Autocorrelation for denoising biomedical signals

Abstract

Photoplethysmography (PPG) has become a standard method for assessment of blood volume changes in clinical care and heart rate in home care [1]. Besides the pulse rate, PPG pulse forms carry signatures of diagnostically relevant events in cardiac cycle and can be used to estimate arterial stiffness. Extraction of these features requires removal of noise, motion artifacts and the superimposed slow varying signals, such as that from breathing, from the signal while preserving pulse morphology. However, modern filtering methods often fail to reproduce all signal features. Here, we propose a novel noise–removal method based on autocorrelation. Autocorrelation is a well-known method used in optics, mainly for estimating the duration of ultrashort laser pulses. We used autocorrelation to remove the noise and baseline wander (BLW) from a set of bioelectrical signals, namely electrocardiogram (ECG) and PPG. These signals comprise pulses (or beats) repeated in time but with slight changes. When we record several such beats and by averaging them get a noise-free signal with distorted morphology. However, taking a few steps further, namely subtracting the average from the original signal and filtering the difference in the frequency domain, enables the noise and BLW extraction from the original signal and reproduction of a faithful noise-free signal. We tested this method on the private ECG database, where added BLW component is from public MIT-NST database, and on the private PPG signals. The results show the superiority of our approach compared to the conventional cubic spline (CSP) method.