bci2000.org BBS

Hi,
I have some questions/confusions in mind please help. Apologies for excess of questions.

1) Does somebody know, what Mathematical algorithms/methods and theory are used to generate Feature plots, topographis and waveforms?

2) In feature plots for p300 potential (time Vs Channels) we are required to look for high concentration values of r(squared). Also in mu Rhythm...Does some body know the theoratical background of r(squared)?.

3) Regarding BCI2000 filters pipe...In linear classifier, feature weights are generated using P300Classfier for P300 based BCIs. How can we generate feature weights for mu rhythms?

4) Also, linear classifier uses Step wise linear discriminant analysis for classification purposes. Does some body know about any other classification algorithms that has been used in BCI2000 filter pipe?

Please share any information. Many Thanks.

ad 1)
The methods used differ between P300 and mu rhythm data.

For P300 data, data are separated into epochs beginning at stimulus presentation, and extending 500 to 600 ms from there. These epochs are then averaged; this is done separately for attended and non-attended stimuli. Finally, the two average waveforms are subtracted from each to obtain data for topographies. In waveform displays, both waveforms are displayed.

For mu rhyhtm data, amplitude spectra are estimated using Burg's MEM algorithm (as described, e.g., in Numerical Recipes in C, 2nd edition, ch 13.6 and 13.7, which is available online at http://www.nrbook.com/a/bookcpdf.php). Spectra are then averaged separately for task conditions, and subtracted to obtain topography data.

ad 2)
r^2 is the squared correlation between data and task condition. Another name for it is "coefficient of determination". Its value is between 0 and 1, and is a measure for how strongly the amplitude at a given frequency or time varies with the task condition (i.e. type of stimulus, or movement imagery). For further information, please see http://www.google.com/search?q=coeffici ... ermination.

ad 3)
For the mu rhythm, you typically use a Laplacian or Common Average spatial filter, which will reasonably extract the mu rhythm signal. Then, simple weight values such as +1 or -1 will be sufficient in the linear classifier. For a procedure how to set up mu rhythm experiments, please see the tutorial on mu rhythm experiments at http://www.bci2000.org/wiki/index.php/U ... I_Tutorial.

ad 4)
The BCI2000 LinearClassifier filter may use weights computed by any optimization method, e.g. LDA. However, there are no tools available for other methods yet.

--Juergen