Abstract
Automatic arrhythmia detection using 12-lead electrocardiogram (ECG) signals plays a critical role in the early prevention and diagnosis of cardiovascular diseases. Traditional methods often treat 12-lead ECGs as a single matrix, capturing the integrity of the signal (cross-lead interactions) but neglecting the diversity of lead-specific features. This paper proposes 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) dataset, MLBF-Net achieved a state-of-the-art average $F_1$ score of 0.855, demonstrating a promising solution for multi-lead ECG analysis.
Introduction: The Challenge of Multi-Lead ECG
A standard clinical ECG records the heart's electrical activity from 12 different spatial angles. Each lead provides a unique anatomical perspective; for example, certain leads are more sensitive to atrial issues, while others are better for ventricular diagnostics.
Most existing deep learning models treat these 12 leads as a single multi-channel input. While this approach captures how leads interact (integrity), it often "washes out" the subtle, specific features of individual leads (diversity). The authors of MLBF-Net argue that to maximize information learning, a model must explicitly account for both lead-specific nuances and holistic relationships.
The MLBF-Net Architecture
The proposed MLBF-Net architecture is built on three main pillars designed to balance diversity and integrity:
- Lead-Specific Branches (BranchNet): The model employs 12 parallel branches, each dedicated to a single lead. Every branch consists of a 15-layer Convolutional Neural Subnetwork for spatial features, a Bidirectional GRU (BiGRU) for temporal dependencies, and an Attention Module to highlight informative segments like the P-wave or QRS complex.
- Cross-Lead Feature Fusion: The output feature maps from all 12 branches are concatenated and passed through a final attention layer, allowing the network to learn global relationships across the entire 12-lead signal.
- Multi-Loss Co-Optimization: This unique strategy assigns a dedicated loss function to each of the 12 individual branches, plus one for the final fused network. This forces each branch to become a specialist in its specific lead, ensuring that "diversity" information is preserved during training.
Despite its multi-branch structure, MLBF-Net is highly efficient, containing only 0.42 million parameters—significantly fewer than standard architectures like ResNet-34 (~10M parameters).
Experimental Results and Visualization
The model was rigorously tested on the CPSC 2018 dataset, which includes 6,877 public records across nine categories, including Normal rhythm and eight types of arrhythmias. MLBF-Net achieved an average $F_1$ score of 0.855, outperforming previous top-ranked models from the challenge.
Ablation studies highlighted the importance of the proposed components:
- The multi-branch approach improved the $F_1$ score by 3.7% over a standard single-branch framework.
- The multi-loss strategy provided an additional 2.4% boost compared to using a single global loss.
To confirm that the model successfully learned lead-specific features, the authors used t-SNE to visualize the feature spaces of individual branches. These visualizations show distinct clustering for different arrhythmias, even for traditionally difficult classes like ST-segment elevation (STE).
Conclusion
MLBF-Net introduces a robust framework for ECG analysis by treating the 12-lead signal as both a unified whole and a collection of distinct perspectives. By optimizing for both diversity and integrity through a multi-loss strategy, the model achieves superior classification performance while remaining computationally efficient. Future work will focus on addressing class imbalances, particularly for rarer arrhythmias like STE, and exploring the model's potential for real-time monitoring in clinical and wearable settings.
Summary updated