[Eeglablist] ICA analysis

[Makoto Miyakoshi]

Dear Robert,

I'd like to invite you to separate thread. Questions in this thread are
focused on technical aspect of it in application, not general validity of
it.

Makoto

On Wed, Jul 5, 2017 at 12:24 PM, Robert Thatcher <rwthatcher2 at yahoo.com>
wrote:

> Dear Sonia and Makoto,
>    ICA reconstruction only gives a visual appearance of artifact
> correction.   Actually, based on mathematics and empirical proof, ICA
> actually creates artifact by replacing the original time series with an
> artifcial time series that has altered all of the phase differences between
> all combinations of channels at each and every time point.   This fact has
> been admitted by Arno and Georges and Robert Lawson and Geert and many
> others based on multiple analytic proofs.  Simple delection of the parts of
> the recording that contain artifact followed by test re-test reliability
> measures to verify that the original dataset is valid and reliable and
> repeatable has been the standard since the 1960s and has been used in over
> 100,000 peer reviewed studies as cited in the National Library of Medicine
> (e.g., Pubmed).   If there is too much artifact then it is best to improve
> one's recording hygiene methodology to minimize artifact or re-test the
> subject.
>
> The premise or assumption that EEG is 100% artifact is false on its face
> and can easily be proven to be false as evidenced by
> intacellular/extracellular correlations to the surface EEG and by many
> other methods including high test retest reliability and significant
> clinical correlations and cross-validation studies such as with Diffusion
> Tensor Imaging or MRI T2 relation times and cross-validated discriminant
> analyses with sensitivity and specificity > 0.95 and connectivity studies
> such as directed coherence, phase lock and shift duration and phase slope
> index and phase-amplitude studies, etc.  If EEG was only artifact or noise
> or volume conduction or not related to underlying sources then none of the
> above EEG studies would have been published and NIH would not have bothered
> spending millions of dollars funding EEG studies, including developing the
> 1st 128 channel EEG system.
>
> Robert
>
>
> On Wednesday, July 5, 2017, 11:55:52 AM PDT, Makoto Miyakoshi <
> mmiyakoshi at ucsd.edu> wrote:
>
>
> Dear Sonia,
>
> > There should be no need to reject bad epoch, since we will be removing
> bad ICA components right? This should give us artifacts free epochs
>
> Rejecting/correcting bad chunk of time-series data is one thing, and
> rejecting artifact subspace is another. In theory, performing the former
> and run ICA will increase quality of decomposition (especially for normal
> informax; AMICA has its own datapoint rejection function). This is because
> such window rejection increases data stationarity (i.e. probability
> distribution function of one part of data is similar to that of other part
> of data). So you should still clean the time-series data before ICA.
>
> Very good point!
>
> Makoto
>
>
> On Fri, Jun 30, 2017 at 11:42 PM, Dr. Sonia Baloni <sbaloni at cbcs.ac.in>
> wrote:
>
> Dear Makoto,
>  Thanks a lot for the information, although that creates one more doubt
> for me:
>
> Yes. As long as ICA detects data rank correctly, you don't need to do
> anything. However, sometimes it fails for unknown reason, so it's safer to
> explicitly use the data rank if you know it (how many channels you
> interpolated etc.) But I don't do the second ICA these days though, it
> usually makes little to no difference if your first ICA is good enough. If
> you want to drastically change the ICA results, you should reject 30-50% of
> data!
>
>
> If first ICA is good enough and second ICA is not required then:
>  There should be no need to reject bad epoch, since we will be removing
> bad ICA components right? This should give us artifacts free epochs
>
> Best
> Sonia
>
>
>
> On 01-Jul-2017, at 06:50, Makoto Miyakoshi <mmiyakoshi at ucsd.edu> wrote:
>
> Dear Sonia,
>
> Sorry for belated response. To add some info to Isaiah's nice reply,
>
> > 1. So if we are performing two ICAs, we should interpolate these
> channels after second ICA?
>
> Yes. As long as ICA detects data rank correctly, you don't need to do
> anything. However, sometimes it fails for unknown reason, so it's safer to
> explicitly use the data rank if you know it (how many channels you
> interpolated etc.) But I don't do the second ICA these days though, it
> usually makes little to no difference if your first ICA is good enough. If
> you want to drastically change the ICA results, you should reject 30-50% of
> data!
>
> > 2. I assume rank of the data (used in ICA analysis) would be equivalent
> to the number of channels in data after running clean line function.
>
> If some of your channels were bridged, you have less rank than the number
> of channels.
>
> > If there are for example 128 channels with which data was collected and
> 10 channels were removed with clean line function then the rank of the data
> would be 118. If we now want to run ICA on this data do we need to add “
> ‘pca’,117 “ in command line option - next to ‘extended’, 1 ? as suggested
> in  the tutotrial : https://sccn.ucsd.edu/wiki/
> Chapter_09:_Decomposing_Data_ Using_ICA
> <https://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA>.
>
> If 128-10, then 'pca', 118 is correct. If 117, one component is wasted
> (which is trivial in this case).
>
> Makoto
>
>
>
> On Tue, Apr 4, 2017 at 7:00 AM, Isaiah Innis <isainnis at indiana.edu> wrote:
>
>  Dr. Baloni:
>
>
> 1. Yes, you would interpolate after the second ICA. I personally
> interpolate as the last pre-processing step. Also I assume you are
> referring to "clean raw data" and not cleanline, as the latter only removes
> line noise.
>
> 2. Yes, your rank would be 128 -10 = 118. If the channels were deleted
> from the dataset entirely ICA should detect this and run fine without
> additional options. If not then you should use the 'pca' command. Are you
> average referencing your data before ICA? In that case your rank would be
> 117 (at least as I understand it - someone on the list please correct me if
> this is inaccurate), otherwise it would be the number of channels remaining
> = 118.
>
> Remember that if you remove any components in the 1st ICA the 2nd should
> be run on the original number of components minus the number removed (e.g.
> 117-25 = 92).
>
>
> Thank you,
>
> On Thu, Mar 30, 2017 at 6:42 AM, Dr. Sonia Baloni <sbaloni at cbcs.ac.in> wr
> ote:
>
> Hi All,
>
>     I am new to EEGlab and trying to work with ICA analysis. I have been
> reading the eeglab list mails on ICA topic. I have gather few footnotes
>  and few question which I would like to ask:
>
>
> 1. Cleanline algorithm removes bad channels from the data. Posts in
> EEGlablist suggests that interpolation should be done after ICA.The
> tutorial suggests that two ICAs should be performed on the same data-set.
> So if we are performing two ICAs, we should interpolate these channels
> after second ICA?
>
> 2. I assume rank of the data (used in ICA analysis) would be equivalent to
> the number of channels in data after running clean line function. If there
> are for example 128 channels with which data was collected and 10 channels
> were removed with clean line function then the rank of the data would be
> 118. If we now want to run ICA on this data do we need to add “ ‘pca’,117 “
> in command line option - next to ‘extended’, 1 ? as suggested in  the
> tutotrial : https://sccn.ucsd.edu/wiki/C hapter_09:_Decomposing_Data_Us
> ing_ICA
> <https://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA>.
>
>
> Thanks a lot.
>
> Best
> Sonia
>
>
>
> ______________________________ _________________
> Eeglablist page: http://sccn.ucsd.edu/eeg lab/eeglabmail.html
> <http://sccn.ucsd.edu/eeglab/eeglabmail.html>
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn
> .ucsd.edu <eeglablist-unsubscribe at sccn.ucsd.edu>
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucs
> d.edu <eeglablist-request at sccn.ucsd.edu>
>
>
>
> --
> Isaiah Innis
> Indiana University '13
> EEG Technician, IUB IRF
>
>
> ______________________________ _________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/ee glabmail.html
> <http://sccn.ucsd.edu/eeglab/eeglabmail.html>
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.uc
> sd.edu <eeglablist-unsubscribe at sccn.ucsd.edu>
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.e
> du <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
> _______________________________________________
> 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/20170706/a072019f/attachment.html>