On the blind recovery of cardiac and respiratory sounds

We present a method for smart auscultation by proposing a novel blind recovery of the original cardiac and respiratory sounds from a single observation mixture, in the framework of nonnegative matrix factorization (NMF). The method learns the basis spectra of the mixing sources in unsupervised or semisupervised fashion depending upon the applications. A modified NMF technique is proposed, which enforces the spectral structure of the target sources in mixture factorization, resulting in good separation of target sources, even in the presence of nonstationary noise. Moreover, data is processed in small batches which 1) enables dynamic bases spectra update technique to mitigate the spectral variations of the mixing sources, and 2) reduces computational complexity. The analytical work is verified through simulations using synthetic as well as actual clinical data collected from different subjects in different clinical sittings. The proposed smart auscultation method demonstrates excellent results even in noisy clinical environments.