[Homer-users] HOMER question

[thuppert]

**************************************************
Original email:
  Well, we have the continuous record of the fNIRS signal of a group of
subjects while they perform a cognitive task lasting 20 seconds. We pretend
to get the OxyHb and DeoxyHb averages of this period of time, and compare
between different subgroups of subjects. So, after making one nirs file for
each subject, we select and load all subjects with the "Import data | to
current Session" option. After that, we set LPF to 0.8 and HPF to 0.1., and
click in "update all files". After that, we tick the "Use DPF correction"
and "Part. vol. correction" options in the advanced filtering options
(keeping the default values). Then we go to the "Averaging" tab, set the
"preTime" in 0, and the "postTime" in 20, and click on "calculate average
all". When is done, we uncheck the "display Average all" (if it is checked,
we have no figure in the upper side, and a blank figure in the lower side).
Then, we go through subject by subject rightclicking in the lower figure and
selecting the option "Export all channels to file" (in any moment we have to
check some channel in the probe scheme to avoid some errors). When it's
done, we open those files and average the HbO and HbR variables across time
for each subject and channel, and make between-subjects ANOVAs for each
channel (or group of channels). Are we doing it well? Can we interpret that
there is a different neural processing between subjects if we obtain
significant differences (assuming that confounding variables are controlled
*********************************************************
Response:

That sounds good for the analysis approach.  One detail is that the pre/post
time used to calculate the hemodynamic response should be longer then the
task.  I.e. for a 20-sec task, set the post-time to (i.e.) 30sec and verify
that the response returns to baseline  (maybe this is not actually
necessary).  Also, the HPF at 0.1 will remove any signals which are SLOWER
then 1/10sec.  If your task is 20-sec long, it may have slow temporal
components to it  (I don't know what the task is)... consider using a
smaller number (or even not using the HPF at all).  Otherwise, the averaging
protocols sound fine (although statistics are most correct when no filtering
is done at all).
The ANOVA tests between subjects is the way I would first approach it for a
publication.

You should probably do multi-way ANOVA where the subject variability is
considered an additional dimension  (you might be already doing this in the
analysis).  Possibly consider channels in the region of interest as another
degree of freedom as well.  The variability between subjects (or channels)
will make it harder to reach significance (i.e p<0.05).  If you are able to
meet this criterion- great.  If not, you may need to better control the
factors that are responsible for the  variability of the response amplitude
across subjects (i.e. partial volume/pathlength).

The partial volume/ pathlength factors in the Modified Beer Lambert Law
(i..e the calculation of HbO2 and HbR from optical density) are subject
dependent and are probably variable across the probe (depending on the
thickness of bone (etc) below the optodes).  There are several groups whom
have argued that this prevents meaningful comparisons of the amplitude of
signals across subjects and even across regions of the probe.  I would
agree.  For example, if you were to test the same subject several times, the
results may vary depending on how reproducibly you placed the optode probe.
However, I would argue that for a large enough sample size of subjects, this
averages out (and there is some support for this in the literature).  This
will be controlled for in the multi-way ANOVA.

 However, the result of this is that a) at the minimum this should be a fact
that you consider when you draw conclusions from such an ANOVA tests.  b)
consider experimental ways to control for this.

There aren't really a lot of experimental studies yet, which have addressed
inter-subject ANOVA tests with NIRS for this reason.

Possible examples of a way to control for variability in the partial
pathlength factors:

1)
    Run an experiment with a positive control task that both subject groups
respond to equally in addition to the task that you are testing for
differences in.  For       example,  both subject group A and B are known to
perform working memory task #1 the same (i.e. previous literature supports).
However, you expect (and wish to test) if group A is significantly different
then group B in the performance of task #2 (a different working memory
task).
            Run both group A and B while they perform both tasks.  Then in
the analysis perform your ANOVA or paired T-tests on Group A:(difference of
Task 2 to Task 1) verses Group B:(diff Task 2 to Task 1).  {i.e. normalize
all subjects to their task 1 response}.   This would allow you to normalize
out probe effects on amplitude and directly test if group A's response to
task 2 is different from group B's response.  Of course, this assumes that a
task can be found that both groups respond to equally.

2)
    Consider comparing the timing of the responses rather than their
amplitudes.  I'm not sure the best way to do this in practice.




Ted Huppert, M.Sc.

PhD Candidate
Harvard University:
Graduate Programs in Biophysics
Photon Migration Imaging lab
Massachusetts General Hospital
Tel: (617)726-9338

thuppert at nmr.mgh.harvard.edu



-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/homer-users/attachments/20060825/b1391274/attachment.html