Artificial Intelligence Deep Learning System for Predicting Chronic Kidney Disease from Retinal Images

Abstract

Purpose : The retina is accessible to non-invasive imaging and a retinal image may provide information on systemic vascular and metabolic diseases. Chronic kidney disease (CKD) is a major health condition associated with significant morbidity, cardiovascular disease and mortality. Screening and diagnosis of CKD are limited in the community because of the need to obtain serum creatinine levels. To address this, we developed and validated a deep learning system (DLS) with convolutional neural network (CNN) using retinal images to detect CKD.

Methods : We developed the DLS using population-based data from 5717 adults who attended the baseline (2004-2011) and follow-up (2011-2017) visit of the Singapore Epidemiology of Eye Diseases study. CKD (prevalent CKD at both visits without overlap) was defined as an estimated glomerular filtration rate (eGFR)<60 mL/min/1.73m². Controls (no CKD) were defined as eGFR≥60+absence of albuminuria. Each image was associated with a diagnostic label of 0 or 1 indicating control or case. The input was 4 concatenated retinal images (1 OD and 1 Mac-centered image per eye per participant) and the output was two nodes classifying the CKD status. 80% of cases and controls were randomly selected for training and 20% for testing. Three separate models were trained: 1) image only 2) clinical data only (age, sex, ethnicity, smoking, diabetes, hypertension etc.) and 3) combination of 1 and 2. Model performances were evaluated using the area under the receiver operating characteristic curve analysis (AUC-ROC). External validation was performed in the Singapore Prospective Study Program (SP2, n=2904, aged ≥25 years).

Results : We used 22,868 retinal images (1362 cases, mean age=68.4y; 4355 controls, mean age=56.5y) to develop and evaluate the DLS. The AUC for images (0.9473) was similar to that of clinical data (0.9317) and improved further combining the two models (0.9549). Corresponding estimates in the SP2 cohort (n=185 cases; 2719 controls) were 0.7339 (image), 0.7554 (clinical) and 0.7621 (combined).

Conclusions : Our results show the feasibility of deep learning to screen for CKD using retinal images at the population-level, however, validation in multiple balanced datasets from diverse settings are needed to achieve more robust performance.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

AUCs in development

 

AUCs in validation cohort