Multi-Modal Deep Learning for Cardiovascular Disease Diagnosis and Risk Prediction Using Retinal Imaging

Cardiovascular diseases (CVDs) remain a major global health concern. Early identification of individuals at risk of CVDs is essential for effective preventive care, reducing healthcare costs, and improving patient outcomes. Retinal imaging has recently emerged as a noninvasive method of detecting microvascular alterations that might enable earlier identification and targeting of at-risk patients. This thesis leverages retinal imaging from the UK Biobank, combined with advanced deep learning techniques, to develop predictive models of CVDs risk and investigate causal links between cardiovascular outcomes and retinal features.

First, we investigated the use of optical coherence tomography (OCT), combined with minimal clinical data, to estimate CVDs risk by developing a convolutional variational autoencoder and a random forest framework. A novel explainability method was proposed to identify clinically interpretable retinal biomarkers. Second, we demonstrated the synergistic value of multimodal retinal imaging through a multi-channel variational autoencoder and a transformer-based classifier architecture that jointly analyses OCT and fundus photographs. An explainability model was applied to highlight the most relevant features from both retinal imaging modalities for classification tasks. Finally, through genome-wide association studies of nnU-Net-derived OCT phenotypes and Mendelian randomization analyses, we identified genetic variants and established causal relationships between cardiovascular traits and specific retinal layer alterations.

These findings collectively advance our understanding of retinal-cardiovascular interactions, providing, computational evidence that multimodal retinal imaging reveals biomarkers linked to systemic vascular health, methodological frameworks for multimodal ophthalmic data integration, and genetic evidence supporting causal links between cardiovascular traits and retinal alterations. The study bridges artificial intelligence and retina imaging data, offering novel insights into the estimation of CVD risk while highlighting the retina's potential as a window to cardiovascular health.