[Eeglablist] Using a laplacian filter with ICA

[Makoto Miyakoshi]

Dear Sean,

Thank you for your comment. Your explanation is very clear and reasonable.
I agree that every signal processing has different assumption, merit, and
demerit. But I recognize that ICA has a singular status among them, IF the
claim of Onton & Makeig (2006) is correct. On the other hand I know its
potential problem as well, which I want to solve some day!

Makoto



On Tue, Feb 14, 2017 at 3:02 AM, Sean Fitzgibbon <
sean.fitzgibbon at ndcn.ox.ac.uk> wrote:

> Hi Makato
>
> It is more than an assumption, however I think your wording is perhaps a
> bit strong.  The surface Laplacian is most sensitive to “*superficial
> radial dipoles*” (Srinivasan, 2007) and less sensitive to deep and/or
> tangential dipoles.  Thus tangential dipoles are disadvantaged both by
> their orientation and because they are typically deeper than radial
> dipoles.  However EEG measurement is also more sensitive to superficial
> radial sources …
>
> Obviously, whenever you manipulate your data you need to consider the
> downstream implications both good and bad.  The surface Laplacian will be
> less sensitive to sulcal and deep sources and will attenuate signal with a
> broad spatial distribution.  However, it is reference free, reduces effects
> of volume conduction, does not require a head conductivity model, and we
> argue, is less sensitive to persistent muscle contamination.
>
> Cheers, Sean
>
> Srinivasan, R. (2006). Anatomical constraints on source models for
> high-resolution EEG and MEG derived from MRI. Technology in Cancer Research
> & Treatment, 5(4), 389–399.
>
>
>
> On 25 Jan 2017, at 00:08, Makoto Miyakoshi <mmiyakoshi at ucsd.edu> wrote:
>
> Dear Sean,
>
> It's a great honor to have the comment from the first author of the paper
> directly!
>
> Is it true that an assumption of using Laplacian is that all EEG sources
> be radial and cannot be tangential? That assumption sounds generally
> reasonable since tangential (i.e. sulcus) sources would suffer a lot from
> cancellation by the facing cortex. However, when we see ICA results, there
> are quite a bit of seemingly tangential sources. Actually, the tangential
> (sulcus) sources may be able to survive if the facing cortical patches are
> less active and/or activated independently. Otherwise, what are these
> results we see in ICA results? Too beautiful to be noise.
>
> Makoto
>
> On Thu, Jan 12, 2017 at 11:43 PM, Sean Fitzgibbon <
> sean.fitzgibbon at ndcn.ox.ac.uk> wrote:
>
>> Dear Abby
>>
>> The first thing to note is that this method is not designed to deal with
>> *transient* muscle activity defined as “*transient contractions of
>> facial, cranial, and neck muscles [that] generate high amplitude electrical
>> signals... Typically [transient] EMG contamination is of sufficient
>> amplitude that it can be easily identified (visually or algorithmically)
>> and it is common practice to excise and discard contaminated periods of EEG*”.
>> We would still advocate excising these segments.  Rather, the method seeks
>> to rectify *persistent* muscle contamination defined as "*sustained
>> non-forceful contractions (e.g. tone) of these muscles [that] generate
>> persistent low amplitude contamination*” that is very difficult to
>> identify in the scalp EEG.  The motivation for our interest in persistent
>> EMG is “*our experiments in paralysed conscious humans have demonstrated
>> that persistent EMG strength exceeds EEG strength from frequencies as low
>> as 10–20 Hz, more so peripherally than centrally, and that by 100 Hz, EMG
>> can be more than 200 times larger than EEG*”.
>>
>> The method, in brief, is to decompose the EEG using ICA, identify and
>> remove EMG components, project back to scalp space, and then apply the
>> laplacian.  The average reference is not applied.  We have shown in earlier
>> papers, consistent with Makato’s comment below, that ICA cannot completely
>> decompose EMG, although it can improve SNR.  However, using paralysed
>> (muscle-free) EEG as the gold standard, we have demonstrated empirically
>> that “*the combination of the ICA procedure and the surface Laplacian
>> results in a strong suppression of [persistent] EMG contamination at all
>> scalp sites and frequencies*” and "*does not impair the detection of
>> well-known, cerebral responses*”.
>>
>> Anyway, to actually respond to your question, it is commonly stated that
>> high spatial density EEG (~128 channel) is required for Lacplacian.
>> However this is challenged by Kayser and Tenke (2006; full ref below), at
>> least for ERPs.  Our own experience is that we get reasonable
>> approximations at lower densities, consistent with Kayser and Tenke.  The
>> 25-channels you have is quite low, but might still be worth trying.  I
>> cannot comment on whether is is useful to reverse the order and do
>> Laplacian before ICA as we have not ever tried that.  However, "*Tenke
>> and Kayser (2005) proposed a method called fPCA, which … demonstrate[s]
>> compelling removal of EMG contamination … [and] requires application of the
>> surface Laplacian to the EEG to estimate CSD, calculation of the
>> amplitude spectra from the CSD, and then application of principal component
>> analysis (PCA) to decompose the spectra*” which sounds similar’ish to
>> what you are proposing and may be a good starting point for you…
>>
>> Cheers, Sean
>>
>> Kayser, J., & Tenke, C. E. (2006). Principal components analysis of
>> Laplacian waveforms as a generic method for identifying ERP generator
>> patterns: II. Adequacy of low-density estimates. *Clinical
>> Neurophysiology*, *117*(2), 369–380. http://doi.org/10.1016/j.clinp
>> h.2005.08.033
>>
>> Tenke, C. E., & KAYSER, J. (2005). Reference-free quantification of EEG
>> spectra: Combining current source density (CSD) and frequency principal
>> components analysis (fPCA). *Clinical Neurophysiology*, *116*(12),
>> 2826–2846. http://doi.org/10.1016/j.clinph.2005.08.007
>>
>>
>>
>>
>> _______________________________________________
>>
>> *Dr Sean Fitzgibbon * BSc PhD
>> Postdoctoral Researcher
>> FMRIB <http://www.fmrib.ox.ac.uk/> Centre, University of Oxford
>> Oxford, United Kingdom
>>
>> On 13 Jan 2017, at 06:55, Makoto Miyakoshi <mmiyakoshi at ucsd.edu> wrote:
>>
>> Dear Abby,
>>
>> That's a new method for me and I don't know how it works. From the
>> abstract, it seems you apply ICA, reject muscle components, then Laplacian?
>> If so, the question is when you want to perform the average reference.
>> Probably you want to average reference after Laplacian so that the computed
>> average reference is EMG free (not exactly free, but very much removed?)
>>
>> By the way, ICA cannot decompose EMG becaues EMG source spreads along
>> with muscle fibers, so my former colleague told me. Empirically, I also
>> experienced that muscle is harder to decomposed compared with other typical
>> artifacts such as eyes, which makes sense.
>>
>> From Lufthansa455 to Frankfurt,
>>
>> Makoto
>>
>>
>>
>> On Tue, Jan 10, 2017 at 10:28 AM, Dickinson, Abigail <
>> ADickinson at mednet.ucla.edu> wrote:
>>
>>> Hi everyone,
>>>
>>>
>>> I jwondered if anyone could help me with a question I had regarding
>>> using a laplacian filter, and at which stage I should implement the
>>>  filter. I currently have a processing pipeline which effectively cleans
>>> spontaneous data collected from low-functioning children (EGI 128 channel,
>>> 500Hz). Due to the nature of our sample I am working with a few
>>> constraints, including a high amount of EMG and relatively short recording
>>> lengths (1-2 minutes).
>>>
>>>
>>> The processing pipeline I currently use involves:
>>>
>>> -FIR filter (high pass:1Hz, low pass:100Hz)
>>>
>>> -remove bad channels
>>>
>>> -down-sampling to the 10-20 system 25 channel montage (in order to have
>>> an adequate k-factor to run ICA)
>>>
>>> -remove bad segments of data
>>>
>>> -Run ICA
>>>
>>> -Remove artifactual components
>>>
>>> -Re-reference to average
>>>
>>>
>>> I recently saw a paper which advocated the use of a laplacian filter
>>> along with ICA to remove EMG (https://www.ncbi.nlm.nih.gov/
>>> pubmed/25455426), and also I would like to be able to run cohernece
>>> analyses at a channel level, so wanted to re-process this data with a
>>> laplacian filter.
>>>
>>>
>>> However, I cannot apply the laplacian after ICA, as at that point I only
>>> have 25 channels. I wondered if anyone could comment on whether it would be
>>> appropriate to apply the laplacian filter either on all 128 channels, or
>>> after removing bad channels, and then continuing with the rest of the
>>> processing pipeline detailed above?
>>>
>>>
>>> Thanks in advance!
>>>
>>>
>>> Best wishes,
>>>
>>>
>>> Abby
>>>
>>>
>>>
>>> ------------------------------
>>>
>>> UCLA HEALTH SCIENCES IMPORTANT WARNING: This email (and any attachments)
>>> is only intended for the use of the person or entity to which it is
>>> addressed, and may contain information that is privileged and confidential.
>>> You, the recipient, are obligated to maintain it in a safe, secure and
>>> confidential manner. Unauthorized redisclosure or failure to maintain
>>> confidentiality may subject you to federal and state penalties. If you are
>>> not the intended recipient, please immediately notify us by return email,
>>> and delete this message from your computer.
>>>
>>> _______________________________________________
>>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>>> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.uc
>>> sd.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.uc
>> sd.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
>
>
>


-- 
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/20170217/726aeb1c/attachment.html>