What About Using Source Estimation to Interpolate Bad M/EEG Channels?

đź’¬ Support & Discussions
1

Hi MNE Community,

I hope this message finds you well!

I’ve been exploring methods for interpolating malfunctioning MEG channels and wanted to share an idea for feedback. Currently, MNE-Python provides two main approaches for interpolating bad channels:

  1. Spherical splines based on [Perrin et al., 1989]—primarily for EEG data.
  2. Field interpolation methods applicable to both MEG and EEG data.

While these methods are effective, they do not incorporate structural information about the human brain. This led me to wonder—could we leverage source-estimation-informed methods to improve interpolation? Note that this interpolation is designed specifically for channel-level analyses, ensuring consistent channel counts across runs and sessions.

The proposed workflow would look something like this:

  1. Mark bad channels in the sensor data.
  2. Perform source reconstruction with data from good sensors.
  3. Project the reconstructed source-level data back to the sensor space and interpolate the bad channels. (maybe via mne.simulation.simulate_raw?)

I’m curious to know if such an approach has been explored before or if there are any major pitfalls to consider. If successful, this method might not only facilitate bad channel interpolation but also have broader applications—for example, aligning sensor-level data across multiple scanning sessions where head movements might introduce misalignment.

Looking forward to hearing your thoughts and insights!

2

In principle, the idea is nice. However, the pitfall that I foresee is the signal’s dependency on the source localization method. Since the inverse problem gives way to an infinite amount of solutions there is no way to tell whether your interpolated signal is actually close to the truth. (for example, an L1 type of solution might give zero, while a minimum norm type of solution might give a signal)

The second issue is that by backprojecting the sources to the field you are in a way doing field interpolation because in both cases (EEG and MEG) the measured field is relatively smooth due to volume conduction.

It’s a nice thought experiment, but I think that it would not be feasible in practice.

3

Thank you for your detailed response! Your points about source localization dependency and field interpolation are very insightful!