WavelNet: A novel convolutional neural network architecture for arrhythmia classification from electrocardiograms

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automated electrocardiogram (ECG) analysis aids cardiovascular disease treatment.
• Current deep learning models for arrhythmia detection lack interpretability and reproducibility.
• Novel approaches are needed for accurate and transparent ECG interpretation.

Purpose of the Study:

• To propose an accurate, interpretable, and reproducible ECG arrhythmia classification model.
• To introduce WavelNet, a novel CNN architecture optimized for ECG analysis.
• To evaluate WavelNet's performance against existing methods.

Main Methods:

• Developed WavelNet, a CNN inspired by SincNet, utilizing wavelet transform-based spectral analysis.
• Trained WavelNet models on a five-class ECG arrhythmia dataset from MIT-BIH.
• Compared WavelNet with vanilla CNN and SincNet models, repeating evaluations for reproducibility.

Main Results:

• WavelNet models demonstrated superior performance in classifying ectopic beats due to adaptive spectral analysis.
• A Symlet 4 wavelet-based WavelNet achieved nearly 90% overall accuracy and high sensitivity.
• Results were reproducible and comparable to state-of-the-art arrhythmia detection models.

Conclusions:

• The WavelNet model provides remarkable, interpretable, and reproducible arrhythmia classification performance.
• Its clinical reasonableness and reproducibility support its integration into precision healthcare systems.
• WavelNet advances automated ECG analysis for cardiovascular disease management.