ECG-AuxNet: A Dual-Branch Spatial-Temporal Feature Fusion Framework With Auxiliary Learning for Enhanced Cardiac...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automated electrocardiogram (ECG) analysis faces limitations in feature integration, interpretability, generalization, and class imbalance.
• Current methods struggle to comprehensively capture both spatial and temporal information from ECG signals.

Purpose of the Study:

• To develop a novel dual-branch framework, ECG-AuxNet, for enhanced cardiac disease diagnosis.
• To address limitations in automated ECG analysis by integrating spatial-temporal features and improving model interpretability and generalizability.

Main Methods:

• ECG-AuxNet utilizes a Multi-scale Transformer Attention CNN for spatial features and a GRU network for temporal features.
• A Dual-stage Cross-Attention Fusion module integrates features, and a Feature Space Reconstruction (FSR) auxiliary task enhances feature discrimination.
• The framework was evaluated on the PTB-XL dataset and validated on the SXMU-2k clinical dataset.

Main Results:

• ECG-AuxNet achieved high F1-scores (78.34% on PTB-XL, 82.63% on SXMU-2k) for class-imbalanced disease recognition, outperforming 9 baseline models.
• The FSR auxiliary task significantly improved feature discrimination by 11.7%, enhancing classification accuracy.
• Grad-CAM analysis demonstrated that attention patterns align with cardiologists' diagnostic focus.

Conclusions:

• ECG-AuxNet effectively integrates spatial-temporal ECG features using auxiliary learning.
• The framework demonstrates robust generalizability and interpretability in cardiac disease diagnosis, aligning with clinical expertise.