Source estimation with few good electrodes after bad channel marking – effect on results?

Discussions
,
1

Hi Brainstorm team,

I’m running an experiment with two experimental conditions and computing source estimation using Minimum Norm Imaging (dSPM, unconstrained).

To compute the noise covariance and the sources, I included all the trials from both conditions together. However, I noticed that in the resulting shared source file, some subjects have as few as 10–20 good channels left.

This can be because in one or a few trials, I marked many electrodes as bad — and Brainstorm only uses the channels that are good in all selected trials for computing the sources?

My questions are:

  1. Does this small number of good channels affect the quality of the source estimation?
  2. What is the best practice to avoid this?

Thanks in advance for your help!

2

Yes, the source estimation improves (i.e. the source localization error decreases) with more electrodes. Though this absolute improvement is less significant for larger electrode numbers. Here few references on this:

  • Having more trials would help, so removing trials with bad channels would not be as impactful
    (It may not be useful in your current situation)

  • Are the artifacts in those bad channels something that can be alleviated with artifact removal algorithms such as ICA?

  • As per [1], small error in source localization can be achieved with a small number of electrodes by optimizing the electrode locations. However, to optimize, the ground-of-truth (or a strong assumption of the source origin) is needed, which in most experimental setups is not known.

3

Thank you for the clarification and the helpful references!

To deal with the issue of bad electrodes, here's what I did:

  • I computed the noise covariance based on the cleaned data, where I had removed the bad electrodes using the channel flags.
  • Then, since source estimation depends on the head model and noise covariance, I went to the common files > 3-shell sphere and selected Compute sources.
  • When the window popped up saying that some electrodes were marked as bad and will not be taken into account, I delete them so that all 64 electrodes would be included in the source estimation step.

My understanding is that since the Imaging Kernel is computed based on the head model and noise covariance (already cleaned), and the final source estimation uses the full sensor data, this approach would allow me to make use of all the electrode positions for localization while relying on clean covariance estimation.

Does this sound like a valid approach, or is there a risk that reintroducing the previously bad electrodes might reintroduce noise into the source results?