[FieldTrip] dipole time course vs. virtual sensor using SVD

[Mehmet-Akif Coskun]

Dear Fieldtrippers,


I have been running a simple analysis for two different subject groups on the dipole time course and a virtual sensor using Singular Value Decomposition. The analysis for the two signals yielded different results. I would like to get comments on this issue.


The computation of the two signals are as below:


dataset: I have 248 channels x 122 time points data (averaged over a number of trials). 


1) Virtual sensor using SVD: I isolated the data from 30-100ms (isolated_data: 248 x 22 timepoints) and apply SVD as [U,S,V]=svd(isolated_data). The first column of U (248x1) accounts for the largest proportion of variance and is a vector of weights assigned to the signal recorded from each of the 248 sensors. So i computed the virtual sensors as virt_sens=U(:,1)'*dataset which is a 1x122 signal.


2)Dipole time course: I fitted 1 dipole (ft_dipolefitting) to the data between 30 and 100ms, computed leadfields (ft_compute_leadfield) for the dipole and projected the pinv'ed leadfields on to the dataset to obtain the dipole time course.


What may cause the differences in the results for the two approaches? What kind of mathematical assumptions does SVD and dipole fitting does? what are the weakness or the strengths of the two approaches and which one has to be trusted more?


I will really appreciate any comments and suggestions.


Best regards,
Mehmet Coskun
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20110414/2de978c6/attachment-0001.html>