Predicting progression to referable diabetic retinopathy from retinal images and screening data using deep learning

Abstract

Purpose : Prior studies report impressive diabetic retinopathy (DR) detection performance using deep learning systems (DLS). However, the utility of DLS for predicting DR and maculopathy progression is unknown. Predicting DR progression could enable individualised, risk-based follow-up or interventions via the early identification of high and low-risk individuals. We aimed to develop multimodal DLS to predict progression to referable DR and maculopathy over 1, 2 and 3 years using curated, two-field retinal images and screening data.

Methods : From 202 928 eyes of 102 446 patients attending the Southeast London Diabetic Eye Screening Programme (Sept 2013 to Dec 2019), 124 418, 89 360 and 71 125 eyes with 1, 2 or 3-year follow-up data were eligible. DLS outcomes were incident (1) referable DR, (2) referable maculopathy or (3) either during follow-up. Detection of any DR at baseline was an auxiliary training task. Inputs were macula/nasal images and screening data (age, gender, ethnicity, diabetes type, diabetes duration, visual acuity and deprivation rank). Data from 68 980 eyes without eligible follow-up were used for DLS pretraining. Image and screening data DLS were developed independently and ensembled for test predictions.

Results : Screening data DLS area-under-the receiver operating characteristic curve (AUROC) for referable DR, maculopathy or either at 2 years were 0.760 (0.684-0.836), 0.733 (0.705-0.761) and 0.733 (0.705-0.760). Image DLS AUROC for referable DR, maculopathy or either at 2 years were 0.898 (0.846-0.951), 0.820 (0.793-0.847) and 0.824 (0.798-0.849). Multimodal DLS AUROC for referable DR, maculopathy or either at 2 years were 0.916 (0.873-0.959), 0.842 (0.819-0.866) and 0.845 (0.823-0.867). Generally, DLS had lower AUROC for all outcomes if initialised without pretraining. Compared to 2-year outcomes, DLS AUROC was higher for 1-year but lower for 3-year intervals.

Conclusions : DLS accurately predicts progression to referable DR and maculopathy using retinal images with modest improvements with additional screening data. Developed DLS for predicting DR or maculopathy progression could enable individualised, risk-based follow-up or interventions with significant time/cost savings for DR screening services/low-risk patients and more timely referrals for high-risk patients.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Data flowchart and methods

 

Predicting progression to referable diabetic retinopathy and maculopathy