A Hybrid Adaptive Multiscale ECG/PCG Fusion Network with Bidirectional Cross-Modal Attention for Automated Cardiac Disease Detection on Wearable Devices

Cardiovascular diseases (CVDs) remain a leading cause of morbidity and mortality worldwide, underscoring the need for accurate and early diagnostic tools. Electrocardiogram (ECG) and phonocardiogram (PCG) signals are routinely acquired in clinical practice and provide complementary electrophysiological and mechanical information about cardia function. However, most existing automated diagnostic systems rely on a single modality, or employ fixed fusion strategies, limiting their robustness across diverse cardiac conditions. This paper proposes a Hybrid Adaptive Multiscale Cardiac Fusion Network (HAM-CFN) for multi-class cardiac disease classification by jointly modeling ECG and PCG signals. The network integrates multiscale temporal and time–frequency feature extraction, bidirectional cross-modal attention for physiological interaction, and an adaptive dynamic fusion module that assigns sample-specific modality weights. Experiments on the PhysioNet/CinC Challenge 2016 dataset demonstrate that the proposed method outperforms conventional single-modality and traditional fusion approaches with a lightweight architecture, achieving an accuracy of 97.60%, a precision of 97.66%, a specificity of 99.39%, a recall of 97.69%, and an F1-score of 97.67%. These results suggest that adaptive multimodal fusion of ECG and PCG signals can enhance automated cardiac diagnosis in resource constrained and realworld clinical settings