Segmentation-Based Adaptive Feature Extraction Combined With Mahalanobis Distance Classification Criterion for Heart Sound Diagnostic System

Abstract

To utilize heart sound features that may vary according to their suitability for segmentation, automatic adaptive feature extraction combined with the Mahalanobis distance classification criterion is proposed to construct an innovative, heart sound-based system for diagnosing heart diseases. The innovation of this system is primarily reflected in the automatic segmentation and extraction of the first complex heart sound $(\textit {CS}_{1})$ and second complex heart sound $(\textit {CS}_{2})$ or each cardiac sound $(\textit {CS})$ , automatic extraction of the segmentation-based frequency feature ${\mathit {FF}}_{1}$ or $\textit {FF}_{2}$ , determination of the diagnostic features [ $\gamma _{11}$ , $\gamma _{12}$ ] and [ $\gamma _{21}$ , $\gamma _{22}$ , $\gamma _{23}$ ], and the development of a classifier model with adjustable sizes corresponding to the given desired confidence levels (denoted as $\beta $ ). Three stages corresponding to the implementation of the novel diagnostic system are summarized as follows. In stage 1, the time intervals between two sequential peaks are automatically calculated and statistically analyzed, and the result is used to determine whether a given heart sound can be segmented. Stage 2 involves automatic extraction of segmentation-based adaptive features for adapting the heart sound to the frequency domain. Finally, the Gaussian mixture model ( GMM )-based objective function $f_{\textit {et}}(\mathbf {x})$ is generated, and the $k^{th}$ component’s confidence region is determined by adjusting the optimal confidence level $\beta _{k}$ and subsequently used as the classification criterion to diagnose a given heart sound. The performance evaluation was validated with sounds from online heart sound databases and sounds from clinical heart databases. Compared with the state-of-the-art diagnostic methods, the overall accuracy $OA$ of 98.8%, $F_{1}$ of 99.27%, and $\kappa $ of 98.6% are much higher.