Explanations of Augmentation Methods for Deep Learning ECG Classification

Recent progress in deep learning has sparked widespread interest in their development and adoption of diverse applications. The effectiveness of deep neural networks, particularly in extracting meaningful patterns from multimedia data, has resulted in their extensive use across various domains such as healthcare. Augmentation using synthetic data and transformation of real world data provides possible solutions to address prevalent data challenge and to consequently improve model robustness, improve generalization and reduce overfitting in development of high-performance DL models. While it has been documented that the choice of appropriate traditional augmentation method is task and data dependent, there is a dearth of explanations on how these augmentations affect the model learning and decision outcomes. This problem is exacerbated in healthcare domain where black-box methods are met with resistance in acceptance and for 1D time-series signals where some augmentation methods can be detrimental to performance. In this paper, we investigate and explain the effect of various augmentation methods for the modality of 1D time-series healthcare Electrocardiogram (ECG) signals. Our results demonstrate that various label in-variant augmentation methods can not only lead to improvement or degradation in model performance but also skew the activation features learned by the model.