A Survey of Multimodal Deep Learning Frameworks for Cardiovascular Disease Prediction and Personalized Risk Stratification

Cardiovascular diseases (CVDs) are the leading cause of mortality worldwide, demanding early diagnosis, accurate classification, and personalized risk stratification. Despite progress in artificial intelligence (AI), existing models remain limited by unimodal data usage and lack of clinical interpretability. This survey presents a comprehensive review of recent advancements in multi-objective, multimodal deep learning frameworks tailored for CVD prediction using heterogeneous data types such as electrocardiograms (ECG), electronic health records (EHR), and multimodal imaging (CT, ultrasound). We explore three research fronts: temporal fusion networks for early CVD detection, graph neural networks (GNNs) with neighborhood optimization for structured data classification, and transformer-CNN fusion architectures for imaging-based risk stratification. We high-light the integration challenges across modalities and emphasize the role of explainable AI (XAI) using SHAP and Grad-CAM. Additionally, we present mathematical formulations for graph-based learning, attention mechanisms, and multimodal fusion. The objective of this review is to synthesize current methods, identify key research gaps, and outline future directions towards a unified, interpretable, and clinically deployable AI system for cardiovascular healthcare.