[pymvpa] Event-related designs - Is the modeling of the HRF not recommened?

Tue Nov 18 18:37:12 UTC 2014

I thought I'd add my impressions on this topic. My opinion is that HRFs
aren't used as much as you might think, because they haven't proven
especially effective at improving decoding performance. We did a
parameter study a couple of years back that suggested this (see link
below), and conversations with people in the area (also those with cog.
neuroscience background for which using HRF would be a default) suggest
others have found something similar, but haven't reported it because the
results aren't very decisive either way.

That may be surprising, since it is clear that some function is
obscuring the underlying patterns of neural activity that we're trying
to learn about. But the basic assumptions of conventional HRF modelling
are that a) there is a standard invariant HRF, and b) the underlying
neural activity has a "square wave" temporal profile. The standard HRF
used in a lot of work originates in particular stimulus paradigms for
low-level vision, and since then it has become clear that HRFs vary from
location to location, and from person to person (though how is not
especially well understood). And also *if* we're talking about higher
cognition it would be surprising if neural processes had the square "all
or nothing" temporal profile.

So practically, I think you needn't worry too much about it. If you use
a HRF you'll probably get good classification performance and satisfy a
certain inclination of reviewer. Whereas if you want to get optimal
results you should do some search or other discovery procedure to
(approximately) learn the HRF from your data.

Brian

http://journal.frontiersin.org/Journal/10.3389/fninf.2012.00024
... and excerpt:
> discrepancies in the shape and timing of the BOLD responses across
> participants (Aguirre et al., 1998; Duann et al., 2002; Handwerker et
> al., 2004) and sessions (McGonigle et al., 2000; Smith et al., 2005). 

On Fri, 2014-11-14 at 07:12 +0000, Hanson, Gavin Keith wrote:
> Check out Misaki, Kim, Bandettini, & Kriegeskorte, 2010 (Comparison of
> multivariate classifiers and response normalizations for pattern
> information fMRI. NeuroImage). They mention that for block or slow ER
> designs, beta estimates and raw data (averaged across images in a
> block for a block design, or the single time point for slow ER design)
> basically amount to the same thing (p. 105). They ultimately recommend
> calculating t-statistics by dividing beta estimates by std. err
> estimates, because this tends to de-weight noisy voxels and improved
> classification in their analysis (this has been my experience as
> well).
> 
> In the case of a slow-ER design, there is not yet any clear consensus
> as to the necessity of a HRF model in the analysis. In my opinion it
> is recommended, because it can really help model out noise – most
> method papers I’ve read perform a GLM to model the HRF as a standard
> preprocessing step. It is certainly not less robust, and has been
> found to increase classification performance in at least the one
> paper. But there are a large number of ways to handle slow ER data
> (concatenate spatiotemporal data together into a single feature set,
> average raw data across the ISI, boxcar), and they each have
> advantages and disadvantages. It’s not bad to just take raw data. For
> example, if you have a large number of stimuli, then individually
> modeling the HRF to each stimulus can be computationally intensive,
> and is unnecessary because the large number of exemplars will serve to
> render the occasional noisy voxel irrelevant. 
> 
> Also, the use of beta maps from a GLM (often performed outside PyMVPA,
> in SPM, FSL or AFNI to take advantage of better HR functions and all
> that) is identical to using the output of the GLM step outlined in the
> PyMVPA tutorial, I believe.
> 
> As for ER design, the larger the ISI, the clearer the patterns of
> activity would be, as they are less influenced by pervious cortical
> activity. You rarely see a fast-ER design in MVPA for just that
> reason. 8-10 seconds is certainly common, and anything too much
> shorter would absolutely require a HRF modeling/GLM step do help
> separate temporally overlapping HR activity, as is mentioned in the
> event-related PyMVPA tutorial.
> 
> Hope this helps, and I’ll be interested to see what other people weigh
> in with,
> 
> - Gavin
> 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Gavin Hanson, B.S.
> Research Assistant
> Department of Psychology
> University of Kansas
> 1415 Jayhawk Blvd., 534 Fraser Hall 
> Lawrence, KS 66045
> 
> On Nov 13, 2014, at 5:40 PM, Michael Notter
> <michaelnotter at hotmail.com> wrote:
> 
> > Hi,
> > 
> > I'm used to the univariate approach of analysing neuroimaging data
> > and therefore always try to put an HRF through my data. The modeling
> > of an HRF also allows to analyse (rapid) event-related designs and
> > not just block designs. The PyMVPA tutorial shows
> > underhttp://www.pymvpa.org/tutorial_eventrelated.html also very
> > nicely how this can be done.
> > 
> > This leads me to my opinion that the modeling of such an HRF
> > regressors should be the way to go, if the stimuli are presented by
> > onself. In other words, if you have an event-related design with
> > clearly separated stimuli, the modeling of the HRF is the way to go,
> > correct?
> > 
> > But even though this is the case for many MVPA studies, most of them
> > chose to take another approach. They either only take volumes that
> > are expected to contain the peak of the BOLD response, average
> > volumes around the peak to one volume or use a GLM (in which the HRF
> > is modeled) to create beta maps on which they than apply the MVPA.
> > Even though they would have enough volumes (ISI of 6-12s) to model
> > the HRF correctly.
> > 
> > Is there a specific reason why this is the case? Is an MVPA analysis
> > that models the HRF or also considers the time dimension less robust
> > and not recommended? Has it been shown to be less accurate or is the
> > usage of only peak BOLD volumes the best way to go?
> > 
> > On this note: Is it better to present stimuli separated by an ISI of
> > at least 8s? What are pitfalls if I want to record an event-related
> > design.
> > 
> > Thanks,
> > Michael
> > 
> > _______________________________________________
> > Pkg-ExpPsy-PyMVPA mailing list
> > Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
> 
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa

-- 
Dr. Brian Murphy
Lecturer (Assistant Professor)
Knowledge & Data Engineering (EEECS)
Queen's University Belfast
brian.murphy at qub.ac.uk