search

mne-tools / mne-python

MNE: Magnetoencephalography (MEG) and Electroencephalography (EEG) in Python
https://mne.tools
BSD 3-Clause "New" or "Revised" License
2.66k stars 1.31k forks source link

Re-reference methods about sEEG data #8481

Open zhengliuer opened 3 years ago

zhengliuer commented 3 years ago

Describe the new feature or enhancement In sEEG data analysis, reference is not alike in EEG. Many methods are used to reference sEEG data, linear and non-linear, such as Common Average Reference(CAR), Gray-white Matter Reference(GWR), Electrode Shaft Reference(ESR), Bipolar Reference, Laplacian Reference and Monopolar Reference. Describe your proposed implementation I read the paper to know the reference in sEEG analysis. Here is the link:https://www.sciencedirect.com/science/article/abs/pii/S1053811918307183

larsoner commented 3 years ago

Which do you need for your analysis? ESR? Is this just CAR for subsets of electrodes?

It would help if you could guide us specifically about what's needed here. Papers are helpful but it takes time to read them -- if others already have expertise and knowledge things will proceed faster. Maybe @adam2392 could comment as well.

larsoner commented 3 years ago

And just a general plan of action, we should not implement every method available. We should implement the ones will benefit a large proportion of users. So if 90% of sEEG folks use CAR and monopolar (which we already support) and ESR (which we do not), we should just implement ESR and not worry about GWR and Laplacian.

zhengliuer commented 3 years ago

Actually, I don't have some specific methods. I am asked to implement them all by my mentor. But CAR is not recommended in most cases in sEEG analysis. While monopolar is widely used, especially in the hospital, well, at least in China. And ESR and Laplacian is quite the same for they just delete different channels' data in the same shaft. And as to GWR, it is a little complex for you need to know what channels are in the white or gray zones for sEEG is very Individually.

adam2392 commented 3 years ago

My impression is that for sEEG:

are supported in mne. These other references are possible but need to happen outside mne. E.g. for gray/white, one must have a list of the wm channels and then gm channels to then use set_eeg_reference(wm_chs).

ESR would be desirable altho I've never used it (becuz mne didn't have it before :p)... This can be done in sEEG by taking the unique letters for each electrode and then referencing the average signal of those with the channels that are part of that electrode. Although a user can do that themselves, I see this as being implementable in the mne-python codebase!

WDYT @larsoner

zhengliuer commented 3 years ago

And a little tip on ESR, some channels' name is very tricky for some shaft named A, and some named A' To compute the Laplacian, each channel was re-referenced to the mean value of its two adjacent contacts along the electrode shaft. This seems similar to Bipolar. I said it wrong before.

adam2392 commented 3 years ago

Laplacian can generalize based on a distance metric defined, but... idk if that's desirable.

@BarryLiu-97 do you want to implement ESR in a PR?

larsoner commented 3 years ago

This can be done in sEEG by taking the unique letters for each electrode and then referencing the average signal of those with the channels that are part of that electrode.

And a little tip on ESR, some channels' name is very tricky for some shaft named A, and some named A'

Agreed, I would not write channel-letter-based referencing stuff. I would make people give the channel names they want to use, it allows them to deal with their own naming schemes.

Really the question is what API do we need to support ESR then I think. Currently the relevant API is:

raw.set_eeg_reference(ref_channels='average', ch_type='auto', ...)

If we were to add the argument picks, which would supersede / ignore ch_type (it actually could replace ch_type entirely but probably isn't worth the deprecation to do so?), then you could for example do:

refs = [ch_name for ch_name in raw.ch_names if ch_name.endswith("A'")]
picks = refs + [ch_name for ch_name in raw.ch_names if ch_name.endswith("A")
raw.set_eeg_reference(refs, picks=picks)

This would reference all picks channels (i.e., those ending in A or A') based on the average of all channels in refs (i.e., those ending in A').

Laplacian can generalize based on a distance metric defined, but... idk if that's desirable.

Let's wait to do this until we see if set_bipolar_reference is insufficient for most users