Dear brainstorm team and users,
I am working on a study using EEG and time-frequency analysis and I would like your opinion on a method I used for source reconstruction.

Without going into detail of my study design, I examined the differences between Condition 1 (C1) and 2 (C2) on EEG signal in different frequency bands, on 65 electrodes. I found a difference in the form of a statistically significant cluster in the alpha frequency band, during a specific time window and involving several electrodes. To better interpret my result, I tried to perform source reconstruction for this cluster.

Here are the steps I undertook with Brainstorm (v 2021-03-17) for this:

First, I imported my clean (ICA-preprocessed and artefacted trials rejected) EEG signal previously filtered in the alpha frequency band (time x chanel x trial) for both C1 and C2.

I computed my BEMhead model with OpenMEEG based on a template MRI normalized in the MNI system (MNI/Colin27; since I don’t have individual MRI) with the 65 electrodes.

For C1, I computed my covariance matrix using the baseline period of all corresponding trials. Then I kept in my dataset only the electrodes that are included in my sensor-level cluster and I computed a source reconstruction (inverse solution) per trial with the covariance matrice computed just before with all the electrodes. I used minimum norm imaging method with current density map measure and unconstrained dipole (in total 45000 dipoles). Finally, I averaged the sources across trials and then across the time window of my cluster.

I applied the same procedure for C2.

At the end, I contrasted the 2 conditions (C1-C2) without statistical test to avoid the problem of double dipping.

I obtained a result without any error and that seems to be consistent and should represent the source of the significant cluster I observed at the sensor level. Yet, I would like to check: Is it OK to “filter” (that is, to restrict) the source reconstruction computation based on some electrodes, and do you think the method I used is valid?

Your feedback on this question will be really appreciated. Thank you in advance for your help.

If you keep only the cluster of electrodes that shows the effect you want to see, the minimum norm solution will show this effect everywhere on the brain, as it reconstructs sources everywhere with whatever input it has. Extreme case, you keep only one electrode, you would obtain the same signal everywhere in your source space (with only a different scaling factor at each location).
This source estimation technique is supposed to be used only if you have a decent coverage of the head. You should use as many sensors as possible to increase the coverage and the quality of the reconstruction. In the case of dipole fitting from MEG recordings, it could make sense to restrict the number of sensors to a smaller subset, but not for whole-brain distributed source imaging from EEG.

So no, I would not consider what you describe here as a valid approach.

I would not recommend to run source localization on a restricted set of electrodes that showed some significant effect in sensor space. When you reduce the set of electrodes you increase spatial uncertainty.

I understand that the method I used is not valid. But is there a way to represent the source coming from the electrodes of the significant cluster at the sensor level ? My goal is to help the interpretation of my results by a source reconstruction without going through source level analyses where I lose power or where double dipping may be blamed on me.

You may have articles or tutorials that could help me in this way, or some suggestions of pipeline analyses ?

Thanks again for your help and your reactivity,
Théophile

When an effect is significant at the sensor level with all the proper corrections for multiple comparisons and the hypothesis is about the existence of an effect rather than about a specific area being involved, it could be acceptable to only report the peaks of a statistical map at the source level without requiring correction for multiple comparisons over the whole brain.

The more sensors, the better defined your cluster.

Note that switching to source space can hardly help you in you statistical analysis. It re-interprets the information carried in a few tens of signals as thousands of heavily correlated time series.