SAR-CardioNet: A Network for Heart Valve Disease Detection From PCG Signal Based on Split-Self Attention With Residual Paths

Abstract

Analyzing phonocardiogram (PCG) signals is getting popularity because of its noninvasive test procedures for the early detection of heart-valve-related diseases. Unlike traditional disease detection methods, an efficient and properly trained deep neural model may become a first-hand disease detection tool that can save any unwanted cardiac failure due to heart valve defects (HVDs). In this article, split-self attention with residual paths is used to design a deep learning network for automatic HVDs’ detection directly from the PCG signals with higher accuracy. In the proposed split-self attention block, a given input feature vector is divided into two equal portions, where the second portion is used to generate attention on the first portion. Moreover, multipath feature extractors are introduced in the proposed network, where the depth of convolution is varied in different paths. It is observed that the use of an attention block at a deeper convolutional layer helps improve the classification performance. Finally, two residual paths are added to avoid overfitting and vanishing gradient problems. The performance of the network has been tested with two publicly available datasets through extensive experimentation. We obtained 96.37% and 99.25% accuracy with an average area under the curve of 98% and 100%, respectively, for the datasets. The comparison with the existing models demonstrates that our network is a very competitive candidate for HVDs’ detection.