Heart-lung sound separation by nonnegative matrix factorization and deep learning

Abstract

Auscultation is an ordinary way to diagnose cardiovascular and respiratory diseases. The heart-lung sound (HLS) from the stethoscope is a mixture of heart sound (HS) and lung sound (LS). The first and most crucial step is to differentiate between HS and LS from the mixture signal. This paper proposes an unsupervised single-channel blind source separation (BSS) algorithm based on nonnegative matrix factorization (NMF) and deep learning (DL). Firstly, HS signals are extracted by the multi-constrained NMF (MCNMF) decomposition algorithm and K-means & SVM (KS) clustering method proposed in this paper, and different types of HS signals are classified by a convolutional neural network (CNN). Then, the HLS is separated by the embedding space centroid network (ECNet) proposed in our study. The method outperforms the known state-of-the-art techniques in all evaluation metrics SNR, SDR, SIR, SAR, and correlation coefficient. Our algorithm has two advantages: 1) this study is the first to combine NMF and DL for the HLS separation. MCNMF with selected constraints can have a more substantial sparsity effect that improves the decomposition accuracy. KS clustering algorithm can distinguish the HS component from the LS component in two clustering results. By mapping classified HLS into an embedding space, the ECNet demonstrates excellent separation results. 2) HS extraction and classification first, then HLS separation. Because our method exploits more detailed information from HS types in HLS before BSS, the ECNet can outperform other HLS separation methods.