[FieldTrip] Optimal MEG data structure for LDA

[Sander van Bree (PGR)]

Dear FieldTrippers,

I've got an MEG dataset from an Elekta Neuromag system, and I would like to apply an LDA multivariate pattern classifier on it. In the paradigm, participants first encode and later retrieve associations between words and images. My training set is two seconds of the encoding part of the experiment, and my testing set is two seconds of the retrieval part. For both sets, the classifier learns the difference between images of objects (class 1) and scenes (class 2). We predict that the memory dynamics are orchestrated by theta oscillations in medial temporal regions, which we predict cause fluctuations in classification performance.

How would you go about feeding this data into the classifier? Should I (1) use the magnetometers, (2) gradiometers, or (3) gradiometers combined using ft_combineplanar? As a bonus question, how would you Z-score the data across the 3 dimensions (trials x channels x time)? Is it wise to normalize across trials for each timepoint separately or would that destroy the oscillatory structure in the data?

Thank you beforehand and best wishes,
Sander van Bree
PhD Student
University of Glasgow
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20210119/5fddf017/attachment.htm>