1

Noninvasive Study of the Human Heart usingIndependent Component Analysis

Yi Zhu, Tong Lee Chen, Wanping Zhang, Tzyy-Ping Jung, Jeng-Ren Duann,

and Chung-Kuan Cheng, Fellow, IEEE

Abstract—We develop a new approach to study the hu-man heart using independent component analysis. Electro-cardiogram (ECG) is an important tool for heart diseasediagnosis. However, the normal 12-lead ECG can only ob-tain limited heart signals and mostly rely on trained andexperienced medical doctors to perform the analysis. Wehave designed experiments so that high spatial of electronicsignals can be recorded from human subjects. Indepen-dent component analysis is applied to the recorded signalsto separate different components from the recorded waves.The various separated components are further analyzed byback-projecting their activities to the surface montage to ex-amine the properties of components. Experimental resultsshow that this is a promising approach and is able to beextended to perform more sophisticated heart simulations.

I. Introduction

A computer model that can simulate the heart is a com-plex topic that medical researchers have been working foron decades. Being able to model a heart provides the abil-ity to diagnose heart diseases efficiently, allowing doctors toeasily locate the problem or perhaps point out which partof heart wall might be failing. One ideal for the model isits ability to take measurements in a noninvasive manner.Not only is it more cost effective for patients, it is alsomuch simpler and faster to prepare, setup and take mea-surements than the invasive counterpart in going throughsurgery. Although electrocardiogram (ECG) can take mea-surements of the heart noninvasively to record the amountof electrical force at a given point on the body, a healthysubject could have abnormal heart rhythm while a knownheart diseased subject could have normal heart rhythm.Thus merely studying ECG readings is not enough in di-agnosis. To get a better understanding of how the heartworks would require the ability to separate out the sourcesof the heart and understand how each source contributesto a pulsating heart.

In this paper, we propose an approach to make use ofIndependent Component Analysis (ICA) techniques to an-alyze electronic heart signals, which are obtained by ourexperiments. ICA refers to a family of related algorithmsthat performs blind source separation when statistical in-dependence are taken into account[1]. Although multiple

Yi Zhu, Wanping Zhang and Chung-Kuan Cheng are with theDepartment of Computer Science and Engineering, University ofCalifornia, San Diego, La Jolla, CA, U.S.A. 92093–0404. Phone:+1 858 534 6184, e-mail: {y2zhu,w7zhang,kuan}@ucsd.edu

Tong Lee Chen is with Intuit, Inc.Tzyy-Ping Jung and Jeng-Ren Duann are with the Swartz Center

for Computational Neuroscience, Institute for Neural Computation,University of California, San Diego, 4150 Regents Park Row, Suite320, La Jolla, CA, U.S.A. 92037. Phone: + 858 458 1927, email:{jung,duann}@sccn.ucsd.edu

channel recordings were used by other researchers, suchas [2][3][4], to study human heart activities, there are notmany previous works trying to separate and study the dif-ferent independent components embedded in the recordedheart waves by high density electrodes. As we mentionedabove, it is of importance to identify different underlyingsources that generate heart waves, which is approached byICA algorithms in this work. The experimental resultsshow that different components of the P-wave, QRS com-plex and T-wave, can be clearly identified and their cor-responding back projections demonstrate different proper-ties, which is a clear sign to show that it is a more effectiveway to locate the underlying problems of human heartsthan normal ECG.

The rest of the paper is organized as follows: in next sec-tion, the ICA concept and algorithm will be presented; Sec-tion 3 will introduce how our experiments are conductedand their results; future work will be discussed in Section4.

II. Independent Component Analysis

Our methods are based on the theory of ICA, whichwas originally proposed to solve the blind source separa-tion problem [5]. At that time, Comon was hoping that adefined mathematical framework would allow a baseline infurther development of the ICA concept. The research ef-fort in ICA was considered small scale until the mid 1990swhere it gained more attraction and popularity from Belland Sejnowski’s infomax principle [1]. Jung et al. appliedthe ICA technique on many different types of biomedicalsignals including ECG, electroencephalogram(EEG), MEG(the magnetic counterpart to EEG), and functional mag-netic resonance imaging (fMRI) scans [6][7]. In the ECGstudy, Jung et al. took eight channels of measurementson the surface of a mother’s chest and abdomen. Thechannels were then processed using ICA to separate thematernal and fetal heart beats. According to their study,Jung et al. discovered that, although biomedical signalsprovide abundant information about the physiological pro-cess, they are often contaminated by distortions caused bysmall movements of the electrical contacts known as arti-facts. ICA, on the other hand, shows promise in separatingartifacts from source signals, and perhaps can further sep-arate components into more fine grained subcomponents.There are also other previous works [8][9][10][11] that triedto use ICA to remove artifacts in the ECG recording. Inthis work, instead, we design more elaborate experimentsto collect high density heart signals and use ICA to an-alyze underlying heart activities. We shall first introduce

2

the ICA concept and algorithm, which is the core techniquewe use.

N source signals s = {s1(t), . . . , sN (t)} can be linearlymixed by multiplying a mixing matrix A and produce Nmixture signals x = {x1(t), . . . , xN (t)} = As. When themixture signals x are known, we would like to recover aversion, u = Wx, of the original sources, s, identical savefor scaling and permutation, by finding a square matrix,W . The key assumption used in ICA to solve this problemis that the source signals are as statistically independentas possible during the time courses. Statistical indepen-dence means the joint probability density function (pdf)of the output sources can be factorized to the product ofthe marginal pdfs of each source:

p(u) =

N∏

i=1

pi(ui) (1)

There are quite a few ICA algorithms, such as Hy-varien’s FastICA and Cardoso’s fourth-order algorithmJADE [12][13]. They have similar performance when gen-erally compared. However, when applied to actual datasets, they may produce difference solutions and the sig-nificance are hard to measure. The review papers [14][15]compared various ICA algorithms and their interrelation-ships. Here we employ the gradient ascent algorithm im-plemented by Makeig et al. [16], which is based on theinfomax ICA algorithm and has been proven to be effec-tive in analyzing biomedical signals. In the following webriefly sketch the infomax ICA algorithm and the gradientascent approach.

The objective of the infomax ICA algorithm is to mini-mize redundancy between the outputs, which is the follow-ing mutual information

I(u) =

∫

p(u) logp(u)

∏

N

i=1pi(ui)

du (2)

This redundancy measure has value 0 when the pdf p(u)can be decomposed as in Equation (1), and is a difficultfunction to minimize directly. Therefore the infomax algo-rithm relates this function to the joint entropy, H(g(u)),where g(u) is the cumulative density function (cdf):

I(u) = −H(g(u)) + E

[

N∑

i=1

log|g′

i(ui)|

pi(ui)

]

(3)

If the absolute values of the slopes of the cdfs, |g′

i(ui)| are

the same as the independent component pdfs, pi(ui) thenmaximizing the joint entropy of the g(u) vector will beequivalent to minimizing the redundancy in the u vector.

Based on this idea, the unmixing matrix W can be up-dated iteratively using the following equation[16]:

∆W = ε∂H(g(u))

∂WW T W = ε(I + yuT )W (4)

where ε is the learning rate and vector y has the elementsyi = (∂/∂ui) log(∂gi(ui)/∂ui). The (W T W ) “natural gra-dient” term in the update equation avoids matrix inver-sions and speeds convergence by normalizing the variance

in all directions[16]. Thus, W is first initialized to theidentity matrix I and iteratively updated using the aboveequation until the change of the matrix is less than a cer-tain threshold.

III. Experiments

A. Equipments & Setup

Our experiments are based on BioSemi’s ActiveTwoBase system (http://www.biosemi.com/ ). A list of themain necessary equipments is as follows:

• 16× 8-channel amplifier/converter modules• 4× 32 pin-type electrodes• Packer Signa electrode gel• 128 electrode holders• CMS/DRL electrodes• LabView software• adhesive pads

The system consists of 128 pin-type electrodes that are de-signed for use with BioSemi’s headcap using the electrodeholders (At most 101 of them are used in our experiments).The electrode holders are attached to the skin of the chestand back of the subject by adhesive pads, each by each.The Ag/AgCl electrodes do not require skin preparation,but does require electrode gel to act as a conductor betweenthe skin and the electrodes themselves. The electrodes areconnected to a 256 channel AD-box that is powered bythe battery box. The AD-box is the multi-channel map-ping system and requires one amplifer/converter moduleper 8 channels in use by the AD-box. The AD-box thenfeeds data measurements through the USB2 receiver thatis connected to a computer. The computer can then savethe data via the LabView software for further analysis. Weset the sampling rate to be 256 points per second.

B. Procedures

The preliminary experiments are performed on two sub-jects. For each subject, the preparation procedure consistsof the following steps:

1. Attach electrode holders to the skin of the subject byadhesive pads. They are attached to the chest andback of the subject respectively, forming two identicalsize matrices. In our two experiments, we chose thesizes of 6× 6 and 7× 7 respectively.

2. Inject adequate amount of Parker Signa electrode gelin the electrode holders using a syringe and blunt nee-dle.

3. Plug in electrodes to the holders.4. Place three electrodes on the subject’s left arm, right

arm and left leg as the unipolar limb leads to formthe Wilson Central Terminal and place the electrodesCMS/DRL on the subject’s waist as the groundingelectrodes.

5. Connect electrodes to the AD-box and set up the restof BioSemi’s hardware.

When the subject is ready for measurement taking, heis in a relaxed standing position. Although a supine posi-tion is ideal in ECG recording, it is not desirable in thiscase because there is a substantial risk to damaging the

NONINVASIVE STUDY OF THE HUMAN HEART USING INDEPENDENT COMPONENT ANALYSIS 3

electrodes that are placed on the back of the subject. Thesubject is asked to do the following four kinds of activities:

1. Standing still and breathing normally for 90 seconds.It is used as a baseline for comparison to the rest ofthe activities.

2. Breathing and holding the breath for intervals of 10seconds alternatively for a period of 90 seconds start-ing with the breathing interval. It is used to tire outthe cardiac muscles so that the contractions of thevarious muscles will start to separate as shown in thewave forms.

3. Horse stance for a certain period for 60 seconds andrecord the waveforms after that. Its function is similarto the above but in a more extreme measure.

4. Leaning to different orientations. The first subject isasked to lean forward and left and the second subjectis asked to lean in four orientations (forward, back-ward, left and right). In each orientation 90-secondwaveforms will be recorded.

In the analysis, we only take the last 50 seconds recordingof each activity, because the early phase of activities maycontain more noise. We use multiple activities in orderto make the heart show different results under differentconditions. ICA can then pick up the subtle differences anddecompose the waveforms into true source components.

C. Results & Analysis

As we mentioned above, we performed the experimentson two subjects, with 72 electrodes (two 6 × 6 matriceson the chest and the back) and 98 electrodes (two 7× 7matrices on the chest and the back) respectively. The 72electrodes (also called channels) are numbered as shownin Figure 1 — note that the left and right orientation inthe chest and back are reverse since they can form a circlearound the body; in the 7×7 case they are arranged in thesimilar fashion. Figure 2 are the photos taken for the sec-ond subject. For the first subject, we recorded 5 activities:still standing, holding breath, horse stance, leaning forwardand left; for the second subject, we recorded 7 activities,with the additional leaning backward and right activities.The recorded raw data is first processed by two-way least-square FIR (Finite Impulse Response) filtering, with thelow-edge frequency in pass band 0.1 Hz and the high-edgefrequency in pass band 40 Hz. Figure 3 and Figure 4 showthe partial recorded raw mixture signals respectively. Theleft portions of the two figures show the wave forms of thestill standing phase, and the right portions show the waveforms after horse stance. It is observed that the electrodeson the chest receive much stronger signals than the ones onthe back. It is reasonable since the human heart is closer tothe front of the body. Regarding the waveforms of the twosubjects, they have different characteristics. For instance,the first subject seems to have stronger T-waves and thesecond one contains P-waves. For each subject, comparethe waveforms recorded in different activities, we find thatthey appear to have different characteristics as well, suchas the rate of the beat and also the shape of the QRS andT-waves.

The ICA algorithm is applied to the above two wave-forms. By the definition, W is a square matrix, thereforethere are 72 and 98 components are produced respectively.However, only a few of them has relatively large ampli-tude and have meaningful waveforms. We believe the restof them are generated by breath and noise thus filter themout. Figure 5 and Figure 6 show the meaningful compo-nents produced by the ICA algorithm for the two subjectsrespectively.

From the results, we have the following observations:

• The ICA algorithm is able to separate P-wave, QRScomplex and T-wave. In two experiments, T-wave canbe clearly separated. Also, P-wave is successfully sep-arated in the second subject.

• Multiple activities are essential to perform ICA suc-cessfully. We have tried different combinations of thewaveforms by each activities. The experimental re-sults show that at least three activities are needed todecompose the T-waves. And more activities guaran-tee better separation. This is because different sourceswould have different behaviors under various circum-stances; for example, the shape of the T-waves and itsdistance to the QRS complex are different in differentactivities, as shown in Figure 3 and Figure 4.

• The ICA algorithm can detect around 7-8 differentcomponents from the original waveforms. Especially,there are multiple components which belong to theQRS components. This phenomenon indicates thatthe QRS complex wave may be generated by variousunderlying sources. For subject 2, even the T-waveis decomposed to two components, which shows thateven T-wave is probably not a simple activity.

• Note that the components that form the QRS com-plex have different peak time (they have been sortedin the ascending order of peak time). They could rep-resent a sequence of wave propagation. Thus, it isuseful to further analyze these components and de-tect the physical positions of the sources that producethese components. We will present the analysis in thefollowing.

To further analyze the components, we use the back pro-jection technique to plot the location maps for each com-ponent. From the ICA algorithm, we obtain the unmixingmatrix, W , then the mixing matrix will be W−1. Thus wehave x = W−1u. Let W−1(:, i) denote the ith column ofW−1, and u(i, :) denote the ith row of u, then back projec-tion for component i, pi can be computed as

pi = W−1(:, i) × u(i, :) (5)

In fact, the column vector W−1(:, i) represents the weightof each channel that contributes to the ith decomposedcomponent. Since we have known the physical locationof each channel, it’s possible to plot the “potential maps”for each component according to the corresponding vector.The maps for subject 1 and subject 2 are shown in Figure 7and Figure 8 respectively. The blue color denotes lowerweights and the red color denotes higher weights.

4

In most of the maps, the weights are concentrated in theright part of the front chest. This is understandable sinceit is the closest to the human heart. Furthermore, the mostconcentrated part, which is highly possible to be the sourceplaces that produce these components, is different for dif-ferent components. Generally, the P-wave (the first mapof the second subject) source occupies the upper portion ofthe maps. The positions of different QRS components aremoving downward with the peak waves moving in the timecourse. And the T-wave sources are located in the lowerportion of the maps. This is consistent with the definitionsof these waves: P-wave is initiated by the SA node, QRScomplex represents the depolization of the ventricle, andT-wave indicates the repolization of the ventricle. Sincethe depolization needs to spread from the AV node to allparts of the ventricles, the sources are moving downward asreflected in the maps of different components of the QRScomplex. Further computation could be done to computethe propagation speed of the waves and therefore help doc-tors decide whether the heart is healthy.

In addition, we also try to estimate the direction ofdipoles according to the maps, by assuming that the vec-tors are pointing from the most negative location to themost positive location. We illustrate four of the vectors inthe maps (a), (b), (d) and (f) of subject 2, which representsthe components of the P-wave, early portion of the QRScomplex, late portion of the QRS complex, and early por-tion of the T-wave respectively. The vector directions aredrawn by white arrow lines, as shown in Figure 8. Theyhave very different directions: the vector direction of theP-wave (Component 1) is pointing from the bottom rightof the chest to the top left of the chest; that of the earlyQRS component (Component 2) is almost parallel to theground, from the back to the chest (remember the map isan extended version); that of the late component (Compo-nent 4) is pointing from the bottom left to the center of thechest; and that of the T-wave (Component 6) is pointingfrom the top right of the back to the center of the chest.However, this estimation is very rough and source local-ization techniques should be used to calculate the origins,direction and magnitude of these vectors later.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

Chest Back

Subject right Subject left Subject left Subject right

up

down

Fig. 1. Layout of Electrodes

IV. Discussion & Future Work

The successful ICA application to electroencephalo-graphic (EEG) signals inspired us to design experiments

to collect heart signals from multiple channels and analyzethem by ICA. However, it is not easy to keep the tight con-tact between skins and electrode holders, especially whenthe subjects are being asked to do various activities. Thecurrent experiment setup is able to achieve this goal but ittakes time to attach the holders. We are still seeking for abetter way to quickly set up the experiments with reliableattachments of holders.

It is an ongoing research project and only our prelim-inary experimental results are presented here. Currentlywe are carrying on this project in the following several di-rections:

• We are taking more subjects to perform the experi-ments in order to discover the stability of the ICA al-gorithm. Also, we are going to use Computed Tomog-raphy (CT) or Magnetic Resonance Imaging (MRI) todetect the physical location of each subject’s heart sothat the sources of components can be more accuratelylocated.

• We are calculating the wave propagation speed ac-cording to the QRS components and trying to verifywhether it is consistent with the physiological obser-vations. If this could be done, it would be helpful fordoctors to detect heart diseases.

• We are also seeking for a better ICA algorithm thatwill take the features of ECG waveforms into account.The current assumption is that all the components aretotally independent. However, the sources are locatedin one human heart and therefore they are somehowrelated during the time course. We are targeting fordesign a new separation algorithm that can take thisproperty into consideration and obtain more accuratecomponent waveforms to perform the analysis.

References

[1] A. Hyvarinen, J. Karhunen, and E. Oja, Independent Compo-nent Analysis, John Wiley & Sons, Inc., 2001.

[2] C. Ramanathan, R.N. Ghanem, P. Jia, K. Ryu, and Y. Ruby,“Noninvasive electrocardiographic imaging for cardiac electro-physiology and arrhythmia,” Nature Publishing Group of Na-ture Medicine Advance Online Publication, 2004.

[3] G. Li and B. He, “Localization of the site of origin of cardiac ac-tivation by means of a heart-model-based electrocardiographicimaging approach,” IEEE Transactions on Biomedical Engi-neering, vol. 48, no. 9, pp. 660–669, June 2001.

[4] G. Li, X. Zhang, J. Lian, and B. He, “Noninvasive localizationof the site of origin of paced cardiac activation in human bymeans of a 3-d heart model,” IEEE Transactions on BiomedicalEngineering, vol. 50, no. 9, pp. 1117–1120, September 2003.

[5] P. Comon, “Independent component analysis, a new concept?,”Signal Processing, vol. 36, pp. 287–314, 1994.

[6] T.P. Jung, S. Makeig, T.W. Lee, M.J. Mckeown, G. Brown, A.J.Bell, and T.J. Sejnowski, “Independent component analysis ofbiomedical signals,” in 2nd International Workshop on Inde-pendent Component Analysis and Signal Separation, 2000.

[7] T.P. Jung, S. Makeig, M. Mackeown, A. Bell, T. Lee, and T. Se-jnowski, “Imaging brain dynamics using independent compo-nent analysis,” Proceedings of The IEEE, vol. 89, no. 7, 2001.

[8] J. O. Wisbeck, A. K. Barros, and R. Ojeda, “Application ofica in the separation of artefacts in ecg signals,” in Proceedingsof International Conference on Neural Information Processing,1998.

[9] A. K. Barros, A. Mansour, and N. Ohnishi, “Removing artefactsfrom electrocardiographic signals using independent componentanalysis,” Neurocomputing, vol. 22, pp. 173–186, 1998.

NONINVASIVE STUDY OF THE HUMAN HEART USING INDEPENDENT COMPONENT ANALYSIS 5

(a) Chest Side

(b) Back Side

Fig. 2. Photos of Electrodes

(a) Still Standing (b) Horse Stance

Fig. 3. Subject 1 Mixture Waveforms

(a) Still Standing (b) Horse Stance

Fig. 4. Subject 2 Mixture Waveforms

[10] S. Tong, A. Bezerianos, and J. Paul et al., “Removal of ecg inter-ference from the eeg recordings in smal animals using indepen-dent component analysis,” Journal of Neuroscience Methods,vol. 108, pp. 11–17, 2001.

[11] T. He, G. Clifford, and L. Tarassenko, “Application of ica inremoving artefacts from the ecg,” Neuro Computing and Appli-cations, vol. 12, 2002.

[12] J. Cardoso and A. Soloumiac, “Blind beamforming for non-gaussian signals,” IEE Proceedings, vol. 140, no. 46, pp. 362–370, 1993.

[13] J. Cardoso, “High-order contrasts for independent componentanalysis,” Neural Computation, vol. 11, no. 1, pp. 157–192,1999.

[14] T.W. Lee, M. Girolami, A.J. Bell, and T.J. Sejnowski, “A unify-ing information-theoretic framework for independent componentanalysis,” Computers and Mathematics with Applications, vol.39, pp. 1–21, 2000.

[15] A. Hyvaerinen, “Survey on independent component analysis,”Neural Computation Survey, vol. 2, pp. 94–128, 1999.

[16] S. Makeig, T.P. Jung, A.J. Bell, D. Ghahremani, and T.J. Se-jnowski, “Blind separation of auditory event-related brain re-sponses into independent component analysis,” Proceeding ofNational Academy of Science USA, vol. 94, pp. 10979–10984,1997.

6

Fig. 5. Source Components for Subject 1

Fig. 6. Source Components for Subject 2

NONINVASIVE STUDY OF THE HUMAN HEART USING INDEPENDENT COMPONENT ANALYSIS 7

(a) Component 1 Map (b) Component 2 Map

(c) Component 3 Map (d) Component 4 Map

(e) Component 5 Map (f) Component 6 Map

(g) Component 7 Map

Fig. 7. Subject 1 Back Projection Maps

8

(a) Component 1 Map (b) Component 2 Map

(c) Component 3 Map (d) Component 4 Map

(e) Component 5 Map (f) Component 6 Map

(g) Component 7 Map

Fig. 8. Subject 2 Back Projection Maps