Arrhythmia Classification using CGAN-augmented ECG Signals
ECG databases are usually highly imbalanced due to the abundance of Normal ECG and scarcity of abnormal cases. As such, deep learning classifiers trained on imbalanced datasets usually perform poorly, especially on minor classes. One solution is to generate realistic synthetic ECG signals using Generative Adversarial Networks (GAN) to augment imbalanced datasets. In this study, we combined conditional GAN with WGAN-GP and developed AC-WGAN-GP in 1D form for the first time to be applied on MIT-BIH Arrhythmia dataset. We investigated the impact of data augmentation on arrhythmia classification. We employed two models for ECG generation: (i) unconditional GAN; Wasserstein GAN with gradient penalty (WGAN-GP) is trained on each class individually; (ii) conditional GAN; one Auxiliary Classifier WGAN-GP (AC-WGAN-GP) model is trained on all classes and then used to generate synthetic beats in all classes. Two scenarios are defined for each case: (a) unscreened; all the generated synthetic beats were used, and (b) screened; only a portion of generated beats are selected and used, based on their Dynamic Time Warping (DTW) distance to a designated template. A state-of-the-art ResNet classifier (EcgResNet34) is trained on each of the augmented datasets and the performance metrics (precision/recall/F1-Score micro- and macro-averaged, confusion matrices, multiclass precision-recall curves) were compared with those of the unaugmented imbalanced case. We also used a simple metric Net Improvement. All the three metrics show consistently that net improvement (total and minor-class), unconditional GAN with raw generated data (not screened) creates the best improvements.