[Eeglablist] Different methods to automatically remove faulty channels

[Makoto Miyakoshi]

Dear Katia,

You found the hidden truth in the EEG data cleaning?
Welcome to the club.

Regarding EEGLAB's old epoch rejection functions, their performance is ok.
I recommend you apply
1. simple amplitude thresholding (+/- 1000 etc) to remove outliers.
2. After rejecting the epochs, improbability test again for the second
rejection.
so that your stat calculated is not affected by outliers.

> Which method should I trust?

If you make yourself comfortable with it, I still recommend clean_rawdata()
with ASR on the continuous data.
There are now many related publications.
https://sccn.ucsd.edu/wiki/Artifact_Subspace_Reconstruction_(ASR)#Reference_.2807.2F09.2F2020_update.29

Also, just recently I found that the ASR algorithm seems basically the same
as popular data imputation methods as as this one
https://urldefense.com/v3/__https://www.researchgate.net/publication/320015038_Missing_Data_Imputation_Toolbox_for_MATLA__;!!Mih3wA!XQ2QKItWaeupyKuQ4G_kEyLhS3nWWjrVXAEd4U75tX-N0srsUtKOk7AHlLq2Kc5VidJ8Hg$ 

Makoto

On Thu, Jul 9, 2020 at 10:24 AM Katarzyna Dudzikowska <
k.a.dudzikowska at gmail.com> wrote:

> Hi all,
>
> I am preprocessing EEG datasets for the first time ever and trying to
> figure out how to take care of any potential unsound channels. Just
> scrolling through the data none of the channels really jump out as bad, but
> that is probably down to my lack of expertise. So to get closer to some
> level of understanding I ran and took notes on all the methods of automatic
> rejection that I am aware of: using probability, kurtosis and spectrum as a
> measure, as well as channel rejection encompassed in clean_raw_data
> function.
> It took me a good moment, but I am just as confused as I was before,
> because there is no coherence whatsoever in the results of these four
> methods.
> There is some overlap between rejection based on Kurtosis and Probability
> (though Probability measure usually suggests to reject only one or two
> channels of several suggested through Kurtosis measure). Results of
> Spectrum measure are usually null, sometimes one of the channels picked up
> through other measures also gets labeled for rejection. The algorithm used
> by clean_raw_data always comes up with multiple channels for rejection,
> many more than any other method, but they virtually never overlap with
> channels labelled by other methods.
>
> In my naivete I kind of thought that there would be at least some
> agreement: if a channel is in some way faulty wouldn't it at least
> sometimes be captured by different methods? In a couple datasets I also
> tried to run automatic channel rejection after doing some artifact
> rejection with ASR and got completely different results to those from
> before artifact rejection which confused me too. I am applying the
> functions to continuous data, after downsampling (to 250 Hz), low pass
> filtering at 0.5Hz and high pass filtering at 45Hz. The channel locations
> are imported and right mastoid channel was added as a reference. Am I
> missing something?
>
> I am really curious why there is such a massive discrepancy between
> different methods. On a more practical side could you give me some advice
> on how to decide on the best method to choose (I have watched/read several
> tutorials, but I am still really confused). Which method should I trust? My
> end goal is to calculate frontal alpha asymmetry, in case that makes any
> difference: is any of the methods more suitable for frequency analysis?
>
> I would be really grateful for some advice.
>
> All the best,
> Katia
> _______________________________________________
> 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
>