[Eeglablist] Enough data points for ICA

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Sun Jun 7 17:15:35 PDT 2015

Dear Yamil,

ICA would work well even with relatively less number of datapoints if data
are stationary (i.e. data from repeated trials and nothing else) with no
too large amplitude.

You mean you have 128ch but only 7-min long data? Hmm that is quite a
concern for me. You may try 'pca', 40 (or around) to see if this dimension
reduction improves the results... if it does, then you'd better use pca
option for the final analysis.

The formula of (ch^2)x30 was empirically derived, and there is no
quantitative experiment on it. So the issue is ambiguous.

I recently updated my wiki page about this issue. It is my personal
opinion. In short, "Someone please investigate this issue."



"...By the way, how many datapoints do we need to obtain good ICA results?
We have been suggesting that there needs to be (number of channel)^2 x 20
to 30 data points for the case of running ICA on 32 channels, and the
number ’20 to 30’ should increase as the number of channels increase.
However, we have not performed a systematic investigation on the minimum
number of data points required for ICA. I personally downsample the data
(with more than 128 channels) to 128Hz, particularly when I need to give up
gamma due to too strong 50/60Hz for cleanline, only for ICA just to obtain
the decomposition matrix. In this case, the number of datapoints is
absolutely lower than the number suggested by the formula, but I have not
encountered any problem so far. Some of my colleagues even told me that the
aggressive downsampling before ICA enhanced the decomposition quality. This
makes sense because most interesting EEG phenomena occur below 50 Hz
anyways. I can easily imagine it is nonsense to use 192kHz sampling rate to
record 1-sec long 64-ch EEG data and expect a good ICA results out of it...
so probably it is not the absolute number of data points that determines
the quality of decomposition. Similarly, it is probably true that ICA on
512Hz-sampled data is not necessarily better than ICA on the same data
downsampled to 256Hz simply because the former has twice as many
datapoints. Someone please investigate this issue."

On Thu, Jun 4, 2015 at 1:42 AM, Yamil Vidal Dos Santos <
hvidaldossantos at gmail.com> wrote:

> Hi all,
> I have seen in this list that the recommended minimum data points to run
> ICA is channel^2 x 30.
> If I have a 128 electrode montage, and I sample at 250Hz, this would mean
> a minimum of 32 minutes of recording. Is this correct? Would it make a
> difference if I use PCA with ICA?
> I'm asking this because in one of my experiments the entire recording
> lasts about 7 minutes. I wanted to use ICA to clean this data, because the
> experiment doesn't have a trial structure, but is instead a continues
> exposure to a speech stream. I have run ICA on this data before and the
> results looked decent.
> Any recommendations?
> Thanks,
> Yamil
> _______________________________________________
> 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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20150607/7babfc84/attachment.html>

More information about the eeglablist mailing list