Quasi-Periodic Non-Negative Matrix Factorization for Phonocardiographic Signals Denoising

Abstract

Mechanical cardiac activity may be monitored with phonocardiographic (PCG) signals giving access to cardiac sounds. However, many noises interfere with cardiac information in raw signals and denoising such signals is necessary before interpretation. Non-negative Matrix Factorization (NMF) is of interest for time-frequency representations to separate noise and signal components. In this paper, to exploit the quasi-periodicity of the PCG, a quasi-periodic NMF (QP-NMF), based on multiplicative updates derived from a Majoration-Minimization algorithm, is proposed to decompose the PCG spectrograms. Numerical simulations show the good behavior of the proposed method to separate quasi-periodic components from the others. Finally, applied on real noisy PCG signals, QP-NMF shows its interest compared to an unsupervised NMF to denoise PCG signals.