Title:Performance of using Mel-Frequency Cepstrum Based Features in Nonlinear Classifiers for Phonocardiography Recordings

Abstract:Cardiovascular system diseases can be identified by using a specialized diagnostic process utilizing a digital stethoscope. Digital stethoscopes provide phonocardiography (PCG) recordings for further inspection, besides filtering and amplification of heart sounds. In this paper, a framework that is useful to develop feature extraction and classification of PCG recordings is presented. This framework is built upon a previously proposed segmentation algorithm that processes a feature vector produced by the agglutinate application of Mel-frequency cepstrum and discrete wavelet transform (DWT). The performance of the segmentation algorithm is also tested on a new data set and compared to the previously reported results. After identifying the fundamental heart sounds and segmenting the PCG recordings, five principal features are extracted from the time domain signal and Mel-Frequency cepstral coefficients (MFCC) of each cardiac cycle. Classification outcomes are reported for three nonlinear models: k nearest neighbor (k-NN), support vector machine (SVM), and multilayer perceptrons (MLP) classifiers in comparison with a linear approach, namely Mahalanobis distance linear classifier. The results underline that although neural networks and linear classifier show compatible performance in basic classification problems, with the increase in the nonlinearity of the classification problem their performance significantly vary.