Pay More Attention with Fewer Parameters: A Novel 1-D Convolutional Neural Network for Heart Sounds Classification

Abstract

The cardiovascular disease (CVD) is one of the major causes of mortality worldwide. Auscultation of heart sounds or phonocardiograms (PCGs) analysis, which is an efficient and non-invasive way, has been shown to be promising and played an important role in preliminary CVD diagnosis. In this study, a deep learning-based PCG classification method is proposed, which is mainly comprised three steps: pre-processing, PCG patches classification using a novel 1-D deep convolutional neural network (CNN), and final predicting of PCG recordings based on the patch-level results. In order to maximize the information flow within the CNN, a block-stacked style architecture with clique blocks is employed, and in each clique block a bidirectional connection structure is utilized. Using the stacked blocks, the proposed CNN achieves both spatial and channel attention, which leads a superior classification performance. Besides, a novel separable convolution with inverted bottleneck is introduced to efficiently decouple features' dependency between spatial and channel-wise dependency of features. Experiments on PhysioNet/CinC 2016 reveal a superior classification performance and the advantage in parameter efficiency of the proposed method comparing to state-of-the-art methods.