Deep learning algorithm to predict diabetic retinopathy (DR) progression on the individual patient level

Abstract

Purpose : Diabetic eye disease is a major cause of vision loss and its global burden is expected to continue to worsen over the next few decades. The insidious nature of DR progression can leave the disease undetected until it threatens vision, therefore early identification of patients at risk is important. The steps on the Early Treatment DR Study (ETDRS) DR severity scale (DRSS) indicate the risk for progression in a group of patients with similar baseline severity, but the tools to effectively triage patients into fast vs slow DR progressors at the individual level are limited. As part of Roche’s comprehensive initiative in Ophthalmology Personalized Healthcare, we developed an automated deep-learning (DL) algorithm to predict individuals that are likely to experience significant DR progression over the next 2 years.

Methods : Stereoscopic 7-field color fundus photographs and ETDRS DRSS scores from sham-treated study eyes and fellow eyes (n=683, n=682, n=645 eyes at months [M] 6, 12, and 24 respectively) of DR patients with macular edema from RIDE/RISE (NCT00473382/NCT00473330), were used for training and validation of DL algorithms with the goal to predict ≥2-step DRSS worsening at M6, M12, and M24 in individual patients. The area under the curve (AUC) was calculated and 5 times repeated 5-folds cross validation was performed to compute statistics of DL model performance.

Results : At baseline, DRSS in sham-treated study eyes and fellow eyes ranged from 10 (absent) to 71 (high-risk proliferative DR). The manually detected rates of ≥2-step worsening in sham study and fellow eyes at M24 were 9.6% and 11.7%, respectively. The best DL algorithm was able to predict ≥2 step ETDRS DRSS worsening at M6, M12, and M24 at an AUC of 0.68 ± 0.13 (sensitivity 66% ± 23% and specificity 77% ± 12%), 0.79 ± 0.05 (sensitivity 91% ± 8% and specificity 65% ± 12%) and 0.77 ± 0.04 (sensitivity 79% ± 12% and specificity 72% ± 14%), respectively.

Conclusions : Our pilot work established the feasibility of developing an automated algorithm for predicting patients with significant DR worsening over a 2-year period. Our data generated on eyes with a broad range of DR severities suggest a possible presence of predictive signals that precede microvascular abnormalities detectable to human graders. Validation in a real world setting is needed to make the algorithm generalizable to the overall population with diabetes.

This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.