The Impact of Using Data Augmentation Techniques for Automatic Detection of Arrhythmia With a Deep Convolutional Neural Network Model

Addressing class imbalance through data augmentation helps in building more robust and fair classifiers that accurately represent the underlying distribution of classes in the dataset. In this paper, we use and compare two commonly known data augmentation techniques which are Synthetic Minority Over-sampling (SMOTE) and Adaptive Synthetic Sampling (ADASYN) to address class imbalance in MIT-BIH dataset. In fact, in this medical dataset, abnormal heart episodes are drastically less frequent than the normal ones. This impacts negatively the classification results due to skew data issues. Thus, we compare the performances of a CNN classification algorithm, first, without any data augmentation, then when applying the SMOTE then ADASYN techniques. The experimental findings indicate that the overall performance of the classification algorithm was boosted. ADASYN oversampling data yields the best accuracy of 95.78%.