[Eeglablist] Poor ICA decompostion

Fri Nov 9 15:13:07 PST 2018

Thank you both for your input!

Scott, I will try your suggestions and will experiment with using the
ICLabel plugin and the multimodal AMICA. I'm actually at the workshop now,
and it's very helpful to learn more about these tools.

Jason, I did try this on a subset of subjects, and didn't see much
difference in the decomposition, but it could have been that I wasn't
looking at things the right way. I'm hoping to continue learning more about
working with components at the workshop this week.

Thank you both again!
Kelly

On Tue, Nov 6, 2018 at 5:27 PM Jason Palmer <japalmer29 at gmail.com> wrote:

> Hi Kelly,
>
> It may help to run Amica on the data without the interpolated channels
> since Amica will just remove these dependent dimensions anyway. Then you
> can interpolate the ICA maps as though they are time points, and thrn
> backproject with the interpolated maps.
>
> Best,
> Jason
>
> On Nov 5, 2018 8:13 AM, "Scott Makeig" <smakeig at ucsd.edu> wrote:
>
> Kelly -
>
> Your 10-26 screen shot of IC maps does look good. I suggest you try
> downloading and applying the ICLabel plug-in, which may suggest several ICs
> accounting for EMG activity in neck/scalp muscles (etc.). Again, ICA
> attempts to separate your data into separate *sources of information.* Each
> scalp/neck muscle has its own source of information to provide, and
> individual-channel noise may be distinctive enough to 'win' an IC filter,
> etc.
>
> I do suggest you experiment with decomposing more data, not just the data
> epochs - though removing data segments with movement noise. After all, most
> of the effective EEG sources are also active between trials, and more
> (clean) data is better.
>
> Another option you might explore is to apply multi-modal AMICA to the
> whole data (both sessions -- IF the electrode positions on the head do not
> change). See this paper in press for examples:
>
> Hsu, S.H., Pion-Tonachini, L., Palmer, J., Miyakoshi, M., Makeig, S. and
> Jung, T.P., 2018. Modeling brain dynamic state changes with adaptive
> mixture independent component analysis. *NeuroImage*, *183*, pp.47-61.
>
> Re PCA vs lower channel count:  The problem with using PCA for dimension
> reduction is that all the independent sources included in the original
> (e.g.) 128 channels load on to the retained PCA subspace - in fact, PCA
> attempts to include as many ICs as possible in each PC. Omitting channels
> also leaves projections of all the independent sources in the data, though
> somewhat less markedly (in principle). It could be a worthwhile experiment
> to try this: stochastically remove (various combinations of) channels,
> perform ICA decomposition, and then cluster the resulting ICs to look for
> e.g. brain IC clusters of interest...  - though exploring this would be a
> small research project in itself.
>
> Scott Makeig
>
> On Fri, Oct 26, 2018 at 7:53 AM Kelly Michaelis <kcmichaelis at gmail.com>
> wrote:
>
>> Hi Arno, Marius and Makoto,
>>
>> Thank you all for the helpful advice!
>>
>> To answer Arno's question about other ICA algorithms, I tried runica, and
>> I found AMICA produced better results. I did also try re-running it with
>> just removing the channels and not interpolating before ICA, and found that
>> it didn't really make a difference.
>>
>> As for the amount of data - I have two recording sessions per subject,
>> and the sessions are split into two .set files (I reapplied saline to lower
>> impedances halfway through the session). Each .set file is submitted to ICA
>> independently. Each file is about 30 minutes of data, but once that gets
>> epoched to 1sec epochs before ICA, it's trimmed considerably, so that
>> before ICA I end up with (data = 128x250x2052), depending on the # of
>> epochs removed. So perhaps this is part of why things don't look as pretty
>> as they could?
>>
>> My data are fairly noisy - lots of EMG and non-repetitive artifacts - but
>> this being my first major EEG study I didn't have much to compare my ICA
>> decompositions to except tutorials online. After hearing from you all,
>> perhaps what I'm getting isn't all that abnormal for 128 channels and
>> somewhat noisy data. Following Makoto's advice, I have been checking the
>> component scroll data for each file, and it looks fine.
>>
>> I've already gone through half the subjects and rejected artifactual
>> components (I'm rejecting about 3-10 components per file - this may be too
>> many? I'm careful to only reject things that are clearly blinks, cardiac
>> artifacts, saccades, or what looks like single bad channels or bad EMG). This
>> screenshot
>> <https://georgetown.box.com/s/78olhxp6wx5k0x3q8u02nhypsd1cgfnc> shows
>> another subject with a few components pulled up - #3 which seems to be
>> clearly brain, and #27, which would be an EMG artifact that I might reject
>> along with the clear blink/cardiac components (1,2,7,9,14)
>>
>> If this still looks poor, I can re-run everything through ICA and try the
>> 1.5Hz filter, reducing the number of components, and subtracting the extra
>> channel (FYI, when referencing, I did use the method where I add a dummy
>> channel, re-ref, then remove the dummy channel - so maybe this is
>> unnecessary?).
>>
>> I'm curious to know what you all think. Thank you!
>>
>> Kelly
>>
>>
>>
>>
>>
>> On Thu, Oct 25, 2018 at 2:47 PM Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
>> wrote:
>>
>>> Dear Kelly,
>>>
>>> I like Marius's advice. Try 1.5Hz high-pass. Also, your data look fine
>>> with me.
>>>
>>> See the pictures shown here. If you screw up in copying weight matrices,
>>> you'll NEVER know it by looking at scalp maps. You MUST check EEG.icaact
>>> which is ICA activation on the scroll data.
>>>
>>> https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline#Tips_for_copying_ICA_results_to_other_datasets_.2806.2F26.2F2018_updated.29
>>>
>>> > I got one suggestion to reduce the number of components down to
>>> something like 64, but this article by Fiorenzo, Delorme, Makeig recommends
>>> against that.
>>>
>>> I have been very worried that Fiorenzo's message was widely
>>> misunderstood. He did not say using PCA is unconditionally bad. Instead,
>>> his message was that do not use PCA to save time. PCA is a lossy
>>> compression, so you do lose information. Also, usually up to 20 principal
>>> components explains >95% of data, but most of the variance here is
>>> artifacts (eye blinks etc). Although 'keeping 95% variance' is a
>>> widely-used engineering standard practice, it does not hold for EEG.
>>>
>>> Using PCA is TOTALLY VALID if you know you need to reduce data
>>> dimension. There is no evidence, for example, that channel rejection is
>>> better than PAC dimension reduction.
>>>
>>> Makoto
>>>
>>>
>>>
>>> On Tue, Oct 16, 2018 at 8:18 AM Kelly Michaelis <kcmichaelis at gmail.com>
>>> wrote:
>>>
>>>> Hi everyone,
>>>>
>>>> I'm wondering if anyone can help shed some light on why I'm getting
>>>> such poor ICA decomposition and what to do about it. I've tried a number of
>>>> pipelines and methods, and each one is about this bad (The link below has
>>>> pictures of the scalp maps from two files below). I'm using a 128 channel
>>>> EGI system. Here is my pipeline:
>>>>
>>>> 1. Import, low pass filter at 40Hz, resample to 250Hz, high pass filter
>>>> at 1Hz
>>>> 2. Remove bad channels and interpolate, then re-reference to average ref
>>>> 3. Epoch to 1s epochs, remove bad epochs using joint probability
>>>> 4. run AMICA using PCA keep to reduce components to 128-#chans
>>>> interpolated
>>>> 5. Load raw data, filter same as above, resample, remove bad chans,
>>>> interpolate, re-reference
>>>> 6. Apply ICA weights to continuous, pre-processed data
>>>> 7. Do component rejection
>>>>
>>>> What am I missing? Does anyone see any glaring errors here? My data are
>>>> a bit on the noisy side, and while I do capture things like blinks and
>>>> cardiac artifacts pretty clearly, I get the artifacts loading on a lot of
>>>> components, and I'm not getting many clear brain components. I got one
>>>> suggestion to reduce the number of components down to something like 64,
>>>> but this article by Fiorenzo, Delorme, Makeig recommends against that.
>>>>
>>>> Any ideas?
>>>>
>>>> Thanks,
>>>> Kelly
>>>>
>>>> Scalp maps:
>>>> https://georgetown.box.com/s/1dv1n5fhv1uqgn1qc59lmssnh1387sud
>>>>
>>>> On Thu, Oct 11, 2018 at 11:10 AM Kelly Michaelis <kcmichaelis at gmail.com>
>>>> wrote:
>>>>
>>>>> Hi everyone,
>>>>>
>>>>> I'm wondering if anyone can help shed some light on why I'm getting
>>>>> such poor ICA decomposition and what to do about it. I've tried a number of
>>>>> pipelines and methods, and each one is about this bad (I've attached
>>>>> pictures of the scalp maps from two files below). I'm using a 128 channel
>>>>> EGI system. Here is my pipeline:
>>>>>
>>>>> 1. Import, low pass filter at 40Hz, resample to 250Hz, high pass
>>>>> filter at 1Hz
>>>>> 2. Remove bad channels and interpolate, then re-reference to average
>>>>> ref
>>>>> 3. Epoch to 1s epochs, remove bad epochs using joint probability
>>>>> 4. run AMICA using PCA keep to reduce components to 128-#chans
>>>>> interpolated
>>>>> 5. Load raw data, filter same as above, resample, remove bad chans,
>>>>> interpolate, re-reference
>>>>> 6. Apply ICA weights to continuous, pre-processed data
>>>>> 7. Do component rejection
>>>>>
>>>>> What am I missing? Does anyone see any glaring errors here? My data
>>>>> are a bit on the noisy side, and while I do capture things like blinks and
>>>>> cardiac artifacts pretty clearly, I get the artifacts loading on a lot of
>>>>> components, and I'm not getting many clear brain components. I got one
>>>>> suggestion to reduce the number of components down to something like 64,
>>>>> but this article by Fiorenzo, Delorme, Makeig recommends against that.
>>>>>
>>>>> Any ideas?
>>>>>
>>>>> Thanks,
>>>>> Kelly
>>>>>
>>>>>
>>>>> _______________________________________________
>>>> 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
>
>
>
> --
> Scott Makeig, Research Scientist and Director, Swartz Center for
> Computational Neuroscience, Institute for Neural Computation, University of
> California San Diego, La Jolla CA 92093-0961, http://sccn.ucsd.edu/~scott
> _______________________________________________
> 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/20181109/7de804cb/attachment.html>