[Eeglablist] ICA component
Tarik S Bel-Bahar
tarikbelbahar at gmail.com
Tue Oct 27 15:15:32 PDT 2015
Addendum below, cheers!
If the trials over your 2 minute recording periods are the same kind of
trials, then you can join them to get a representative ICA decomposition. I
think researchers tend to take all "similar cognitive periods or tasks" and
put them together for ICA. When and if you get good components out, you'll
then be able to re-apply to the continuous data, segment the different
trials, and then check out component dynamics between conditions.
You should find at least an eye-artifact IC in decompositions of brief
periods (like just 5 minutes total given to ICA), that have been cleaned
properly, but where eye blinks and eye artifacts occur multiple times in
the data and are retained in the data. If you're using ICA strictly for
cleaning out eye-artifacts, then your approach will likely yield one or two
known artifactual ICs that you can remove with some confidence. It's best
to run the analysis by cleaning in a traditional way and an ICA way, and
comparing them for yourself.
See also several articles over last 5 years summarizing various EEG
cleaning approaches, including specific foci on eye-artifact and
muscle-artifact detection. See also recent tools like the PREP and TAPEEG
pipeline toolboxes for eeglab and matlab, respectively.
At 20 channels you are at the lowest (or below) the limit of channels
usually used for ICA-EEG approaches. You may want to consider a traditional
non-ICA based approach. I think several past posts on eeglab list have
talked about the appropriateness of ICA for sparse EEG.
20 channels with a lot of time should work to give you interpetable
artifactual and "true/cortical/cognitive" ICs. you may even want to
downsample in time. Though we're pro-ICA, you should take a look at
microstate decompositions which I think have been done a lot more with <32
channel data sets, and "produce" scalp maps that are similar to some of the
main alphabet of ICs found with ICA.
You should also see what you get, and definitely play with tools like
SASICA or others. Note that there are at least ~5 to dozen ICs that show up
regularly in most datasets, across most studies that use ICA. see the
onton/makeig chapter, or other work, for those exemplars.
You may want search "cleaning approaches for sparse EEG" on google scholar.
Techniques from ~1950 to 2000 dealt primarily with sparse EEG, and there
are plenty of matlab implementations (e.g., Gratton's regression based
On Tue, Oct 27, 2015 at 5:18 PM, Dorian Grelli <dorian.grelli at gmail.com>
> Thank you again Stephan and thank you Tarik. Just for being sure..."124"
> is the seconds of my dataset? Maybe you meant 120 Stephan.
> My epochs are 4 seconds long and my datasets are parts of a single session
> where we perform 2' minutes recording every 10' minutes no-recording so, at
> the end, we got 6 datasets per each subject. Would be ok to join the 6
> datasets in a single file in order to get a more reliable ICA? In this
> case, after removing artifactual components, I will have to split again the
> data because we are interested in time course and we want to analyse each
> recording phase.
> I sent my first email because I really want to understand if it makes
> sense or not running something like SASICA for removing artifacts. In order
> to do that ICA has to be reliable.
> Subjects were forced to keep eyes closed during the recording phases and,
> maybe, I don't have too many artifacts. Maybe. I already reject bad epochs
> which passed a treshold or the amplitude.
> The problem is that I am self thought in this field (like many of you I
> suppose) so I am quite ok in some aspects and very bad in other things
> (like trails = epochs). However thank you for the suggestions Tarik!
> Il 27/Ott/2015 18:13, "Stephen Politzer-Ahles" <
> stephen.politzer-ahles at ling-phil.ox.ac.uk> ha scritto:
>> In that case it sounds like this is probably ok. Using the formula listed
>> above, you have (124*500)/(20^2) = 155 points per weight, which is much
>> more than the sample dataset. Also now that I think of it, 2 minutes of
>> epoched data is still a decent amount (that's e.g. 120 one-second epochs,
>> I've done ICA on comparable sizes of data before).
>> Stephen Politzer-Ahles
>> University of Oxford
>> Language and Brain Lab, Faculty of Linguistics, Phonetics & Philology
>> On Tue, Oct 27, 2015 at 4:56 PM, Dorian Grelli <dorian.grelli at gmail.com>
>>> Thank you Stephan for claryfing many points. My sampling rate is 500 hz.
>>> Il 27/Ott/2015 17:53, "Stephen Politzer-Ahles" <
>>> stephen.politzer-ahles at ling-phil.ox.ac.uk> ha scritto:
>>>> Hello Dorian,
>>>> Regarding your last question, you get as many independent components as
>>>> you have channels; you had 20 channels, which is why you got 20 components.
>>>> Examples with 256 components would have come from 256-channel caps.
>>>> As for just how much data is enough, other people on the list can
>>>> probably answer that better than me. 2 minutes does sounds very short. But
>>>> this also depends on your sampling rate (as mentioned in the paragraph you
>>>> quoted); 2 minutes of 1000 Hz data (i.e., sampled every millisecond) is a
>>>> lot more than 2 minutes of 250 Hz data (i.e., sampled once every four
>>>> Also, a trial is the same thing as an epoch.
>>>> Stephen Politzer-Ahles
>>>> University of Oxford
>>>> Language and Brain Lab, Faculty of Linguistics, Phonetics & Philology
>>>> On Tue, Oct 27, 2015 at 11:02 AM, Dorian Grelli <
>>>> dorian.grelli at gmail.com> wrote:
>>>>> Hi guys,
>>>>> I am very new with eeg data analysis and it would be great to have
>>>>> some support from you!
>>>>> I found the quotation below from this tutorial:
>>>>> *"Very important note: We usually run ICA using many more trials that
>>>>> the sample decomposition presented here. As a general rule, finding Nstable
>>>>> components (from N-channel data) typically requires more than kN^2 data
>>>>> sample points (at each channel), where N^2 is the number of weights in the
>>>>> unmixing matrix that ICA is trying to learn and k is a multiplier. In our
>>>>> experience, the value of k increases as the number of channels increases.
>>>>> In our example using 32 channels, we have 30800 data points, giving
>>>>> 30800/32^2 = 30 pts/weight points. However, to find 256 components, it
>>>>> appears that even 30 points per weight is not enough data. In general, it
>>>>> is important to give ICA as much data as possible for successful training.
>>>>> Can you use too much data? This would only occur when data from radically
>>>>> different EEG states, from different electrode placements, or containing
>>>>> non-stereotypic noise were concatenated, increasing the number of scalp
>>>>> maps associated with independent time courses and forcing ICA to mixture
>>>>> together dissimilar activations into the N output components. The bottom
>>>>> line is: ICA works best when given a large amount of basically similar and
>>>>> mostly clean data. When the number of channels (N) is large (>>32) then a
>>>>> very large amount of data may be required to find N components. When
>>>>> insufficient data are available, then using the 'pca' option to jader.m
>>>>> <http://sccn.ucsd.edu/eeglab/locatefile.php?file=jader.m>** to find
>>>>> fewer than N components may be the only good option.*"
>>>>> I don't know if each of my datasets has enough datapoints for
>>>>> performing an ICA. Each dataset has 20 channels, last 2 minutes and is 4
>>>>> seconds epoched, baseline corrected and pass band filtered. I also reject
>>>>> bad epochs.
>>>>> Which is the meaning of "trials" in the quotation above? Would be
>>>>> better to have longer registrations?
>>>>> When I run ICA I got 20 components. Why are there some examples with
>>>>> 256 components?
>>>>> 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...
More information about the eeglablist