Wavelet Transform Generated Inherent Noise Reference for Adaptive Filtering to De-noise Pulse Oximeter Signals

Abstract

As exemplified during the COVID-19 pandemic and in post-operative intensive care units, monitoring blood oxygen saturation (SpO₂) levels is crucial in terms of assessing a patient’s health condition. Due to random movements of the subject, a pulse-oximeter-driven photoplethysmographic (PPG) signal becomes noisy while recording, with motion artefacts (MAs), which will disturb the morphological features, leading to incorrect SpO₂ levels. The MA noise may contain either low- or high-frequency components, resulting in a scenario with inband and out-of-band noise. The reduction of in-band noise with an adaptive filter requires a reference signal, and an additional sensor such as an accelerometer is normally used in addition to the PPG sensor to capture the MAs. The present work focuses on the generation of a reference for inherent noise using a wavelet transform (WT), thereby eliminating the need for an external sensor. The computed values of the correlation coefficient and magnitude squared coherence are used to establish the validity of the generated inherent noise reference. Our WT-driven adaptive filtering method reduces MAs, simplifies the correct approximation of the SpO2 and heart rate, and also restores the respiratory components. The de-noised PPG signals presented here and a corresponding numerical study prove the usefulness of the proposed method, which has a worst-case accuracy of 0.5% in regard to SpO₂ estimations.