Purchase this article with an account.
E Simon Barriga, Jeremy Benson, Gilberto Zamora, Javier Lozano, Sheila C Nemeth, Peter Soliz; Real-World Use of Artificial Intelligence to Screen for Diabetic Retinopathy at Diabetes Care Clinics. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5462.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present clinical performance results of an automatic diabetic retinopathy (DR) tele-screening system operating for the past two years at a network of comprehensive diabetic care clinics.
VisionQuest has introduced EyeStar, an artificial intelligence (AI) DR screening system which is being used at thirteen comprehensive diabetes care clinics in Monterrey, Mexico. Since September 2016, a total of 12,427 patients with diabetes have been screened for the presence of DR using the EyeStar system. Retinal images are captured by a nurse or medical technician using Canon CR-2 or Volk’s Pictor Plus retinal cameras. Images are uploaded to a central server and are processed and returned by EyeStar to the clinics within 5 minutes. The EyeStar system returns an output of “non-refer” for cases with no DR, Mild non-proliferative DR (NPDR), or Moderate NPDR; or “refer to ophthalmology” for cases with Severe NPDR, proliferative DR (PDR), or suspect for clinically significant macular edema.To evaluate the performance of EyeStar, we sample 20% of the cases which are then over-read by a certified retinal grader. The performance of EyeStar is measured in terms of sensitivity, specificity, and predictive values.
N=11,951 patients (96%) were imaged with Canon cameras and N=476 (4%) with Volk cameras. Sensitivity of EyeStar for referable DR is 98% with a specificity of 77%. EyeStar’ negative predictive value (NPV) is 99.9% and its positive predictive value (PPV) is 21.2%. Workload reduction, defined as the percent of total cases that did not need referral to an ophthalmologist, is 77%.
The application of AI to DR screening has been a topic of extensive research, however, very few examples of real-world applications have been reported so far. The EyeStar system demonstrated clinically safe and effective performance for the detection of diabetic retinopathy in a clinical setting for the past two years. The software screens out 77% of patients who do not need a referral to an ophthalmologist while maintaining high levels of safety. Resource scarcity and limited access to care make automation a viable alternative to screen their diabetic population.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only