Heart sound cancellation from respiratory sound using a filter bank based semi-blind source extraction algorithm

This paper introduces a novel semi-blind single-channel source extraction algorithm to solve the problem of heart sounds (HS) cancellation from single channel respiratory sounds (RS) recordings. Underlying RS are extracted from those HS contaminated segments in the recorded signal by the proposed algorithm which incorporates shelving filter, filter bank, and template matching using FIR filters. The algorithm has been tested on simulated data of different types of RS (normal RS and abnormal adventitious sounds) and real RS data, respectively. The average power difference between the power spectral density of the original and reconstructed RS signals is found to be maximally 0.728 plusmn 0.0008 dB, showing the excellent similarity of the two. The algorithm has also been validated qualitatively by experienced clinicians performing subjective listening test of the reconstructed signals.