Proceedings of the 2005 IEEE
Engineering in Medicine and Biology 27th Annual Conference
Shanghai, China, September 1-4, 2005
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
critical when medical sensors are used in wearable monitoring
scenarios. Previous research has demonstrated that independ-
ent component analysis (ICA) can be applied to pulse oximeter
signals to separate photoplethysmographic (PPG) data from
motion artifacts, ambient light, and other interference in low-
motion environments. However, ICA assumes that all source
signal component pairs are mutually independent. It is impor-
tant to assess the statistical independence of the source compo-
nents in PPG data, especially if ICA is to be applied in ambula-
tory monitoring environments, where motion artifacts can have
a substantial effect on the quality of data received from light-
based sensors. This paper addresses the statistical relationship
between motion artifacts and PPG data by calculating the cor-
relation coefficients between arterial volume variations and
motion over a range of stationary to high-motion conditions.
Analyses indicate that motion significantly affects arterial flow,
so care must be taken when applying ICA to light-based sensor
data acquired from wearable platforms.
Keywords—Motion artifacts, independent component
analysis, pulse oximetry, wearable sensors
I. INTRODUCTION
Wearable medical devices are becoming popular in
home healthcare, telemedicine, rehabilitation, and athletic
training due to advances in sensors, device miniaturization,
power consumption, computation speed, and wireless com-
munication. These devices surpass the usefulness of their
traditional desktop counterparts by continuously monitoring
vital signs in “real-world” usage scenarios. However, mo-
tion artifact reduction and separation are a more serious
concern in wearable environments, prompting the attention
of researchers [1-5]. Recently, increased emphasis has been
placed on Independent Component Analysis (ICA) for the
separation of motion artifacts from desired quantities [1, 2,
6, 7]. ICA is especially attractive because it does not require
prior knowledge of the system. (ICA is consequently re-
ferred to as blind source separation.) ICA methods can
separate mixed signals with multiple source components
when multiple observation sets can be acquired.
Application of ICA is based on the assumption that all
source signal components are mutually independent. It is
therefore necessary to examine the relationship between
each source signal pair to make sure that they satisfy this
prerequisite. This paper addresses the statistical independ-
ence of two signal sources in photoplethysmographic (PPG)
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.
0-7803-8740-6/05/$20.00 ©2005 IEEE. 3585
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.
pulse oximeters: arterial volume variation and motion arti-
fact.
The following two sections discuss the promise of ICA
in pulse oximetry and provide the theoretical basis for this
approach. The Methods section then describes how data
were collected and how the statistical independence between
arterial volume variation and motion was assessed. The Re-
sults section summarizes these analyses and notes that the
application of ICA to PPG data should be performed with
caution.
II. BACKGROUND AND MOTIVATION
Independent component analysis is usually denoted as
x(t ) = M ⋅ s (t ) (1)
m×n
where M ∈R is the mixing matrix, x(t ) ∈ R m
is the
observed mixed signal vector, and s (t ) ∈ R is the source
n
component vector. This approach has been applied to the
well-known “cocktail party” problem, where multiple (n)
speakers’ (singers’) voices are recorded (mixed) by multiple
(m) microphones. The task of ICA is to recover each
speaker’s original voice s (t ) from the mixture x (t ) .
ICA has found application in the biomedical field, having
been successfully utilized in human electroencephalograms
[8, 9] and functional magnetic resonance imaging [10]. Re-
cently, researchers realized that ICA might be suitable for
the reduction and separation of both motion artifacts and
ambient light interference in photoplethysmographic pulse
oximeter data [1, 2, 6, 7]. In this context, s (t ) ∈ R refers
n
to the n source signal components: the volume fraction of
arterial blood, venous blood, bones, and other non-blood
tissues. The mixing matrix, M , represents the attenuation
of multi-wavelength optical signals by the n fractional vol-
umes when traveling through blood and tissues. Finally, the
mixture, x (t ) , is the m signals received from the optical
detector corresponding to the m wavelengths. In a real sys-
tem, the mixture is usually a current signal that has been
conditioned with analog and digital processing elements.
In practice, the number of independent components con-
tributing to the mixture signal is unclear. A typical optical
pulse oximeter can, with minimal hardware and software
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

III. THEORY In (5), dV is considered to be zero in a stationary condi-

tion, so
A. Optical Attenuation Through Tissue
rs
With pigmented epidermis and embedded bones ne- dA s = − (6)
glected, a finger tip can be considered a homogenous mix-
µa
ture of arterial blood, venous blood, and tissue. Given this In motion conditions, dA can be calculated using signals
assumption, light intensity in tissue is given by the Beer- for two light sources:
Lambert law [12]:
µ v , R ⋅ r m, IR − µ v , IR ⋅ r m, R
( )( )(
I t = I 0 e − µ t T e − µ vV e − µ a A ) (2) dA m =
µ a , R ⋅ µ v , IR − µ a , IR ⋅ µ v , R
where I t is the light intensity detected by the photodiode; (7)
Given arterial volume variations in both conditions, one
I 0 is the incident light intensity; µt , µv and µ a are wave- can now test how motion affects arterial volume variations
length-dependent absorption coefficients for bloodless tis- by examining the correlation between dA s and dA m . The
sue, venous blood, and arterial blood, respectively, in cm-1; greater the correlation, the less the arterial volume variation
and T , V , and A denote the volume fraction of tissue, is affected by motion artifacts.
venous blood, and arterial blood.
IV. METHODS
Blood is the greatest contributor to light attenuation in tis-
sue. To simplify analysis, tissue volume variations are usu- To study the statistical relationship between source com-
ally ignored due to their small contribution to optical at- ponents, we developed an experiment to collect mixtures
tenuation. The heart’s pumping action generates arterial (sets of motion-corrupted data) when the object (a finger) is
pulsations which create relative changes in arterial blood kept stationary or moved in different ways. We primarily
volume, represented by dA . Arterial blood pressure (e.g., > address two components: motion artifacts and arterial vol-

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.

Figure 3. PPG data in stationary (upper graph) and mo-

tion conditions (lower graph).

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
for stationary and motion conditions are not acquired simul-
taneously, there is no absolute time correspondence between
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
the MATLAB built-in script xcov.m), then the maximum for
each pair is chosen to represent the statistical relationship
between Avvs and Avvm.
V. RESULTS AND DISCUSSION
Fig. 4 illustrates the statistical relationship between Avv
in stationary versus movement conditions. In all cases, the
correlation coefficient between Avvs and Avvm is less than
0.5, implying the two are not significantly correlated. In
other words, motion seriously affects arterial volume varia-
tions and, subsequently, the photo-plethysmographic sig-
nals.
Looking at this figure, one can see that rotating the wrist
and stretching or bending the elbow affect Avv the most,
while rotating the shoulder joint and swinging the finger
horizontally affect Avv the least. Arbitrary movements
combining all patterns have the worst effect. This is sensi-
ble, given that these movements most greatly affect the
hemodynamics in the arterial system.
Figure 4. Correlation between arterial volume variations
in a stationary condition versus movement conditions
that exhibit different motion patterns.
VI. CONCLUSION
Recently published work regarding the use of ICA meth-
ods to separate motion artifact from light-based sensor sig-
nals assumes that motion artifact correlates only to changes
in venous volume variations. This effort addressed the
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.
tically independent from motion. Caution is therefore pru-
dent when ICA methods are applied to pulse oximeter sig-
nals.
REFERENCES
1. Yoo, S.K., Photoplethysmography (PPG) device and the
method thereof. 2005: U.S. patent.
2. Stetson, P.F. Independent Component Analysis of Pulse
Oximetry Signals Based on Derivative Skew. in the Fifth
International Conference on Independent Component
Analysis. 2004. Granada, Spain: Springer. p. 1072-1078
3. Yao, J. and S. Warren. A Novel Algorithm to Separate
Motion Artifacts from the Photoplethysmographic Signal
for a Pulse Oximeter. in 4th Joint EMBS-BMES Conf.
2004. San Francisco, CA. p. 2153-2156.
4. Hayes, M.J. and P.R. Smith, A New Method for Pulse
Oximetry Possessing Inherent Insensitivity to Artifact.
IEEE Transactions on Biomedical Engineering, 2001.
48(4): p. 452-461.
5. Rhee, S., B.-H. Yang, and H. H. Asada, Artifact-
resistant power-efficient design of finger-ring plethys-
mographic sensors. IEEE Transactions on Biomedical
Engineering, 2001. 48(7): p. 795-805.
6. Stetson, P.F., Blind source separation of pulse oximetry
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.
13. "Technical Bulletin - Signal Extraction Technology®,"
http://www.masimo.com/IMAGES/technology/lab1035.
pdf
14. Rnceus. "Hemodynamic Measurement Terminology,"
http://www.rnceus.com/hemo/term.htm
15. "Callisto II (TM) BluetoothTM RS-232 Cable Re-
placement Evaluation Kit,"
http://www.flextronicssemi.com/FlextronicsCallistoIIPr
oductBrief.pdf