Abstract
Purpose :
Patients with Diabetic Retinopathy (DR) can reduce chances of vision loss via regular screening and timely treatment. However, the current standard of care is based on stratifying patients into a small number of risk groups. More accurate quantification of progression risk could allow improved care via personalized disease management. In this work, we developed and validated a deep learning (DL) algorithm using primary field color fundus photographs (CFPs) to predict 6, 12, and 24 month progression to DR.
Methods :
A DL algorithm was developed using a longitudinal dataset of 664,622 CFPs retrospectively collected across 367,146 visits from 574,860 eyes of 289,826 subjects with diabetes (mean age: 54 yrs, 60% women). The images were independently graded for DR (26% prevalence). The resultant algorithm was evaluated using an independent validation dataset with 166,661 CFPs across 91,942 visits from 144,003 eyes of 72,457 diabetic patients (mean age: 54 yrs, 61% women, 26% DR prevalence).
Results :
For predicting the progression to DR, the algorithm had an area under the receiver operating characteristic curve (AUC) of 0.73 (95% confidence interval (CI), 0.68-0.78), 0.71 (95% CI, 0.69-0.74), 0.72 (95% CI, 0.71-0.74) at 6, 12, and 24 months respectively (Figure 1). Kaplan-Meier plots also show stratification of low and high risk groups for progression to DR (Figure 2).
Conclusions :
In this study, our DL model predicted the future risk of progression to DR using primary field CFPs. Further research is necessary to determine the feasibility of applying this algorithm in the real world clinical setting for optimizing diagnosis and management of DR.
This is a 2020 ARVO Annual Meeting abstract.