Hi Sepehr,

This depends on the aim behind your study. However, for each subject, people sometimes tend to preprocess the data first, and epoch it after that, then reject epochs (for cleaning). Finally, run ICA on the resulting data. 

HTH.

Samer

Thank you very much for your helpful responses. I really need to keep data as much as possible and correct artifactual epochs instead of rejecting them. After correcting artifacts I want to calculate ERPs. Because I have 40 trials for averaging for each subject, if I reject artifactual epochs, the number of trials decrease significantly and the SNR of ERPs reduces. So I need the best possible answer I can get from ICA  in artifact correction. 

ICA algorithms typically take sample*channel matrices. So epoching itself will not have any effect on the results of the algorithm because the data is re-concatenated in most packages/programs (e.g. EEGLAB)  before submission to the algorithm. 

Removing data samples in the form of entire epochs will have an effect on the results of ICA algorithms. Especially if those epochs contain large, unique movement artifacts. ICA algorithms will typically place unique movement artifacts in single components.

It would probably be best to submit all epochs to the algorithm unless there are some egregious movement artifacts in the recording. No EEG will be able to be recovered during periods of large movement artifact.

Was this helpful? I've attached a book chapter for reference. Section 6 contains the relevant discussion. 

Chapter Electroencephalography (EEG): neurophysics, experimental met...

Thank you very much Michael. It was a very helpful answer because my EEG data contains very large movement artifacts in some segments. I was wondering if ICA can recover EEG from these kind of artifacts, which based on your answer it is better to reject them before applying ICA on whole data. In addition, thank you very much for sending your interesting book chapter. 

I agree with Epoched data is more practical for artifacts detection. You can try with a 20 s length window if you dont want to segment your data in small windows. However you can trust what jason says in this link https://sccn.ucsd.edu/pipermail/eeglablist/2011/004048.html

Thank you very much Ricardo.

In my opinion, the best choice is to apply ICA to the 5 min EEG recording. In this way, you will remove the same component (or components) from all the epochs. Indeed, I use to do both steps independently:

1. Apply ICA to the raw data and remove the components (in EEG, tipically cardiac and oculographic ones).

2. Perform a visual inspection to the raw data to detect muscle activity. As muscle activity usually affect to all EEG channels, so ICA cannot isolote that artifact in one component.

I also would go for epoch by epoch application of ICA instead of whole dataset. However in such case processing time would be a great concern. Please note that ICA can only handle global artifacts (appearing on multiple channels in the same time window). Another concern would be the length of the EEG epoch. If any artifact falls on the consecutive epochs (overlapping in the nearby time windows) then it might not give the best result.

Hope it helps.

Thanks for raising this query, the responses it's getting are very insights and useful. Thus I am also following. 

Thanks a lot guys, 

All of answers was so useful and I realized I was a little wrong about the way I should use ICA and the mentioned points have pivotal role in my simulations. 

since different parts of continuous eegdata might have artifacts of different sources, running ICA to each epoch seems sensible but this epochs shouldn't be of a very short length or ICA couldn't do its work so well. 

with a 32 channel-eegdata, a segment of at least 256000 datapoints is necessary for running ICA. 

but don't forget to drop bad channels before running ICA and interpolate them after that.

good luck.

Dear Zhaleh, 

Thanks for your answer. However, I didn't understand how you calculated the number of datapoints required for a 32 channel recording. Why 256000? What is the role of sampling frequency here? 

I want to epoch the continuous EEG data of 8 minutes long into non overlapping epoch window in EEGLAB without any event information .When I select the 'Extracting Epoch " the GUI will pop up to set the Event information first .After epoch of the data I want to run ICA on the data to get the clean data .Kindly help me in this regard

I have an idea for this problem but I don't know whether it works properly. You can create a text file of events manually (I did this in an ERP experiment (attached here). But you may write the time of events in accordance to your epoching strategy). Then, you can add this file to your data in EEGLAB, from File > Import event info > From Matlab array or ASCII file. Try it and let me know if it works properly.

Muhammad, you can try eeg_regepochs() function. It adds dummy events.

The developers of EEGLab generally advise running infomax on larger data sets (i.e continuous data sets).. with a sufficient number of trials, epoching the data before can be done. you would probably be best off running ICA on both the epoched vs.continuous data sets and making a simple comparison of the quality of the decompositions (which generally tend to show better performance for continuous data sets with 64+ electrodes) see: https://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA