EW-GAT: Edge-Weighted Graph Attention Framework for Patient-Aware Heart Disease Classification Using Feature-to-Feature Similarity Graphs

Abstract

Heart disease remains an important cause of global death, requiring accurate and clinically interpretable prediction models. However, traditional Machine Learning (ML) methods often fail to capture meaningful dependencies among clinical features and lack robustness across diverse patient populations. To overcome these issues, this research introduces an Edge-Weighted Graph Attention Network (EW-GAT) for patient-aware heart disease classification. The approach first models feature-to-feature correlations to identify clinical dependencies, which are then used as a structural prior to improve patient-to-patient similarity during graph construction. An adaptive edge-weighting mechanism further enhances graph representation by integrating patient-specific variations. These weights are incorporated into the attention mechanism to emphasize clinically relevant relationships during message passing. Finally, a Multi-Layer Perceptron (MLP) layer predicts disease outcomes from refined patient embeddings. From the results, the proposed EW-GAT model attains higher performance, achieving 99.92% accuracy on the Cardiovascular dataset when compared to the existing Denoising Autoencoder (DAE) model. The improvements are supported by strict cross-validation and statistically significant analysis, confirming the model's robustness and generalization capabilities.