Eyes on the Heart: Detecting Cardiovascular Disease Using Retinal Images with Explainability

Cardiovascular diseases (CVDs) are a leading cause of death worldwide, creating a critical need for accurate and accessible early risk detection methods. Recent studies have demonstrated that retinal biomarkers—such as the optic disc, optic cup, and blood vessels—can be used to predict cardiovascular risk through painless and accessible retinal imaging, with deep learning models like ResNet, InceptionV3, and GANs achieving successful results. Building on these insights, this study introduces a multi-input convolutional neural network (CNN) that processes four types of inputs: the original retinal image and segmented maps of the optic cup, optic disc, and blood vessels. Each of these inputs is passed through an individual pre-trained InceptionV3 backbone with shared configuration. The extracted features from each branch are then concatenated into a unified feature vector to capture both textural and anatomical patterns. The model, trained and evaluated on the EyePACS dataset, achieved a test accuracy of 94%, with precision, recall, and F1-score all at 94%, and an AUC of 96%. K-fold cross-validation was used to ensure robust performance across data splits. Grad-CAM was applied to visualize the most influential retinal regions contributing to each prediction, reinforcing the model's interpretability and potential for use in real-world cardiovascular risk screening.