I had a question about using beamforming methods on eeg task data( using stop task data) . I am using the DICS method given I am interested in time frequency responses. However, quite a few eeg files have noisy channels and I have removed them and averaged referenced. I was wondering if it is better to interpolate before referencing , or not interpolate at all ? What if I don’t interpolate as more than 1/3 of channels are noisy , and if it is worth interpolating at all ? what happens to DICS if I use fewer channels ?
Hi @apoorva6262
it’s not recommended to interpolate many channels as this leads to rank-deficient covariance matrices (or CSD matrices). This then poses a problem to the inversion step of the beamforming, especially if many channels are interpolated. The reason is that you are adding dimensions (channels) but not information (because the interpolated channels are a linear combination of already existing channels).
Generally, beamforming needs a higher number of channels to be estimated nicely. A good rule of thumb might be 60+ channels.
You can have a look if the noise might be suppressed by the beamformer, leaving the channels in. Otherwise I would recommend looking at dipole fitting or minimum norm estimation if you are left with too little channels or if they are not approx. evenly distributed over the head.
Some more background on the rank-deficient covariance matrices can be found here: A unified view on beamformers for M/EEG source reconstruction - ScienceDirect , section 3. Practical considerations and best practices in beamformer source reconstruction; 3.1. Estimation and inversion of covariance matrices.
HTH,
Britta
Thanks @britta-wstnr , I have 64 channel eeg, and I am interpolating about 2-3 channels. I was wondering if it is even worth interpolating them because those new channels are not adding in any new information ?
Hi @apoorva6262
for beamforming it indeed does not give any benefits to interpolate the channels (rather the opposite as it makes your data rank deficient). Of course you might still choose to interpolate them for any channel-level analyses you might be doing as well.
In any case, it can help to compare beamformer output to MNE output given the number of channels.
BTW, also note that beamforming needs a very precise forward model, especially with EEG
HTH,
Britta