Deep Learning Retinal Image Analysis for the Detection of CKD and Cardiovascular Risk Factors in the General Population

Abstract

Background: Retinal blood vessel patterns provide an opportunity to personalize an individuals risk assessment for CKD and cardiovascular risk factors (CVRF). In this study we propose a deep learning (DL) based prediction tool that uses retinal images from the Irish Longitudinal Study on Ageing (TILDA) to detect the existence of CKD in community dwelling individuals aged 50 years and over. Methods: TILDA is a stratified random sample of the general population of Ireland. N=4569 participants underwent a detailed health assessment including retinal photography. We developed a convolutional neural network architecture inputting a single retinal image per participant for the prediction of CKD & CVRF. Binary cross entropy was used as a loss function. Analyses were conducted on the FRAILMatics HPC “Tinney”. Results: See Table 1 & Image 1 for results. Conclusion: A DL retinal image algorithm has good discrimination for CKD, eGFR and CVRF in community dwelling individuals. The prediction emphasis of our DL algorithm focuses on slightly different structures within the retinal image to predict serum creatinine versus serum cystatin.