Dear Farrikh, 

I am not so sure about that. Please consider infact that:

1. In general, repeating twice the same procedure is a sign of inefficiency;

2. Ideal filter or procedures does not exist even in the numerical world: every computation you do introduces plenty of noise terms related, first of all, to the non-linear behavior of the phase in the filter response. 

Anyhow I suggest you to verify how much is this non-linear distortion and evaluate if it is negligible or not. Have a try is always a good idea. 

Hope it helps,

Ciao! 

387

well for the ICA part , i can suggest it will be same , just because it project your data to components which are independent  of each other with no correlation as well , so if you follow that process again you will be getting the same answer as there is no random or fuzzy variable involved.

thank you mr @Mohammad Sarfraz

Dear Farrikh, 

I am not so sure about that. Please consider infact that:

1. In general, repeating twice the same procedure is a sign of inefficiency;

2. Ideal filter or procedures does not exist even in the numerical world: every computation you do introduces plenty of noise terms related, first of all, to the non-linear behavior of the phase in the filter response. 

Anyhow I suggest you to verify how much is this non-linear distortion and evaluate if it is negligible or not. Have a try is always a good idea. 

Hope it helps,

Ciao! 

thank you mr Annese

I suggest you to quantify the results of your artifact removal by yourself that will give you the answer you wanted. Now question arises: How to quantitatively evaluate? 

Well, you can first of all extract the visually obvious artifact regions and then take the envelope of that data segment to get artifact templates. Also select a visually artifact-free clean EEG segment (around 10 sec at least). Add the artifact template with the clean data and now perform/apply your artifact removal algorithm for two cases: 1) one-time denoising  2) two-time denoising

Each time you calculate the signal distortion in RMSE (you have clean signal to compare with the artifactual and your resultant reconstructed signal both, so you can quantify both) and also the improve in SNR for both cases. Then compare the SNR and RMSE for signals before and after artifact removal. Now you know what is the answer!

FYI, to data synthesis you can take a look at one of my articles here

https://www.researchgate.net/publication/270757310_A_Wavelet-Based_Artifact_Reduction_from_Scalp_EEG_for_Epileptic_Seizure_Detection?ev=prf_pub

Article A Wavelet-Based Artifact Reduction From Scalp EEG for Epilep...

thank you mr islam

First, the definition of clean data is ambiguous because it depends on the application. In theory, you could have clean data free of noise, free of artifacts or both. A typical procedure in EEG data is to apply some filtering technique to a given frequency to improve the SNR and then to apply a denoising technique to remove the artifacts without affecting as much as possible the quality of the recorded data.

The first suggestion is to look into the signals. I think this is the most reliable way to check your signals.

As you mention, you don´t want to go through your data and you don´t know if your data is already cleaned, you could try ICA and look at the components. However, it could happen at least two things.

1) If ICA was not previously applied you will observe clearly components related to typical artifacts, e.g.: eye artifacts, cardiact artifact.

2) If ICA was already applied and some ICA component were rejected. It means that there was a dimension reduction of the space and then a back-projection of the components. This is done by selecting a given component into the mixing matrix [A], then replacing the choosen  column(s) by zeros and then applying the back-projection into the original data (Check the ICA book of Hyvärinen, Karrhunen, and Oja). If you apply ICA in this situation, you will get a missinterpretation of the data, why? because some ICA algortihms will not reach a correct solution since the number of dimensions was previously decreased to an unknown number.

Try the easiest way, ask to the people that provided you the data, if there was a preprocessing step and/or the acquisition software. Moreover, typically EEG records are accompanied by a protocol and a header file with relevant information (date of the recording, system, number of channels, filters, artifact rejection, patient name, etc.). By the way, to get trusted results for your research, do not try to guess about your data.

thank you for clear explanation Mr fernandez

I forgot to mention that using FastICA algorithm, there is an option for computing independent component in a deflatonary mode. This means that ICs are computed one by one and the algorithm will stop automatically according to a some given threshold. However, this method add some error. Moreover, there is no guarranty that you will get the same number of ICs as a number of channels. Just for checking this can be a quick solution.

The more reliable way is to use FastICA in a parallel computation mode. That is, supposing the number of ICs equal to the number of channels. As I previously mention.