Abstract
Auscultation is an effective noninvasive medical procedure for examining the cardiorespiratory system. However, the cardiac and respiratory acoustic sounds interfere in time as well as in spectral contents, which hampers the diagnostibility of the classical stethoscope. We propose a method for smart auscultation by blindly recovering the original cardiac and respiratory sounds from a single observation mixture. We decompose the spectrogram of the mixture into independent, non-redundant components, by employing non-negative matrix factorization (NMF). To group the decomposed components into original sources, a new unsupervised technique is proposed. Time-frequency masking is used to recover the original sources. This smart auscultation method is successfully applied to actual data collected from different subjects in different clinical settings. Our method demonstrates excellent results even in noisy clinical environments.