[FieldTrip] normalizing EEG power
f.roux at bcbl.eu
Tue Jan 15 09:36:19 CET 2013
if the shapes of your power spectra look qualitatively similar and are
simply shifted up or down, you could normalize the power on a
frequency by frequenyc basis by computing z-scores
for each individual within their respective gender
z(i,f) = (x(i,f) - x_males(f))./sd_males(f); % for males
z(i,f) = (x(i,f) - x_females(f))./sd_females(f); % for females
with x(i,f) being the individual power spectrum, x_males(f) and x_females(f)
being the group mean power spectra for males and females, and sd_males(f) and sd_females(f)
being the standard deviations of the power spectra for each group.
At least this should help to get rid of gender specific differences and you wouldn't
have to normalize your values using total power etc.
----- Original Message -----
From: "Ingrid Nieuwenhuis" <inieuwenhuis at berkeley.edu>
To: "FieldTrip discussion list" <fieldtrip at science.ru.nl>
Sent: Monday, January 14, 2013 10:46:21 PM
Subject: [FieldTrip] normalizing EEG power
This is more a data analysis question than FieldTrip question, but I
hope I may still pick your brains for this:
Does anyone know of a good method to normalize EEG power? In general,
EEG power is higher in female than male participants (probably due to
gender differences in skull thickness or other non-brain-physiology
related gender differences). When correlating EEG power to some
behavioral measure for instance, this can interfere. I know that people
often calculate relative power in these cases - dividing the power in
each frequency band by the total power -, but there are some obvious
problems with this measure: Total power both contains power due to
unspecific (and for instance gender specific) contributions, but also
power due to the very brain activity of interest. This brain activity of
interest can be very different depending on the current cognitive state.
Thus dividing by total power can result in:
1) 'mixing' in band specific brain results to all other bands
2) not getting rid of the unspecific (for instance gender related) power
In my data both 1 & 2 happen. The topo plots of the relative power look
horrible, some brain related power gets totally divided out (all becomes
quite uninterpretable), and still the female participants have more
power than the male in most bands.
I also tried normalizing by using only the power in higher frequencies
(45 - 57Hz), with the underlying idea that in these high frequencies
power is more due to unspecific noise and less to brain (so I divided by
power in 45 - 57Hz instead of total power). Now my topoplots look much
better (very much like the absolute power), while the non specific
gender related power difference is decreased. However not totally gone
yet. Also, this approach is tricky when subjects have high muscle tone /
muscle artifacts, because this high band does contain muscle related power.
So what I need is to get an reliable and stable estimate of the non
brain dependent power, and use that to normalize the power spectrum by.
Something I thought of was: getting rid of the power at frequencies
where there is a clear peak on top of the 1/f (in the spectrum), then
spline interpolate those frequencies back in, and use that as total
power? Anyone know of such an approach being used by anyone? Makes
sense? Maybe fitting the 1/f in some way (possibly after removing peaks
at specific freqs first). And which frequencies should be used for total
power? Include the lowest frequencies? The lowest possible frequency
contributes most to total power (due to 1/f drop off), but is also
estimated least well (due to having least amount of cycles). Does it
make sense to leave those low frequencies out?
I also prefer to do this on the whole (cleaned) data, to get a reliable
estimate, and then use the same value to normalize all data (so just
scale subjects differently in respect to each other). Because that's
another problem of normalizing by total power, the total power changes,
so you always divide by a different value. I'm working on sleep data,
and the powerspec changes a lot depending on sleep stage. Using one
value for all data is valid under the assumption that this gender
difference in power is stable over time, which I would assume it is. I
could for instance use the wake data to estimate the total power, and
normalize all sleep data by it?
Probably someone is going to suggest to go to source space: But I was
wondering if this would even help, because normally one does not have
different forward models for male and female participants, right? Also
this is a lot of work, I'd rather not... But is people have used this
and it worked without side effects, I'd love to hear the details!
So in summary:
1) are there other ways to normalize power except for using relative
power (dividing by total power)?
2) Is it valid (and wise) to leave out the power of the lowest freqs for
the total power?
3) Is it valid to use high freqs instead of total power?
4) is it valid to estimate high freq power (of total power) over all
data (or even other data in same run), and use that to normalize all
data with (so always divide by same value)?
5) refs or FT code welcome!
Ingrid Nieuwenhuis PhD
Sleep and Neuroimaging Laboratory
Department of Psychology
University of California, Berkeley
Tolman Hall, room 5305
fieldtrip mailing list
fieldtrip at donders.ru.nl
More information about the fieldtrip