[Eeglablist] Pipeline of processing to optimize ICA for artrifact removal

Thu Feb 23 20:33:49 PST 2012

Some additional >> comments,   Scott Makeig

On Thu, Feb 23, 2012 at 9:04 AM, Arnaud Delorme <arno at ucsd.edu> wrote:

> Dear Mosdestino,
>
> 1. Difficulties with ICA. When removing ICA components, one of the main
> concern is the quality of your decomposition. We are currently working on
> tools to assess this quality although this can be tricky because of the
> large inter-subject variability. In the meantime, if you have multiple
> components for each type of artifacts, this is usually a sign that the
> quality of your decomposition is poor. One of the main factor to increase
> quality is to increase the amount of data and also high pass filtering if
> you have large offset in your data channels.
>
> >> Modestino, you do not mention the data length he is using. An important
value is #timepoints/(#channels)^2 ... this should be much more than 1
(even as high as 30 for 256 channels, in our (limited) experience -- I hope
we will run a numerical experiment on this soon...  Running ICA on
insufficient length data is the most common problem in applying ICA
successfully.

> 2. The reference should not alter the decomposition because it is a linear
> transformation. This said, it does in some cases, because of numerical
> implementation of ICA algorithms. I have found myself that average
> reference tend to perform better than single reference. However, this is a
> qualitative judgment and I could be proven wrong in the future. In general,
> ICA works irrespective of the reference you use.
>
> 3. Yes, you should include the EOG channels if you have recorded them
> using the same reference as the other channels. Otherwise you should
> exclude them.
>

>> See my more extensive comment to this list a few days back...

> 4. It is true that high pass filtering above 0.5 Hz tends to remove slow
> drifts. ICA is sensitive to these drifts. Some users even like to filter
> above 2 Hz for noisy datasets, but then you might loose interesting data.
> However, again there has not been a clear demonstration that the quality of
> ICA increases with filtering. When the analog filter of the amplifier is
> sufficient, and data channels do not exhibit strong offsets, ICA performs
> well on raw data. Note that you should not apply a decomposition obtained
> on filtered data on an unfiltered dataset (the large offset in your data
> channels which have been removed by filtering will be used to compute ICA
> component activity -- your ICA component activities will be ruined).
>
> >> I think Arno's 'ruined' above is a bit strong. If you have not trained
ICA on some portion of the data, then applying an ICA unmixing matrix
trained on some *other* data obviously may not give meaningful results -- *
unless* the new data has the *same* spatial source structure. In this case,
does it??

> 5. You can detrend if you want. It should not affect much the data for
> long records. Most likely, the ICA decomposition will remain unchanged.
>

>> Here I disagree -- large, long term offsets and linear trends will
certainly affect the ICA decomposition (hence the practice of high pass
filtering the data before ICA decomposition).

6. ICA signal amplitude is -- unless you unselect the option -- scaled to
> microvolt. The default value for thresholding might not be optimal so
> adjust it manually.
>

>> To be more clear:  ICA decomposes the data into a product of fixed
component scalp maps by maximally independent component time courses. When
the scalp maps are normalized to RMS=1 amplitude (as generally in EEGLAB),
then the RMS of the component time series must be the RMS contribution
(across all the scalp channels) to the channel data.   -Scott Makeig

Best,
>
> Arno
>
> On Feb 21, 2012, at 6:20 AM, Modestino, Edward J *HS wrote:
>
> Hello Experts,****
> I am trying to come up with a pipeline of my own processing including
> using ICA to remove artifacts.  I am hoping people will be generous enough
> to share with me their knowledge on this topic. ****
> ** **
> First, in using ICA, I have found many difficulties.  I have attempted to
> remove artifacts like lateral eye movements and eye blinks using ICA.  In
> some cases it has appears that certain components which match
> topographically the known presentation of eye movements, when removed,
> appear to be taking some of the cortical EEG with it.  This is consistent
> with Lindsen and Bhattacharya (2010).  In other cases, it appears that
> residue of the artifact still remains in the EEG after component removal.
> This is consistent with Shackman, McMenamin, Maxwell, Greischar and
> Davidson (2010) and at least two other people reporting this to the email
> listserv only within the last week.  I have used various EEGLAB plugins to
> choose which components should be considered artifacts based on statistics
> and still found some components that were clearly artifacts not earmarked
> for removal as artifacts.   ****
> ** **
> Second, I am working with Biosemi and another system with data that is
> unreferenced.  I am wondering if it would be easier to perform ICA on
> unreferenced data to avoid mixing the sources even further with a common
> average reference. ****
> ** **
> Third, should I include bipolar EOG channels in ICA.  It does not seem
> like a good idea as they are referenced differently than the EEG channels.
> ****
> ** **
> Fourth, I know that Arno has mentioned that ICA performs better on data
> that has been high-pass filtered at 0.5 or even1 Hz to remove skin
> conductance and other slow signals that are not of cortical origin.  Some
> others have suggested doing ICA on data filtered this way and then applying
> the ICA weighting to the unfiltered raw data.  I am not sure how this would
> work and if it can be justified. ****
> ** **
> Fifth, should I detrend the data (remove the mean) from the continuous
> data?  Or should I detrend at the epoch level or merely baseline correct
> (remove the baseline mean from only the baseline period)?  If I am doing
> spectral analysis, I believe detrending is necessary.  In such a case,
> should this be done on the continuous data or on the epoched data?****
> ** **
> Sixth, I feel confident using a thresholding technique to remove
> excursions greater than +/-60 microvolts as shown by Fu et al. (2001) as
> artifacts.  I can set this up in the automated artifact removal in EEGLAB.
> If I were to use  the default options on the thresholding or any other
> statistical means of classifying artifacts, often all of my data is gone.
> I thought using the automated and default settings might be good to
> preprocess the data BEFORE using ICA decomposition.  However, as there is
> often no data left, or very little, perhaps 20-40% of the epoched data if I
> am lucky, if I am unlucky none, this has not worked out well.****
> ** **
> I have been trying various combinations of all of these and still feel
> like I need some input from others as to the order of these things (it may
> not matter if these are linear processes). ****
> ** **
> My feeling is all (or some) of these things may have an effect on the
> outcome of ICA decomposition and subsequently artifact removal.  As I am
> finding residual artifacts left after component removal and sometimes even
> what appears to be EEG cortical signal removed, I thought it might be a
> good idea to get input on these things for an optimal pipeline.  *Please
> let me know your thoughts on a pipepline you prefer, what steps I should
> use and perhaps even an order of steps. *
> ** **
> Thanks in advance 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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120223/ab76f96d/attachment.html>