Hierarchical fusion of electrocardiogram and phonocardiogram data improves heart failure detection

Heart failure with reduced ejection fraction (HFrEF) is a clinical syndrome with high morbidity and mortality. Developing cost-effective and easily deployable screening methods is crucial for improving early diagnosis and management. We propose LGC²-Net, a hierarchical fusion network that leverages multi-channel electrocardiograms (ECGs) and phonocardiograms (PCGs) for HFrEF detection. LGC²-Net simultaneously exploits complementarity across modalities and channels with its channel-specific and channel-shared branches. Each channel-specific branch employs a local-to-global hierarchical attention mechanism to capture both local and global semantic information within each ECG-PCG pair. The channel-shared branch further aligns and aggregates features from all channels, enabling effective modeling of inter-channel correlations. We established a new multi-channel ECG-PCG dataset with 2,480 synchronized recordings collected from 620 subjects using a digital stethoscope. Experiments demonstrate that LGC²-Net surpasses existing methods by 7.42% in average accuracy, highlighting its potential as an accurate, non-invasive, and scalable tool for HFrEF screening.