[Eeglablist] Channels for ICA

Wed Aug 1 11:02:29 PDT 2018

Dear Fabio,

I will disappear from the list until November (others in the lab will
answer instead), so this is my last response to you on this issue.

> Can I run ICA on the 10-minute continuous recording (even if different
"states" happen every 2 minutes) so that I have enough data points
(sqrt(500 Hz x 600 sec / 30) = 100), and *then *do my chunking?

Technically yes.

> Or is this approach invalid because of the variance in the tasks that are
asked every 2 min?

VERY good point! Such variance difference across task (blocks) is called
non-stationarity. ICA's assumption is data be stationary. So theoretically
speaking you are violating (better to say 'undermining') the assumption of
stationarity. Addressing data non-stationarity is one of the most difficult
problem in data processing. As far as I know, there is no established ICA
preprorcessing pipeline there, although the fully-functional code of
multi-model adaptive-mixture ICA has been there for ten years! We just
don't know how to integrate the multiple models in a useful way.

In most cases, however, using single-model (i.e., standard) ICA across
different tasks is fine. My explanation for this is that in EEG occipital
alpha is always loud (show high amplitude), so is central mu, etc... these
loud member's behaviors are more or less the same across many types of
cognitive experiments because they respond to very general cognitive
function such as attention and moving/not moving etc. You also want to
remember that ICA is biased to high-amplitude activities.

In practice, if I were you, yes I would concatenate all the blocks to run a
single ICA, then make comparison on both across the block-separated
conditions and within-block conditions.

Makoto

On Tue, Jul 31, 2018 at 9:51 PM Giatsidis, Fabio <fabio_giatsidis at brown.edu>
wrote:

> Thank you Tarik and Makoto for your precious suggestions!
>
> Actually, I should have clarified that the 2-minutes recordings are
> segments from a single 10-minute EEG recording, and the patient was asked
> to do different things every 2 minutes (keep eyes open, keep eyes closed,
> etc.). I thought that chunking the continuous recording in small,
> task-specific sub-recording before start cleaning each sub-recording would
> have made better sense. But given your answers, a new question comes to my
> mind:
> - let's assume I first cut down arbitrarily the number of channels to
> ~100 or less as per Tarik's suggestion (i.e. excluding those around the
> ears, etc. that are of little help anyway). Can I run ICA on the 10-minute
> continuous recording (even if different "states" happen every 2 minutes) so
> that I have enough data points (sqrt(500 Hz x 600 sec / 30) = 100), and *then
> *do my chunking? Or is this approach invalid because of the variance in
> the tasks that are asked every 2 min?
>
> Sorry for the confusion, I'm just trying to make the most sense out of
> what I have and to not make missteps! Thank you immensely for all your help!
> -Fabio
>
> --------------------------------
> *Fabio Giatsidis, M.D.*
> Resident in Neurology - University of Rome "Tor Vergata" - Rome, Italy
> Post-doctoral research fellow - Brown University - Providence, RI, USA
>
>
> On Wed, Jul 25, 2018 at 3:18 PM, Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> wrote:
>
>> Dear Fabio,
>>
>> > I have a (probably naive) question regarding ICA.
>>
>> I'll give you my naive answers, not intentionally but due to my
>> limitations. Follow them at your own risk.
>>
>> > - whether I should or not choose the whole 128 channel set for running
>> ICA
>>
>> Yes, unless your data is too short for that. Remember, (number of
>> channels)^2 x 30 data points at 256 Hz sampling rate is a rule of thumb for
>> running ICA.
>>
>> > - consequently, how I should *a priori* decide which channels to
>> consider for ICA and which not.
>>
>> You can determine a priori which anatomical regions you are going to
>> analyze. Then, if ICA gives you the '(stationary) effective source
>> locations' that overlap /are close enough to those pre-selected regions,
>> pick them up for the final analysis.
>>
>> ICA is a hypothesis-free approach, but that does not mean you cannot have
>> a hypothesis.
>> You might enjoy reading the classic discussion between ICA pioneers and
>> Karl Friston about how ICA could be used in neuroscience data analysis.
>> Friston KJ. Modes or models: a critique on independent component
>> analysis for fMRI. Trends Cogn Sci.  1998. Oct 01; 2(10) 373-375
>>
>> > The recordings are ~2 minutes long and the sampling rate is 1000 Hz.
>>
>> If you have only 2 min, you definitely cannot perform >100ch ICA.
>> sqrt(250 Hz x 120 sec / 30) is about 31, so you want to use 'pca' option
>> to perform dimension reduction to obtain 31 ICs.
>>
>> > I would like to keep as many channels as possible during
>> pre-processing, and afterwards discard the ones I realize are not useful to
>> my analysis - if this approach seems reasonable.
>>
>> Record longer. 2-min EEG is too short if you want to use ICA.
>>
>> Makoto
>>
>> On Thu, Jul 19, 2018 at 3:26 AM Giatsidis, Fabio <
>> fabio_giatsidis at brown.edu> wrote:
>>
>>> Hello EEGLAB list,
>>>
>>> I have a (probably naive) question regarding ICA.
>>> I have been using an EGI 128-channel system to record resting states. I
>>> have been reading a bit about ICA, but it is still not clear to me:
>>> - whether I should or not choose the whole 128 channel set for running
>>> ICA, and
>>> - consequently, how I should *a priori* decide which channels to
>>> consider for ICA and which not.
>>>
>>> The recordings are ~2 minutes long and the sampling rate is 1000 Hz. I
>>> would like to keep as many channels as possible during pre-processing, and
>>> afterwards discard the ones I realize are not useful to my analysis - if
>>> this approach seems reasonable.
>>>
>>> Also:
>>> - if as a very first step I delete some clearly bad channels and then
>>> interpolate them to repopulate the original channel set, is it legit to
>>> include such interpolated channels during ICA?
>>>
>>> Thank you very much!
>>> Best,
>>> -Fabio
>>>
>>> --------------------------------
>>> *Fabio Giatsidis, M.D.*
>>> Resident in Neurology - University of Rome "Tor Vergata" - Rome, Italy
>>> Post-doctoral research fellow - Brown University - Providence, RI, USA
>>> _______________________________________________
>>> 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
>>
>
>

-- 
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/20180801/588b3dfb/attachment.html>