[Paper Review] A Study of Deep Feature Fusion based Methods for Classifying Multi-lead ECG

Bin Chen, Wei Guo+5 more|arXiv (Cornell University)|Aug 6, 2018

ECG Monitoring and Analysis14 references3 citations

TL;DR

This paper proposes MBCRNet, a multi-branch convolutional residual network that fuses features from multiple ECG leads using three distinct fusion strategies to improve classification of normal and abnormal ECG rhythms. Evaluated on the Chinese Cardiovascular Disease Database using 10-fold cross-validation, the method achieves 87.04% accuracy and 89.93% sensitivity, outperforming prior approaches by leveraging multi-lead synchronization and orthogonality.

ABSTRACT

An automatic classification method has been studied to effectively detect and recognize Electrocardiogram (ECG). Based on the synchronizing and orthogonal relationships of multiple leads, we propose a Multi-branch Convolution and Residual Network (MBCRNet) with three kinds of feature fusion methods for automatic detection of normal and abnormal ECG signals. Experiments are conducted on the Chinese Cardiovascular Disease Database (CCDD). Through 10-fold cross-validation, we achieve an average accuracy of 87.04% and a sensitivity of 89.93%, which outperforms previous methods under the same database. It is also shown that the multi-lead feature fusion network can improve the classification accuracy over the network only with the single lead features.

Motivation & Objective

• To improve automatic ECG classification by exploiting the synchronizing and orthogonal relationships between multiple ECG leads.
• To investigate the effectiveness of deep feature fusion strategies in enhancing classification performance over single-lead models.
• To develop a multi-branch convolutional residual network (MBCRNet) tailored for multi-lead ECG signal processing.
• To evaluate the proposed method on the Chinese Cardiovascular Disease Database (CCDD) under rigorous 10-fold cross-validation.
• To demonstrate that multi-lead feature fusion consistently improves classification accuracy compared to single-lead baselines.

Proposed method

• Design a multi-branch convolutional neural network (MBCRNet) where each branch processes a single ECG lead independently.
• Integrate residual learning blocks to stabilize training and improve gradient flow in deep architectures.
• Apply three distinct feature fusion strategies—early, late, and hybrid fusion—post-convolutional feature extraction.
• Use 1D convolutional layers followed by batch normalization and ReLU activation for temporal feature extraction from each lead.
• Fuse features at different stages: early fusion combines raw signals before convolution, late fusion combines final features, and hybrid fusion combines intermediate features.
• Train the model using cross-entropy loss with stochastic gradient descent and apply 10-fold cross-validation for robust evaluation.

Experimental results

Research questions

• RQ1 How does multi-lead feature fusion improve ECG classification accuracy compared to single-lead models?
• RQ2 Which feature fusion strategy—early, late, or hybrid—yields the best performance in classifying normal and abnormal ECG rhythms?
• RQ3 Can a multi-branch residual network effectively capture lead-specific and cross-lead dependencies in ECG signals?
• RQ4 Does the proposed MBCRNet achieve state-of-the-art performance on the Chinese Cardiovascular Disease Database (CCDD) for multi-lead ECG classification?
• RQ5 To what extent do the synchronizing and orthogonal relationships among ECG leads enhance classification performance when properly modeled?

Key findings

• The proposed MBCRNet achieved an average classification accuracy of 87.04% on the CCDD dataset using 10-fold cross-validation.
• The model demonstrated a sensitivity of 89.93%, indicating strong detection capability for abnormal ECG rhythms.
• Feature fusion strategies significantly improved performance over single-lead models, confirming the value of multi-lead information integration.
• The hybrid fusion strategy outperformed both early and late fusion methods in terms of accuracy and robustness.
• The residual learning components contributed to stable training and improved convergence, especially in deeper network configurations.
• The results show that MBCRNet outperforms previous methods on the same CCDD benchmark, establishing a new performance baseline.