Principal component conditional generative adversarial networks for imbalanced ECG classification enhancement.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Electrocardiogram (ECG) diagnostics are crucial for cardiovascular disease management.
• Increasing volumes of diverse, long-term ECG data overwhelm traditional manual annotation.
• Challenges in ECG classification include imbalanced data, individual variability, and long sequence analysis.

Purpose of the Study:

• To address limitations in ECG data augmentation for imbalanced datasets.
• To develop a novel method for generating high-fidelity ECG principal component features.
• To improve the accuracy of arrhythmia classification in diverse patient populations.

Main Methods:

• Proposed a Principal Component Analysis-based Conditional Generative Adversarial Network (PCA-CGAN).
• Shifted data augmentation from waveform generation to principal component feature generation.
• Implemented a two-stage conditional encoding-decoding architecture with Transformer's global attention mechanism.

Main Results:

• PCA-CGAN achieved stable convergence on a large-scale heterogeneous ECG dataset.
• Successfully resolved the 'dilution effect' in data augmentation, balancing Precision and Recall.
• Augmented data significantly improved ResNet model's F1 score, especially for rare arrhythmias like atrial premature beats.

Conclusions:

• PCA-CGAN offers a systematic solution for ECG data imbalance, redefining signal generation objectives.
• The method effectively handles waveform jitter and heterogeneity, improving diagnostic feature capture.
• Established a theoretical foundation for applying ECG-assisted diagnostic systems in clinical settings.