Noninvasive Study of the Human Heart using Independent Component Analysis
Y. Zhu, T-L Chen, W. Zhang, T-P Jung, J-R Duann, S. Makeig and C-K Cheng
University of California, San Diego
Oct 18, 2006
Outline
Background Independent Component Analysis Experiments
Equipments & Procedures Results
– components, back projection maps Summary & Future Work
Background
Objective of heart simulation Diagnose heart diseases efficiently Help doctors easily locate the problem
Advantage of noninvasive measurement More cost effective Much simpler and faster to prepare,
setup and take measurements
12-lead ECG shortcomings Too few information to separate
different sources A heart disease may be caused by
multiple conditions E.g. myocardial infarction may
happen in multiple locations Need more channels to detect ECG
waveforms
Contributions Design noninvasive experiments to
collect heart signals from around 100 channels
Analyze the data using Independent Component Analysis (ICA)
Successfully identify different components of P-wave, QRS-complex and T-wave
Previous works on ICA Originally proposed by to solve blind source
separation problem by Camon [1] in 1994 Gained more attraction and popularity fro
m Bell and Sejnowski’s infomax principle [2]
Jung et al. applied ICA to ECG, EEG, MEG and fMRI [3][4]
Separate maternal and fetal heart beats and remove artifacts
ICA definition
N source signals s = {s1,s2,…,sN} linearly mixed: x = {x1,x2,…,xN} = As
If x is known, recover sources as u = Wx
u is only different from s in scaling and permutation
ICA definition
Objective is to find a square matrix W
Key assumption: the source signals are statistically independent
ICA definition Joint probability: the probability of two o
r more things happening together Statistical independence: the joint proba
bility density function (pdf) can be factorized to the product of individual probabilities of each source
( , ) ( ) ( )p x y p x p y
ICA algorithms Gradient descent by infomax principle
[2] Hyvarien’s FastICA [2] Cardoso’s 4th-order algorithms JADE
[5][6] Many others [7] They may produce difference solutions
and the significance is hard to measure
Gradient descent approach Has been proven to effective in analyzing bi
omedical signals Objective is to minimize the redundancy
Equivalent to maximizing the joint entropy of the cumulative density function (cdf)
duup
upupuI N
i ii1)(
)(log)()(
Gradient descent approach W can be updated using the following itera
tive equation
(cdf) (entropy)
: learning rate
WWW
ugHW T
))((
( ) ( )u
g u p x dx
1
1( ( )) ( ) log ( )
N
i ii
H g u g u g uN
Gradient descent approach W is first initialized to the identity
matrix and iteratively updated until the change is sufficiently small
Main Parameters when using the package: Learning rate: 10-4
Stopping threshold: 10-7
Maximum steps: 103
Experiments equipments BioSemi’s ActiveTwo Base system Main components:
4x32 pin-type active electrodes Collecting signals and remove common mode noise in real ti
me 128 electrode holders
Fix the electrodes Electrode gel
Conductor between electrodes and skin Adhesive pads
Fix the holders on skin 16x8 channel amplifier/converter modules LabView Software
ICA Package: EEGLAB
Experiments setup prodcures 1. Attach electrode holders to the
skin by adhesive pads, forming two identical matrices on the chest and back
2. Inject gel in the holders 3. Plug in electrodes
Experiments setup procedures (cont’d)
4. Place 3 electrodes on the left arm, right arm and left leg as the unipolar limb leads and place the electrodes CMS/DRL on the waist as the grounding electrodes
Connect electrodes to the AD-box
Experiments setup
Experiments setup
Experiment Phases
Actions Description
Action I Stand and breath normally
Action II Breath and hold breath for intervals of 10 seconds
Action III Hold horse stance for a certain period and record after that
Action IV Lean to forward, backward, left and right (4 poses)
Purposes for multiple phases
Create different conditions so that different waveforms can be generated
The distances between P-wave, QRS complex and T-wave vary in different circumstance
Enable ICA algorithm to separate them
Characteristics of recorded waves
The electrodes on the chest receive much stronger signals Heart is closer to the front
Waves in different activities have different characteristics Heart beat rates Shapes of QRS complexes and T-
waves
Recorded waves for subject 1 (Action I - standing)
1500
+-
Scale
0 1 2 3 4 5
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Recorded waves for subject 1 (Action III - horse stance)
816
+-
Scale
105 106 107 108 109 110
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Recorded waves for subject 2 (Action I - standing)
800
+-
Scale
0 1 2 3 4 5
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Recorded waves for subject 2 (Action III - horse stance)
770
+-
Scale
100 101 102 103 104 105
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Characteristics of ICA results QRS complex and T-wave can be clearly
separated for subject 1 P-wave, QRS complex and T-wave can
be clearly separated for subject 2 QRS complex is decomposed into
several components with different peak time Maybe a sequence of wave propagation
Multiple activities are essential to perform ICA successfully At least 3, more are better
Separated components for subject 1
8
+-
Scale
85 86 87 88 89
7
6
5
4
3
2
1
Separated components for subject 2
10
+-
Scale
5 6 7 8 9
7
6
5
4
3
2
1
Back projection W is obtained unmixing matrix,
is mixing matrix The i-th column of represents
the weight of each channel that contributes to the i-th decomposed component
According to physical location of each channel, we can plot potential maps for each component
1W
1W
Characteristics of back projection maps Weights are concentrated in the left
part of the front chest P-wave source occupies upper
portion Sources are moving downward from
QRS components to T-waves Estimate the dipoles according to
the maps – from the most negative to most positive locations
Illustration of electrodes locations
1
2
3
4
5
6
8
9
10
11
12
13
15
16
17
18
19
20
22
23
24
25
26
27
29
30
31
32
33
34
36
37
38
39
40
41
50
51
52
53
54
55
57
58
59
60
61
62
64
65
66
67
68
69
71
72
73
74
75
76
78
79
80
81
82
83
85
86
87
88
89
90
Chest Back
Subject right Subject left Subject left Subject right
up
down
7 14 21 28 35 42
43
44
45
46
47
48
49 56 63 70 77 84 91 98
92
93
94
95
96
97
Subject 1QRS component 1
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1QRS component 2
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1QRS component 3
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1QRS component 4
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1QRS component 5
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1QRS component 6
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 1T-wave component
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 2P-wave map
Subject 2QRS component 1
QRS component 2Chest
Back
Subject 2QRS component 3
Subject 2QRS component 4
Chest
Back
Subjectright
Subjectleft
Subjectleft
Subjectright
Chest Back
Subject 2T-wave component 1
Subject 2T-wave component 2
Chest
Back
Summary
Design experiments to collect stable heart signals from multiple channels for analysis
Apply ICA techniques to find out meaningful heart wave components
Plot back projection maps to discover the properties of each component
Future work Experiment on more subjects Calculate wave propagation speed
according to the QRS components; verify the consistency with physiological observations
Seek for better ICA algorithms with the consideration on heart wave characteristics
References [1] P. Camon. Independent component analaysis, a new concept? Si
gnal Processing, 36:287-314, 1994 [2] A. Hyvaerinen, J. Karhunen and E. Oja. Independent Component
Analysis. John Wiley & Sons, Inc. 2001 [3] T.P. Jung et al. Independent component analysis of biomedical s
ignals. In 2nd International Workshop on Independent Component Analysis and Signal Separation
[4] T.P. Jung et al. Imaging brain dynamics using independent component analysis. Proceeding of the IEEE, 89(7), 2001
[5] J. Cardoso and A. Soloumiac. Blind beamforming for non-gaussian signals. IEE proceedings, 140(46):362-370, 1993
[6] J. Cardoso. High-order contrasts for independent component anlysis. Neural Computation, 11(1):157-192, 1999
[7] A. Hyvarinen. Survey on independent component analysis. Neural Computation Survey, 2:94-128, 1999
