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

The detection of cardiac arrhythmias from electrocardiogram (ECG) signals is essential for preventing sudden cardiac deaths. However, model performance on minority classes, such as supraventricular (S) and fusion (F) beats, which make up less than 3% of samples, is severely hampered by severe class imbalance in benchmark datasets like MIT-BIH. Conventional methods, such as basic GAN-based augmentation, oversampling, and undersampling, frequently result in low-fidelity synthetic samples or fail to preserve important ECG morphological features, leading to less-than-ideal F1-scores for uncommon arrhythmias. In this paper, a novel framework for high-fidelity minority class augmentation using conditional Wasserstein GAN with gradient penalty (cWGAN-GP) is proposed. It is integrated with a hierarchical multi-stream ResNet34 classifier that combines raw, Parzen-filtered, and similarity map features. In order to generate realistic ECG beats, the cWGAN-GP conditions generation on class labels. Mode collapse is prevented by enforcing gradient penalties and Wasserstein distance. Extensive experiments on the MIT-BIH arrhythmia database (5 AAMI classes) demonstrate significant improvements: minority S and F classes achieve F1-score increases of up to 28%, while the macro F1-score improves from 0.78 (baseline) to 0.94. The approach is computationally efficient for possible wearable device deployment and outperforms state-of-the-art GAN-augmented models in minority class detection.