Online classification no more than 40%

[trinsma]

Hello,

We have been trying to set up a P300 speller for the past few weeks, but we seem to be unable to get the online classification performance higher than 40%. Usually, it picks the right column or row, but not both. In the offline analysis using P300Classifier we have achieved accuracies of 60% (for some subjects) up to 100% for one subject, which was not the case during their online session. Could there be a difference in how the offline analysis works versus the online classification due to settings that we have overlooked? We are not sure about whether technical settings like the SourceChGain have been set properly, so could it have something to do with that?

It would be nice if someone could take a look at one of our data files and the corresponding parameter file, which I will upload as soon as I get permission from the administrator.

We are using the following setup:

- 128 channel net
- 59 channels are included in the analysis
- Windows XP with BCI2000
- Mac OS 10.5 with NetStation
- AmpServerPro module
- Flatscreen LCD
- We are aware of the issues with LCD screens, but the CRT we tried was only capable of 60 Hz, which produced flickering.

Any help is greatly appreciated!

Regards,

Thijs

[mellinger]

Could there be a difference in how the offline analysis works versus the online classification due to settings that we have overlooked?

You may test this by specifying your original training data files as a training set, and your online files as a testing data set in P300Classifier, and using "Apply Feature Weights" after "Generate Weights". In the Details pane, you will then see how the classifier works with online data. If you get results very different from your actual online results, this might be due to settings in the online system. E.g., did you enable spatial filtering in the online system but not in P3Classifier? Also, P3Classifier only allows "raw" (i.e. none) and "CAR" as spatial filters, while there are more options in the online system.
Differences could also be due to bad generalization of the classifier, i.e. it performs well on its training data but poorly on unseen data. E.g., you might have a certain type of artifact in your training data but not in your online data. The trained classifier might be good at identifying that artifact instead of the actual P300 response, so it will perform poorly on your training data.

which I will upload as soon as I get permission from the administrator.

Uploading files is generally enabled for the board. What kind of error message do you get?

Juergen