ft_appenddata and denoise_pca causing problems

jan-mathijs schoffelen jan.schoffelen at DONDERS.RU.NL
Mon Nov 8 09:07:37 CET 2010


Dear Luisa,

>
> Hi,
> I'm sorry if this has been discussed before and if it has, could you
> please direct me to the relevant emails? Thanks.
>
Don't worry. This has not yet been discussed here before, and it's an
interesting issue. Yet, probably most people wouldn't know what you
are talking about, because the function denoise_pca is not yet part of
the FieldTrip release version ;o). At the moment, it's only the CCNi
and the fcdc who have it available.

> I am encountering a problem when using ft_appenddata and
> denoise_pca. My MEG experiment  (4D) consists of 10 blocks per
> session, for each of which I have a separate data set.
>
> During preprocessing (after artifact rejection), I concatenate these
> data sets to find denoising weights per session using the function
> denoise_pca for 4D (I do this on shorter 400 ms epochs to save
> computational power). Afterwards, I apply these weights per block,
> downsample the data and concatenate the resulting data again in
> order to do an ICA per session to remove heart beat artifacts. I
> downsample because I use longer epochs to do that (1.5 sec).
> However, when I do an ICA on the denoised and concatenated session,
> I get lots of high variance noise components (first figure):
> Trying to find the root of this, I went back and did this analysis
> for one block only (denoising for one block, no concatenating) and
> the result looks much better (second figure):
> Then, as a test, I tried to concatenate two blocks, find the weights
> for those two concatenated blocks, apply them to the individual
> blocks, concatenate again and do the ICA on these two blocks and I
> get this (last figure):
> So, it seems like when denoising per session (the concatenated
> blocks), I introduce noise when I apply the pca weights to the
> individual blocks. Whereas the whole point of the exercise was to
> reduce noise by denoising per session rather than block.

Denoise_pca indeed tries to reduce the noise in the magnetometer data
by computing a set of balancing coefficients, which are used to
subtract a weighted combination of the signals measured at the
reference sensors from the signals measured at the magnetometer coils.
As such, it is assumed that the signals picked up by the references
reflect purely environmental noise. If there is sensor specific noise,
e.g. a jump in one of the references, the denoising algorithm will
actually inject noise into the data. I suspect that in some of the
blocks there is some unaccounted noise in your references, which both
deteriorates the results in the concatenated block case, and also
leads to suboptimal results when denoise_pca operates on data that
contains the 'noisy' block.

> Why is this? Am I doing something fundamentally wrong? I'm wondering
> because a colleague of mine has done an almost identical analysis
> step a while ago and she doesn't get problems with high variance
> noise components. I'm not sure whether the problem lies with
> append_data or denoise-pca, but I know that append_data has been
> changed recently, so this might be one possible source of the problem.

I don't think ft_appenddata would be the cause of this, because this
function only concatenates data structures.

As a diagnostic I would check the quality of the reference channel
data first, and see whether there are anomalies across blocks there.

Good luck,
Jan-Mathijs


Dr. J.M. (Jan-Mathijs) Schoffelen
Donders Institute for Brain, Cognition and Behaviour,
Centre for Cognitive Neuroimaging,
Radboud University Nijmegen, The Netherlands
J.Schoffelen at donders.ru.nl
Telephone: 0031-24-3614793

---------------------------------------------------------------------------
You are receiving this message because you are subscribed to
the  FieldTrip list. The aim of this list is to facilitate the discussion
between  users of the FieldTrip  toolbox, to share experiences
and to discuss  new ideas for MEG and EEG analysis.
See also http://listserv.surfnet.nl/archives/fieldtrip.html
and http://www.ru.nl/neuroimaging/fieldtrip.
---------------------------------------------------------------------------



More information about the fieldtrip mailing list