Hierarchical and Contrastive Learning for Multilateral Personalized Cardiovascular Detection Using Retinal and Cardiovascular Biomarkers

Cardiovascular diseases, such as heart attacks, kill millions of people each year. This is a major health problem around the world. Current ‎tools like the Framingham Risk Score fail in many cases. These tools do not work well for all types of people. This happens because individual differences are not considered. This paper presents a new model. It uses deep learning to predict heart attacks. The model uses data ‎from two sources: the eye and the heart. It combines features from retinal images and cardiovascular signals. Retinal images show blood ‎vessels in the eye. These vessels give clues about the body's microvascular health. Heart signals, including heart rate variability, reflect the ‎condition of the heart. The model uses several advanced parts. First, a graph transformer looks at blood vessel shapes in the retina. Then, ‎another transformer processes the heart signals over time. These features are fused using a special fusion technique. It finds the most useful ‎combined features from both sources. Then, a genetic algorithm selects the best features. It keeps the features that are most useful and easiest to ‎understand. Finally, a contrastive classifier sorts patients into risk groups. The model achieves very high accuracy. The model gives results ‎that doctors understand. It uses SHAP and Grad-CAM to show which features matter most. This helps doctors trust the model. The system ‎is easy to use in real clinics. It does not need blood tests or invasive tools. Just eye images and heart signals are needed. This model is powerful, fast, and easy to use. It helps doctors find heart risks early. It works well for many different people. It is a good tool for preventing ‎heart attacks. The approach suggests it has the potential to transform heart screening practices‎.