[Eeglablist] ICA in short duration recordings

[Scott Makeig]

Paul -

For ICA decomposition is best to use as much data as you have - e.g, the
entire recording session containing the 'epochs' of interest - and to
retain the full sampling rate (ICA can use info in higher frequencies as
well). Of course, against this is the concern about stationarity - *of the
source model*, meaning source projection patterns (scalp maps) AND source
probability density functions (pdfs, component voltage histograms). Thus,
rejecting data containing signficant !#$? muscle/movement artifact is
important, as is considering rejecting data when the subject is in another
'state' (of consciousness, attention,...). In particular, it is inadvisable
to use the data interpolation features of ASR (UNLESS you absolutely need
near real-time processing), though with conservative cutoffs its data
rejection features can be useful.

A (currently) advanced option is to use (mm)AMICA decomposition in its
multi-model mode to automatically separate the data into 2 or more subsets
with distinct, though typically not hugely different, ICA decompositions.
There are 'Post-AMICA tools' available to assess which portions of the data
are segregated by mmAMICA, either across continuous time or within an
erpimage showing event-related model dynamics across event-related epochs
extracted from the data.

Another option, when the total number of available time points is
insufficient, is to reduce the number of channels decomposed . Nima built a
command line function (in the misc func EEGLAB folder) to select from an
electrode montage any given number of electrodes with max. evenly spaced
locations (for example, when asked giving a near-evenly spaced subset of 43
of the 64 electrodes). But by applying the thinking above you may be able
to avoid this for your data. [An advanced option here would be to select
and decompose many different channel subsets, then interpolate the
resulting IC maps and cluster on maps and time courses - I don't know
anyone who has tried this yet and would be interested to hear from anyone
who has].

I don't get a clear enough picture of your data recording to apply the
principles above to your data ... How would you do so?

Scott

On Thu, Jan 20, 2022 at 1:22 PM Beach, Paul <paul.anthony.beach at emory.edu>
wrote:

> Scott,
>
> Thanks for responding.
>
> It’s 64 channel EEG. If my ’time points per epoch’ is the same as ’number
> of time samples’, then after downsampling this result is ~47,000 for a
> typical subject with ~180s of recording. For 360s recording ’time points
> per epoch’ is ~84,000.
>
> So if I’m interpreting things correctly
> 3min recording: k ~ 11
> 5min recording: k ~ 20
>
> If I’m interpreting Makoto’s description of this question correctly ("minimum
> of (number of channel)^2 x 20 to 30 data points to perform ICA”), I
> theoretically need that ‘k’ to be between 20-30. So it seems as though a
> minimum of 5 min or so would be necessary for ICA, which I think I’ve seen
> others describe.
>
> If that’s the case, better to go off just ASR, I suppose, given most
> subjects have about 3min of recording, correct?
>
>
>
>
> Paul Beach D.O., Ph.D.
> Movement Disorders Fellow, PGY6
> Emory University School of Medicine
>
>
> On Jan 20, 2022, at 11:58 AM, Scott Makeig <smakeig at gmail.com> wrote:
>
> 
> Paul -
>
> You omit a crucial variable ... how many channels are you decomposing? The
> complexity of the ICA model of the data is proportional to the square of
> the number of channels, since ICA learns a square (channels x channels)
> unmixing matrix. As the data length becomes shorter, the number of channels
> that can be well-decomposed becomes smaller.  What is k for you? where
>
> k = number_of_time_samples / number_of_channels^2
>
> Scott
>
> On Thu, Jan 20, 2022 at 11:54 AM Beach, Paul via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> EEGLAB team – and other EEG’ers,
>>
>> I am currently working with a number of datasets with relatively short
>> durations – 3-5 minutes, generally. My goal is to examine endogenous ERPs
>> from heartbeat processing (heartbeat evoked potential) in controls and a
>> clinical group. I utilize a processing protocol whose artifact rejection
>> mainly involves ASR, epoching to heartbeats, and then use of ICA to reject
>> components at the single subject level.
>>
>> I almost always run into an issue whereby I only get a few components in
>> ICA (usually out of 50-60) that are *clearly* neural in origin and maybe a
>> few where it’s a more difficult call. My understanding is that, in general,
>> the majority of components will not be necessary to keep/won’t be neural
>> (particularly ICs after the first 20 or so). However, talking with other
>> EEG’ers I find that they will typically have 10 or so.
>>
>> Notwithstanding that EEGLAB does recommend not rejecting components until
>> group analysis, I wanted to query thoughts on the ‘appropriateness’ of even
>> using ICA on recordings of such relatively short duration. I’ve read the
>> brief discussion in Makoto’s processing page on this and ASR does a great
>> job removing lots of things like eye blinks and even clear cardiac field
>> artifact
>>
>> SO: should one just rely on ASR and perhaps use ICA for eye components
>> that are missed (and keep everything else)? Should one avoid ICA completely
>> for shorter duration recordings? Should I continue what I’m already doing?
>> Some other middle ground? Or should I just forget about single subject
>> level IC rejection completely and rely on its use in group-level analysis?
>>
>> Thanks for your time and thoughts.
>> --
>> Paul Beach DO, PhD
>> PGY6, Movement Disorder Fellow
>> Department of Neurology
>> Emory University School of Medicine
>>
>>
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__nam11.safelinks.protection.outlook.com_-3Furl-3Dhttp-253A-252F-252Fsccn.ucsd.edu-252Feeglab-252Feeglabmail.html-26data-3D04-257C01-257Cpaul.anthony.beach-2540emory.edu-257C180f439303e6499882f408d9dc360bb4-257Ce004fb9cb0a4424fbcd0322606d5df38-257C0-257C0-257C637782946950139475-257CUnknown-257CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0-253D-257C3000-26sdata-3DnhsKmSkwX8I3OMux-252B-252FFgBK9ul-252F5LjaJwoGV-252Fa2gdgXk-253D-26reserved-3D0&d=DwIFaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=42pfKCwA6GD7WxLsb3JhDl4vnvazNnqKkE-yalxbj6vNVWMLJt8w_C6slAQ9kWOQ&s=WQG1gYQy-76tRACibRjTR4wAZkDEA81_nEoqrYw_F6Y&e= >
>> 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 92093-0559, http://sccn.ucsd.edu/~scott
> <https://urldefense.proofpoint.com/v2/url?u=https-3A__nam11.safelinks.protection.outlook.com_-3Furl-3Dhttp-253A-252F-252Fsccn.ucsd.edu-252F-7Escott-26data-3D04-257C01-257Cpaul.anthony.beach-2540emory.edu-257C180f439303e6499882f408d9dc360bb4-257Ce004fb9cb0a4424fbcd0322606d5df38-257C0-257C0-257C637782946950295709-257CUnknown-257CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0-253D-257C3000-26sdata-3D3zxo2Gvgnndy7V47VaF9sn-252B7K7hbRMlIhbj9ElOeNHQ-253D-26reserved-3D0&d=DwIFaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=42pfKCwA6GD7WxLsb3JhDl4vnvazNnqKkE-yalxbj6vNVWMLJt8w_C6slAQ9kWOQ&s=NgkNEgd-wPS0tCPH2_OOsP_1NNDvk2z_3y8Y2a3iDAw&e= >
>
>

-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott