Supervised Classification of Large Database of Heart Sound Records based on Noise Content

Mohammad K. Zia, Benjamin Griffel, and John L. Semmlow

Keywords

Noise Detection, Biomedical Signal Processing, Spectral Subtraction, Coronary Artery Disease

Abstract

Coronary artery disease (CAD) is a leading cause of death in the United States. Although drug treatment can control the progression of CAD, it is usually detected in advance stages where drug treatment is ineffective and invasive treatment is required. While current CAD diagnostic tools are very accurate, they are either too invasive or too expensive to be used as a regular screening tool to detect CAD in early stages. We are currently developing a noninvasive and cost-effective system to detect CAD using heart sounds. A major weakness of this method is extreme sensitivity to noises commonly encountered in a clinical environment. Because the auditory CAD correlates are faint, these noises can easily mask the CAD signal. Since we have a large database of records of normal and CAD patients, we cannot manually screen the database to reject records that contain noise. In this paper, we propose a supervised classification method using k-nearest neighbors to classify records based on noise content. Our method achieves an overall classification accuracy of 88.94%.

Important Links:

Go Back