MRG: Add interpolation for NIRS signals

[rob-luke]

Reference issue

Ongoing fNIRS improvements #7057

What does this implement/fix?

Adds the most basic type of "fixing" of bad channels, simply replaces them by the closest good channel.

Additional information

Maybe interpolation isn't the right word for what I wrote, but it seemed inline with previous functions. And I would like to expand to something smarter like actual interpolation

[larsoner]

Is there an established correct way to do this?

If not, I wonder if the spherical spline interpolation used in EEG would be good enough. Maybe the thing to do is to get a clean dataset that's good for testing, e.g., with five sensors arranged in an X pattern (maybe you already have this?). Then you could try interpolating the middle sensor just by taking the mean of the other four, or by using the edges of a Delaunay triangulation, or by spherical spline interpolation (what we use for EEG) and see which one explains the most variance of the original sensor data. If they are all close then we can just use the EEG code path.

[rob-luke]

Is there an established correct way to do this?

Absolutely not. You are massively overestimating the state of the field. In most of what I read people discard channels with bad scalp coupling index then do not discuss the impact of this. From talking to people, most studies create a grand average with different number of averages per channel (depending on how many channels were discarded).

What prompted me to write this code is that in the group analysis MNE will take only the channels that have no individual recording marked as bad. It is more common to have poor channels in NIRS than EEG, as the optodes don't have that nice pin like ending to get through peoples hair that EEG has. So when analysing a group of 10 recordings, you end up with no channels left.

If not, I wonder if the spherical spline interpolation used in EEG would be good enough.

Wonderful idea. I was actually hoping you would come up with a better solution than mine, and that's why I included the method parameter in the function definition. I can take a recording like this in the next few weeks.

This code to find the closest channel will also be useful when we write the short channel correction code, which in the basic approaches uses the closest short channel to each long channel to remove systemic responses.

@larsoner so do we keep this closest approach and then add a spline interpolation method later? Or hold off on this until I have time to collect the data and look at the spline method?

[larsoner]

so do we keep this closest approach and then add a spline interpolation method later? Or hold off on this until I have time to collect the data and look at the spline method?

Let's wait until 0.20 is merged released to merge any PR that implements a method, then we have ~6 mo to decide what the default should be without needing a deprecation cycle. I'm then okay with merging nearest to start and keep PRs manageable.

[rob-luke]

Great. Sounds like a plan. I will keep working on this and ping you when it needs another look. Thanks

[codecov]

Codecov Report

Merging #7428 into master will increase coverage by 0.00%.
The diff coverage is 93.87%.

@@           Coverage Diff           @@
##           master    #7428   +/-   ##
=======================================
  Coverage   90.12%   90.12%           
=======================================
  Files         452      452           
  Lines       82800    82846   +46     
  Branches    13083    13093   +10     
=======================================
+ Hits        74626    74669   +43     
- Misses       5349     5350    +1     
- Partials     2825     2827    +2     

[rob-luke]

@larsoner can you take another look at merging this now that 0.2 is released. Thanks

[rob-luke]

@larsoner could you please review again. I believe I have made the changes you requested. And the CI failure seems unrelated

[larsoner]

Also needs latest.inc update but you might want to wait a few minutes until other PRs are merged so you can rebase first

Also feel free to change the variable name as well as the string if you want (I only pushed a commit to change the string), and feel free to force push over my change if you want, too

[rob-luke]

@larsoner my attempt at git magic has removed your commit!!! 😢

The code change you made is still there, but you credit via commit history has disappeared.

[larsoner]

No problem about the lost commit.

Just a couple more small things I noticed, don't need to be done here if you want to move on

[rob-luke]

Thanks I will fix these here

[rob-luke]

@larsoner and @agramfort could you please review. I believe the failing test is unrelated.

[rob-luke]

Thanks for the feedback @agramfort

[larsoner]

Thanks @rob-luke