Automatic detection of heart valve disorders using Teager-Kaiser energy operator, rational-dilation wavelet transform and convolutional neural networks with PCG signals

The heart sound signals (phonocardiogram, PCG) have been for decades a fundamental diagnostic tool to detect a potential cardiovascular pathology in clinical practice. Due to the lack of proper training, the medical practitioners may sometimes overlook some of the vital information either during the auscultation of the heart sound or visual inspection of the PCG signal. Therefore the need for an automated and accurate anomaly detection method based on PCG becomes urgent. The aim of this work is to design a reliable algorithm for the automatic detection of heart valve disorders (HVDs) without any segmentation of the heart sound signals. Teager-Kaiser energy operator (TKEO) and rational dilation wavelet transform (RDWT) are utilized to extract representative features in order to detect abnormal patterns in PCG signals with the employment of deep learning model. First, TKEO is used to extract the instantaneous energy of the source that generates the PCG signal rather than the energy of the signal itself. Then, RDWT is employed to decompose the instantaneous energy of the PCG signal into different sub-bands. The oscillatory characteristics of PCG has been retained in these sub-bands, which are served as discriminant features. Third, these features are fed to one-dimensional (1D) convolutional neural networks (CNN) for classification. Finally, experiments including two types of classification named binary classification (normal vs. abnormal) and multi-class classification (normal vs. aortic stenosis vs. mitral regurgitation vs. mitral stenosis vs. mitral valve prolapse), are carried out on a well-known and publicly available PCG database to verify the effectiveness of the proposed method. The overall average accuracy for binary (5 cases), four-class and five-class classification is reported to be 100%, 99.00%, 99.75%, 99.75%, 99.60%, 98.87% and 98.10%, respectively. The proposed method has obtained superior accuracy in comparison to most of the state-of-the-art approaches using the same database. It can serve as a potential diagnostic tool for the automated detection of HVDs in routine auscultation examination.