Combination of Canonical Correlation Analysis and Empirical Mode Decomposition Applied to Denoising the Labor Electrohysterogram

Abstract

The electrohysterogram (EHG) is often corrupted by electronic and electromagnetic noise as well as movement artifacts, skeletal electromyogram, and ECGs from both mother and fetus. The interfering signals are sporadic and/or have spectra overlapping the spectra of the signals of interest rendering classical filtering ineffective. In the absence of efficient methods for denoising the monopolar EHG signal, bipolar methods are usually used. In this paper, we propose a novel combination of blind source separation using canonical correlation analysis (BSS_CCA) and empirical mode decomposition (EMD) methods to denoise monopolar EHG. We first extract the uterine bursts by using BSS_CCA then the biggest part of any residual noise is removed from the bursts by EMD. Our algorithm, called CCA_EMD, was compared with wavelet filtering and independent component analysis. We also compared CCA_EMD with the corresponding bipolar signals to demonstrate that the new method gives signals that have not been degraded by the new method. The proposed method successfully removed artifacts from the signal without altering the underlying uterine activity as observed by bipolar methods. The CCA_EMD algorithm performed considerably better than the comparison methods.

Full text

Page 1

 1  
 
Abstract?  the  electrohysterogram  (EHG)  is  often  corrupted  by  
electronic   and   electromagnetic   noise   as   well   as   movement  
artifacts,   skeletal   EMG   and   electrocardiograms   from   both  
mother   and   fetus.   The   interfering   signals   are   sporadic   and/or  
have   spectra   overlapping   the   spectra   of   the   signals   of   interest  
rendering  classical  filtering  ineffective.  In  the  absence  of  efficient  
methods   for   denoising   the   monopolar   EHG   signal,   bipolar  
methods   are   usually   used.   In   this   paper,   we   propose   a   novel  
combination   of   Blind   Source   Separation   using   Canonical  
Correlation  Analysis  (BSS_CCA)  
Decomposition  (EMD)  methods  to  denoise  monopolar  EHG.  We  
first   extract   the   uterine   bursts   by   using   BSS_CCA   then   the  
biggest  part  of  any  residual  noise  is  removed  from  the  bursts  by  
EMD.   Our   algorithm,   called   CCA_EMD,   was   compared   with  
wavelet  filtering  and  independent  component  analysis  (ICA).  We  
also   compared   CCA_EMD   with   the   corresponding   bipolar  
signals   to   demonstrate   that   the   new   method   gives   signals   that  
have   not   been   degraded   by   the   new   method.   The   proposed  
method   successfully   removed   artifacts   from   the   signal   without  
altering  the  underlying   uterine  activity  as   observed   by   bipolar  
methods.   The   CCA_EMD   algorithm   performed   considerably  
better  than  the  comparison  methods.    
 
Index  Terms?  uterine  EMG;;  canonical  correlation  analysis;;  
empirical  mode  decomposition;;  preterm  labor.    
and  Empirical  Mode  
I.   INTRODUCTION  
lectrohysterogram   (EHG)   is   the   uterine   electromyogram  
signal   recorded   externally   on   pregnant   women.   This  
signal  has  long  been  expected  to  have  important  clinical  
applications,   especially   in   relation   to   predicting   and  
preventing  pre-­term  labor.  The  EHG  is  a  problematic  signal  
that   has   very   low   frequency   and   is   very   low   amplitude   as  
compared  to  various  sources  of  contaminating  noise.  In  this  
paper  we  describe  a  large  step  in  helping  to  isolate  the  signal  
from  noise,  and  thus  to  overcome  important  barriers  that  have  
prevented   exploiting   the   potential   of   the   signal   to   provide  
information   on   the   genesis   and   evolution   of   human   labor.        
During   recordings,   the   EHG   is   frequently   contaminated   by  
different   artifacts   which   may   often   complicate   the  
interpretation   of   the   EHG.   The   typical   noise   origins   in   the  
EHG   are   maternal   and   fetal   electrocardiograms   (ECG),  
abdominal  muscle  electromyogram  (EMG),  maternal  and  fetal  
movement  artifacts  and  electronic  noise  from  the  surrounding  
electronic  devices.    
The  EHG  is  a  noisy  signal  even  recorded  by  bipolar  electrodes  
that  are  fairly  close  together.  Wavelet  filtering  has  been  used  
successfully  on  bipolar  EHG  for  removing  maternal  and  fetal  
ECG  as  well  as  stationary  electric  noises  [1].  Most  techniques  
of  wavelet  filtering  assume  that  the  noise  is  of  low  amplitude  
and   stationary   when   compared   to   the   signal   of   interest.  
However,  the  noise  in  monopolar  EHG  is  not  stationary  and  
usually   of   higher   amplitude   than   the   signal   of   interest.   In  
addition,   many   sources   of   noise   have   main   frequency  
components  that  are  close  to  the  frequency  of  the  signal  of  
interest.   The   EHG   has   most   of   its   energy   in   the   frequency  
band   of   0.1   to   1.5   Hz,   as   compared   to   maternal   cardiac  
frequency   (typically   1.2   Hz),   maternal   respiration   (typically  
0.2  Hz)  and  fetal  respiration  (typically  1  Hz).  As  the  noise  is  
in  the  same  frequency  bands  and/or  is  of  high  amplitude  and  
sporadic,  it  cannot  be  rejected  by  classical  or  wavelet  filters.    
Recently,   researchers   have   focused   their   efforts   on  
multichannel   EHG   in   the   hope   of   applying   propagation  
analysis  to  predict  preterm  labor.  An  example  of  this  type  of  
approach  is  the  work  on  signals  obtained  from  a  16  electrode  
matrix  that  have  been  recorded  in  Iceland  in  the  last  few  years  
[2].  All  the  related  studies  have  been  done  by  using  bipolar  
signals  to  reduce  noise  effect  [3-­5].  Spatial  resolution  is  very  
important   for   signal   propagation   analysis   and   the   use   of   a  
bipolar  signal  reduces  this  resolution  by  almost  half.  Another  
important  limitation  in  using  bipolar  electrodes  in  a  matrix  is  
that   any   analysis   of   the   direction   of   propagation   becomes  
meaningless  or  biased,  as  the  direction  of  propagation  along  
the   line   between   the   two   bipolar   electrodes   is   privileged.  
Bipolar  measurements  of  electric  phenomena  are  basically  a  
way  of  rejecting  the  part  of  the  signal  that  is  measured  on  both  
electrodes  and  keeping  only  the  part  that  is  dissimilar  to  the  
two  electrodes.  The  assumption  underlying  this  technique  is  
????? ???? ???????? ????????al   activity   is   dissimilar   to   the   two  
electrodes  and  that  the  common  part  is  from  further  afield  and  
thus  not  relevant.  It  is  highly  uncertain  that  this  holds  for  the  
Combination  of  Canonical  Correlation  Analysis  and  
Empirical  Mode  Decomposition  applied  to  denoising  the  
labor  Electrohysterogram  
Hassan  M.,  Boudaoud  S.,  Terrien  J.,  Karlsson  B.  and  Marque  C.  
 
E  

Page 2

 2  
case   of   the   EHG.   It   is   therefore   likely   that   important  
information  is  lost  by  using  bipolar  signals  only.  
Blind  Source  Separation  (BSS)  methods  such  as  Independent  
Component   Analysis   (ICA)   and   Principal   Component  
Analysis   (PCA)   are   increasingly   being   used   in   biomedical  
signal  processing  involving  analysis  of  multivariate  time  series  
data  such  as  EEG  [6,  7].  
Recently,   a   new   method   for   muscle   artifact   elimination   in  
scalp  EEG  has  been  developed  that  does  not  have  some  of  the  
disadvantages   of   ICA   [8].   The   method   is   based   on   the  
statistical   Canonical   correlation   analysis   (CCA)   method  
applied  as  a  blind  source  separation  (BSS)  technique,  called  as  
BSS_CCA.   This   method   has   shown   considerably   better  
performance  than  ICA  in  some  applications  [8,  9].  
In  this  paper  the  aim  of  CCA  is  the  extraction  of  the  uterine  
bursts   based   on   the   hypothesis   that   the   bursts   have   higher  
autocorrelation  coefficients  than  the  noise.  Thus  the  expected  
result  from  applying  CCA  is  the  extraction  of  uterine  burst.  
The  bursts  observed  after  applying  CCA  only,  contain  fewer  
artifacts  than  the  original  signals  but  still  contain  the  artifacts  
that  have  high  autocorrelation.  Empirical  mode  decomposition  
(EMD)  was  chosen  as  a  second  step  to  remove  this  noise  from  
the  extracted  burst.  EMD  technique  was  introduced  by  Huang  
et   al.   [10]   to   analyze   nonstationary   and   nonlinear   signals.  
EMD  has  become  a  very  important  tool  to  analyze  biomedical  
signals.   The   use   of   EMD   for   analyzing   esophageal  
manometric  data  in  gastro  esophageal  reflux  disease  indicated  
good   performance   in   removing   different   kind   of   artifacts  
(respiratory,  motion...)  from  electrogastrogram  (EGG)  signals  
[11,  12].  The  EMD  approach  also  proved   to  be  efficient  in  
removing   artifacts   from   ECG   signals   [13].   For   EEG   data,  
EMD  algorithm  has  also  been  employed  [14].  In  this  work  it  
was  demonstrated  how  reconstruction  using  the  Hilbert-­Huang  
transform  can  successfully  be  applied  to  contaminated  EEG  
data  for  the  purposes  of  removing  unwanted  ocular  artifacts.  
More   recently,   Wu   and   Huang   have   introduced   a   noise  
assisted   version   of   the   EMD   method,   called   Ensemble  
Empirical  mode  decomposition  (EEMD)  [15].    This  method  
has   shown   better   performance   than   EMD   as   it   extracts   the  
IMFs  in  a  manner  so  that  the  mode  mixing  disadvantage  of  the  
EMD  method  is  corrected.  We  did  not  implement  it  in  this  
work,  as  EEMD  is  much  slower  than  EMD  and  has  not  been  
used,   to   our   knowledge,   in   signal   processing   for   clinical  
applications.    
Wavelet   transform   (WT)   denoising   methods   could   also   be  
good   candidates   for   the   second   step   of   the   processing  
presented  here.  However,  in  this  first  attempt  we  chose  EMD  
rather  than  WT  methods  for  two  reasons:  i)  EMD  is  a  data-­
driven  algorithm:  it  decomposes  the  signal  in  a  natural  way  
without  prior  knowledge  about  the  signal  of  interest  embedded  
in  the  noise  which  is  not  the  case  the  wavelet  transform;;  ii)  
EMD   demonstrated   better   performance   than   WT   when  
combined  with  ICA  in  denoising  EEG  signals  [16].      
The  aim  of  this  work  is  to  combine  the  use  of  BSS_CCA  and  
EMD  algorithms  to  design  a  new  tool,  called  CCA_EMD,  to  
remove   the   main   interferences   embedded   in   monopolar  
abdominal   EHG   recordings.   The   method   is   applied   to   real  
signals   recorded   on   women   during   pregnancy   and   labor.   A  
comparison  to  bipolar  signals  obtained  in  the  same  matrix  is  
made.   A   quantitative   comparison   between   CCA_EMD,  
wavelet  filtering,  bipolar  signals  and  ICA  is  also  presented.    
II.   MATERIELS  AND  METHODS  
A.   Recording  protocol  
The  EHG  signal  recording  was  performed  using  a  16-­channel  
multi-­purpose  physiological  signal  recorder,  most  commonly  
used  for  investigating  sleep  disorders  (Embla  A10).  We  used  
reusable   Ag/AgCl   electrodes.   The   measurements   were  
performed   at   the   Landspitali   University   hospital   in   Iceland,  
following   a   protocol   approved   by   the   relevant   ethical  
committee  (VSN  02-­0006-­V2).  
The  signal  sampling  rate  was  200  Hz  on  16  bits.  The  recording  
device   has   an   anti-­aliasing   filter   with   a   low   pass   cut-­off  
frequency  of  100  Hz.  The  concurrent  tocodynamometer  paper  
trace  (TOCO)  was  digitized  in  order  to  ease  the  identification  
of  contractions.  Recordings  from  six  women  during  labor  are  
used  in  this  work.    
B.   Canonical  Correlation  Aanalysis  (CCA)  
In   BSS   approach,   the   observed   multichannel   signals   are  
assumed   to   reflect   a   linear   combination   of   several   sources  
which  are  associated  with  underlying  physiological  processes,  
artifacts  and  noise.  The  BSS  approach  aims  to  recover  a  set  of  
unobserved   source   signals   by   using   only   a   set   of   observed  
mixtures   of   sources.   The   observed   time   series   X(t)   =[x1(t);;  
x2(t);;  .  .  .  ;;  xK(t)]T  ,  is  the  result  of  an  unknown  mixture  of  a  set  
of  unknown  source  signals  S(t)  =  [s1(t);;  s2(t);;  .  .  .  ;;  sK(t)]T  with  
t  =  1;;  .  .  .;;N,  where  N  the  number  of  samples,  K  the  number  of  
sensors   and   T   is   the   transpose   operator.   The   mixing   is  
assumed  to  be  linear,  thus  reducing  the   mixing  to  a   matrix  
multiplication  
( )X t
?
where  A  is  the  unknown  mixing  matrix.  The  aim  is  to  estimate  
the  mixing  matrix  and  recover  the  original  source  signals  S(t).  
This  could  be  done  by  introducing  the  de-­mixing  matrix   W  
such  that  it  approximates  the  unknown  source  signals  in  S(t),  
by  a  scaling  factor:  
( )Z t WX t
?
Ideally,  W  is  the  inverse  of  the  unknown  mixing  matrix  A,  up  
to  scaling  and  permutation.  There  are  many  ways  to  solve  the  
BSS   problem   depending   on   the   definition   of   contrast  
functions.   The   ICA   method   tries   to   make   the   estimated  
( )AS t
 
( )
 

Page 3

 3  
sources  as  non-­Gaussian  as  possible.  However,  in  PCA  and  
most  of  the  ICA  algorithms,  the  temporal  correlations  are  not  
taken  into  consideration  for  solving  contrast  functions.  CCA  
solves   this   BSS   problem   by   forcing   the   sources   to   be  
maximally  autocorrelated  and  mutually  uncorrelated,  while  the  
mixing  matrix  is  assumed  to  be  square  [17].  
Consider   the   observed   data   matrix   X(t)   and   its   temporally  
delayed  version  Y(t)  =  X(t-­1).  The  CCA  method  obtains  two  
sets  of  basis  vectors,  one  for  x  and  the  other  for  y,  such  that  the  
correlations   between   the   projections   of   the   variables   onto  
these  basis  vectors  are  mutually  maximized.  
The  total  covariance  matrix  is  given  by    
 
 
T
XX XY
YX YY
C
C
C
C
X
Y
X
Y
CE
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
??
??
??
?
?
?
??
 
where  CXX  and  CYY  are  the  auto-­covariance  matrices  of  X  and  Y  
respectively,  CXY  is  the  cross-­  covariance  matrix  (CXY=  CT
and  E  is  the  expectation  operation.  
The  canonical  correlation  between  X  and  Y  can  be  calculated  
by  solving  these  equations:  
YX),  
112
112
XX
XXXYYY YX
YY
YYYXXX XY
C C C Cww
C C C Cww
?
?
??
??
?
?
 
with   the   canonical   correlation   coefficient   ?   as   the   square  
root  of  the  eigen-­value,  and  wX  and  wY  as  eigenvectors.  Since  
the  solutions  are  related,  only  one  of  the  eigen-­value  equations  
needs  to  be  solved  to  get  the  demixing  matrix  w.  The  CCA  
gives  the  source  signals  that  are  uncorrelated  with  each  other,  
maximally   autocorrelated   and   ordered   by   decreasing  
autocorrelation.  
When   BSS-­CCA   is   applied   to   the   EHG,   the   sources  
contributing  to  the  EHG  and  noise  are  obtained.  The  artifacts  
can   be   removed   by   setting   equal   to   zero,   the   columns  
representing  the  activations  of  the  related  sources,  before  the  
( )
Denoised
Xt :  
( )
Denoised
Xt
?
reconstruction  of  
( )
Denoised
AZ t
 
with  Z(t)  the  sources  obtained  by  BSS-­CCA,  and  
Denoised
A
the  
mixing  matrix,  with  its  columns  related  to  artifact  sources,  set  
to  zero.  
C.   Empirical  Mode  Decomposition  (EMD)  
The  empirical  mode  decomposition  (EMD)  was  proposed  by  
Huang   et   al.   as   a   new   signal   decomposition   method   for  
nonlinear   and   nonstationary   signals   [10].   The   EMD  
decomposes   a   signal   into   a   collection   of   oscillatory   modes,  
called  intrinsic  mode  functions  (IMF),  which  represent  fast  to  
slow  oscillations  in  the  signal.  Each  IMF  can  be  viewed  as  a  
subband  of  a  signal.  Therefore,  the  EMD  can  be  viewed  as  
subband  signal  decomposition.  
Given  a  signal  x(t),  the  effective  algorithm  of  EMD  can  be  
summarized  as  follows  [10]:  
1.   Identify  all  extrema  of  x(t)  
2.   Interpolate   along   the   point   of   x(t)   identified   in   the  
first  step,  in  order  to  form  an  upper  emax(t)  and  lower  
envelope  emin(t).  
3.   Compute  the  mean  m(t)=(  emin(t)  +  emax(t))/2  
4.   Extract  the  detail  d(t)=x(t)-­m(t)  
5.   Iterate  on  the  residual  m(t)  
In  practice,  the  above  procedure  has  to  be  refined  by  a  sifting  
process  [10]  which  amounts  to  first  iterating  steps  1  to  4  upon  
the  detail  signal  d(t),  until  this  latter  can  be  considered  as  zero-­
mean   according   to   some   stopping   criterion.   Once   this   is  
achieved,   the   detail   is   referred   to   as   an   Intrinsic   Mode  
Function  (IMF),  the  corresponding  residual  is  computed  and  
step   5   applies.   By   construction,   the   number   of   extrema   is  
decreased  when  going  from  one  residual  to  the  next,  and  the  
whole   decomposition   is   guaranteed   to   be   completed   with   a  
finite  number  of  modes.  
Denoising   by   EMD   is   in   general   carried   out   by   partial  
signal   reconstruction,   which   is   based   on   the   fact   that   noise  
components  lie  in  the  first  several  IMFs.    
D.   CCA/EMD  combination  
 We  assume  that  the  BSS  is  the  best  way  to  extract  the  uterine  
bursts  and  based  on  the  hypothesis  that  the  sources  of  uterine  
bursts   have   higher   autocorrelation   than   the   sources  
corresponding  to  the  artifacts,  we  choose  the  CCA  method  as  a  
way   to   extract   the   uterine   bursts   and   in   the   same   time  
eliminate   all   the   low   autocorrelated   noise.     The   sources   of  
device  noise  (electronic  artifacts)  are  high  autocorrelated  and  
then  it  is  not  possible  to  remove  it  by  using  only  the  CCA  
method.   It   has   been   demonstrated   that   EMD   shows   good  
performance  in  removing  this  kind  of  noise  [12].  For  this  and  
other  reasons,  we  chose  to  use  EMD  as  the  complementary  
tool   to   remove   the   residual   electronic   noise.   We   call   this  
combination  the  CCA_EMD  algorithm.  
E.   Comparative  study  
For  performance  evaluation  purposes,  the  proposed  technique  
is  compared  with  two  other  techniques  that  we  have  already  
used  for  denoising  EHG  signals  in  our  group.  
?  
Wavelet  filtering:  The  specific  algorithm  which  is  based  
on   the   redundant   wavelet   packet   transform   was  
developed  by  Leman  et  al.  and  has  already  been  used  to  
remove  artifacts  from  bipolar  EHG  signal.  By  applying  
this  algorithm,  the  corrupting  ECG  seems  to  be  totally  
removed   while   all   other   artifacts   (fetal   movements,  
electronic  noise)  are  still  presented  [1].    
?  
FastICA:   This   is   a   BSS   approach   that   uses   similar  
techniques   as   CCA_EMD   and   is   therefore   useful   for  
comparison.   The   selection   of   the   ICA   components  

Page 4

 4  
accounting  for  the  artifact  removing  was  based  on  visual  
inspection  and  the  signal  was  reconstructed,  excluding  
the  components  related  to  the  artifact.  
For  a  quantitative  comparison  between  the  different  methods,  
signal  to  noise  ratio  (SNR)  was  used  as  a  criterion.  For  each  
contraction  and  each  EHG  channel,  the  SNR  was  estimated  by  
computing  the  energy  of  the  burst  over  the  energy  of  the  base  
lines   present   before   and   after   the   EHG   burst.   The   SNRs  
obtained  by  CCA_EMD  were  compared  with  those  obtained  
with   bipolar   signal   (vertical   differentiation),   ICA   and   the  
wavelet  method.  For  statistical  comparison,  we  used  the  two  
tailed  sign  test  with  a  minimal  significant  level  of  0.05.  
III.   RESULTS  
Figure  1A  shows  a  typical  example  of  raw  monopolar  EHG  
signals  recorded  from  a  woman  during  labor.  The  canonical  
components  (CC)  obtained  by  applying  the  BSS_CCA  method  
are  shown  in  Figure  1B.  The  fetal   movements  pump  spikes  
(located  in  the  16th  CC)  and  part  of  electronic  noises  (located  
in  the  14th  and  the  15th  CC)  are  well  distinguished  from  the  
components  related  to  the  uterine  activity.    
A  very  important  step  to  take  into  account  here  is  the  choice  of  
the  CCs  corresponding  to  the  artifacts,  in  order  to  then  remove  
them   before   signal   reconstruction.   We   should   detect   the  
?????????? ?????????????? ??? ???? ??????????? ????? ?????????
?????????????????????????????????????????????????????  
The  methodology  we  propose  to  choose  the  optimal  threshold  
value  can  be  described  as  following:  
?  
Calculate   the   CCA   components   and   the   associated  
autocorrelation  coefficients.    
?  
We  choose  a  threshold  ranging  between  0  and  1  (with  
0.1  steps),  then  remove  the  CCs  below  this  value  and  
reconstruct  the  signals.  
?  
We  compute  the  original  bipolar  signals  (BipOrg)  from  
two   raw   channels   X   and   Y,   and   the   bipolar   signal  
obtained  from  the  same  channels  after  the  two  preceding  
processing  steps  (BipDen).  We  then  get  two  versions  of  
the   bipolar   signals,   one   created   from   the   raw   signals  
directly  (BipORG)  and  the  other  by  eliminating  all  the  
CCs   below   the   given   threshold   of   autocorrellation.  
(BipDen)    
?  
We  then  compute  the  correlation  between  BipOrg  and  
BipDen.  
?  
We  repeat  these  steps  for  the  20  contractions  from  six  
women  and  then  we  calculate  the  average  and  standard  
deviation  at  each  autocorrelation  value,  for  each  given  
threshold.    
 
 
 
 
 
Fig.1.  A)  Original  raw  signals  from  a  woman  in  labor  B)  corresponding  CCA  
components    
 
The   curve   in   figure   2A   indicates   that   0.5   is   the   optimal  
threshold   value   for   eliminating   noisy   CCs   as   we   got   the  
highest  correlation  between  BipOrg  and  BipDen  for  this  value.  
Computation  of  the  mean  square  error  (MSE)  instead  of  the  
correlation  confirms  this,  as  the  lowest  MSE  is  observed  at  0.5  
(results  not  shown).  
Figure   2B   presents   the   autocorrelation   coefficients   curve  
computed   for   the   20   contractions   from   six   women   during  
labor,  and  the  autocorrelation  coefficients  curve  for  the  CC  of  
the  contraction  presented  in  figure  1.  The  threshold  (presented  
as  a  dashed  horizontal  line)  is  fixed  at  0.5.  All  the  CCs  below  
this  value  are  excluded.  For  the  presented  contraction  (solid  
line)   we   reject   the   last   three   CCs.   We   can   notice   that   this  
threshold   corresponds   to   the   last   four   CCs   on   the   curve  
representing  the  autocorrelation  coefficients  averaged  over  the  
6  women  during  labor  (Figure  2  dashed  curve).    
 
20406080100120 140160
16
15
14
13
12
11
10
9  
8  
7  
6  
5  
4  
3  
2  
1  
Time(s)
20 4060 80100120140160
16
15
14
13
12
11
10
9  
8  
7  
6  
5  
4  
3  
2  
1  
Time(s)
A)  
B)  

Page 5

 5  
 
 
 
Fig.2.  A)  The  correlation  coefficient  between  Bip  Org  (raw  bipolar  signal)  and  
Bipden  (processed  bipolar  signal)  in  function  of  the  autocorrelation  threshold  
B)   Solid  curve:   autocorrelation  coefficients   from  the  contraction   shown   in  
Figures  1.  Dashed  curve:  averaged  autocorrelation  curve  obtained  from  the  6  
women  in  labor.  The  horizontal  dashed  line  represent  the  threshold  value.  
 
By  using  a  threshold  value  equal  to  0.5  and  removing  before  
reconstruction  of  the  EHG  the  components  below  this  value,  
we  obtain   the  intermediate  denoised  EHGs   shown  in   Fig.3.  
Notice  that  all  the  fetal  movements,  maternal/fetal  ECG  and  
part  of  electronic  noises  have  been  removed  from  the  original  
signal.   After  BSS_CCA  denoising,   some   noise  remains  that  
presents   high   autocorrelation   coefficient   (specially   the  
electronic  noise  coming  from  the  devices).  These  artifacts  are  
not  completely  removed  by  BSS_CCA.  To  remove  this  noise  
we   apply   EMD   to   the   signals   previously   denoised   by  
BSS_CCA.  
 
Fig.3.  Uterine  bursts  extracted  by  CCA  only.  The  different  curves  represent  
the  16  simultaneous  recorded  channels  for  one  contraction.  
 
Figure  4  shows  the  final  signals  after  CCA_EMD.  Here,  based  
on   visual   inspection,   partial   reconstruction   is   applied   by  
removing   the   first   three   IMFs   that   we   consider   to   be   high  
frequency   noise.   In   figure   4,   we   can   clearly   differentiate  
between  the  baseline  and  the  uterine  bursts,  which  will  greatly  
ease  the  segmentation  of  uterine  bursts.  
 
Fig.4.  Signals  denoised  by  CCA  _EMD.  The  two  vertical  lines  indicate  the  
start  and  the  end  of  the  uterine  bursts  based  on  the  TOCO  trace.  
 
Figure  5  presents  an  example  of  one  of  the  channels  extracted  
from  the  signals  shown  in  Figures  1,  3  and  4,  in  order  to  better  
evidence   the   way   each   step   of   CCA_EMD   improves   the  
signal,  from  a  very  noisy  monopolar  signal  with  various  types  
of   artifacts,   to   a   clearly   visible   uterine   burst.   This   result   is  
based  on  a  threshold  autocorrelation  value  of  0.5  as  described  
above.    
00.10.20.3 0.40.50.60.7 0.8 0.91
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Autocorrelation  threshold
Correlation
246
CCA  Component
81012 1416
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Autocorrelation
 
 
Over  6  women
presented  woman
20406080 100120140 160
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Time(s)
20 406080100120140160
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Time(sec)
B)  

Page 6

 6  
 
Fig.5.  Top:  Original  signal;;  Middle:   signal  after   CCA  processing;;   Bottom:  
denoised  signal  after  CCA_EMD.  
 
Some   residual   noise   from   pump   spikes   and   high   frequency  
electronic  noise  is  clearly  visible  on  the  middle  panel  of  figure  
5.  This  residual  noise  is  then  completely  erased  from  the  burst  
by   applying   the   EMD   method   (figure   5   bottom).   As   the  
bipolar  signal  is  our  only  available  reference,  we  then  compare  
the  raw  bipolar  signal  (computed  as  the  difference  between  the  
raw  channel-­1  and  raw  channel-­2,  BipOrg)  and  the  denoised  
bipolar  signal  (computed  as  the  difference  between  denoised  
channel-­1  and  denoised  channel-­2,  BipDen).    
In  figure  6  we  present  a  typical  example  that  indicates  how  
CCA_EMD  enhances  the  uterine  activity  information  at  least  
as   well   as   the   bipolar   method.   The   two   bipolar   signals  
(original   and   denoised)   are   quite   similar   and   there   is   no  
distortion   (RMS   error=10-­4,   correlation   coefficient=0.74).  
Time-­frequency   representations   (Scalograms)   show   that   the  
CCA_EMD   has   not   visibly   affected   the   uterine   burst  
frequency  content  when  compared  to  bipolar  signals.    
Figure  7  shows  the  quantitative  difference  between  the  SNR  of  
monopolar,   bipolar,   CCA_EMD,   FastICA   and   the   Wavelet  
algorithm  described  in  [1].  The  values  represent  the  median  of  
SNR  over  the  16  channels  for  the  six  women  during  labor.  The  
monopolar  raw  signals  present  a  poor  SNR  (median  SNR  =  -­
3.4  dB).  The  results  indicate  clearly  that  CCA_EMD  has  the  
highest  median  SNR  equal  to  6.74  dB  compared  to  3.16  dB  for  
wavelet,  5.02  dB  for  FastICA  and  5.2  dB  for  bipolar  signals.  
 
 
 
 
 
 
 
 
 
 
Fig.6.   From   top   to   bottom:   monopolar   channel-­1;;   monopolar   channel-­2;;  
corresponding  bipolar  signals:  raw   data  (Left)  and  denoised  data  (Right)..;;  
scalograms  of  the  bipolar  signal  computed  from  raw  data  (Left)  and  Denoised  
data  (Right).  
 
Fig.7.  The  SNR  for  monopolar  signal,  wavelet,  ICA,  bipolar  and  CCA_EMD  
methods.    
204060 80100 120 140160
-­0.5
0
0.5
Original  signal
2040 60 80100120 140160
-­0.1
-­0.05
0
0.05
0.1
CCA
2040 60 80100 120 140160
-­0.1
-­0.05
0
0.05
0.1
CCA  EMD
Time(sec)
20 4060 80100 120140160
-­0.8
-­0.6
-­0.4
-­0.2
0
0.2
0.4
0.6
2040 6080 100120140160
-­0.05
0
0.05
204060 80100 120140160
-­0.8
-­0.6
-­0.4
-­0.2
0
0.2
0.4
0.6
204060 80100120140160
-­0.05
0
0.05
20 4060 80 100120 140160
-­0.06
-­0.04
-­0.02
0
0.02
0.04
0.06
Time(sec)
2040 6080 100120140 160
-­0.06
-­0.04
-­0.02
0
0.02
0.04
0.06
Time(sec)
Time(sec)
Frequency(Hz)
 
 
20406080 100 120140 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.005
0.01
0.015
0.02
Time(sec)
Frequency(Hz)
 
 
204060 80 100120 140 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
-­8
-­6
-­4
-­2
0
2
4
6
8
MonopolarWavelet ICA Bipolar CCA_EMD
SNR

Page 7

 7  
IV.   DISCUSSION  AND  CONCLUSION    
In  this  work  a  novel  combination  of  two  important  methods  
has  been  used  in  order  to  completely  remove  artifacts  from  a  
monopolar   EHG   signal.   The   method   consists   of   two   steps.  
First,  BSS_CCA   is  used  to  extract  the  uterine  bursts  in   the  
presence  of  high  intensity  noise   with  overlapping  frequency  
band.  Then  the  signals  are  given  a  final  cleaning  by  applying  
EMD.  
In  addition,  the  method  described  is  fast  and  computationally  
economical   and   can   be   used   as   a   preprocessing   step   to  
facilitate   segmentation  of  uterine  EMG  bursts.  Furthermore,  
the  noise  activities  are  present  in  the  least  autocorrelated  CCA  
components,   which   indicates   that   it   is   possible   with   this  
technique   to   separate   noise   sources   (small   autocorrelation)  
from  the  uterine  activity  (higher  autocorrelation).  
Concerning  the  CCA  method,  we  have  in  this  work  proposed  a  
method  that  analyses  the  autocorrelation  coefficient  curve  to  
estimate   the   ones   that   have   to   be   eliminated,   based   on   a  
similarity   to   a   reference   signal   (BipOrg).   In   addition,   the  
estimation   of   the   threshold   value   by   using   criteria   such   as  
Akaike   Information   Criteria   could   be   an   important   parallel  
way  to  optimize  the  threshold  selection.  
The   removal   of   artifacts   by   EMD   was   based   on   visual  
inspection  of  the  IMFs.    For  this  method  to  be  useful  in  real  
time  and/or  in  a  clinical  setting,  the  selection  of  the  proper  
IMFs   has   to   be   automated   based   also   on   the   use   of   some  
computable  criterion.  
This  paper  introduces  the  first  combination  between  CCA  and  
another  method  applied  to  a  particular  signal.  In  future  work,  
comparison   between   CCA_EMD   and   other   combination  
method  could  be  interesting,  i.e.  the  use  of  wavelet  transform  
instead   of   EMD,   testing   EEMD   instead   of   EMD   and   the  
comparison   with   other   existing   combination   such   as  
???????????????????????  
We  are  presently  working  on  the  comparison  of  bipolar  and  
denoised   monopolar   EHG   signals   to   investigate   the  
propagation   of   the   uterine   electrical   activity.   Preliminary  
results  show  a  clear  superiority  of  monopolar  analysis  as  it  has  
better  spatial  resolution  and  it  does  not  bias  the  measure  of    
propagation  directions  as  bipolar  methods  usually  do.    
Finally,  we  conclude  that  this  work  has  opened  the  door  for  
the  use  of  monopolar  EHG  recordings  to  investigate  uterine  
contractions.   This   will,   in   our   opinion,   have   important  
applications   in   studying   the   genesis   of   human   labor   and   to  
develop  ways  of  predicting  preterm  labor.    
V.   ACKNOWLEDGMENT  
This  project  is  financed  by  the  Icelandic  centre  for  research  
RANNÍS  and  the  French  National  Center  for  University  and  
School  (CNOUS).  The  authors  would  like  to  especially  thank  
Thora   Steingrimsdottir   and   Ásgeir   Alexandersson   for   their  
help  in  the  acquisition  of  EHG  signals.  
VI.   REFERENCES  
[1]   H.   Leman   and   C.   Marque,   "Rejection   of   the   maternal  
electrocardiogram  in  the  electrohysterogram  signal,"  IEEE  Trans  
Biomed  Eng,  vol.  47,  pp.  1010-­7,  Aug  2000.  
B.  Karlsson,  J.  Terrien,  V.  Guðmundsson,  T.  Steingrímsdóttir,  and  
C.   Marque,   "Abdominal   EHG   on   a   4   by   4   grid:   mapping   and  
presenting   the   propagation   of   uterine   contractions,"   in   11th  
Mediterranean  Conference  on  Medical  and  Biological  Engineering  
and  Computing,  Ljubljana,  Slovenia,  2007,  pp.  139-­143.  
M.   Hassan,   J.   Terrien,   B.   Karlsson,   and   C.   Marque,   "Spatial  
analysis   of   uterine   EMG   signals:   evidence   of   increased   in  
synchronization  with  term,"  in  Conf  Proc  IEEE  Eng  Med  Biol  Soc,  
2009,  pp.  6296-­9.  
M.  Hassan,  J.  Terrien,  B.  Karlsson,  and  C.  Marque,  "Interactions  
between   Uterine   EMG   at   Different   Sites   Investigated   Using  
Wavelet   Analysis:   Comparison   of   Pregnancy   and   Labor  
Contractions,"   EURASIP   Journal   on   Advances   in   Signal  
Processing,  vol.  2010,  p.  9,  2010.  
J.   Terrien,   T.   Steingrimsdottir,   B.   Karlsson,   and   C.   Marque,  
"Synchronization   between   EMG   at   Different   Uterine   Locations  
Investigated   Using   Time-­Frequency   Ridge   Reconstruction:  
Comparison   of   Pregnancy   and   Labor   Contractions,"   EURASIP  
Journal  on  Advances  in  Signal  Processing,  vol.  2010,  p.  10,  2010.  
P.  LeVan,  E.  Urrestarazu,  and  J.  Gotman,  "A  system  for  automatic  
artifact   removal   in   ictal   scalp   EEG   based   on   independent  
component  analysis  and   Bayesian  
Neurophysiol,  vol.  117,  pp.  912-­27,  Apr  2006.  
C.  J.  James  and  O.  J.  Gibson,  "Temporally  constrained  ICA:  an  
application   to   artifact   rejection   in   electromagnetic   brain   signal  
analysis,"   IEEE   Trans   Biomed   Eng,   vol.   50,   pp.   1108-­16,   Sep  
2003.  
W.  De  Clercq,  A.  Vergult,  B.  Vanrumste,  W.  Van  Paesschen,  and  
S.  Van  Huffel,  "Canonical  correlation  analysis  applied  to  remove  
muscle   artifacts   from   the   electroencephalogram,"   IEEE   Trans  
Biomed  Eng,  vol.  53,  pp.  2583-­7,  Dec  2006.  
M.  De  Vos,  A.  Vergult,  L.  De  Lathauwer,  W.  De  Clercq,  S.  Van  
Huffel,  P.  Dupont,  A.  Palmini,  and  W.  Van  Paesschen,  "Canonical  
decomposition  of  ictal  scalp  EEG  reliably  detects  the  seizure  onset  
zone,"  Neuroimage,  vol.  37,  pp.  844-­54,  Sep  1  2007.  
N.  E.  Huang,  Z.  Shen,  S.  R.  Long,  M.  C.  Wu,  H.  H.  Shin,  Q.  
Zheng,  N.  C.  Yen,  C.  C.  Tung,  and  H.  H.  Liu,  "The  Empirical  
Mode  Decomposition  and  the  Hilbert  spectrum  for  nonlinear  and  
non-­stationary  time  series  analysis,"  Proc.  Royal  Soc.  ,  vol.  454,  
pp.  903-­995,  1998.  
H.   Liang,   Z.   Lin,   and   R.   W.   McCallum,   "Artifact   reduction   in  
electrogastrogram   based   on   empirical   mode   decomposition  
method,"  Med  Biol  Eng  Comput,  vol.  38,  pp.  35-­41,  Jan  2000.  
H.  Liang,  Q.  H.  Lin,  and  J.  D.  Chen,  "Application  of  the  empirical  
mode   decomposition   to   the   analysis   of   esophageal   manometric  
data  in  gastroesophageal  reflux  disease,"  IEEE  Trans  Biomed  Eng,  
vol.  52,  pp.  1692-­701,  Oct  2005.  
M.   Blanco-­Velasco,   B.   Weng,   and   K.   E.   Barner,   "ECG   signal  
denoising  and  baseline  wander  correction  based  on  the  empirical  
mode  decomposition,"  Computers  in  Biology  and  Medicine,  vol.  
38,  pp.  1-­13,  2008.  
D.   Looney,   L.   Li,   T.   M.   Rutkowski,   D.   P.   Mandic,   and   A.  
Cichocki,   "Artifacts   Removal   from   EEG   Using   EMD,"   in  
International   Conference   on   Cognitive   Neurodynamics.   ICCN,  
2007.  
Z.  Wu  and  N.  Huang,  "Ensemble  Empirical  Mode  Decomposition:  
a  Noise-­Assisted  Data  Analysis  Method,"   Advances  in  Adaptive  
Data  Analysis,  vol.  1,  pp.  1-­41,  2009.  
B.  Mijovic,  M.  De  Vos,  I.  Gligorijevic,  J.  Taelman,  and  S.  Van  
Huffel,   "Source   separation   from   single-­channel   recordings   by  
combining   empirical-­mode   decomposition   and   independent  
component  analysis,"  IEEE  Trans  Biomed  Eng,  vol.  57,  pp.  2188-­
96,  Sep.  
O.  Friman,  M.  Borga,  P.  Lundberg,  and  H.  Knutsson,  "Exploratory  
fMRI  analysis  by  autocorrelation  maximization,"  Neuroimage,  vol.  
16,  pp.  454-­64,  Jun  2002.  
[2]  
[3]  
[4]  
[5]  
[6]  
classification,"  Clin  
[7]  
[8]  
[9]  
[10]  
[11]  
[12]  
[13]  
[14]  
[15]  
[16]  
[17]