A multiscale mean shift localization approach for robust extraction of heart sounds in respiratory signals

This paper addresses the problem of heart sound (HS) extraction in different types of single-channel respiratory sound (RS) signals by proposing a multiscale mean shift localization approach. First, the incoming respiratory signal (RS) are identified into linear/nonlinear portions by using third-order cumulant. Second, the identified linear and nonlinear portions are processed separately to tackle the large variations in the signal characteristics of adventitious sounds. The time-varying mean-shifts of the weighted log likelihood ratios of wavelet features are then calculated to capture the signal dynamics of various noisy RS signals. The proposed approach provides promising results giving an overall false localization rate as low as (1.8 ± 1.8)% for normal lung sound (LS) and (0.1 ± 1.7)% for adventitious sound signals. Therefore, the presented approach successfully attempts to solve the key clinical challenges faced by the existing localization methods in terms of respiratory ailments.