Cumulant-based basis selection and feature extraction to improve heart sound classification

Cardiac auscultation, the direct hearing and interpreting the heart sounds, is a fundamental clinical skill that requires years to develop and refine. The interpretations are commonly prone to variations resulting in highly subjective diagnosis. Alternative technologies such as magnetic resonance imaging (MRI) and echocardiography are on the rise. However, these are expensive, and instead technologies to support or automate cardiac auscultation are becoming important and are currently widely being researched. The accuracy of cardiac auscultation could be improved through extracting objective information from phonocardiography (PCG) signals to be used for automated heart sound classification. This study focuses on the classification of new features extracted from PCG signals represented by wavelet packet transform (WPT). A wavelet packet tree is constructed for each PCG signal, and higher-order cumulants (HOC) of the wavelet packet coefficients (WPC) are extracted and used as features, named hoc_WPC features. With the features, merits of time-frequency analysis of WPT and statistical analysis of HOC are exploited. PCG signals have been classified successfully using hoc_WPC features. An improvement of 3.02% sensitivity and 0.19% specificity have been achieved in differentiating normal heart sounds and regurgitations. The hoc_WPC features are further capable to classify heart sounds into normal, mitral regurgitation, aortic regurgitation, and aortic stenosis, with 96.95% accuracy. Basis selection is another issue in analysis signals by WPT. For basis selection, an approach is proposed to reduce the initial search space from the entire tree to a trapezoidal sub-tree of it, and then four basis selection methods are proposed: i) multilevel basis selection (MLBS); ii) cumulant-based trapezoidal multi-level basis selection (CT_MBS); iii) cumulant-based trapezoidal best basis selection (CT_BBS);and iv) cumulant-based trapezoidal local discriminant basis (CT_LDB). With MLBS an energy-based information measure is used to select the best nodes of the three bottom levels of a wavelet packet tree for feature extraction. With cumulantbased trapezoidal basis selection methods, HOC are used to define information measure. This is based on the feature extraction experiment whereby the ability of HOC to represent the information laid throughout a wavelet packet tree has been shown. CT_MBS is an extension of the MLBS, whereby cumulant measure is used to prune the wavelet packet tree instead of energy. CT_BBS and CT_LDB are the extensions of the commonly used basis selection methods, which are best basis selection (BBS) and local discriminant analysis (LDB). The best classification accuracy of 98.17% was achieved by CT_LDB in classifying different types of heart sounds of this study.

Preview PDF FSKTM 2013 9R.pdf Download (1MB) | Preview

Actions (login required)

View Item