Prediction of future cardiovascular diseases from retinal images using a deep-learning-based hybrid model

Abstract

Purpose : Deep learning models using retinal photos can predict coronary artery calcium score and stratify the risk of cardiovascular diseases (CVD). However, whether future CVD events can be foretold from ocular images is unknown. Therefore, we aim to develop a deep-learning-based hybrid system that directly predicts future CVD events using retinal photos without clinical information.

Methods : A total of 51,444 participants who had taken retinal photos and had no previous history of CVD in the UK Biobank dataset were included in this study. In the follow-up period of 10 to 14 years, 1,644 participants were diagnosed with CVD. There were 3,882 unreadable photos removed from training and validation datasets (1,138 with CVD; 2,744 without CVD). We split the remaining dataset into training and validation sets (70% and 30%, respectively). The unreadable images were further used as blank references. A convolutional neural network (CNN) that extracts feature vectors from the retinal images was trained in a multi-task scheme to predict age, sex, blood pressure, pulse pressure, BMI, diabetes, renal diseases, alcohol, and smoking habits. Finally, the hybrid model applies the Linear Discriminant Analysis (LDA) method to the extracted CNN vectors to find the maximal difference between the samples with future CVD events and those without.

Results : Our hybrid model achieved mean sensitivity of 99.8% (95% confidence interval, CI, range from 99.7% to 99.9%), mean specificity of 85.7% (95%CI range from 80.6% to 90.6%), and mean area under the curve (AUC) of 0.927 (95%CI range from 0.908 to 0.944) predicting the CVD events on the validation dataset. In comparison, the model performed poorly on unreadable images with a mean sensitivity of 12.7% (95% CI range from 10.0% to 95.6% range) and a mean specificity of 87.2% (95% CI range from 84.4% to 90.0%).

Conclusions : Our algorithm can accurately predict CVD events from retinal images. The people identified with future CVD can be referred to cardiologists and treated early.

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