Abstract
Purpose :
To develop and validate deep learning algorithms (DLAs) using retinal images for predicting DKD progression in Asian populations with diabetes.
Methods :
We utilized data from the Singapore Integrated Diabetic Retinopathy Program (SiDRP, 2010-2019), a primary care cohort with diabetic patients with estimated glomerular filtration rate (eGFR) <90mL/min/1.73m2 at baseline. DKD progression was defined as an eGFR decline ≥30% from baseline, or an annual eGFR decline ≥5mL/min/1.73m2 in 3 years. Controls were defined as: 1) total eGFR decline <10% from baseline, 2) annual eGFR decline ≤5mL/min/1.73m2, and 3) baseline eGFR≥60 mL/min/1.73m2. We developed three DLAs: 1) image-only model using two macular-centred images per participant (one image per eye); 2) Risk factor (RF) - only model including age, sex, ethnicity, duration of diabetes, HbA1c, and systolic blood pressure; 3) hybrid model combining both. Model performance was assessed through 5-fold cross-validation using metrics including area under the curve (AUC), sensitivity (Sn), and specificity (Sp) at the optimal threshold (where Sn = Sp). External validation utilized data from diabetic participants in the Singapore Epidemiology of Eye Diseases study, a population-based cohort study with 4-6 years of DKD progression.
Results :
13,964 retinal images from 731 cases and 6251 controls were used to train and validate the models. In internal validation, the performance of the image model was slightly lower than the RF and hybrid models (both p<0.05) with AUC (95% confidence interval) 0.712 (0.692-0.732), 0.764 (0.746-0.783) and 0.770 (0.752-0.787). The sensitivity (Sn) and specificity (Sp) for the image-only model were 67%, whereas the RF and hybrid models had 71% each. In external validation (136 cases: 911 controls), AUCs were 0.755 (0.712-0.799) for image-only, 0.791 (0.752-0.830) for RF-only, and 0.800 (0.762-0.838) for the hybrid with Sn/Sp of 68%, 71%, and 73%, respectively
Conclusions :
Our results demonstrate that, while the image-only model's performance is slightly lower, it remains comparable to that of the RF and hybrid models in both internal and external validation. Notably, the external validation results underscore the robustness of our image model, reaffirming its effectiveness in diverse settings.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.