averaging in sensor space

Ole Jensen ole.jensen at FCDONDERS.RU.NL
Fri Aug 12 12:13:44 CEST 2005


Regarding the discussion on planar gradients and averaging over subjects:

As Tom suggests it is always best to average over subjects in source
space. However, sometimes we do not have high enough signal-to-noise in
order to reliably identify the sources while we still observe consistent
effects at the sensor level. I think it is an important discussion since
it pertains to how we approach MEG data in general. Therefore I would
like to provide an argument for why I believe it is sensible to average
over subjects in sensor space:

Say that a dataset has been analyzed in the frequency domain and we have
representations of power for each sensor. I find it convenient to apply
the combined planar gradient for power representations since this will
tend to give the strongest power in sensors above the source. An axial
gradient representation will result in two regions of strong power at
each side of the source (i.e. where the fields exit and enter). We have
had quite good luck averaging combined planar gradient power
representations over subjects at the sensors level both with and without
realignment. The FieldTrip has statistical methods implemented for
dealing with the multiple comparison problem (multiple sensors). Also,
for instance the group in Tuebingen has several publications were power
(albeit using axial gradients) where they average over subjects at the
sensor level. The approach is quite similar to that applied by the
ERD/ERS community on EEG data. Note that differences in head size,
position etc might diminish a given effect - however, I do not believe
that averaging over subjects in sensor space will result in false
positives.

With respect to event related fields one should be very careful
averaging the axial gradient over subjects at the sensor level. The
orientation of sources from subject to subject might be different thus
partially canceling a given effect. One solution is to calculate the RMS
of the two orientations of the planar gradient and then average over
subjects. This has for instance been done convincingly by the BRU/LTL
Helsinki group in N400m studies. We have also good experiences with that
approach at the Donders. The main disadvantage of averaging combined
planar gradients of ERFs is that we loose information about source
orientation.

I would advice against doing source modeling on MEG data which has been
averaged over subjects (this is often done for EEG data). Here it is
appropriate to average either a current estimate of "beamformed" power
estimate in source space after realigning the source representations to
a standard brain.

We are often using the following approach

1) Calculate time-frequency representations (TFRs) of power for planar
gradients
2) Combined the planar gradient for each orientation
3) Average over subjects in sensors space
4) Use randomization statistics to identify clusters of difference
5) Use the beamformer to estimate power in source space for time and
freq tiles identified in 4) in individual subjects
6) Morph the results of beamformed power to a standard brain
7) Average morphed power representations in source space


Bests,

Ole






> Oh, as long as I'm here:
>
> Personally I wouldn't bother averaging across subjects, because even
> "aligned", averaging subjects together in sensor space is nearly
> meaningless.  (Though an interesting project, by analogy to Talairach
> or MNI space, is to "warp" the sensor space data into a common space,
> based on "anatomical" landmarks such as the peaks of the AEF and/or
> SEF.  I still wouldn't do that, though.)
>


--
Ole Jensen
Principal Investigator
F.C. Donders Centre for Cognitive Neuroimaging
P.O. Box 9101
NL-6500 HB Nijmegen
The Netherlands

Office  : +31 24 36 10884
MEG lab : +31 24 36 10988

Fax     : +31 24 36 10989

e-mail : ole.jensen at fcdonders.ru.nl
URL    : http://oase.uci.ru.nl/~olejen



More information about the fieldtrip mailing list