1. 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

2. Outline Background Independent Component Analysis Experiments
Equipments & Procedures
Results – components, back projection maps
Summary & Future Work

3. 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

4. 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

5. 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

6. Previous works on ICA
Originally proposed by to solve blind source separation problem by Camon [1] in 1994
Gained more attraction and popularity from 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

7. 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

8. ICA definition Objective is to find a square matrix W
Key assumption: the source signals are statistically independent

9. ICA definition
Joint probability: the probability of two or more things happening together
Statistical independence: the joint probability density function (pdf) can be factorized to the product of individual probabilities of each source

10. 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

11. Gradient descent approach
Has been proven to effective in analyzing biomedical signals
Objective is to minimize the redundancy
Equivalent to maximizing the joint entropy of the cumulative density function (cdf)

12. Gradient descent approach
W can be updated using the following iterative equation (cdf) (entropy)
: learning rate

13. 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

14. Experiments equipments
BioSemi’s ActiveTwo Base system
Main components:
4x32 pin-type active electrodes
Collecting signals and remove common mode noise in real time
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

15. 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

16. 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

17. Experiments setup

18. Experiments setup

19. 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)

20. 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

21. 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

22. Recorded waves for subject 1 (Action I - standing)

23. Recorded waves for subject 1 (Action III - horse stance)

24. Recorded waves for subject 2 (Action I - standing)

25. Recorded waves for subject 2 (Action III - horse stance)

26. 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

27. Separated components for subject 1

28. Separated components for subject 2

29. 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

30. 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

31. Illustration of electrodes locations

32. Back
Subject 1 QRS component 1
Chest
Subject
right
left
Chest
Back

33. Back
Subject 1 QRS component 2
Chest

34. Back
Subject 1 QRS component 3
Chest

35. Back
Subject 1 QRS component 4
Chest

36. Back
Subject 1 QRS component 5
Chest

37. Back
Subject 1 QRS component 6
Chest

38. Subject 1 T-wave component
Back
Subject 1 T-wave component
Chest

39. Subject 2 P-wave map

40. Subject 2 QRS component 1

41. Back
QRS component 2
Chest

42. Subject 2 QRS component 3

43. Back
Subject 2 QRS component 4
Chest

44. Subject 2 T-wave component 1

45. Subject 2 T-wave component 2
Back
Subject 2 T-wave component 2
Chest

46. 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

47. 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

48. References
[1] P. Camon. Independent component analaysis, a new concept? Signal Processing, 36: , 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 signals. 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): , 1993
[6] J. Cardoso. High-order contrasts for independent component anlysis. Neural Computation, 11(1): , 1999
[7] A. Hyvarinen. Survey on independent component analysis. Neural Computation Survey, 2:94-128, 1999