Extract qualified heart sound in varying environment using parallel-training lms algorithm

Abstract

For cardiovascular disease (CVD) prevention, intelligent heart sound (HS) auscultation using wearable medical devices (WMDs) is a convenient and effective way. However, ambient noise often covers vital HS variants inevitably during HS acquisition and hinders further diagnosis. In order to extract qualified HS signal by source-limited WMD and generate real-time diagnostic conclusion in a noisy and varying environment, an effective de-noising method with low computational complexity and high robustness is demanded. Nevertheless, most existing de-noising methods are PCbased with high computational complexity, and some of them are specified to certain noises or rather sensitive to varying environment. To tackle such bottleneck problems, inspired by up-to-date low-footprint least-mean square (LMS) algorithm, a novel parallel-training LMS (PTLMS) algorithm is proposed in this paper, which utilizes the adaptive filter input and estimation error more effectively and efficiently, and thus outperforms 1) updating adaptive filter coefficients (AFCs) twice within one update-cycle thus enhances its convergence speed and property, and 2) carrying less N multiplications than LMS on calculating double-update of AFCs. Consequently, when extracting qualified HS signal in a noisy and varying environment, the adaptive filter convergence speed of PTLMS is faster with good convergence property thereby being high robustness and its computational complexity is lower than LMS meanwhile. The both theoretical analysis and testing experiment show that it keeps not only HS signal's pathological morphology unchanged after de-noising, but also overwhelms the LMS algorithm significantly.