[Eeglablist] Auditory steady-state response and ERP analysis pre-processing steps

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Wed Sep 2 10:26:05 PDT 2015


Dear Michael,

> What are the pros and cons of each?

I personally recommend that you reject channels to keep your data full rank
because it is simpler for those who don't know what rank is. If you have a
practice to run rank() function to check the data rank before running ICA,
channel rejection is unnecessary. But be careful with the matlab rank()
function, sometimes it returns strange number (especially when you provide
long data).

Makoto

On Wed, Aug 26, 2015 at 8:29 AM, Michael Boyle <mrboyle at live.unc.edu> wrote:

> Thanks for the links to the Nima's paper and your pipeline Makoto, these
> were extremely helpful. I have started using the PREP pipeline and I like
> it so far. I assume for subsequent ICA analysis that interpolated channels
> must be excluded for proper ICA component extraction, correct? I was
> wondering about the case where no channels are excluded. Since the data
> have been re-referenced to an average reference the data dimensionality
> must be reduced by 1 to keep the dataset at full rank. I have found in the
> EEGLAB wiki that with the runica function, using the flag 'pca' followed by
> number of principal components to keep (n-1) is a good way to reduce data
> dimension, but I have also seen suggestions to just remove a data channel
> and not do a reconstruction from a principal component subspace. What are
> the pros and cons of each? I'd guess that for the dimension reduction when
> the 'pca' flag is thrown, runica would project the data in the PC subspace
> consisting of the eigenvectors with the maximum eigenvalues up to the
> number of eigenvectors specified in the function argument, but I'm not
> really sure if this is true.
>
> Thanks!
> Michael
>
> On Tue, Aug 25, 2015 at 6:00 AM Andreas Widmann <widmann at uni-leipzig.de>
> wrote:
>
>> Dear Makoto, Michael, and list,
>>
>> > Standard filtering policy does not go well with variations of
>> (multivariate) Granger Causality analysis. Many says 1-Hz high-pass filter
>> we do is no good for analyzing averaged ERP. High-pass filter should be
>> done before CleanLine.
>> Nima introduced a clever solution for the Cleanline vs. highpass filter
>> issue in the prep-pipeline/paper: The low frequency signal components are
>> preserved during highpass filtering (i.e., the difference between filtered
>> and unfiltered data) and re-added to the cleaned data after applying
>> Cleanline.
>>
>> Best,
>> Andreas
>>
>> > > 0. Start with 128-channel EEG from EGI HydroCel Geodesic Sensor Net,
>> with flags indicating events and epochs around events predefined (-0.1s to
>> 0.9s around the events).
>> >
>> > Use -1 to 2 sec to epoch. This is because you want to use 3-Hz 3-cycle
>> wavelet later. It does not matter even if this causes overlap of the
>> trials; it always happen as long as you use sliding windows.
>> >
>> > Also, you may want to use continuous data as long as it is cleaned.
>> >
>> > > 1. Keep all valid (even if lots of line noise) channels in dataset,
>> only remove channels which clearly have no physiological information.
>> >
>> > I agree, unless you discard all the channels.
>> >
>> > > 2. Epoch the data, reject bad epochs by eye (significant movement or
>> rare, not-stereotyped artifacts)
>> >
>> > No, make every effort not to reject epochs due to eye activity
>> (saccade, blink, etc) This is because ICA is best capable of identifying
>> them. Save your trials.
>> >
>> > > 3. Run ICA on epoched data.
>> >
>> > Make sure that you clean your data before running ICA.
>> >
>> > > 4. Reject components related to blinks, stereotypical muscle
>> activity, and 60 Hz noise, and EKG artifacts.
>> >
>> > No, do not reject anything. You'll create STUDY later, in doing which
>> ICs with 15% residual variance and those with outside-brain dipoles are
>> kicked out. Usually this will reject around 70% of your ICs (depending on
>> your data).
>> >
>> > > 5. Reconstruct sensor-level signals from remaining ICA weights
>> >
>> > It's a forward projection. Fine.
>> >
>> > > 6. Re-reference all cleaned EEG channels to the average reference
>> >
>> > Nima emphasized that when you use average reference you should choose
>> only clean channels to include. See his paper for details. Jason also said
>> average reference should be done before ICA, but it will cause rank
>> reduction (and does not change the results except for zero-centering the
>> scalp maps) so I would recommend you do it after ICA.
>> >
>> > > 7. Interpolate (using spherical splines from neighboring electrodes)
>> any channels that were rejected in step 1.
>> >
>> > No it's not necessary if you stick to ICA-centric analysis to the end.
>> >
>> > > 8. Calculate ERPs, time-frequency analysis, source localization
>> >
>> > Use STUDY for the group-level analysis.
>> >
>> > Additional info can be found here.
>> > http://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline
>> >
>> > Makoto
>> >
>> > On Tue, Aug 18, 2015 at 11:54 AM, Michael Boyle <mrboyle at live.unc.edu>
>> wrote:
>> > Dear EEGLABers,
>> >
>> > I would like to know whether or not the particular pre-processing steps
>> listed are recommended for ERP analysis, time-frequency analysis (wavelet
>> amplitude spectrograms or DFT spectrograms, ITPC), and source localization
>> analysis.
>> >
>> > In particular, I am uncertain about the ordering of steps such as
>> interpolating bad channels before or after re-referencing or ICA, or if
>> particular steps work for some analysis strategies but not others (e.g. if
>> a preprocessing step would work fine for time-frequency analysis but not be
>> appropriate for source localization). Any feedback is much appreciated!
>> >
>> > 0. Start with 128-channel EEG from EGI HydroCel Geodesic Sensor Net,
>> with flags indicating events and epochs around events predefined (-0.1s to
>> 0.9s around the events).
>> > 1. Keep all valid (even if lots of line noise) channels in dataset,
>> only remove channels which clearly have no physiological information.
>> > 2. Epoch the data, reject bad epochs by eye (significant movement or
>> rare, not-stereotyped artifacts)
>> > 3. Run ICA on epoched data.
>> > 4. Reject components related to blinks, stereotypical muscle activity,
>> and 60 Hz noise, and EKG artifacts.
>> > 5. Reconstruct sensor-level signals from remaining ICA weights
>> > 6. Re-reference all cleaned EEG channels to the average reference
>> > 7. Interpolate (using spherical splines from neighboring electrodes)
>> any channels that were rejected in step 1.
>> > 8. Calculate ERPs, time-frequency analysis, source localization
>> >
>> > Best and many thanks,
>> > Michael
>> >
>> > _______________________________________________
>> > 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
>>
>>


-- 
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/20150902/f38cddcf/attachment.html>


More information about the eeglablist mailing list