Application of ICA in Removing Artifacts from the ECG


By Tal Dahan and Astar Sade
Supervised by Eldad Klaiman



Abstract



Action potential propagation through the heart generates electrical currents thet can be detected at the surface of the body. A recrding of these electrical changes is called ECG
Analyse of ECG provides us with information in variety of domains:
1.The anatomic orientation of the heart.
2.The relative size of the atrials and the ventriculars.
3.Different heart disorders .
4.The size and the location of the blood vessels leading to the heart and the perturbation of those blood vessels.
5.The impact of different medicines of those blood vessels

The problem


In routinely recorded ECG's , many types of noise and artifact are present (e.g.muscle movements or electrode movements).Since any detail in the ECG's signal us very significant it is very important to remove these noises.Analog or digital filters are widely used to reduce the influence of interference superimposed on the ECG. However , many of these methods to filter out the noise and artifacts from ECG are only partially successful. On the one hand, the filters often lead to a reduction in the amplitudes of the component waves. On the other hand, some of the noise and artifacts are random in nature and have a wide rang of frequency. Hence the filters fail to remove the interference when itis whithin the same frequency range as the cardiac signal

The solution



Assuming that the ECG measured signal is a linear mixture of the actual cardiac signal and the noise of all sources,and that these components are statistically independent,we perform Independent Component Analysis (ICA). Under the assumptions above, we can describe the 12 lead ECG signal as a vector, X. The mixing model is then written as x=As, or s=Wx, where s represents the independent sources, and A (or inv(W)) represents the linear mixing of sources. We then find a matrix W that maximizes the nongaussianity, relying on the Central Limit Theorem, saying that that the distribution of a sum of independent random variables tends toward a gaussian distribution. The mixing matrix giving the most nongaussian s vector, is also the one separating x to its independent components.
We then tried this method on four individuals to separate the ECG's signal from the atrefact and found statistical criterions that characeristic only the ECG's signal. Here is the result of one of them:




The orginal signal:




The signal after cleaning with ICA



Acknowledgment



We are especially grateful to our project supervisor Eldad Klaiman.
We would also like to express our gratitude to the Ollendorf Minerva Center for supporting this project.


Related Documentation:

Project report
Final presentation
Matlab code
Code documentation