I am working on getting some source estimates from my resting state MEG data using the DICS beamformer. In order to do that, I am computing the CSD matrix and the DICS spatial filter. I read in a few papers that computed the DICS beamformer using only the gradiometers and I found some information that if we were to mix sensors, we would need to use a covariance matrix. Is this accurate? And if I were to use only the gradiometers, would it be better/faster to select the gradiometer data when computing the CSD matrix?
you are correct about mixing sensors. Magnetometers and gradiometers measure in different scales, hence numerically, one signal is several order of magnitudes larger than the other, biasing the DICS filters. To counteract this, you need to normalize the signals, which is typically done by whitening the signal using a ?noise? CSD matrix. The ?noise? CSD is computed using a stretch of signal that serves as a recording of the size of the signals during rest. I?m actually in the process of implementing this whitening operation, but it will be some time before I finish it and get it merged into MNE-Python. In the meanwhile, just using the gradiometers should give no problems. Theoretically, it does not matter whether you select these during CSD computation, forward computation or DICS filter computation. However, the slowest part of the pipeline would be the CSD computation, so restricting that computation to grads only would have the largest speedup effect.
Just a follow up question. When the whitening pipeline is implemented in mne, I would still like to use it in my resting state MEG data. What would be the "noise" CSD in resting data?
Also, in the future I would like to combine not only the sensors in MEG data, but EEG with MEG. Would the same apply here?
resting state data does not really have a baseline ?noise? period. You can try computing the CSD over the entire recording (use the `decimate` parameter!) and use that to whiten the signal, but I don?t know if that is accepted procedure in the literature. I?m not a resting-state researcher, so I don?t really know.
With combining EEG and MEG, the same principle holds. Whitening must be applied, but afterwards, all is good. It?s the same situation as with using MNE for source estimation.