Development of a deep learning algorithm to detect early-stage chronic kidney disease from retinal images

Abstract

Purpose : Chronic kidney disease (CKD) is a major public health problem associated with cardiovascular disease, renal failure, and deaths. We previously developed a retinal image-based deep learning algorithm (DLA) to detect moderate/above CKD (stage 3-5). Detection of CKD in earlier stages is important for preventing progression to advanced stages. Therefore, we aimed to develop a DLA for detecting early-CKD (stage 1-2) from retinal images.

Methods : In this case-control study, we utilized retrospective data collected from adults aged ≥40 years who attended the baseline (2004-2011) or the 6- or 12-year follow-up (2011-2020) visits of the population-based Singapore Epidemiology of Eye Diseases study. Early-CKD (prevalent CKD at any visit) was defined as an estimated glomerular filtration rate (eGFR) ≥60 mL/min/1.73m² and presence of albuminuria (urinary albumin-to-creatinine ratio [UACR] ≥30 mg/g). Controls (no CKD in at least 2 of the 3 visits) were defined as eGFR ≥60 mL/min/1.72m² and absence of albuminuria (UACR <30 mg/g). We developed the image-based DLA (Image-only) using two macular-centred images (1 image per eye for each participant) and compared its performance to a risk factor model (RF-only) including age, sex, ethnicity, diabetes, and hypertension, and a hybrid model combining the two models. Models were evaluated using 5-fold cross validation by metrics including the area under the receiver operating characteristic curve (AUC), sensitivity (Sn) and specificity (Sp).

Results : 9,556 retinal images from 2265 cases and 2513 controls were used to train and validate the models. In internal validation, the image-only model and the hybrid model performed better than the RF model (Figure) with AUC (95% CI) of 0.839 for image-only, 0.792 (0.779-0.804) for RF and 0.866 (0.856-0.877) for the hybrid. The Sn and Sp at optimum threshold (defined as Sn = Sp) for the 3 models were: 75% for image-only, 71% for RF-only, and 78% for the Hybrid.

Conclusions : Our results show that a retinal image-based DLA can detect early-CKD with good accuracy and combining image and traditional risk factors could enhance the performance additionally. Further validation in external datasets from diverse settings are warranted to evaluate the generalisability and scalability of this non-invasive screening algorithm.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.