Improving Deep Learning-based Cardiac Abnormality Detection in 12-Lead ECG with Data Augmentation.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Automated electrocardiogram (ECG) classification relies on deep learning models.
• Training these models necessitates large, professionally annotated datasets, which are costly and time-consuming to acquire.
• Data augmentation offers a cost-effective solution for expanding existing ECG datasets.

Purpose of the Study:

• To investigate the efficacy of ECG augmentation techniques for improving deep learning-based automated ECG classification.
• To introduce and evaluate three novel ECG augmentation methods: Limb Electrode Move, Chest Electrode Move, and Heart Vector Transform.
• To assess the performance gains achieved by combining novel and existing time series augmentation methods.

Main Methods:

• Development of three novel ECG augmentation techniques simulating electrode misplacement and altered heart axis.
• Integration of these novel methods with nine established time series augmentation techniques.
• Evaluation of augmented datasets using deep learning models on four benchmark ECG datasets (ICBEB, PTB-XL Diagnostic, PTB-XL Rhythm, PTB-XL Form).

Main Results:

• Models trained with data augmentation demonstrated improved performance across all tested datasets.
• Area Under the ROC Curve (AUC) increased by 3.5% (ICBEB), 1.7% (PTB-XL Diagnostic), 1.4% (PTB-XL Rhythm), and 3.5% (PTB-XL Form) compared to models without augmentation.
• Individual analysis confirmed the effectiveness of the three proposed novel augmentation techniques.

Conclusions:

• ECG data augmentation significantly enhances the performance of deep learning models for automated ECG classification.
• The novel augmentation techniques (Limb Electrode Move, Chest Electrode Move, Heart Vector Transform) are effective additions to existing augmentation strategies.
• Data augmentation presents a viable and low-cost method for enriching ECG datasets, thereby improving classification accuracy and reducing reliance on extensive manual annotation.