading center—was 86%.
And the images that were readable without dilation were actually 87% of the time and for the eyes that then were later dilated in order for the EyeArt system to grade them—there were only a few eyes—the system functioned so well before that the sensitivity and specificity remained the same whether you did it dilated or not.