Professional Documents
Culture Documents
A Short Study to Assess the Potential of Independent Component Analysis for Mo-
tion Artifact Separation in Wearable Pulse Oximeter Signals
Jianchu Yao1 and Steve Warren2
1
Department of Technology Systems, East Carolina University, Greenville, NC, USA
2
Department of Electrical & Computer Engineering, Kansas State University, Manhattan, KS, USA
Abstract—Motion artifact reduction and separation become pulse oximeters: arterial volume variation and motion arti-
critical when medical sensors are used in wearable monitoring fact.
scenarios. Previous research has demonstrated that independ- The following two sections discuss the promise of ICA
ent component analysis (ICA) can be applied to pulse oximeter in pulse oximetry and provide the theoretical basis for this
signals to separate photoplethysmographic (PPG) data from approach. The Methods section then describes how data
motion artifacts, ambient light, and other interference in low-
were collected and how the statistical independence between
motion environments. However, ICA assumes that all source
signal component pairs are mutually independent. It is impor- arterial volume variation and motion was assessed. The Re-
tant to assess the statistical independence of the source compo- sults section summarizes these analyses and notes that the
nents in PPG data, especially if ICA is to be applied in ambula- application of ICA to PPG data should be performed with
tory monitoring environments, where motion artifacts can have caution.
a substantial effect on the quality of data received from light- II. BACKGROUND AND MOTIVATION
based sensors. This paper addresses the statistical relationship
between motion artifacts and PPG data by calculating the cor- Independent component analysis is usually denoted as
relation coefficients between arterial volume variations and
motion over a range of stationary to high-motion conditions. x(t ) = M ⋅ s (t ) (1)
Analyses indicate that motion significantly affects arterial flow, m×n
so care must be taken when applying ICA to light-based sensor where M ∈R is the mixing matrix, x(t ) ∈ R m
is the
data acquired from wearable platforms.
observed mixed signal vector, and s (t ) ∈ R is the source
n
This material is based upon work supported by the National Science Foundation under grants BES–0093916. Any opinions,
findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily re-
flect the views of the NSF.
Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on May 25,2010 at 03:49:01 UTC from IEEE Xplore. Restrictions apply.
modifications, be configured with a variable number of light 60 mmHg) is much higher than venous blood pressure (e.g.,
sources to accommodate this uncertainty, which makes the < 10 mmHg), so the blood volume in veins is much more
ICA algorithm more attractive. Furthermore, ICA algo- susceptible to motion artifacts [13, 14]. As a result, we as-
rithms can separately address arterial and venous volume sume that motion primarily affects venous blood volume
variations. As noted in [3], this enables the calculation of variation, dV , and therefore neglect its effect on arterial
both arterial and venous oxygen saturation. volume changes. With these considerations, the detected
To get an early indication of whether ICA methods would light intensity varies with dA and dV as follows:
be appropriate for the extraction of motion artifacts from
pulse oximeter signals, the authors downloaded an ICA ( )( )( )
dI t = − I 0 µ a e − µ t T e − µ vV e − µ a A dA −
( )( )(e )dV
(3)
MATLAB program (fastICA.m) from [11] and used this
I 0 µ v e − µ t T e − µ vV −µa A
algorithm to separate some previously acquired source sig-
nals. SO2 was then derived from these components, and val- Dividing this change by the “dc” (average) value, we
ues in the range of 78%~106% were obtained, which is rea- normalize this derivative to get a ratio that is a linear combi-
sonable. This preliminary work indicated the promise of nation of dA and dV :
ICA methods for this application and encouraged further
I ac dI t
study. r= = = − µ a dA − µv dV
I dc It (4)
While previous work indicated that ICA might be appli-
cable to pulse oximeter signals; the authors felt it was im- where µa and µv are linearly related to arterial and venous
portant to test the ICA assumption that each source signal
oxygen saturation, respectively. Note that (4) is wavelength
component pair is mutually independent from the other
components. More specifically, it is essential to determine dependent. I ac and I dc are acquired time series data. When
whether motion artifact and arterial volume variations are multiple light sources are used, (4) can be modified to the
independent, since this assumption is key to the application following vector format:
of ICA methods in pulse oximetry. Not only would this
analysis help to justify the application of ICA, principle I ac , i dI t , i
ri = = = − µ a , i dA − µ v , i dV
component analysis, and other blind source separation I dc , i I t ,i (5)
methods in this field; it would also help to specify user in- where i = R, IR refers to each light source wavelength.
structions that would avoid faulty applications of these ap-
proaches. This work is presented in the METHODS and RE- B. Approximate Calculation of Arterial Volume Varia-
SULTS sections of this paper.
tion in Stationary and Motion Conditions
3586
Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on May 25,2010 at 03:49:01 UTC from IEEE Xplore. Restrictions apply.
ume variations. Estimates of arterial volume variations in
both stationary and motion-corrupted cases are obtained 1
using the model presented in the THEORY section for the
interaction between optical signals and tissue. Correlations
between the two components reflect how much motion arti-
facts affect arterial volume fractions in different motion pat-
terns and provide information to guide appropriate applica- 2
tion of ICA techniques to pulse oximeter signals.
A. Experimental Setup and Data Collection
3
Data were collected using the wearable pulse oximeter in 4
Fig. 1. This pulse oximeter uses red and near-infrared light
sources with wavelengths of 660 nm and 940 nm respec-
tively. A PIC 16F873 microcontroller modulates the two
LEDs, controls data acquisition, and transmits acquired data
to a PC over a Bluetooth link through an RS-232 serial port.
Callisto-II Bluetooth modules from BrightCom (which are Figure 2. Different types of motion.
now distributed by Flextronics Semiconductor [15]), facili-
tate data exchange in cable-replacement mode using the
Bluetooth serial port profile. The PC saves these data to text
files.
Six sets of data were collected from a finger tip: one sta-
tionary set (no motion) and five sets where different types of
motion were imposed. Care was taken to minimize tissue
fractional volume variations. The five motion patterns were
the following (see Fig. 2):
1. Finger movements (three cases): left-right (swing-
ing), up-down (bending), and arbitrary finger move-
ments while the arm and wrist remained still.
2. Wrist movements: wrist rotations while keeping the
elbow and fingers still.
3. Elbow movements: stretching and bending the elbow
while keeping the wrist and fingers still.
4. Shoulder movements: rotation of the shoulder joint
while keeping the elbow, wrist, and fingers still.
5. Arbitrary movements: combinations of the above.
Two example data sets are shown in Fig. 3.
B. Data analysis
A MATLAB script reads the signal mixture from the test
Figure 1. Wearable pulse oximeter used to collect data. files, calculates arterial volume variation (Avv) in both sta-
tionary (Avvs) and motion conditions (Avvm), and finds the
3587
Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on May 25,2010 at 03:49:01 UTC from IEEE Xplore. Restrictions apply.
correlation between Avvs and each Avvm. Because data sets tically independent from motion. Caution is therefore pru-
for stationary and motion conditions are not acquired simul- dent when ICA methods are applied to pulse oximeter sig-
taneously, there is no absolute time correspondence between nals.
the two series – we assume that heart rate remains constant
between consecutive tests. Therefore, a set of correlation
coefficients with various time shifts is first calculated (using REFERENCES
the MATLAB built-in script xcov.m), then the maximum for 1. Yoo, S.K., Photoplethysmography (PPG) device and the
each pair is chosen to represent the statistical relationship method thereof. 2005: U.S. patent.
between Avvs and Avvm. 2. Stetson, P.F. Independent Component Analysis of Pulse
Oximetry Signals Based on Derivative Skew. in the Fifth
V. RESULTS AND DISCUSSION
International Conference on Independent Component
Analysis. 2004. Granada, Spain: Springer. p. 1072-1078
Fig. 4 illustrates the statistical relationship between Avv
3. Yao, J. and S. Warren. A Novel Algorithm to Separate
in stationary versus movement conditions. In all cases, the
Motion Artifacts from the Photoplethysmographic Signal
correlation coefficient between Avvs and Avvm is less than
for a Pulse Oximeter. in 4th Joint EMBS-BMES Conf.
0.5, implying the two are not significantly correlated. In
2004. San Francisco, CA. p. 2153-2156.
other words, motion seriously affects arterial volume varia-
4. Hayes, M.J. and P.R. Smith, A New Method for Pulse
tions and, subsequently, the photo-plethysmographic sig-
Oximetry Possessing Inherent Insensitivity to Artifact.
nals.
IEEE Transactions on Biomedical Engineering, 2001.
Looking at this figure, one can see that rotating the wrist 48(4): p. 452-461.
and stretching or bending the elbow affect Avv the most, 5. Rhee, S., B.-H. Yang, and H. H. Asada, Artifact-
while rotating the shoulder joint and swinging the finger resistant power-efficient design of finger-ring plethys-
horizontally affect Avv the least. Arbitrary movements mographic sensors. IEEE Transactions on Biomedical
combining all patterns have the worst effect. This is sensi- Engineering, 2001. 48(7): p. 795-805.
ble, given that these movements most greatly affect the 6. Stetson, P.F., Blind source separation of pulse oximetry
hemodynamics in the arterial system. signals. 2004, Nellcor Puritan Bennett Inc.: U.S. patent.
7. Stetson, P.F., Blind source separation of pulse oximetry
signals. 2003, Patent Application Publication
US2003/0088164: United States.
8. Vigario, R., Extraction of ocular artifacts from EEG
using independent component analysis. Electroenceph,
clin. Neurophysiol., 1997. 103(3): p. 395-404.
9. Jung, T.-P., et al., Removing electroencephalographic
artifacts by blind source separation. Psychophysiology,
2000. 37: p. 163-178.
10. Makeig, S., et al. Transiently time-locked fMRI activa-
tions revealed by independent components analysis. in
Proceedings of the National Academy of Sciences. 1997.
p. 803-810.
11. "The FastICA package for MATLAB,"
http://www.cis.hut.fi/projects/ica/fastica/
12. Yoshiya, Y.S. and K. Tanaka, Spectrophotometric moni-
toring of arterial oxygen saturation in the fingertip.
Med. Biol. Eng. Comput., 1980. 18: p. 27-32.
Figure 4. Correlation between arterial volume variations
13. "Technical Bulletin - Signal Extraction Technology®,"
in a stationary condition versus movement conditions
http://www.masimo.com/IMAGES/technology/lab1035.
that exhibit different motion patterns.
pdf
14. Rnceus. "Hemodynamic Measurement Terminology,"
VI. CONCLUSION
http://www.rnceus.com/hemo/term.htm
Recently published work regarding the use of ICA meth-
15. "Callisto II (TM) BluetoothTM RS-232 Cable Re-
ods to separate motion artifact from light-based sensor sig-
placement Evaluation Kit,"
nals assumes that motion artifact correlates only to changes
http://www.flextronicssemi.com/FlextronicsCallistoIIPr
in venous volume variations. This effort addressed the
oductBrief.pdf
statistical independence of motion artifacts and the arterial
and venous components of photo-plethysmographic data.
The results indicate that arterial volume variations are not
statistically independent from motion. Caution is therefore
3588
Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on May 25,2010 at 03:49:01 UTC from IEEE Xplore. Restrictions apply.