[Eeglablist] automatized ICA data correction and channel interpolation

Marius Klug marius.s.klug at gmail.com
Mon Jul 10 02:47:16 PDT 2017


Dear Makoto, Luca, all,

sorry for the late reply! As a quick explanation from my side, I know very
well that the human eye (or brain) is the best classifier system we have at
the moment and I do the IC selection manually as well, since I've also
tested many algorithms and found them to be insufficient. I just wanted to
point out that there are _some_ options, however limited in their
performance, that attempt the automatic classification and if the reviewer
insist, one can point those out. The benefit here would be that for example
with SASICA one can use the pre-made classification and do a manual
inspection in addition based on the measures computed by SASICA, as the
authors have suggested themselves. It's a helper tool for the
not-so-experienced user that also might help the paper get through
reviewing.

Also I know of Luca's project but if I'm not mistaken it's not out there to
be used, is it? I would be more than wiling to test it extensively since,
as I've said, increasing the level of reproducibility is one of my personal
goals in my analysis pipeline!

Marius

2017-07-03 19:31 GMT+02:00 Luca Pion-Tonachini <lpionton at ucsd.edu>:

> Hi Makoto,
>
> Just to clarify, while the process of developing my method does use inputs
> from the ICLabel website, the final product that people will use will
> *only* look at the EEG measures and nothing from the website. Just like
> the other methods, it will not require people to actively contribute labels
> to work once released. The label collection is only for the development of
> the classier.
>
> Luca
>
>
> On 06/30/2017 05:52 PM, Makoto Miyakoshi wrote:
>
> Dear Marius,
>
> Thank you for your opinion.
>
> My colleague Luca has been working on automatic IC labeling using machine
> learning algorithm: but it still uses *user input *as the data to be
> tranied. So it is NOT a solution to develop a better or perfect algorithm
> that judges what is what based on EEG measures (he tested all the kinds of
> algorithms available at the timepoint of last year, and found nothing was
> perfect.) See this page for his data collection scheme (which is also a
> educational tool)
>
> http://reaching.ucsd.edu:8000/auth/login
>
> I have been collaborating with radiologists and neurologists. They
> diagnose patients, particularly neurologists even determines which brain
> tissue to remove. What algorithm do they use? They use eyeballs. After all,
> humans are still the best learning machine today (though it is very
> tempting to make my own criteria for IC labeling for non-aggressive data
> cleaning)
>
> Makoto
>
> On Fri, Jun 30, 2017 at 12:54 AM, Marius Klug <marius.s.klug at gmail.com>
> wrote:
>
>> Hi Gian, Makoto, list
>>
>> the quest for more reproducibility in EEG is one of my personal topics of
>> interest, so here are a few comments from my point of view: I too would
>> consider an automated method for IC classification a valuable tool, and
>> dismissing it is not as easy as Makoto suggests if you ask me. But I agree
>> with Makoto: there are no perfect solutions and manual inspection to date
>> is hard to get around since experienced inspectors are just better than the
>> algorithms. However the reasoning that it's not reproducible is valid, and
>> also maybe a trained eye is better than automated methods, but a novice
>> might not exactly know where to look at and the quality may suffer...
>>
>> You didn't specify if you just want to reject eye-movement and other
>> artifacts or want to go the other way round and just keep ICs that are
>> specifically generated by the brain. In the latter case you might want to
>> try out a few things, because there are some that work okay-ish:
>> First, you can use dipole fitting and check the residual variance for
>> each IC, and have a threshold of <0.15 (typically) for brain ICs. This is
>> not a fail-safe method, however, since eye-ICs and sometimes also very
>> narrow (not spread around the whole head) muscle or other artifactual ICs
>> can have a small residual variance. You can also check if the located
>> dipoles lie inside the brain of the dipfit model.
>> Then there's SASICA, an EEGLAB plug-in also published (Chaumon, Bishop &
>> Busch, 2015). I highly recommend the paper also for a deeper understanding
>> of the IC classification. Playing around with SASICA I found it to be okay
>> - not perfect - for an automated method. You can play around with the
>> different classifiers (it uses several methods in combination - dipfit
>> being one of them, but not mandatory) and their respective thresholds and
>> check if it suits you.
>>
>> Unfortunately, EEG methods do have a high degree of subjectivity in
>> several steps, and the automated methods are usually not as good as manual
>> inspection yet (bad channel detection and time-domain artifacts as well),
>> but it's a trade-off between clear reproducibility and best quality, the
>> topic as a whole needs to be treated with care. I hope you find some
>> valuable information in my suggestions and wish you a successful study and
>> smooth data analysis process! ;-)
>>
>> Best,
>> Marius
>>
>>
>> 2017-06-30 4:02 GMT+02:00 Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
>>
>>> Dear Gian Marco,
>>>
>>> Ah I missed your important email! Sorry for being so late.
>>> Automated ICA rejection is not more reliable than manual inspection,
>>> because the programmer implemented his or her own criteria (or his or her
>>> research results on criteria) to the application. Therefore, for rebuttal
>>> you can say 'Then why do neurologists still use their eyeballs to identify
>>> epileptic spikes, given the great advance of machine learning technology
>>> today' etc... Trained eyes are still one of the best solutions.
>>>
>>> > Another question is about bad channel interpolation. In EEGlab there
>>> is the kurtosis method but it does not work so good. Do you known any other
>>> automatic method that recognize bad channel?
>>>
>>> I recommend either the one implemented in clean_rawdata() plugin or the
>>> one in PREP plugin. The former was developed by Christian Kothe and the
>>> other by Nima Bigdely-Shamlo, both are former SCCN colleagues (now in
>>> Qusp). Let me share a piece of my recent writing.
>>>
>>> We first performed outlier channel detection, rejection, and
>>> interpolation using the clean_rawdata plug-in (contributed by Christian
>>> Kothe) also available through the EEGLAB Extension Manager. This plug-in
>>> calculates each scalp channel signal’s correlation to its random sample
>>> consensus (RANSAC) estimate computed from nearby scalp channel signals in
>>> successive 5-s segments. Channel signals exhibiting low correlation to
>>> signals in neighboring scalp channels (e.g., here r < 0.8 at more than 40%
>>> of the data points) were rejected and then replaced with an interpolated
>>> channel using the spherical option in eeg_interp which makes use of
>>> Legendre polynomials up to degree 7 to calculate unbiased expected channel
>>> values (see Mullen et al., 2015).
>>>
>>> Makoto
>>>
>>>
>>>
>>> On Mon, Mar 27, 2017 at 7:58 AM, Gian Marco Duma <gmduma90 at gmail.com>
>>> wrote:
>>>
>>>> Dear EEGlab community, I'm writing because I need a suggestion. ICA
>>>> works very well for eye blink and eye movements correction, and indeed it
>>>> works very well too for artifacts identification as muscle contraction,
>>>> electrical noise and so on. Thanks to the experience it becomes possible to
>>>> recognize the specific components by visual inspection, even if an
>>>> experimenter must be very careful in components rejection. I submitted a
>>>> pre-registered reports to Cortex journal, and they asked me for an
>>>> automatized ICA components rejection method because visual inspection is
>>>> not considered as a reproducible method. So I'm writing to ask for a
>>>> suggetion about possible automatized components rejection methods,
>>>> specially for eye blink and eye movements.
>>>> Another question is about bad channel interpolation. In EEGlab there is
>>>> the kurtosis method but it does not work so good. Do you known any other
>>>> automatic method that recognize bad channel?
>>>> Thanks for your help
>>>>
>>>> _______________________________________________
>>>> 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
>
>
> _______________________________________________
> 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
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170710/954e5daf/attachment-0001.html>


More information about the eeglablist mailing list