Blind Source Separation of Heart and Lung Sounds Based on Nonnegative Matrix Factorization

Lung sound (LS) brings valuable information for lung status and respiratory analysis. However, the interference of heart sound (HS) usually occurs and raises confusion on pathological state during the LS recording. To solve this question, separation of HS and LS from mixed heart-lung sound (HLS) has become one of major issues in the biomedical research. A novel approach based on nonnegative matrix factorization (NMF) as one of blind source separation (BSS) techniques is proposed. In this paper, the chosen and enhanced mixed HLS signal is brought to the time-frequency domain and forms a multivariate data stationary time series. This multivariate data are then processed for dimension reduction by constant Q transform, which is well known as log-frequency STFT. The result of log-frequency STFT is then used as the input pattern of NMF. The average performance for quantitative evaluation of the proposed NMF-based approach is above 80% and better than the directly applied NMF. Another advantage provided by NMF is it only requires single channel as input signal instead of multichannel which is usually required by other BSS methods.