Skip to main content

  • Reliable AI System for DR Screening

    By Lynda Seminara
    Selected by Richard K. Parrish II, MD
    Add to My Bookmarks

    Journal Highlights

    American Journal of Ophthalmology, December 2022

    Download PDF

    Mehra et al. looked at real-world out­comes after incorporating an artificial intelligence (AI)–based system into an existing primary care telemedicine diabetic retinopathy (DR) screening program. In their single-institution review, conducted at the Mayo Clinic, image gradeability was excellent, and there were no false-negative results.

    For this study, the authors reviewed medical records of 1,052 adult patients who were screened for DR during an 18-month period. They gathered non­mydriatic fundus photographs and had them analyzed by the IDx-DR AI-based system, which was designed to detect greater-than-mild DR via analysis of two fundus photographs from each eye using a nonmydriatic fundus camera, in accordance with the Mayo Clinic DR screening program. For any non­mydriatic image that was not gradable, reflex dilation (1% tropicamide) and mydriatic photography were conducted before repeat AI-based analysis. Patient factors that interfered with image gradeability were analyzed, and all images underwent manual overread per the telemedicine screening protocol.

    Altogether, 965 (91.7%) of the 1,052 patients had AI-gradable fundus photographs, and 580 (55.1%) had AI-gradable nonmydriatic images. Approximately 93% of patients with ungradable nonmydriatic photographs (440 of 472) had reflex dilation. Of the 965 with AI-gradable images, 138 (14.3%) were classified as positive (greater than mild disease) and 827 (85.7%) as negative. Compared with manual overread assessment of greater-than-mild nonproliferative DR, which required a comprehensive dilated eye exam, the sensitivity was 100%, spec­ificity was 89.2%, positive predictive value was 27.5%, and negative predic­tive value was 100%. Few demographic variables affected the gradeability of images. The percentage of gradable images was higher for patients under 70 years of age (93.5% [61.9% nonmydri­atic]) than for older patients (85.3% [31.0% nonmydriatic]) (p < .001).

    The IDx-DR is the first AI-based system approved by the FDA for routine DR screening. Although the nonmydriatic images of older patients have a lower rate of gradeability, the authors believe that select patients of any age may benefit from having their images screened. 

    The original article can be found here.