Parameters for Distinguishing Heart Sounds to Classify Normal and Pathological Cases

Abstract

The phonocardiogram (PCG) signal is a recording of heart sounds with useful and significant mechanical activity of the human heart information. In this work, we employed signal processing methods such as the Fast Fourier Transform (FFT), Continuous Wavelet Transform (CWT), the Bispectral analysis, and MFCCs to handle these data. These methods are used on our already filtered and labelled dataset as either normal or abnormal. S1, S2, and murmurs were segmented using an automatic approach. From these segments, we extracted multiple parameters and combined them in one feature matrix including the maximum amplitude, dominant frequency, weighted centroid, energy, phase entropy, and more. These features served as training features for machine learning algorithms such as the K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) for a binary classification of normal and pathological heart sounds. Furthermore, we produced 2D images from the analysing techniques to feed a CNN-SVM hybrid model, were we reached quite important accuracy results for different classification purposes.