Multimodal Deep Learning for Cardiovascular Risk Stratification: Integrating Retinal Biomarkers and Cardiovascular Signals for Enhanced Heart Attack Prediction

Cardiovascular diseases, particularly myocardial infarction, have remained a significant cause of death around the world. Therefore, dedicated non-invasive risk prediction frameworks are required. Conventional measures for risk prediction among most heterogeneous patients, such as the Framingham Risk Score and ASCVD calculator, are often less precise and poorly generalizable as they fail to capture individual variations in the mechanisms. This study seeks to present a new multimodal deep learning model developed for cardiovascular risk stratification by fusing retinal microvascular features with cardiovascular physiological signals. The pipeline includes a Hierarchical Retinal Vessel Graph Transformer (HRV-GT) for graph-based vascular representation, Spectro-Temporal Cardiovascular Transformer (STC-T) for capturing short- and long-term signal variability, Sparse Manifold Multimodal Fusion (SMM Fusion) for joint alignment of features, Evolutionary Feature Selection with Clinical Prior Constraints (EFS-CP), and a Contrastive Hierarchical Risk Classifier (CHRC-Net) for decision modeling structured. The model was evaluated across the UK Biobank and SEED datasets (n = 55,000), achieving an AUC of 0.97, sensitivity of 91%, and specificity of 89%, with more than 35% better performance than existing models in reducing misclassification. Clinically interpretable biomarkers that combine high computational efficiency (inference time $\approx \; 37$ ms) affirm its design for real-world deployment in preventive cardiovascular screening processes.