[Eeglablist] Data Processing Steps for ICA?

Fri Nov 21 07:29:40 PST 2014

Dear Paul,

> (even scarier than those on the wiki page)

Hmm interesting. Thank you for letting me know about it.
How ICA behaves to the rank-deficient data is more or less the same. The
ghosts showed scalp-map wise near-perfect and time-series wise less perfect
compensatory patterns. I've seen it in both infomax and AMICA and the
patterns were the same.

Makoto

On Fri, Nov 21, 2014 at 1:09 AM, Paul Yu-Chun Chang <Y.Chang at lipp.lmu.de>
wrote:

> Dear Makoto,
>
> Many thanks for your reply. At one time I did see ghost signals (even
> scarier than those on the wiki page) when I didn't remove a channel after
> re-referencing.  Indeed this is profound to me, but thank you again for
> your suggestions.
>
> Best regards,
>
> Paul
>
>
>
> 2014-11-16 23:21 GMT+01:00 Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
>
>> > If I just rereference after ICA, would ICA-matrix transformation also
>> take care of the data rank issue or should I still discard 1 channel?
>>
>> Sometimes it does, sometimes not. Therefore I would say it does not. When
>> we got the weird results shown in that wiki page I directed, even Scott did
>> not know what was going on. In order to avoid this kind of confusion I
>> recommend you reject a channel. You can alternatively use pca option to
>> reduce the rank, but it is less straightforward in my opinion. I believe we
>> should simplify the data as much as possible, so remove any potential
>> source of later confusion.
>>
>> Makoto
>>
>> On Wed, Nov 5, 2014 at 12:29 AM, Paul Yu-Chun Chang <Y.Chang at lipp.lmu.de>
>> wrote:
>>
>>> Dear Arno & Makoto,
>>>
>>> Many thanks for your reply and the suggested pages; they are indeed very
>>> helpful. One small question I still have, then, is that it is suggested in
>>> Makoto's page to discard (any) 1 channel from the data at the step of
>>> average-rereferencing (due to reduction of data rank by 1), before running
>>> ICA. If I just rereference after ICA, would ICA-matrix transformation also
>>> take care of the data rank issue or should I still discard 1 channel?
>>>
>>> Thank you very much,
>>>
>>> Paul
>>>
>>> 2014-11-05 9:03 GMT+01:00 Arnaud Delorme <arno at ucsd.edu>:
>>>
>>>> There is also this page just in case.
>>>>
>>>> http://sccn.ucsd.edu/wiki/Quick_Rejection_Tutorial
>>>>
>>>> Makoto, I have slightly edited your new page.
>>>>
>>>> > 2. When and how to perform re-referencing:
>>>> It seems I should re-reference to average before ICA, but there seems
>>>> to be some counterarguments, too:
>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008309.html
>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008854.html
>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2003/000090.html
>>>>
>>>> Just re-reference to the average reference. Don't forget to reject one
>>>> channel after this.
>>>>
>>>>
>>>> Yes, I would reference to average reference before running ICA as well
>>>> although this is not critical (you can always rereference after running ICA
>>>> and this will transform the ICA matrix as well). After running ICA for
>>>> years, it seems that decomposition using the average reference are slightly
>>>> better. However, there is no formal comparison I know of. In theory, since
>>>> it is all linear, it will not change anything for ICA (although because of
>>>> numerical implementation it might).
>>>>
>>>> > I'm not pretty sure whether to include EOG channels either, as there
>>>> seems to be pros and cons:
>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008854.html
>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2007/001801.html
>>>>
>>>> Do include EOG channels for ICA. I made a wrong answer last time.
>>>>
>>>>
>>>> I would include the EOG channels if you recorded them with the same
>>>> reference as ICA. Otherwise do not include them.
>>>>
>>>> Hope this helps,
>>>>
>>>> Arno
>>>>
>>>>
>>>>
>>>> Makoto
>>>>
>>>> On Sat, Oct 25, 2014 at 2:54 AM, Paul Yu-Chun Chang <
>>>> Y.Chang at lipp.lmu.de> wrote:
>>>>
>>>>> Dear All,
>>>>>
>>>>> I have 32-channel Neuroscan recordings (ca. 1 hr per subject) and am
>>>>> recently trying to use ICA to remove blinks and eye movements from the data
>>>>> (EEGLAB version 12.0.2.4b). After checking out some relevant
>>>>> discussions/wiki pages/tutorials I find myself however still a bit confused
>>>>> (particularly regarding re-referencing, baseline correction, and channels
>>>>> to be included when running ICA) and am having some further questions. I'd
>>>>> really appreciate your help.
>>>>>
>>>>> Some essential steps as I understand now involve (pls correct me if
>>>>> I'm wrong):
>>>>>
>>>>> 1. Reject bad epochs/channels to prune the data.
>>>>> (source:
>>>>> http://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA ;
>>>>> http://sccn.ucsd.edu/eeglab/workshop06/handout/Practicum_3_ICA_Process.pdf
>>>>> )
>>>>> 2. Train ICA firstly on 1 Hz high-pass filtered pruned dataset.
>>>>> 3. Apply ICA weights to the same subject's 0.1 Hz high-pass filtered
>>>>> data and evaluate components.
>>>>> (source: http://sccn.ucsd.edu/pipermail/eeglablist/2011/004424.html)
>>>>> 4. Interpolate bad channels.
>>>>>
>>>>> I'm however not so clear about:
>>>>>
>>>>> 1. When to perform base-line correction:
>>>>> It seems most people suggest not to do it before ICA training based on
>>>>> Groppe's 'split-half' paper:
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2010/003080.html
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008854.html
>>>>> But there are some counterarguments:
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2012/005513.html
>>>>> So I guess it's still better to do it after ICA?
>>>>>
>>>>> 2. When and how to perform re-referencing:
>>>>> It seems I should re-reference to average before ICA, but there seems
>>>>> to be some counterarguments, too:
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008309.html
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008854.html
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2003/000090.html
>>>>>
>>>>> I'm not pretty sure whether to include EOG channels either, as there
>>>>> seems to be pros and cons:
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2014/008854.html
>>>>> http://sccn.ucsd.edu/pipermail/eeglablist/2007/001801.html
>>>>>
>>>>> Also is it all right to use average of M1/M2 (mastoid sites) for
>>>>> re-referencing instead? Should M1/M2 be included for ICA training?
>>>>>
>>>>> Many thanks!
>>>>>
>>>>> Paul
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>>
>>>>> Paul Yu-Chun Chang
>>>>> Graduate School Language & Literature Munich - Class of Language
>>>>> Ludwig-Maximilians-Universität München
>>>>> Schellingstraße 10
>>>>> 80799 München, Deutschland
>>>>> Email: Y.Chang at lipp.lmu.de
>>>>>
>>>>> _______________________________________________
>>>>> 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
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Makoto Miyakoshi
>>>> Swartz Center for Computational Neuroscience
>>>> Institute for Neural Computation, University of California San Diego
>>>>  _______________________________________________
>>>> 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
>>>>
>>>>
>>>>
>>>
>>>
>>> --
>>>
>>> Paul Yu-Chun Chang
>>> Graduate School Language & Literature Munich - Class of Language
>>> Ludwig-Maximilians-Universität München
>>> Schellingstraße 10
>>> 80799 München, Deutschland
>>> Email: Y.Chang at lipp.lmu.de
>>>
>>
>>
>>
>> --
>> Makoto Miyakoshi
>> Swartz Center for Computational Neuroscience
>> Institute for Neural Computation, University of California San Diego
>>
>
>
>
> --
>
> Paul Yu-Chun Chang
> Graduate School Language & Literature Munich - Class of Language
> Ludwig-Maximilians-Universität München
> Schellingstraße 10
> 80799 München, Deutschland
> Email: Y.Chang at lipp.lmu.de
>

-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20141121/54b56332/attachment.html>