[FieldTrip] ICA decomposition & backprojection

Frederic Roux f.roux at bcbl.eu
Tue Nov 20 17:10:24 CET 2012

Dear Patrick,

I just stumbled across your post from a while ago.
Since I am far from being an expert, take anything I say
with a grain of salt.

If you reject components in your raw data at MEG sensor level with ICA
what will happen is that you will end up with a rank deficient covariance matrix.

Assuming that you you reject 3 components, your matrix will have a rank of N-3, 
instead of N (because you just lost 3 linear independent components), where N is the
number of linearly independent time series in your channel x time matrix.

So instead of having a Nx3 leadfield matrix, you end up with a N-3x3 leadfield matrix.

However, since the inverse problem is anyways heavily underdetermined (ie the number
of sensors is way smaller than the possible number of dipoles), this shouldn't
matter much as long as you keep the number of components down to a few (eg ECG and EOG).

For example rejecting n = 100 components would results in a rank N-100,
which I guess could cause some troubles.

So my advice would be to keep the number of rejected IC components to a strict
minimum (eg 1 or 2 ECG components and 1 or 2 EOG components).

You should also compare the results of your beamformer with and without
ICA cleaning to see any changes related to the ICA application.

Best wishes,

----- Original Message -----
From: "Patrick Jung" <patrick.jung at esi-frankfurt.de>
To: fieldtrip at science.ru.nl
Sent: Wednesday, November 14, 2012 6:24:32 PM
Subject: [FieldTrip] ICA decomposition & backprojection

Dear Fieldtrippers, 

if I do ICA in my preprocessing to reject EOG and ECG signals, the new MEG guidelines say that this might alter later source reconstruction and connectivity analyses (which I aim to do). 

So, worried about this, I wonder if applying ft_rejectcomponent on my raw dataset 

1.) automatically modifies the leadfield matrix? 

2.) the unmixing (and mixing?) matrix is saved in the new data structure somewhere (to evtl make use of it at later analysis steps)? 

3.) automatically stores the information how many and what ICA components were subtracted from the data in my new data structure. 

Would you generally discourage to apply ICA in the preprocessing when planning source reconstruction and connectivity analysis? 

It’s a complicated topic for me as a clinician so I hope there is no mixing in my questions ;-) 

Many thanks for your answers, 

fieldtrip mailing list
fieldtrip at donders.ru.nl

More information about the fieldtrip mailing list