End-to-end heart sound segmentation using deep convolutional recurrent network

Abstract Heart sound segmentation (HSS) aims to detect the four stages (first sound, systole, second heart and diastole) from a cycle in phonocardiogram (PCG), which is an essential step automatic auscultation analysis. Traditional HSS methods need manually extract features before dealing with tasks. These artificial highly rely on extraction algorithms, often result poor performance due different operating environments. In addition, high-dimension frequency characteristics of audio also challenge traditional effectively addressing This paper presents novel end-to-end method based convolutional long short-term memory (CLSTM), directly uses recording as input address Particularly, layers are designed meaningful perform downsampling, LSTM developed conduct sequence recognition. Both components collectively improve robustness adaptability processing Furthermore, proposed CLSTM algorithm easily extended other complex annotation tasks, it does not corresponding tasks advance. can be regarded powerful feature tool, integrated into existing models for HSS. Experimental results real-world PCG datasets, through comparisons peer competitors, demonstrate outstanding algorithm.