I am working with a MEG dataset recorded using a 4D-Neuroimaging system with 246 magnetometers (2 bad channels). During preprocessing, I filtered and demeaned the data, and I used regression instead of ICA to remove ocular artifacts, as I intended to use a beamformer for source reconstruction.
However, when I compute the data covariance for the beamformer, the resulting covariance matrix looks unusual as you can see in the figure attached.
Could you help me understand what might be causing this issue and guide me on how to troubleshoot it to pinpoint the root of the problem?
@emarca — it is difficult to know exactly why your data covariance looks this way.
The 4D channel naming/numbering system starts at the top-center of the head and spirals outward and down (inferior temporal lobe channels have the highest numbers). Based on that, you have high covariance at the top of the head. I am assuming a supine data collection because 4D seated is uncomfortable.
Does the underlying data look good? The artifact could be coming from outside of the room and poorly cleaned by the ref chans. It also could be coming from the sensor electronics before they are converted to optical outputs. The electronics are housed in that direction (positive Z axis). There could be some faulty electronics causing noise in the ADC housing that are seen maximally at the top of the head. Also, there are something like 10-20 data acquisition cards - that each handle a certain number of channels. A card can go bad and transmit intermittent noise on the connected channels, making most of the data look normal but then a brief period of highly covariant artifact. A lot of possibilities to produce the result you have.
I would do an ICA on the data and scroll through the time series to see if this helps identify anything. You would likely an ICA with a spatial map that has artifact in that region.
Also - are you dropping bad segments from your data before doing the covariance. You can have a brief artifact that completely dominates the covariance. If you chunk your data (make_fixed_length_epochs) and threshold your epochs and then redo you covariance. See if this helps.
