Attention-Enhanced Multi-Modal Machine Learning for Cardiovascular Disease Diagnosis

Abstract

Cardiovascular diseases (CVDs) continue to be a major contributor to global mortality, emphasizing the pressing need for precise and early diagnostic methods. Machine learning presents promising opportunities; however, existing approaches still struggle with challenges such as multi-modal data integration, feature heterogeneity, and class imbalance. This study aims to build a scalable, interpretable, and high-performing machine learning framework for CVD classification by integrating clinical, demographic, and imaging information. The proposed approach utilizes hybrid feature fusion by combining early and late fusion strategies, incorporates a dynamic attention mechanism to emphasize relevant features, and applies SHAP-based interpretability for transparent reasoning. Its lightweight design and use of transfer learning enhance computational efficiency and adaptability to small datasets. Experiments on a multi-modal dataset achieved superior results with 94.8% accuracy, 92.3% sensitivity, and 96.1% specificity compared to baseline models. SHAP-based analysis further identified key feature contributions, enhancing model transparency. Overall, the framework provides a robust and efficient solution for CVD detection with potential for clinical implementation, though further testing on diverse datasets is advised to strengthen generalizability and clinical relevance.