A multi-branch vision transformer with GA-optimized focal loss for ECG signal classification

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Electrocardiogram (ECG) signal classification is vital for diagnosing cardiac arrhythmias.
• Challenges include data imbalance and capturing complex temporal patterns in ECG data.

Purpose of the Study:

• To propose a novel hybrid deep learning model, MBViT-FocalGA, for enhanced ECG signal classification.
• To address limitations in current arrhythmia detection methods.

Main Methods:

• Developed a Multi-branch Vision Transformer (MBViT) architecture to learn local and global ECG signal patterns.
• Utilized a Genetic Algorithm (GA) to optimize Focal Loss hyperparameters, mitigating class imbalance.
• Evaluated performance on the MIT-BIH Arrhythmia Database in two-class and five-class scenarios.

Main Results:

• The MBViT-FocalGA model achieved 98% overall accuracy and high F1 scores.
• Demonstrated superior performance compared to traditional deep learning models in both classification scenarios.
• Successfully improved the recognition of minority classes in imbalanced ECG datasets.

Conclusions:

• Transformer-based architectures show significant potential for ECG classification.
• Adaptive loss function optimization is effective in improving arrhythmia detection accuracy.
• The MBViT-FocalGA model offers a promising approach for clinical application in cardiac arrhythmia detection.