Motion Artifact Reduction in Photopleythysmography Using Magnitude-based Frequency Domain Independent Component Analysis Rajet Krishnan, Balasubramaniam Natarajan and Steve Warren Electrical and Computer Engineering Department Kansas State University Manhattan, KS 66506 Telephone: (785)-532-5600; E-mail: rajetk@ksu.edu, bala@ksu.edu, swarren@ksu.edu Abstract— Corruption of photopleythsmograms by motion ar- tifacts has been a serious obstacle to the reliable use of pulse oximeters for real-time, continuous state-of-health monitoring. In this work, we propose a motion artifact reduction methodology that is effective even in the case of severe subject movement. The methodology involves an enhanced preprocessing unit consisting of a motion detection unit (MDU), period estimation unit, and a Fourier series reconstruction unit. The MDU aids in identifying clean data frames versus those corrupted with motion artifacts. The period detection unit is used to determine the fundamental frequency of a corrupt frame. The Fourier series reconstruction unit reconstructs the final preprocessed signal. The reconstruc- tion process primarily utilizes the spectrum variability of the pulse waveform. Preprocessed data are then fed to a magnitude- based frequency domain Independent Component Analysis (FD- ICA) unit. This helps reduce motion artifacts present at the frequency components chosen for reconstruction. Experimental results are presented to demonstrate the effectiveness of the proposed motion artifact reduction method. The efficacy of the technique is compared with time domain ICA and complex frequency domain ICA methods. Index Terms - Photoplethysmogram (PPG), Independent Component Analysis (ICA), Motion Detection Unit (MDU), Magnitude-Based Frequency Domain ICA (FD-ICA). I. I NTRODUCTION Photoplethysmography is a non-invasive, optical means to obtain relative blood volume in tissue as a function of time. Photoplethysmographic data can be acquired with reflectance- or transmittance-mode sensors, and multiple excitation wave- lengths allow parameters from time domain PPGs to be converted to values of heart rate and blood oxygen saturation. Blood flow to tissue is modulated by each cardiac cycle of the subject and also by other processes like respiration and subject motion. If PPG data are to be reliably obtained from wearable sensors that are used for real-time, continuous state- of-health monitoring, then effective algorithms for motion artifact reduction must be employed. Various signal processing techniques have been investigated to address the problem of recovering quasiperiodic PPG signals from measurements corrupted with motion artifacts, including wavelet analysis and decomposition techniques [1] and adaptive filters [2]. However, the study in [3] indicates that both wavelet transform and adaptive filter techniques introduce phase distortions in PPG data. Work involving analog filters and moving average techniques is presented in [4]. This artifact extraction problem has also been viewed as a blind source separation problem in [5] and [6]. In [5], an enhanced prepro- cessing unit preceded the ICA block. The preprocessing unit consisted of signal period detection using an autocorrelation method followed by a block-interleaving operation. However, this technique relies on the ability of the autocorrelation technique to correctly detect the period and hence provides erroneous results in the presence of extreme motion artifacts. In [6], an improved preprocessing technique is described that employs extrapolation/truncation of each cardiac cycle to the mean of the measured cardiac cycle followed by ICA. This method is highly prone to errors and inconsistencies, since accurate cardiac cycle measurements become difficult in the presence of extreme disturbances due to motion. In this work, we present a new motion artifact reduction methodology that combines an enhanced signal preprocessing unit and a frequency domain ICA unit. The preprocessing unit incorporates a Fourier series reconstruction of the PPG data that utilizes the spectrum variability and the quasi- periodicity of the pulse waveform. Following this is a novel ICA routine performed on the PPG data in the frequency domain (FD-ICA) that considers only magnitude information. This technique assumes instantaneous mixing of statistically independent sources in the time domain and a constant mixing matrix for the time frame considered. The routine is, however, different from the conventional complex frequency domain ICA described in [7]. A comparison of this technique with the time domain ICA and complex frequency domain ICA techniques shows that the new magnitude-based frequency domain ICA approach is more effective in reducing motion artifact. The paper is organized as follows. The new preprocessing unit and the frequency domain ICA routine are detailed in section II. The efficacy of the proposed methodolgy is 978-1-4244-2390-3/08/$25.00 ©2008 IEEE 1