Improving deep learning in arrhythmia Detection: The application of modular quality and quantity controllers in data augmentation

Among the most prevalent diseases with significant fatality rates are cardiac disorders. In recent years, the application of deep learning in diagnosing various cardiac conditions, namely arrhythmia, has gained widespread attention. Nevertheless, deep neural networks struggle to detect arrhythmia due to skewed datasets and a lack of data in different classes. If used effectively, data augmentation can address this gap by adding further synthetic samples in the correct distribution of the corresponding skewed classes. To achieve this, we have instituted the Modular Distribution and Volume Controller method, abbreviated as MDVC. Our method concentrates on both qualitative and quantitative aspects of data augmentation to elevate efficiency and create a significant and varied amount of synthetic samples. Seven distinct methods of data augmentation are utilized in fusion to synthesize samples. Subsequently, the distribution controller determines the most advantageous distribution of artificial samples for each data augmentation technique, emphasizing the dispersion and collision of different classes. The maximum overall data augmentation volume, the volume of each class, and the volume of each data augmentation technique are defined by the volume controller through the novel x, alpha, and beta parameters. Classifying the 17 classes of the MIT-BIH dataset using the MDVC yielded an accuracy of 98.9 % using a 10-fold cross-validation strategy; thus, we have outperformed state-of-the-art data augmentation techniques such as RandAugment and alpha-trim by 1.3 % and 0.8 %, respectively.