Enhancing ECG classification with continuous wavelet transform and multi-branch transformer

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Accurate electrocardiogram (ECG) signal classification is vital for diagnosing heart diseases.
• Existing methods often involve complex preprocessing and struggle with high-level time-series feature extraction.
• Convolutional Neural Networks (CNNs) have limitations in capturing intricate ECG signal patterns.

Purpose of the Study:

• To develop an advanced ECG classification method.
• To overcome limitations of traditional CNN-based approaches.
• To improve the accuracy and efficiency of automatic heart disease diagnosis.

Main Methods:

• Utilized continuous wavelet transform (CWT) to convert ECG signals into time-series feature maps, reducing preprocessing needs.
• Introduced a multi-branch transformer architecture for enhanced feature extraction.
• Focused on preserving important features while removing redundant information.

Main Results:

• Achieved 98.53% accuracy and 97.57% F1 score on the CPSC 2018 dataset (6877 cases).
• Attained 99.38% accuracy and 98.65% F1 score on the MIT-BIH dataset (47 cases).
• Outperformed most existing ECG classification methods in performance.

Conclusions:

• The proposed method accurately classifies ECG time-series feature maps.
• Demonstrates significant potential for diagnosing cardiac arrhythmias.
• Offers valuable advancements for automatic ECG diagnosis systems.