Abstract
Automatic arrhythmia detection using 12-lead electrocardiogram (ECG) signals is vital for early cardiovascular disease prevention. Traditional deep learning methods often treat 12-lead ECGs as a single matrix, capturing "integrity" (inter-lead interaction) but neglecting "diversity" (lead-specific features). We propose the Multi-Lead-Branch Fusion Network (MLBF-Net), which integrates multi-loss optimization to jointly learn both diversity and integrity. Evaluated on the China Physiological Signal Challenge 2018 (CPSC 2018), MLBF-Net achieved an average $F_1$ score of 0.855, outperforming existing models with a significantly smaller parameter footprint.
Introduction
Standard 12-lead ECGs provide a comprehensive view of cardiac electrical activity from multiple anatomical perspectives. In automated classification, most existing deep learning models concatenate these leads into a single input matrix. While effective at capturing global patterns (integrity), this approach often overlooks the unique, localized information provided by individual leads (diversity). To maximize information extraction, a model must balance these two aspects.
The MLBF-Net Architecture
The proposed MLBF-Net architecture is designed to explicitly model lead-specific features while maintaining a global understanding of the ECG signal. It consists of three main components:
- Lead-Specific Branches (BranchNet): The 12-lead ECG is split into 12 individual signals, each processed by a dedicated "BranchNet." Each branch includes a CNN subnetwork for local morphological features, a Bidirectional GRU (BiGRU) for temporal dependencies, and an attention module to highlight critical signal components.
- Multi-Lead Feature Fusion: The outputs from all 12 branches are concatenated. A final attention module then processes this combined feature map to produce a global classification based on the "integrity" of the full lead set.
- Multi-Loss Co-Optimization: The network is trained using 13 separate loss functions: one for each individual branch ($Loss_1$ to $Loss_{12}$) and one for the fused network ($Loss_c$). This multi-loss strategy ensures that each branch learns discriminative features specific to its lead.
Experimental Evaluation
The model was rigorously tested on two major benchmarks:
- CPSC 2018: An open dataset containing 6,877 12-lead ECG records across 9 classes (Normal + 8 types of arrhythmia).
- PTB-XL: A large clinical dataset containing 21,837 records used to evaluate performance across diagnostic, sub-diagnostic, and rhythm categories.
Key Results
MLBF-Net achieved a state-of-the-art average $F_1$ score of 0.855 on the CPSC 2018 private test set, outperforming the previous challenge-winning model (0.837). Notably, the model is highly efficient, utilizing only 0.42 million parameters—a fraction of the size of standard 34-layer ResNet architectures (~10.47M). This efficiency makes it an ideal candidate for deployment in resource-constrained environments like wearable monitors.
Discussion and Conclusion
The multi-loss optimization proved crucial; the authors found that individual branch losses are essential for the branches to "specialize" in lead-specific features. While the model excels at identifying most arrhythmias, t-SNE visualization indicated some difficulty in distinguishing ST-segment depression (STD) from normal signals due to noise sensitivity. Future work will focus on improving robustness against real-world clinical noise and handling variable-length recordings.
MLBF-Net provides a promising, information-fusion-based solution for multi-lead ECG analysis, balancing high performance with computational efficiency.
Summary updated