Effective Data Augmentation, Filters, and Automation Techniques for Automatic 12-Lead ECG Classification Using Deep...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automatic electrocardiogram (ECG) analysis is vital for diagnosing cardiac conditions.
• Enhancing machine and deep learning model robustness is crucial for accurate cardiac abnormality classification.
• Existing methods require improvements in data handling and automated analysis.

Purpose of the Study:

• To propose novel data augmentation, filtering, and automation techniques for detecting cardiac abnormalities from 12-lead ECGs.
• To evaluate the effectiveness of these techniques using a deep residual neural network (ResNet) model.
• To improve the classification performance of deep learning models for ECG analysis.

Main Methods:

• Developed 15 data augmentation techniques and 6 filters.
• Implemented an end-to-end deep residual neural network (ResNet) model.
• Evaluated performance on the China Physiological Signal Challenge (CPSC) dataset with 9 diagnostic classes.
• Utilized a modified RandAugment technique for random combinations of augmentation and filters.

Main Results:

• Data augmentation (wander addition, dropout, scaling) and sigmoid compression filter significantly improved average F1 scores compared to baseline.
• Sigmoid compression yielded a 2.04% relative improvement in average F1 score.
• Random combinations of selected augmentations/filters via modified RandAugment achieved a 2.54% relative improvement.
• Horizontal and vertical flipping augmentations negatively impacted performance.

Conclusions:

• Proposed data augmentation, filters, and automation techniques effectively enhance ECG classification performance.
• These methods improve deep learning model accuracy without altering model architecture or hyperparameters.
• The developed techniques offer a valuable solution for automated cardiac abnormality detection from ECG recordings.