Accurate Classification of Heart Sounds for Disease Diagnosis by A Single Time-Varying Spectral Feature: Preliminary Results

Early detection of abnormal heart conditions helps reduce sudden cardiac deaths caused by cardiovascular diseases (CVDs). For this purpose, cardiac auscultation is generally preferred as the first step for heart disease diagnosis. The auscultation task involves the use of a stethoscope by physicians to listen and interpret the heart sounds. However, it demands higher expertise from the physicians for accurate evaluation of the heart conditions. In the last few years, there is an increasing interest in automating heart sound analysis process. In automated heart sound analysis, the heart sound signal or a number of features extracted from this signal are processed by a classifier to create a decision regarding the diagnosis of CVDs. It should be clear to observe that the performance of the classifier is critically dependent on the choice of the features. In most cases, more than one feature are needed to obtain an accurate decision and the calculation of these features are usually very complicated. In this paper, we present a simple time-varying spectral feature for heart sound classification. The proposed feature is simple to implement yet it allows the subsequent classifier block to perform its task with very high accuracy. An additional advantage of the proposed classifier is that it enables the classifier to perform the classification by using only one feature. The performance of the proposed feature is tested for both binary and multiclass classification problems and evaluated with two different classifiers. Results show that the proposed feature can be used with any classifier for both two-class and multi-class classification problems and yields excellent results.