[Eeglablist] Decimation prior to ICA and related things

[Baris Demiral]

Of course, if you have two conditions each 40 trials, 2 seconds long,
and recorded with 64 channel, 512Hz sampling rate, -and assuming you
concatenate the condition epochs- and run ICA on those epochs, in the
most optimistic case you hardly reach k=(80*512*2)/64*64=20.

This again comes to the point of applying ICA on epoched versus
continuous data. I assume (I bet) that for the cognitive ICs,
condition epoched data will be more accurate, at least yielding to
better cognitive/condition/event relevant ICs. As soon as you have
clean epochs, with enough length (good sampling rate) you may get good
ICs. I think increasing sampling rate would be a better thing to do
rather than increasing the epoch length but you can do both.

Baris

On Sun, Mar 18, 2012 at 8:55 AM, Scott Makeig <smakeig at gmail.com> wrote:
> The data_length/channels^2 = k > 30 heuristic was based on observing our
> 256-channel data decompositions. For fewer channels, I don't think this (k >
> 30) may be necessary (though a small k will very likely prove a problem!).
>
> But for small numbers of channels, data length is not so much an obstacle --
> for example, decomposing 32 channels with k=30 would require only 30*32^2=
> 32K data points: At a sampling rate of (say) 256 Hz, this would be ~2
> minutes of data. With 128 channels, the equivalent heuristic would be ~32
> minutes.
>
> I hope one of us will be able to do a proper study, decomposing data subsets
> of various lengths and measuring mutual information reduction and dipolarity
> (see Delorme et al., PLoS One 2012). Also, Jason Palmer has been working on
> theoretic lower bounds on ICA accuracy at given data lengths and channel
> numbers. I'll summarize result with him when possible.
>
> Scott Makeig
>
>
> On Sat, Mar 17, 2012 at 5:17 PM, Tarik S Bel-Bahar <tarikbelbahar at gmail.com>
> wrote:
>>
>> 1. not sure if you have to decimate, have seen some people to do this to
>> better meet requirements for good ICA decompositions.
>> 2. don't think higher sampling rate will give you better ICA. I think it's
>> more of a matter that ICA is fed data that gives an accurate and lengthy
>> representation of the whole "Data space".
>> 3. quote from earlier eeglablist post: "the key factor is how much data
>> you have  (timepoints / channels^2). If this is > 30 (or near to it), then
>> we find it preferable to return all possible components (since pca does a
>> rather poor job of separating sources)."
>> So if below this threshold there is some reason for adding more time to
>> the protocol, or reducing channels, or  decimation
>>
>>
>>
>> On Tue, Mar 13, 2012 at 6:00 AM, Modestino, Edward J *HS
>> <EJM9F at hscmail.mcc.virginia.edu> wrote:
>>>
>>> Dear EEGLAB experts,
>>>
>>>
>>>
>>> (1) Is it true that  ICA must be subject, like all the model-based
>>> spectral analysis methods, to a recommendation that one decimate to the
>>> lowest frequency capable of representing the actual signal content of the
>>> data without alienation effects?  Does one NEED to decimate the data before
>>> running ICA?  For instance, we have a data set recorded at 1,000 Hz.  Do we
>>> need to decimate this to approximately 128 or 256?
>>>
>>>
>>>
>>> (2 ) According to the formula that Dr. Makeig gave to determine the
>>> optimal amount of data,  #timepoints/(#channels)^2, it would appear that a
>>> higher sampling rate will give better ICA results.  Is this the case?
>>>
>>>
>>>
>>> (3) Finally, using this formula, #timepoints/(#channels)^2, is there a
>>> threshold or cutoff one needs to be exceeded to have the optimal amount of
>>> data to run ICA. Simply doing the equation without any way to interpret the
>>> output is not helpful.
>>>
>>>
>>>
>>> Thanks for your help,
>>>
>>> Dr. Modestino
>>>
>>>
>>>
>>> Edward Justin Modestino, Ph.D.
>>>
>>> Postdoctoral Research Associate
>>>
>>> Ray Westphal Neuroimaging Laboratory
>>>
>>> Division of Perceptual Studies
>>>
>>> Department of Psychiatry and Neurobehavioral Sciences
>>>
>>> University of Virginia
>>>
>>> Email: ejm9f at virginia.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
>>
>>
>>
>> _______________________________________________
>> 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
>
>
>
>
> --
> Scott Makeig, Research Scientist and Director, Swartz Center for
> Computational Neuroscience, Institute for Neural Computation; Prof. of
> Neurosciences (Adj.), University of California San Diego, La Jolla CA
> 92093-0559, http://sccn.ucsd.edu/~scott
>
> _______________________________________________
> 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



-- 
Ş. Barış Demiral, PhD.
Department of Psychiatry
Washington University
School of Medicine
660 S. Euclid Avenue
Box 8134
Saint Louis, MO 63110
Phone: +1 (314) 7477 1603