Combine gradiometer channels

[AJQuinn]

This is a work-in-progress start on general functions to combine meg gradiometer channels.

Idea is to allow users to combine gradiometer channels in the standard data instances, and eventually to replace the channel combining code used in mne/viz with this more general solution.

Would be great to get some feedback etc - any/all opinions welcome! @drammock @larsoner @agramfort

Some points for next steps...

• modify coil_type and kind to something indicating a combined gradiometer
• modify channel units, depending on combination method
• currently _combine_meg_grads assumes inst only contains 'grads'
• outsource return instance creation in _combine_meg_grads to specific class methods
• need to handle noise_cov? projections?
• addtional combination methods? eg PCA/SVD

Reference issue

Relevant: #1280

What does this implement/fix?

I've implemented a private function in mne.channels.layout and linked it up to epochs.as_type as a first use-case

[larsoner]

Some preliminary comments/ideas!

[larsoner]

kind should still be MEG I think. For coil_type I think we'd maybe want to add a 3016 here:

https://github.com/mne-tools/mne-python/blob/main/mne/io/constants.py#L929

We'll need to make a PR to fif-constants for this. We can do this later, but to make mne/io/tests/test_constants.py happy you can for fif-constants, make a new branch with a commit that adds it, push to your fork, and tell test_constants to use your commit like here:

https://github.com/mne-tools/mne-python/pull/11064/files#diff-d545e84fd98e16f320b83c21e8c35a67ed64d378ded8fb35d3734e9f2196b53eR23-R25

[AJQuinn]

Thanks - I think this is done in 5fa25d1 and 1e87305 referencing this fork of fiff-constants https://github.com/AJQuinn/fiff-constants/tree/grad_combined

Not sure how critical the naming choice is?

[larsoner]

I don't think the units should change, right?

[AJQuinn]

Makes sense for mean and rms - I'll ignore this for now.

[larsoner]

Yes -- both should raise an error as you shouldn't need these. Same if you try to make_forward_solution... but we can tackle these separately because there will be a lot of these "don't do this" paths

[larsoner]

To start you could do inst.copy().pick(...), but in theory this should already live in as_type in some reasonable way

[larsoner]

This should also already live in as_type. It would be good to reuse the code, deduplicating / making shared private functions as necessary

[AJQuinn]

Ok - have some more updates which can be explored with this code.

import mne

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = (sample_data_folder / 'MEG' / 'sample' /
                        'sample_audvis_filt-0-40_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file)
events = mne.find_events(raw, stim_channel='STI 014')

event_dict = {'auditory/left': 1, 'auditory/right': 2, 'visual/left': 3,
              'visual/right': 4, 'smiley': 5, 'buttonpress': 32}
reject_criteria = dict(grad=4000e-13)

epochs = mne.Epochs(raw, events, event_id=event_dict, tmin=-0.2, tmax=0.5,
                    reject=reject_criteria, preload=True, picks='grad')
cmb_epochs = epochs.as_type(ch_type='combined_grad', mode='rms')

epochs['visual'].average().plot_joint()
cmb_epochs['visual'].average().plot_joint()

Producing this standard ERF....

and this combined grads ERF

[AJQuinn]

and a bit cleaner if you remember to baseline correct....

cmb_epochs['visual'].average().apply_baseline().plot_joint()

[agramfort]

just a quick suggestion

[agramfort]

you can do:

Suggested change

	for idx in range(picks.shape[0]):
	for pair in picks:

then do below

[agramfort]

Suggested change

	ch1 = inst.info['chs'][picks[idx, 0]]
	ch2 = inst.info['chs'][picks[idx, 1]]
	ch1 = inst.info['chs'][pair[0]]
	ch2 = inst.info['chs'][pair[1]]

[drammock]

and a bit cleaner if you remember to baseline correct....

something is off about the plot. Since RMS makes the signal all positive, it should be plotting with a one-sided colormap. In our plotting code this will be handled by a variable called norm usually.