[Homer-users] HomER: more questions about filter

[thuppert]

I recieved a few more questions on signal processing in HomER, so I'm
sharing my answers to the list server.  If anyone wants to add more, we
encourage such discussion:



%****************************************
  1.Is it appropriate to choose the propriate LPF to remove heart
rate(around 1 Hz) and breath rate( around 0.3Hz)? E.g. LPF=0.1, far lower
both of the rates.
     I think PCA is a good way to remove physiological noise from the data,
however, the baseline data is necessary. If the baseline data is
unavailable, whether is proper LPF another choice to remove these noise?


	-The low-pass (and high-pass) filters in HomER do a good job of removing
the cardiac signal (at ~1hz) and for low-frequency drifts like slow movement
or blood pressure oscillations.  I will typically use a low-pass of between
.5-.8 hz and a high-pass of ~1/30-1/60 hz.  I usually pick a LPF and use the
same one for all subjects.    Breathing rate is hard to remove with
band-pass filters for the reasons in question 2.




  2. I notice the signal quality can be improved by lowering the value of
LPF. In your old user guide, you used  0.5Hz LFP. If 0.1Hz is used, the
quality will be better, the time course will be smoother. However, is it
O.K.? Is the filtered signal credible? Maybe the channel should be exclude
according to the cardiac cycle, you mentioned in the previous e-mail.


	-Although band-pass filtering is good at removing physiology, you are also
filtering the stimulus response.  The stimulus response is generally
slow-ish, so a LPF of .5 is probably okay in general.  The frequency
spectrum of the response is probably in the range of hundredths to tenths of
hertz.  Of course it depends on what the task is.  For example, if the
person did 2sec finger-tapping, then filtering with a LPF of 0.1 would
definantly affect the calculated response!  For a long breath-hold, you have
to be carefull not to use too high of a filter for the HPF, these are very
slow changes.

When you choice the LP/HP filters, you really need to think about what the
expected response is.  I try to use the least amount of filtering as
possible! I'd rather error on the side of noisier data, then to have
over-filtered and effected the response in some way.

 Often you can get rid of physiology noise by having a lot of repetions of
the stimulus (I like event-related stimulus designs for this reason).
Alternatively, if you have auxillary recorded measures of physiology (i.e.
you record the cardiac cycle with a pulse-oximeter or a respiration belt for
breathing), you can incorporate this information into the analysis by
including these time-courses as regressors in linear modeling
(deconvolution).  This is a feature in the new version of HomER (which we
are about to release- hopefully within the week).  There is an example of
this in sample data that will be provided in that release as well.  I'll
send an email when it goes on the web.







  3. Is PCA filter  a necessary step for data process? In some
articles[1],only  the band pass filter was chosed, in others, only PCA was
done[2]. If I do data process, I will choose both of them. Band pass filter
can remove baseline drift and probably improve the SNR. PCA can remove
motion artifact. How do you think about it?

	-There are 2 types of PCA filtering that you can use in HomER.
		1) Motion correction-  this uses the spatial covariance in the data to
remove motion artifacts.  Motion will 		effect all channels and all
wavelengths (i.e. a sudden, sharp change in intensity all over the data).

		2) Baseline components- this uses the spatial covariance of a baseline
(rest) run to calculate the components 		that approximate physiology.

			Ref: Zhang Y, Brooks DH, Franceschini MA and Boas DA.
Eigenvector-based spatial filtering for reduction of 			physiological
interference in diffuse optical imaging. Journal Biomed Optics 10(1):
011014. 2005.


	As with band-pass filtering, PCA needs to be used sparingly!  It removes
motion (or physiology), but if over-used it 	can start to remove some of
the stimulus response as well.  The strongest eigen-vectors (components) are
those that have alot of spatial covariance.  Since motion or physiology are
(generally) global changes (affecting the entire probe), the PCA filtering
works to remove them.  The more components, the more of the "details" of the
data that get removed.  Think of it analogous to a discete Fouier transform.
If I include the entire continium of frequencies, then I can reproduce any
signal as a sum of the components (sines and cosines).  If I include only a
subset of the frequencies (i.e. only those really low frequencies), then I
can only model slow osccilatory data.  PCA works the same way as a high-pass
filter, only its not sines and cosines, but the priniple components of the
data.  I can completly model the data if I use ALL components.  If I only
use the first components, I can only model the highly spatially covarient
parts of the data.  Or, in the case of filtering, if I use everything EXCEPT
the first components, then I can remove the highly spatially covarient
aspects of the data.  The more components I filter out, the more "details"
of the data are lost.


I typically use both PCA and band-pass filtering.  As you noted, they are
good for different things.  The physiology PCA can remove things like
breathing or Meyer waves, which contain frequencies (~1/10-1/3 hz) that I
would be affraid to try to remove with a band-pass... but for cardiac or
slow drifts, its easier to use the band-pass filters.







 4. Wilcox mentioned further process should be done if motion artifacts were
detected, lthough PCA can remove some motion artifact[2]. PCA can't remove
all motion artifacts, can it? In my data, it seems true. Some short-term
motion artifact still existed after PCA filtering. If so, how can we judge
the motion artifact? Is there some more clear criterions?

	-As mentioned above, PCA removes components that are spatially covarient.
It works best on motion IF the motion affected all channels and wavelengths!
If not, its not going to do much and should not be used.  It certainly
cannot remove all motion- especially localized motion (i.e. if the probe is
really large or flexible).  Also, PCA doesn't work if there are many noisy
channels in the data.  Pure noise has NO spatial covarience.  So a single
really noisy channel (or wavelength) can affect how well the PCA works.
Remove noisy channels from the analysis first.  This can be done in
pre-processing or in HomER by clicking on the probe display untill the
source-detector line is dotted (This data is no longer used in the PCA).
[That's certainly true in the up-coming release of the program... the old
version had problems with that].




[2].Wilcox T, Bortfeld H, Woods R, Wruck E and Boas DA. Using near-infrared
spectroscopy to assess neural activation during object processing in
infants. Journal Biomed Optics 10(1): 011010. 2005.





The last word on signal-processing is that there is no specific recipe I can
prescribe to deal with all types of noise and artifacts.  Some tools work
better in some situations.  NIRs is quite different from fMRI analysis, in
that fMRI tends to be instrumentational noise limited... instrument noise is
white and it can be filtered; motion is easier to deal with because its a
physical body moving through space and you can "see" the edges.  NIRS has
very different noise sources- its physiological.  As the field develops,
we'll solve some of these (hopefully).


Hope this helps-



Ted Huppert, M.Sc.

PhD student-Harvard Univ.
Dept of Biophysics
Photon Migration Imaging lab
Mass General Hospital/CNY

Tele: (617)726-1223
Cell: (617) 869-1205

thuppert at nmr.mgh.harvard.edu