Construction of CNNs for Abnormal Heart Sound Detection using Data Augmentation

Cardiovascular diseases (CVDs) are the main cause of deaths all over the world. To detect the abnormalities of a heart automatically, convolutional neural networks (CNNs) learned by using heart sound signals (i.e., the phonocardiogram or PCG) are proposed. Generally, CNNs need sufficient annotated training data to achieve the high performance. However, the annotated PCGs (i.e. PCGs labelled with abnormal or normal) dataset is not sufficient because of personal information and burden of physicians. Therefore, we need to improve the classification performance of CNNs even when annotated PCGs is insufficient. In this paper, in order to solve above problem we consider two data augmentation (DA) methods, one is Window Slicing with Spectrogram (WSS), which slices single PCG to make multiple signals and transforms the signals into spectrogram data, the other is Synthetic Spectrogram based GANs (SSG), which generates synthetic data using generative adversarial networks (GANs). In order to show the validity of considered two DA methods, we perform some experiments concerning heart sounds detection and discuss the results of experiments in point of the accuracy, the sensitivity and the specificity.

Please read the paper for more details.