Performance Comparison of Variants of LMS Algorithms

Abstract

Photoplethysmography (PPG) is a noninvasive method to measure heart rate which is commonly used in wearable devices. PPG measured at the wrist is corrupted with Motion Artifacts (MA) during physical activities and therefore a certain amount of signal processing is needed before estimating the heart rate. There are several proposed methods to remove MAs from a PPG, such as independent component analysis, sparse spectrum based methods, and Singular Value Decomposition (SVD). These methods have a high complexity compared to the Least Mean Square (LMS) adaptive later technique with acceleration data as the reference signal. In this paper, the performances of several types of LMS liters are computed and compared with the performance of the SVD method. The results show that the general LMS algorithm and the normalized LMS algorithm perform better than the other variants of LMS algorithm and that the general LMS algorithm performs better than the SVD. An autoregressive integrated moving average (ARIMA) model of the PPG signal has also been estimated for future applications of this signal.