[Eeglablist] Baselining and filtering for ICA with epoched data

Fri Aug 17 20:41:15 PDT 2018

Dear Eric,

> If I calculate the ICA weights using data filtered at 1 Hz, by definition
this solution will *not* give maximally independent components when it is
applied to 0.1 Hz filtered data. In the 0.1 Hz filtered data there is low
frequency information in the data that ICA was not trained on and therefore
doesn't "know" anything about.

Correct.

> That low frequency information must, mathematically, end up somewhere in
the IC activations when I multiply the unmixing matrix with the 0.1 Hz
filtered data. What are the consequences and pitfalls of this? Is it
something to worry about? Under what circumstances?

Let's focus on two scenarios I brought up in my previous response.
1. If the < 2Hz activity is temporally correlated with > 2Hz activity ->
will be decomposed altogether with no conflict.
2. If the < 2Hz activity is independent of > 2Hz activity -> depending on
the source of <2Hz activity, the ICs localized nearby got affected.

Imagine the Scenario 2 in this way: You are in the dinner party. On a
buffet table, there are big four plates of salad, steak, salmon, and
desert. You want to take some salad to your place. The moment you reached
your arm to the salad plate--power outage happened. All the lights are off.
It's pitch dark, but you decided to take the salad anyway. You know where
it the salad plate is, and you have a tong and a plate in your hand, so why
not? So you take the salad to your plate in the darkness, enjoying the
unexpected inconvenience. Now the light is back. Guess what you find in
your plate--you have the salad, but it is soaked with mushroom soup!
Probably someone spilled his bowl of mushroom soup over the salad plate by
accident. Now you eat salad unexpectedly mixed with mushroom soup.

This is what happens when > 2 Hz ICA is applied back to > 0.1 Hz data. The
unexpected mushroom sour contamination represents < 2 Hz activities. The
2-Hz highpassed ICA could only see the four plates, and did not know about
the mushroom soup contamination. If you apply ICA again with full data
including the mushroom soup contamination, most likely the mushroom-soup
contaminated foods are distinguished as a new independent component.

> My guess is that the low frequency information gets divided across the
ICs in a somewhat a priori unpredictable way depending on the scalp
distribution of the low frequency information and the scalp distribution of
the ICs.

In the example above, it depends on which plate the mushroom soup
contaminate. It may hit only the salad plate, or it may hit three plates,
etc.

> If I then remove one of the ICs, I may remove some of this low frequency
information with it.

If you replace the salad plate with a new one, then the
mushroom-soup-contaminated salad is also gone.

> If the low frequency information in question is purely noise, this could
lead to some pattern of noise/artifact in the data that is hard to
interpret or move noise to electrodes that didn't originally include it.

Correct.

> One of the reasons for using the 0.1 Hz filter is that part of the
effects I am interested in (e.g., later ERP components) contain information
below 1 Hz (see Tanner et al., 2015).

ERP indeed seems a broadband phenomenon. Broad toward the lower end.

> Can I be confident that ICA does a good job of isolating artifact from
neural activity of interest if part of that activity of interest was not
present in the training dataset?

To determine this, we need a study to determine SNR in the infraslow EEG.
We SCCN people tend to say 'there is more noise in infraslow' but
apparently there are convincing infraslow EEG studies. But I don't know how
much signal and how much noise are in the infraslow range. Another reason
why we SCCN people cut infraslow range is because ICA is biased toward
power. EEG signal has 1/f power spectral density, which means the lowest
frequency tends to have highest impact to ICA results, although we are
*usually* more interested in alpha and theta which are main contributors to
ERPs.

One suggestion you may want to try is that you separate the data into <2Hz
and >2Hz, and perform ICA separately. You have to adjust data rank when in
doing so, the filtering can affect data rank, particularly for < 2Hz data.
Then you can see what you are excluding. You can also see if any components
found in < 2Hz ICA is similar/dissimilar to > 2Hz ICA.

Makoto

On Tue, Feb 27, 2018 at 9:14 AM Eric Fields <eric.fields at bc.edu> wrote:

> Hi Jumana and others,
>
> To be clear, my question isn't about implementation (I know it is
> relatively easy to calculate ICA weights from one dataset and apply them to
> another). My question was about how this effects the data.
>
> If I calculate the ICA weights using data filtered at 1 Hz, by definition
> this solution will *not* give maximally independent components when it is
> applied to 0.1 Hz filtered data. In the 0.1 Hz filtered data there is low
> frequency information in the data that ICA was not trained on and therefore
> doesn't "know" anything about. That low frequency information must,
> mathematically, end up somewhere in the IC activations when I multiply the
> unmixing matrix with the 0.1 Hz filtered data. What are the consequences
> and pitfalls of this? Is it something to worry about? Under what
> circumstances?
>
> My guess is that the low frequency information gets divided across the ICs
> in a somewhat a priori unpredictable way depending on the scalp
> distribution of the low frequency information and the scalp distribution of
> the ICs. If I then remove one of the ICs, I may remove some of this low
> frequency information with it. If so:
>
>    1. If the low frequency information in question is purely noise, this
>    could lead to some pattern of noise/artifact in the data that is hard to
>    interpret or move noise to electrodes that didn't originally include it.
>    2. One of the reasons for using the 0.1 Hz filter is that part of the
>    effects I am interested in (e.g., later ERP components) contain information
>    below 1 Hz (see Tanner et al., 2015). Can I be confident that ICA does a
>    good job of isolating artifact from neural activity of interest if part of
>    that activity of interest was not present in the training dataset?
>
>
> Have these issues been addressed anywhere in the literature or does anyone
> have recommendations?
>
> Eric
>
> -----
> Eric Fields, Ph.D.
> Postdoctoral Fellow
> Cognitive and Affective Neuroscience Laboratory
> <https://www2.bc.edu/elizabeth-kensinger/>, Boston College
> Aging, Culture, and Cognition Laboratory
> <http://www.brandeis.edu/gutchess/>, Brandeis University
> eric.fields at bc.edu
>
> On Wed, Feb 21, 2018 at 1:00 PM, Ahmad, Jumana <jumana.ahmad at kcl.ac.uk>
> wrote:
>
>> It depends if you want to examine component activity below 1Hz. Most
>> artifacts of interest, such as blinks and saccades should be higher
>> frequency etc.
>>
>>
>> *------------------------------------------*
>> *Jumana Ahmad*
>> Post-Doctoral Research Worker in Cognitive Neuroscience
>> *EU-AIMS Longitudinal European Autism Project (LEAP) & SynaG Study*
>> Room M1.26 <https://maps.google.com/?q=M1.26&entry=gmail&source=g>.Department
>> of Forensic and Neurodevelopmental Sciences (PO 23) | Institute of
>> Psychiatry, Psychology & Neuroscience | King’s College London | 16 De
>> Crespigny Park | London SE5 8AF
>> <https://maps.google.com/?q=16+De+Crespigny+Park+%7C+London+SE5+8AF&entry=gmail&source=g>
>>
>> *Phone:* 0207 848 5359| *Email:* jumana.ahmad at kcl.ac.uk
>> <antonia.sanjose at kcl.ac.uk> | *Website:* www.eu-aims.eu | *Facebook:*
>> www.facebook.com/euaims
>>
>> ------------------------------
>> *From:* eeglablist <eeglablist-bounces at sccn.ucsd.edu> on behalf of Eric
>> Fields <eric.fields at bc.edu>
>> *Sent:* 21 February 2018 03:45:30
>> *To:* EEGLAB List
>> *Subject:* [Eeglablist] Baselining and filtering for ICA with epoched
>> data
>>
>> Hi,
>>
>> I know there have been other threads related to this, so I apologize if
>> this has been addressed directly and I missed it.
>>
>> Groppe et al. (2009) showed that ICA gives more reliable results if you
>> use the full epoch instead of the prestimulus period to baseline. The
>> reason generally given for this is that baseline correction changes the
>> scalp distribution of sources depending on what is happening in the
>> baseline period. By this logic, using the full epoch should improve ICA
>> (because longer periods are less affected by random variations), but no
>> baseline correction at all should be even better.
>>
>> Meanwhile, Winkler et al. (2015) have suggested that ICA works best on
>> data high pass filtered at 1-2 Hz.
>>
>> Assuming I prefer to use a 0.1 Hz high pass filter (because of
>> distortions 1 Hz filters can cause in the ERP: Tanner et al., 2015), I have
>> two questions:
>>
>>
>>    1. Does the removal of additional low frequency noise you get from
>>    using a full epoch baseline (vs no baseline) outweigh the downsides of
>>    baseline correction for ICA?
>>    2. Alternatively, is it appropriate to apply a 1 or 2 Hz filter to
>>    the data used for ICA training, and then apply the ICA solution to an
>>    EEGset filtered at 0.1 Hz? Winkler et al. suggest this, but what happens to
>>    the low frequency information in the data when the ICA solution that has
>>    been learned without it is applied? Can this cause problems?
>>
>>
>> Thanks!
>>
>> Eric
>>
>> -----
>> Eric Fields, Ph.D.
>> Postdoctoral Fellow
>> Cognitive and Affective Neuroscience Laboratory
>> <https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww2.bc.edu%2Felizabeth-kensinger%2F&data=01%7C01%7Cjumana.ahmad%40kcl.ac.uk%7C60790510144d4d17f8f608d579548d03%7C8370cf1416f34c16b83c724071654356%7C0&sdata=Vmoa8t0Q6S95V85lnDxKRWHj4Wt%2B2Pw87O2v5nSZE%2BM%3D&reserved=0>,
>> Boston College
>> Aging, Culture, and Cognition Laboratory
>> <https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.brandeis.edu%2Fgutchess%2F&data=01%7C01%7Cjumana.ahmad%40kcl.ac.uk%7C60790510144d4d17f8f608d579548d03%7C8370cf1416f34c16b83c724071654356%7C0&sdata=KkZ5TQyb7BlOduU%2BXU66q2I5qJJXDvjk%2BPguBy3veQI%3D&reserved=0>,
>> Brandeis University
>> eric.fields at bc.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

-- 
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/20180817/c007e1e8/attachment.html>