RetinoCardioNet: Multi-Modal Deep Learning Framework for Cardiovascular Risk Assessment Using Retinal Fundus Imaging

Abstract

Retinal fundus imaging offers a non-invasive window into microvascular health, with growing evidence linking retinal abnormalities to systemic cardiovascular conditions. However, most computational models in this domain rely on isolated image features and fail to incorporate vascular geometry or clinical metadata. This study introduces RetinoCardioNet, a unified multi-modal deep learning framework designed for cardiovascular risk prediction using retinal fundus images, vascular graphs, and structured clinical data. . The proposed system integrates three data modalities: high-resolution retinal images processed through a ResNet-50 encoder with self-supervised SimCLR pretraining, graph neural networks (GCNs) encoding vessel topology, and a clinical metadata encoder. These features are fused via a multiread cross-attention mechanism. The framework was trained on public datasets (EyePACS, Messidor, UK Biobank) and evaluated using a 5 -fold cross-validation protocol. Model optimization used Adam with a learning rate of , cosine annealing, and early stopping based on validation AUC. RetinoCardioNet achieved an AUC of 0.942 , F1-score of 0.916 , precision of 0.906 , and recall of 0.927. Ablation studies showed performance dropped by up to  when removing key components, confirming the contribution of each modality. Visual attention maps further improved interpretability. Conclusion: RetinoCardioNet offers a clinically relevant, interpretable, and scalable framework for noninvasive cardiovascular risk screening, showing potential for deployment in preventive cardiology, especially in resource-limited settings.