Noise detection in phonocardiograms by exploring similarities in spectral features

Abstract Analysis and interpretation of heart sounds (HSs) can be seriously hindered by noise contamination when signals are acquired in noncontrolled environments. Signal processing methodologies are then required in order to robustly analyse HSs collected in different recording settings. Some works already address this problem using complex calculus that are usually dependent on the accurate segmentation of the signals. As such, the aim of the present study is the development of a low-complex automatic algorithm able to discriminate clean from contaminated HS signals (or phonocardiograms) recorded in real-life situations. Spectral features were used to characterize the different behaviours of clean and noisy HSs in phonocardiograms (PCGs) in noisy conditions. In particular, besides the normal interferences associated to the auscultation in a noncontrolled environment, other noisy sounds were purposely simulated and included vocalizations, ambient sounds and also other physiological interferences rather than HSs. The available signals were recorded in 24 healthy volunteers and in eight patients diagnosed with different cardiac disorders. The subjects included in the healthy dataset followed a pre-defined protocol, during which ambient, physiological and vocal interferences were simulated. A total of 288 PCGs, recorded in pulmonary and mitral auscultation positions, comprises the healthy dataset. The pathological dataset includes 16 PCGs, acquired in pulmonary and tricuspid chest locations. The described approach was compared with two existent methodologies developed for the same purpose. Our algorithm was found to return the best performance regarding the different types of interferences, resulting in an average sensitivity and specificity of 88.4% and 85.6%, respectively, for healthy dataset and 84.3% and 85.8%, respectively, for the dataset with pathology. These results correspond to an increase of up to 4.7% in SE and 39.0% in SP when comparing to the two referred methodologies in healthy dataset. Regarding the pathological dataset our approach improved noise detection by up to 27.0% in SE and 34.1% in SP.