Source estimation of time frequency

Discussions
1

Dear experts,

I have conducted time frequency analysis for single trials on the sensor level, and obtained significant differences in alpha band activity between the two conditions. Now, I want to do the source estimation for alpha band activity. I am not sure whether the following alternatives are feasible.

The first one:

Filter the original recordings (artifact free) into 8-13Hz for single trials；

Compute sources for each trial；

Average the sources for each subject under each condition；

Run a t-test. The potential differences of cortical activation between the two conditions indicates the sources of distinct alpha band activity.

The second one:

Compute sources on the recordings (artifact free) for single trials;

Conduct time frequency decomposition based on the sources (we plan to use the DK template) for single trials;

Average the time frequency for each subject under each condition；

Run a t-test for the source-based time frequency. The potential differences between the two conditions indicate the sources of distinct time frequency.

Also, I want to know if it’s reasonable to computed sources for time frequency based on averaged trials. The steps are following:

Compute sources on the ERF for each subject under each condition;

Conduct time frequency decomposition based on the sources;

Run a t-test for the source-based time frequency.

2

Both approaches are possible.
Ideally, both should converge to similar observations.
Note that in the first case, you compare amplitudes with positive and negative values, and in the second case, you compare power values.

The first one:
Filter the original recordings (artifact free) into 8-13Hz for single trials；

To avoid edge effects covering your entire epoch of interest : filter the continuous files instead of the single epochs.
The recordings from which you compute the noise covariance must be filtered in the same way as the recordings of interest.

Also, I want to know if it’s reasonable to computed sources for time frequency based on averaged trials.

No, this doesn't work:
https://neuroimage.usc.edu/brainstorm/Tutorials/TimeFrequency#Things_to_avoid

1 Like
3

Many thanks ! Your suggestions helped a lot.

4

Thanks for this useful question! I am trying to use both approaches for my study. I just want to confirm a few things:

For approach 1:
Once we average sources within each subject and condition, we need to normalize with respect to a baseline (z-score) followed by rectification, in order to conduct comparisons across subjects, is this correct? (since the sign is ambiguous across subjects)

Here, we plot amplitudes of sources, and we have no idea if these sources contributed ERD or ERS in the filtered band. Right?

For approach 2:
After averaging time-frequency for each subject within a condition, we need to baseline normalize using db, ERD/ERS or z-score, right?

Here, we can plot TF heatmaps, and understand whether the activity is ERD or ERS wrt to baseline. Right?