Based on my experience, it is more efficient to use simple methods. The selection of time (interval after the stimuli occur - for the P300 e.g. 250 ms - 600 ms, since the P300 component latency may vary greatly) and channels (typically central and parietal electrodes on the surface of the head) is optimal. Given the sampling rate of e.g. 1 kHz, you would have 350 x features, which is - of course - unacceptable. To reduce the dimensionality, you can band-pass filter and sub-sample the data significantly, because outside of maybe 0.1 Hz - 12 Hz frequency range, there is very little useful data anyway.

Consider reading the following publication:

http://doc.ml.tu-berlin.de/bbci/publications/BlaLemTreHauMue10.pdf

Thanks a lot

There is also currently a growing interest in Covariance matrice (and Riemannian geometry) for BCI data and some variants are taylored for P300 data.

You can find more about this here :

http://arxiv.org/abs/1409.0107