Best options for broadband gamma localization

💬 Support & Discussions
1

Dear MNE experts,

We try to compare high frequency broadband power (70-140 Hz) in V1 (primary visual) during ISIs that follow different kinds of visual stimulation.

We use 306 channels Elekta system (only 204 gradiometers in this case) and apply
LCMV beamformer. We tried different combinations of options:

  • common filter vs condition-specific filters;
  • with/without empty room covariance matrix;
  • with/without NAI normalization.

We found that the using vs not using NAI (with condition-specific filters) had strongest effect on the results. Since NAI is computed on the basis of the smallest eigenvalue of the cross-spectral density matrix, we explored these eigenvalues.

We found that the smallest eigenvalues strongly positively correlate with periodic alpha and especially beta power at the source level. (Alpha and beta power was measured in the primary visual cortex using another method: sLoreta). Does it mean that the smallest eigenvalue of the cross-spectral density matrix is not a good noise estimate and can bias the results?

Does anyone have recommendations for localization of the absolute (we have no baseline) broad-band gamma power (~70 – 140 Hz) ?

Best regards,

Elena

1 Like
2

The post below might be helpful for you. The data is from a MEGIN system and the analysis is of alpha/gamma during visual processing. The jupyter notebook from the post goes through a full beamformer analysis with justified parameters.

https://github.com/Neuronal-Oscillations/FLUX/blob/main/MNEPython/DICS.ipynb

–Jeff

2 Likes
4

Hello Jeff,

Thank you for giving a link to this great notebook!

We try to apply this code to our data with gamma and alpha, but observed that when a DICS filter is created in the alpha analysis, regularization is set to 0:
afilters = make_dics(epochs.info, fwd, csd_acommon.mean() , noise_csd=csd_anoise.mean(), reg=0, pick_ori=‘max-power’, reduce_rank=True, real_filter=True, rank=rank, depth = 0)

…while in the gamma analysis, regularization is set to 0.05:
gfilters = make_dics(epochs.info, fwd, csd_gcommon.mean() , noise_csd=csd_gnoise.mean(), reg=0.05, pick_ori=‘max-power’, reduce_rank=True, real_filter=True, rank=rank, depth = 0)

In the alpha section on the notebook, there is a detailed explanation why regularization can be set to 0 (between ln [31] and ln [32]). And for me it seems that this explanation should work for all frequencies, not only alpha. So it’s not clear why not to set reg=0 for gamma, too? In the gamma section, regularization is just set to 0.05 without further comments… Maybe you have any ideas on this? Would be very thankful for any thoughts!

Viktoria

5

Yes I agree, I found these to be very helpful resources. Based on the text in the notebook - it seems like both regularizations are meant to be 0.05.

@olejen - would be able to confirm either way.

–Jeff

1 Like
6

I would actually think that reg=0 for both alpha and gamma - and that we should change the one for gamma. The reason being that we do the truncated psedo-inverse with rank k (thus implicitly regularizing). But let me check this. Ole

2 Likes