CAR in Spatial Filtering

[aloplop]

Hi again,

Firstly, I would like to wish you a great year 2009.

I would like to ask you one more question about spatial filtering.

I have performed some sessions using these electrodes: C4 F4 T8 P4 Cz,
what seems a Large Laplacian configuration.

However, when I configured BCI2000´s Spatial Filter Block I selected
the CAR method to take just the channel C4 as output, because the
Sparse Matrix didn´t work in an old version I had. (Now it´s fixed).

Some tests have been done using the CAR method. For example, using the AR Filter, I have obtained an average accuracy of 80 % using one bin with 3Hz BinWidth in the classifier and using as application the Cursor Feedback task.

The question is:
do you think not using the Sparse Matrix with Large Laplacian
coefficients have affected too much in my tests ?? or using
CAR is right although I have only used 5 electrodes??

Thanks again.

[gschalk]

Aloplop,

As you point out, The channels you selected correspond to a Large Laplacian configuration. By definition, the Common Average Reference subtracts the average of all electrodes from the signal electrode. A Large Laplacian reference subtracts the average of the next nearest neighbors from the signal electrode. In your case, this is almost the same, except that for the CAR, because you are averaging over all electrodes (including the signal electrode), and because you only have 5 electrodes, you are subtracting a non-negligible portion of the signal from the signal electrode. Thus, a Large Laplacian may be somewhat better. Also, you can easily implement a Large Laplacian with the Full Matrix. The SpatialFilter would be something like

1 -1 -1 -1 -1

assuming you transmit C4 as the first channel in the TransmitChList and the others in channels 2-5.

I hope this helps.

Gerv

[gschalk]

Ooops,

I just realized that it should be

+1 -0.25 -0.25 -0.25 -0.25

and not
1 -1 -1 -1 -1

Gerv

[aloplop]

Hi,

thanks Gerv for your quick response in these special days. I´ll do some
tests using the Sparse Matrix when I return back from holidays.

I forgot to say that using that configuration and selecting CAR
filtering in the Offline Analysis Tool I had obtained an average
r^2 of about 0.24 so I thought it was OK and I continued using it in some
more sessions.

Thanks again and Happy New Year !