[VBMEG-users:00123] Using VBMEG for classification purposes

[Giuseppe Lisi]

Dear VBMEG users,
I need to classify Left and Right tapping data. And I have many options
doing it with VBMEG. The most reasonable approaches that I tried are:

1 approach) train one VBfilter (Left VBfilt) with the Left tapping EEG
and the Left tapping fMRI and Area prior and then train another VBfilter
(Right VBfilt) with the Right tapping EEG and the Right tapping fMRI and
Area prior. And then apply both the filters to the Test EEG in order to
get the 2 X cortical current from which I extract some features. The
performance of this method is 80% which is lower than the 90% of using
EEG alone.

2 approach) merge the fMRI and Area priors of the Left and Right
tappings into one SPM map. Then train only one VBfilt using all the EEG
data and the merged fMRI and Area prior. In this way I apply only one
filter to the Left or Right test data in order to obtain the respective
cortical currents. This approach gives me 70% performance.

The problems of these approaches are:
1 approach) when I apply the Left tapping EEG data to the Right VBfilt I
obtain cortical currents which are not very different from the ones
obtained applying Right tapping EEG to the Rigth VBfilt, and vice versa.
Apparently the VBfilt influences the result very much making the EEG
input less important. Do you have any idea if this problem is solvable?
Is there a way to discover the probability that a VBfilt was trained
with a given EEG data in input?

2 approach) When the Left tapping EEG (or Right tapping EEG) is applied
to the single VBfilt, cortical currents in both Left and Right areas are
strong. While I would expect a more clear activation for the
contralateral area. Do you have any idea about this problem?

Do you have any suggestion in general about the usage of VBMEG in the
specific context of left and right tapping classification? How would you
do it?

I am sorry for the long email, but I am running out of ideas and I don't
know how to proceed.
I thought you might have some suggestions.
Thank you for your help.
Best,
Peppe

_________________________________________________________
VBMEG users mailing list
vbmeg...@lists.sourceforge.jp
Options:
http://lists.sourceforge.jp/mailman/listinfo/vbmeg-users
Archives:
http://groups.google.com/group/vbmeg-users

[taku-y]

Hi Peppe,

> 1 approach) when I apply the Left tapping EEG data to the Right VBfilt I
> obtain cortical currents which are not very different from the ones
> obtained applying Right tapping EEG to the Rigth VBfilt, and vice versa.
> Apparently the VBfilt influences the result very much making the EEG
> input less important. Do you have any idea if this problem is solvable?
> Is there a way to discover the probability that a VBfilt was trained
> with a given EEG data in input?

Although not implemented the current VBMEG, the model probability can be calculated once the prior current variance parameter alpha was obtained. Let denote the model estimated with Left tapping EEG as M_left, including alpha, and the model with the Right tapping EEG as M_right. Given a test EEG data B_test, the joint probability of the EEG signal and cortical currents for M_left and M_right are

P(B_test,J | M_left) = P(B_test | J)P0(J | M_left)
P(B_test,J | M_right) = P(B_test | J)P0(J | M_right)

Since both of P(B_test | J) and P0(J | M_left) (or P0(J | M_right)) are Gaussian, the marginal probabilities P(B_test | M_left) and P(B_test | M_right) can be obtained. You should choose one of the two models M_left or M_right based on these probabilities.

In the case of classification of the left and right tappings, you does not need to further calculation, because you may have already get a model, M_left or M_right, corresponding to the result of decoding. However, if there are four conditions such as left tapping/grasping and right tapping/grasping, combination of model selection written above and pattern classifier with the cortical current estimated with the selected model might work.

> 2 approach) When the Left tapping EEG (or Right tapping EEG) is applied
> to the single VBfilt, cortical currents in both Left and Right areas are
> strong. While I would expect a more clear activation for the
> contralateral area. Do you have any idea about this problem?

Unfortunately, this problem cannot be solved in VBMEG. In my experience, using a fMRI prior created with combining multiple conditions that activates different cortical region (I actually have used four visual quadrants), such false positives in current estimation were likely to occur. So I recommend the 1st approach.

Taku

---
Taku Yoshioka
tak...@atr.jp