A non-negative matrix factorization approach based on spectro-temporal clustering to extract heart sounds

Abstract Auscultation is often the first clinical analysis realized by doctors when patients have any symptoms related to a heart disease. In a typical scenario of auscultation, heart sounds are interfered by lung sounds both frequency and time domain. This fact causes mixtures that are composed of heart and lung sounds. In this paper, a non-negative matrix factorization (NMF) approach to extract heart sounds from mixtures composed of heart and lung sounds is addressed. Specifically, three contributions motivated by the clustering principle are presented: two of these clusterings are based on spectral content and one is based on temporal content in order to discriminate heart and lung sounds. The first spectral clustering measures the spectral similarity between the bases factorized from NMF and those ones belonging to a dictionary created from a training database composed of only isolated heart sounds. The second spectral clustering analyzes how the energy of the bases, provided by NMF, is distributed along the frequency. The temporal clustering is based on the activations provided by NMF in order to find repetitive temporal patterns hidden in these activations using an estimated heart rate of the mixture in the low frequency range. Evaluation shows that the proposed method obtains promising results and outperforms recent non-based-NMF and based-NMF state-of-the-arts methods.