Multimodal Heart Disease Risk Prediction and Pathological Feature Modeling Using Graph Neural Networks

Background Heart disease remains a leading cause of global mortality, underscoring the urgent need for accurate and early risk prediction. Conventional approaches often rely on single-modality data and struggle with limited accuracy, poor generalization, and inadequate integration of heterogeneous medical information. Although machine learning has advanced diagnostic capabilities, existing methods have not fully exploited the relational modeling power of Graph Neural Networks (GNNs) in multimodal clinical settings. Purpose This study aims to develop a novel GNN-based framework for heart disease risk prediction and pathological feature modeling that effectively fuses clinical, imaging, and genomic data. The goal is to enhance predictive performance, model robustness, and clinical interpretability through structured multimodal integration. Methods We propose a graph-based multimodal fusion architecture that unifies diverse patient data into a single graph structure, where nodes represent patients or features and edges encode inter-modal relationships. A Graph Convolutional Network (GCN) is employed to learn node embeddings by aggregating neighborhood information across layers. The model is trained on a self-constructed dataset of 2,000 patients with standardized clinical, cardiac imaging (ECG/MRI), and genomic features. An explainability module based on SHAP values is integrated to provide transparent, feature-level contributions to predictions. Results Our method achieves an accuracy of 91.4% (±1.2), outperforming Support Vector Machines (83.5%), Random Forest (86.2%), and Convolutional Neural Networks (88.1%), notably improving accuracy by 3.3 percentage points over CNNs. It also attains superior recall (89.5%), F1-score (89.9%), and AUC (0.93), with statistically significant gains (p < 0.01). Ablation studies confirm that both the multimodal fusion module and the GNN component are essential, contributing 4.6% and 3.9% accuracy drops when removed, respectively. The model demonstrates strong robustness under noise and multi-task conditions, though performance declines when key modalities (e.g., genomic data) are missing. Conclusions By leveraging graph-structured representation and GCN-based learning, our approach enables more accurate, robust, and interpretable heart disease risk prediction from multimodal data. This work advances the methodology for integrating heterogeneous medical data and provides a promising tool for early screening and personalized cardiovascular care. Future efforts will focus on handling missing data and deepening clinical interpretability through domain-informed attention mechanisms.