Minority class-aware multiclass arrhythmia detection using conditional GAN augmentation.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Cardiac arrhythmias detection from ECG is crucial for preventing sudden cardiac deaths.
• Class imbalance in datasets like MIT-BIH severely impacts minority beat detection (e.g., supraventricular and fusion beats).
• Conventional augmentation methods often yield low-fidelity samples and fail to preserve ECG morphology.

Purpose of the Study:

• To propose a novel framework for high-fidelity minority class augmentation in ECG signals.
• To enhance the detection of uncommon arrhythmias by addressing class imbalance.
• To improve F1-scores for supraventricular (S) and fusion (F) beats.

Main Methods:

• Developed a conditional Wasserstein GAN with gradient penalty (cWGAN-GP) for realistic ECG beat generation, conditioned on class labels.
• Integrated cWGAN-GP with a hierarchical multi-stream ResNet34 classifier utilizing raw, Parzen-filtered, and similarity map features.
• Employed gradient penalties and Wasserstein distance to prevent mode collapse and ensure high-fidelity generation.

Main Results:

• Achieved significant improvements on the MIT-BIH arrhythmia database (5 AAMI classes).
• Minority S and F classes showed F1-score increases up to 28%.
• Macro F1-score improved from 0.78 (baseline) to 0.94, outperforming state-of-the-art GAN-augmented models.

Conclusions:

• The proposed cWGAN-GP framework effectively generates high-fidelity minority ECG beats, significantly improving arrhythmia detection accuracy.
• The approach is computationally efficient, suitable for wearable device deployment.
• This method offers a substantial advancement in accurately identifying rare cardiac arrhythmias.