[Neuroimaging] Effects of motion outliers on HRF model (in sparse acquisition fMRI)

[Christopher J Markiewicz]

Hi all,

I apologize in advance because, as Pythonic as my pipeline is, my issue
here isn't really Python-related. However, the people on this list are
the most likely to have dealt with similar issues (of places I know to
look). If you'd rather I post on NeuroStars, I can, but I'm not sure how
much people are actually using that.

Anyway, my functional data comes from evenly-spaced, sparse acquisitions
(TA=2.25s, TR=3.375s), and I've used artdetect in nipype to tag motion
and intensity outliers. It's a fast, event-related design (one event
every 2 TRs).

In the past, my strategy has been to estimate HRF betas on the full
dataset, and then excluding motion outliers in analysis by removing any
event estimate that had an above-threshold contribution from an outlier
volume. That is, in an NxM design matrix estimating N events from M
scans, if scan j is an outlier, we exclude all events i such that
DM[i,j] > (e.g.) 10% of max(HRF).

Another strategy I'm looking into is to add nuisance regressors for
outlier volumes to the design matrix, and limiting bleed-over into
unrelated events. This is running into problems with "runs" of outliers,
which can leave some events with nothing by which to estimate or only
very small contributions from volumes that are going to be dominated by
other events. I could remove such events, entirely, but for various
reasons (mostly involving maintaining ordering so that off-by-one errors
don't slip into our analysis) I'd like to have some representation of
each event.

The best external-to-our-lab resource I could find was this Gabrieli Lab
protocol
(https://github.com/gablab/mindhive/wiki/Example-of-sparse-fMRI-data-analysis-using-BIPS),
which seems to indicate they've included the full dataset and noted
outliers after estimation.

Does anybody have any experience with outlier exclusion at or before HRF
estimation, or is this the current best practice?

Thanks,
-- 
Christopher J Markiewicz
Ph.D. Candidate, Quantitative Neuroscience Laboratory
Boston University