Segmentation-based heart sound feature extraction combined with classifier models for a VSD diagnosis system

摘要

In this paper, boundary curve models for the diagnostic features [T_(12), T_(11)] and [F_G, F_W] are proposed to diagnose ventricular septal defects (VSD), which are generally divided into 3 types: small VSD (SVSD), moderate VSD (MVSD) and large VSD (LVSD). The VSD diagnosis is accomplished in three steps. First, in the time domain, the diagnostic features [T_(12), T_(11)], which are the time intervals between two adjacent first heart sounds (S1) as well as the interval between S1 and the second heart sound (S2), are extracted from the envelope E_T for the heart sound (HS); in the frequency domain, the envelope E_F for every cardiac cycle sound that the HS is segmented into, based on a moving windowed Hilbert transform (MWHT), is proposed to extract the diagnostic features [F_G, F_W]. which are the center of gravity and the frequency width of the frequency distribution. Second, to evaluate the detection ability of the proposed diagnostic features, a classification boundary method based on the support vector machines (SVM) technique is proposed to determine the classifiers to diagnose the VSD sounds. Furthermore, to simplify these classifiers and make them parameterizable, according to their shapes, the least squares method is employed to build ellipse models for fitting the classification boundary curves. Finally, the numerical results based on the ellipse models are introduced for diagnosis of the VSD. Moreover, to validate the usefulness of the proposed method for sounds besides VSD and normal sounds, aortic regurgitation (AR), atrial fibrillation (AF), aortic stenosis (AS) and mitral stenosis (MS) sounds are used as examples to be detected. As a result, the classification accuracies (CA) achieved is 98.4% for the detection of clinical VSD sounds from normal sounds and are 95.1%, 94.8% and 95.0%, respectively, for the detection of clinical SVSD, MVSD, and LVSD among VSD sounds.