Improving Convolutional Neural Network with Recurrent Model and Generative Adversarial Networks for Imbalance Classification of Wearable ECG Signal Quality Assessment

Abstract

Signal quality assessment is important procedure to assess if a signal can be preprocessed into next step or not. In this study, we proposed an improving of convolutional neural network with recurrent model and generative adversarial network. The experiment shows that enhancing convolutional neural network by using recurrent model or data augmentation via GAN for wearable ECG signal quality assessment is feasible to try. Adding recurrent model to CNN increase its performance by approximately 3% margin in un-augmented dataset. Generating fake samples from “unacceptable” class can enhance CNN performance up to 2% margin. Combination of recurrent model and GAN still not yet matched found. However, It doesn’t mean that the combination is impossible to do.