Analysis of estochastic classificators for effective detection of heart murmurs from phonocardiographic signals

Abstract

Automated diagnostic systems for the detection of cardiac murmurs, described in the literature, recorded in short time intervals, the physiological dynamic from databases of segmented phonocardiography signals, obtaining a multiplicity of samples from the same patient regardless the cardiac dynamics of each individual, which does not guarantee proper identification of cardiac abnormality.
The diagnostic system aided proposed begins with the study of 1060 signals taken from the four areas of cardiac auscultation in 144 patients, mainly grouped into two classes: normal and pathological. The records are grouped by their position in the cardiac period in four types: normal, systolic mumur, diastolic and systolic-diastolic, obtaining a database phonocardiographic signals, it is preprocessing, however, it isn’t segmented for to conserve cardiac dynamics of each individual in the study. Then generate a representation space from the PLP and the cepstral coefficients calculated from the FFT and STFT, developing a stochastic analysis with HMM and GMM classifiers and feature selection techniques to extract relevant information from the physiological dynamics, to permit adequate and efficient training of the system with an adequate rate of classification for clinical diagnostic support.