[FieldTrip] DICS question

Schoffelen, J.M. (Jan Mathijs) jan.schoffelen at donders.ru.nl
Mon Jun 22 09:52:03 CEST 2020


Dear Craig,

Let me forward this e-mail to the discussion list, in order to widen the scope to all of Donostia and beyond.

Indeed, the estimation of spectrally resolved GC requires a full bandwidth signal. I have not extensively played with the different options here, but the scenario that you describe, i.e. compute a DICS spatial filter on a narrow bandwidth-of-interest signal, and subjecting the broadband sensor signals to this spatial filter, will, in my opinion, not be really optimal. The reason for this is that the beamformer then can only suppress interference from that particular bandwidth. Signal components that lie outside this range might still leak into the estimated time courses, which may lead to distortion and invalid GC estimates.

So, if you would be using an LCMV beamformer, I’d always recommend a broadband spatial filter, optionally on mildly highpass-filtered data. Then, the risk could be that the 1/f characteristic of the signals results in a covariance matrix whose signal components are dominated by low-frequency stuff. One could then worry about suboptimal spatial filtering of the higher frequency components. This is probably true, and I have worried about this myself, but in my experience this is the least of the worries one should have. (as a side note: I have played around with spectral whitening of the data prior to the beamforming. The reasoning behind this is that the individual variance components in the data are more evenly balanced across the frequency range, but the long story short is that this does not seem to influence the final GC estimate).

A larger worry would be the fact that source-estimated GC spectra very often have a canonical shape, unless you use extremely well-defined data such a visual attention experiment that induces strong gamma band oscillations. If you are into applying connectivity analysis in a more ethereal context, such as a cognitive or language experiment, you will see that source-level GC spectra will be always dominated by the low-frequency components (i.e. most of the time you’ll find a large peak at low frequencies). This of course can reflect genuine neuronal coupling, but my suspicion is that this is most of the time due to ‘weak signal asymmetries’, in other words due to asymmetries in SNR and location specific leakage of power. This can be further investigated using time-flipping and an appropriate statistical evaluation. For inspiration, you can of course always consult my 2017 PNAS paper :).

Another worry of course would be the fact that you’d be working with Elekta data :).

Practically, what I typically do, is to compute the broadband spatial filter with LCMV, and apply these filters (as is, or dimensionality reduced after a parcellation step) to the sensor level cross-spectral density data (from DC to Nyquist). The spectral factorization algorithm should be applied to the source-level data, which in my experience is most stable if you do this pairwise. I know that this is suboptimal in the context of an interpretation of direct versus indirect interactions, but I have always found this an argument that is part of a non-discussion at this stage of the analysis. One should first trust the estimates overall, before a discussion about direct versus indirect coupling becomes relevant.

Best wishes,
JM








On 19 Jun 2020, at 18:06, Craig Richter <craiggrichter at gmail.com<mailto:craiggrichter at gmail.com>> wrote:

Hey JM,

How’s it going? Covid treating you ok? Been a bit crazy here, but normalcy is resuming.

I have a question about DICS for you… :-/... Anyhow, so for coherence at a given frequency DICS is superior, since the spatial-filter is specific to that frequency, which is about the same as a building a noise-covariance matrix from a narrow-band filtered signal? But to do source-level GC, we need the broadband time-series, so we’re forced to use the unfiltered data and LCMV. But… what if we know the frequency we want to analyse, so we build the spatial filter on this frequency, or narrow-band filtered data, then compute virtual time series from the raw data, and compute GC. Then we simply look at our frequency of interest in the spectral results. Would this potentially generate a better GC estimate at the frequency of interest, or would it just mess everything up?

PS - I’ve cc’ed some other interested folks.

Thanks!

C.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20200622/1a7e6793/attachment.htm>


More information about the fieldtrip mailing list