Purpose: The prediction of atherosclerosis using retinal fundus images and deep learning has not been shown possible. The purpose of this study was to develop a deep learning model which predicted atherosclerosis by using retinal fundus images and to verify its clinical implications by conducting a retrospective cohort analysis.
Design: Retrospective cohort study.
Methods: The database at the Health Promotion Center of Seoul National University Hospital (HPC-SNUH) was used. The deep learning model was trained using 15,408 images to predict carotid artery atherosclerosis, which was named the deep-learning funduscopic atherosclerosis score (DL-FAS). A retrospective cohort was constructed of participants 30-80 years old who had completed elective health examinations at HPC-SNUH. Using DL-FAS as the main exposure, participants were followed for the primary outcome of death due to CVD until Dec. 31, 2017.
Results: For predicting carotid artery atherosclerosis among subjects, the model achieved an area under receiver operating curve (AUROC) and area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, positive and negative predictive values of 0.713, 0.569, 0.583, 0.891, 0.404, 0.465, and 0.865 respectively. The cohort consisted of 32,227 participants, 78 cardiovascular disease (CVD) deaths, and 7.6-year median follow-up visits. Those with DL-FAS greater than 0.66 had an increased risk of CVD deaths compared to those with DL-FAS <0.33 (hazard ratio: 8.33; 95% confidence interval [CI], 3.16-24.7). Risk association was significant among intermediate and high Framingham risk score (FRS) subgroups. The DL-FAS improved the concordance by 0.0266 (95% CI, 0.0043-0.0489) over the FRS-only model. The relative integrated discrimination index was 20.45% and net reclassification index was 29.5%.
Conclusions: A deep learning model was developed which could predict atherosclerosis from retinal fundus images. The resulting DL-FAS was an independent predictor of CVD deaths when adjusted for FRS and added predictive value over FRS.
Copyright © 2020 Elsevier Inc. All rights reserved.