[Eeglablist] Artifact detection and ICA on data with multiple epoch types

[Arnaud Delorme]

The command line function eeg_epoch2continuous also allows to transform EEG epochs into a continuous dataset.
Best,

Arno

On 5 May 2013, at 16:06, Mikołaj Magnuski wrote:

> It seems I didn't send my last message to eeglablist, it can be found at the end of this one.
> 
> Anyway, Becky, I have found and uploaded the beforementioned function
> to https://sites.google.com/site/analizasygnalow/zasoby/nasze-narzedzia
> the function name is epoch_linearize.m and it is called this way:
> newEEG = epoch_linearize(EEG);
> 
> it does not retain events, but you don't need these for ICA.
> For linearizing multiple files you can use linearize_all_eeg
> (also on the page) like this:
> ALLEEG = linearize_all_eeg(files, folder);
> 'files' is assumed to be a cell matrix of file names
> (for example: {'subj01.set', 'subj02.set'}), while 'folder'
> a string defining filepath (like: 'C:\DATA\Exp1_2012\').
> 
> You would still have to save these files and later
> join relevant ones. I had a script for this too
> where you could specify with regular expressions
> what files should be linearized and then concatenated, 
> but I didn't transform it into a more universal function, 
> I'll try to do this soon if you would find it useful.
> 
> If you have any troubles using these functions -
> let me know!
> 
> 
> You can also transform all your subject subfiles (epochs of different length) into continuous signals (instead of epochs you can place 'boundary' events, although it is not necessary for ICA) - such that you have one file per subject. Then you would perform ICA and copy ica weights to your original epoched files. I think there is a function in EEGlab to 'de-epoch' the data (Arno, Makoto, am I correct?). Anyway - I wrote such function some time ago, so I'll see if I can still find it. The function did not retain events from the original dataset, but you do not need this for ICA. 
> (there is yet another option - you could epoch the data into small chunks of 1-s long or so, reject epochs containing artifacts or signal of no interest, then perform ICA and copy ica weights to your normally epoched data. I too have a function for this, which allows for several useful options like automatic rejection of windows (these small epochs) that are too far from cerain events or event sequences - so that you wouldn't have to manually click and reject windows ocurring in breaks or during a task you are not interested in. Unfortunatelly the function help is still in polish - but the basic options are easy to explain in one short mail - that is if you think you will find such function useful)
> 30 kwi 2013 19:24, "Becky Prince" <becky.prince at york.ac.uk> napisał(a):
> Dear EEGLAB users,
> 
> I have a question about the best artifact detection and ICA procedures for data sets with multiple epoch types of different lengths.  
> 
> Each trial in my EEG task is made up of 3 non-overlapping epochs of interest: 1) 4 seconds, 2) either 1.5 or 2.5 seconds, 3) 2.5 seconds.  For my previous ERP experiments I've used the following data processing procedure: epoch the data, reject bad epochs, run ICA on remaining epochs, reject ICA components, analyze data in channel space.  For my current experiment (with 3 epochs per trial), I haven't been able to run ICA on all of the epoched data at once because, as soon as I epoch the data, it is split into a separate data set.  I'm fairly sure that these 3 separate epoched data sets cannot be merged/appended for a single ICA run because the epochs are of differing lengths.
> 
> I've come up with 2 different protocols for processing the individual data sets.  I'm not sure which is better or if I'm overlooking other options.
> 
> Solution 1: artifact rejection and ICA on continuous data
> 1) reject artifacts in continuous data
> 2) run ICA on continuous data
> 3) reject ICA components
> 4) epoch the data (creates 3 separate data sets)
> 5) analyze data in channel space
> 
> Solution 2: artifact rejection and ICA on epoched data
> 1) epoch the data (creates 3 separate data sets)
> Then for each of the 3 epoched data sets: 
> 2) reject epochs
> 3) run ICA on remaining epochs
> 4) reject ICA components
> 5) analyze data in channel space   
> 
> My problem with solution 1 is that it is much more time consuming to detect the artifactual sections of continuous data than it is to detect and reject bad epochs.  Also, ICA seems to run more slowly in this case (even after taking into account the differences in the size of the continuous and epoched data sets).  My problem with solution 2 is that it doesn't seem right to run ICA separately for each epoched data set for the same subject/session.  Because I am doing within-subject comparisons among the 3 trial segments, I'm concerned about producing artificial differences among these epochs as a result of differences in the ICA decomposition and component rejection.
> 
> Sorry for the long question.  Can anyone tell me what they would recommend?  Any suggestions and/or criticisms of my current processing steps would be very helpful!
> 
> Many thanks,
> Becky
> PhD Candidate
> Department of Psychology
> University of York
> 
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20130506/67799879/attachment.html>