Page 1 of 1

about P300GUI

Posted: 03 Mar 2009, 00:57
by Yueqing Li
Hi,

I met some problems in building the classifier with P300GUI. I had 5 runs in the training session. I wanted to use four of them to build the classifier and the left to test the classifier. However, I can't build a good classifier with the training data. The accuracy is always very low in the testing run (less than 25%). This happened with several participants with motor disabilities (ALS and CP), but never appeared with healthy participants. However, from the accuracy of the training data, there is no obvious difference.

Is there anything special I should do in get the training data with participants with motor disabilities? Thanks a lot in advance.

P300

Posted: 03 Mar 2009, 06:42
by gschalk
Yueqing,

The accuracy on the training data is not a good indicator for performance. The P300 response will be smaller and more variable in people with disabilities, and thus performance will be reduced. What may help is to collect more training data (I don't know how many characters there were in total in the 5 training runs), so that the classifier is more stable. Other than that, you will have to optimize system parameters (e.g., flash intensity, duration, size of the matrix, etc.), to optimize results.

Gerv

Posted: 03 Mar 2009, 23:44
by Yueqing Li
Dr. Schalk:

Thanks for your reply. I used 10 characters in each run. So, there were totally 50 characters in the training session. Is that enough for participant with motor disabilities?

I have one question for testing the training classifier weight. What is a good classifier? It seems that the accuracy produced in the command window is the only criterion. Is there any exact standard, such as 90%, 100%? Thanks.

ALS patients

Posted: 04 Mar 2009, 08:53
by gschalk
Yueqing,

You seem to be having enough training data.

In general, only very few BCI labs in the world have been working with ALS patients. My colleagues have quite some experience and also nice success with this population, but it is definitely a challenge. You will need to simply work with the population and learn what combination of BCI parameters, distance to screen, matrix size, etc., optimize the results.

Sorry I can't give you a direct recipe for success.


Gerv