feat: implement ASR

[Randomidous]

Describe the solution you'd like
Pick up implementation of ASR and continuous-DSS for inclusion in the package.

Original work:

• meegkit's ASR
• seminal ASR paper
• presentation on ASR
• matlab implementation

Adaptations:

• Adaptive ASR (deprecated)
• Riemannian ASR (deprecated)
• PyASR (closed)
• Juggler's ASR

Describe the solution you'd like
A robust, quick, and tested implementation of the Artifact Subspace Reconstruction method

[sappelhoff]

Adaptations:

There is also https://github.com/nbara/python-meegkit/blob/master/meegkit/asr.py of course :)

[BabaSanfour]

Thank you for opening the issue @Randomidous. I think continuous-DSS and ASR dev should be seperate. I would say continuous-DSS is more urgent to finalize the DSS suit and have it in the first "official" release of the package.

For ASR:

• There was discussions on MNE forum to include it mne, like we discussed on discord, I think this would be a nice place to have it (I think @larsoner and @drammock would agree).

• Implementation wise, I think asrpy is the python version of the EEGLAB ASR, however it is very data hungry. meegkit's version is a bit different, plus it have the translation of Rienmannian ASR from Matlab. However when we played with it, it took a lot of time to run. So I think the issue here is not the implementation itself, but having a robust, fast, and memory efficient implementation. @snesmaeili had already started exploring this; and we have the code in a private repo. What we can do on our end, is create a new branch here for ASR and share the code we have and then we move the discussion from there. You can play with python implementations of ASR to see where they fail, that might give you some ideas to explore.

• We are also interested in exploring Juggler's ASR.

[drammock]

xref to prior discussions:

• EEG noise reduction with ASR mne-python#7479
• Artifact Subspace Reconstruction mne-python#9302

Agree on doing continuous DSS in a separate PR from ASR. FWIW, I've recently spoken to @nbara who said he's happy to have meegkit code adapted into mne-denoise. But: we do of course want fast, memory-light implementations if at all possible, so if @BabaSanfour and @snesmaeili tested various existing implementations and are not satisfied with the performance, then no objection to waiting for a re-implementation.

[Randomidous]

Thanks for the input everyone :)

I have updated the description and will treat ASR and continuous DSS seperately. (How could I forget the meegkit ASR implementation? smh..)
@BabaSanfour, your suggestion sounds great, so let's roll with that!

[BabaSanfour]

@drammock I have tested both with large clinical datasets and either memory explodes or it takes forever to run; I think @snesmaeili tested it with mobile eeg data. So our idea was to find ways to make it better. We could eventually re-use some code from meegkit also if found useful or asrpy.

@Randomidous this week we are a bit busy, we will create the ASR branch and share the code sina worked on the next week and we can move it from there; meanwhile feel free to tackle the rest of the points we discussed on discord.