From ETDRS to EviRed: Eyeing a new artificial intelligence-based algorithm
The development of a novel AI-based algorithm aims to improve risk prediction for diabetic eye disease.
This article was reviewed by Ramin Tadayoni, MD, PhD.
The first diabetic retinopathy (DR) classification, the Airlie House system, was developed more than 50 years ago by an international, multidisciplinary consortium of experts and subsequently modified and improved by the Early Treatment Diabetic Retinopathy Study (ETDRS) Research Group.
Now, Intelligent Evaluation of Diabetic Retinopathy (EviRed), a national public-private consortium in France, aims to develop and validate a new and improved system by taking advantage of the information available using modern imaging technology and the power of artificial intelligence, said Ramin Tadayoni, MD, PhD, who is leading the EviRed project.
“The ETDRS classification is basically an estimation of the risk of progression to proliferative DR based on findings from stereoscopic 7-field fundus photographs. It was a wonderful system when it was first developed, which was at a time when nothing else was available besides fundus photography, but it provides insufficient prediction precision for modern patient care” said Tadayoni, professor of ophthalmology, University of Paris, Paris, France.
“EviRed aims to develop a new artificial intelligence-based algorithm that is able to integrate data from the medical record and different sources of imaging in order to generate a better prediction of the risk of progression to DR and its complications. The goals of this project are to help doctors with their decision making so that we can offer better care to our patients and be better at preventing blindness.”
Advantages of modern imaging
Modern retinal imaging tools provide much more information than ETDRS standard 7-field fundus images for determining the severity of DR. For example, lesions outside the 7-field imaging area can now be identified using ultrawide field imaging.
“This information will probably help us to improve risk prediction because we know that many DR lesions lie outside the 7-field imaging area. The risk of progression to proliferative DR may be different for two patients with ETDRS level 55 DR if one has neovessels outside of the 7-field area and the other patient doesn’t,” Tadayoni explained.
Findings from OCT angiography could also help to improve risk prediction. As described in the ETDRS Report #13, features found on fluorescein angiography, including leakage, capillary loss and dilatation, and various arteriolar abnormalities, were associated with retinopathy severity and risk of progression to DR. Today, OCT angiography provides a better way to characterize the vasculature in the diabetic retina.
“OCT angiography is a non-invasive test for imaging the 7-field area in just minutes, and it seems even superior to fluorescein angiography for identifying areas of nonperfusion and even retinal neovessels,” Tadayoni said (Figure 1).
(Figure 1) Montage from 2 15- ×15-mm OCTA images (PlexElite, Carl Zeiss Meditec) of the fundus of a diabetic patient. An example of the high level of information that can be acquired today in a short time during follow-up of a diabetic patient that needs an exceedingly long time of analysis unless helped and improved with algorithms under development by EviRed. (Image courtesy of Ramin Tadayoni, MD, PhD)
Use of OCT by itself might also be found to be able to predict the development of diabetic macular edema, which is not provided by the ETDRS classification. In addition, it is now known that various demographic and medical factors that are not included in the ETDRS classification, such as blood pressure and glycemic control, affect the risk of DR progression. Including the non-imaging variables could also help to improve the precision and accuracy of a prognostic algorithm.
Building and validating the algorithm
A clinical trial is now underway in France that aims to recruit a cohort of 5000 patients with diabetes. Participants will be followed for an average of 2 years and the collected data will be used to train and validate the EviRed prediction system.
Eligible patients are age 18 years and older with type 2 diabetes or a >10-year history of type 1 disease and who have no history of vitrectomy. Imaging will include ultra-widefield photography, OCT and OCT angiography, and other background and medical data. The primary outcome measure for the study is progression to complicated retinopathy.
The study group will be randomly split into a 4000-person training cohort and a 1000-person validation cohort. The main objective of the study is to validate the EviRed prognostic tool and evaluate its accuracy for predicting progression to severe retinopathy in the following year. Analyses will also be conducted to evaluate its accuracy for assessing DR severity and individual components of DR complications and to compare its performance with the predictions made by ophthalmologists using the ETDRS classification system.
