Mining discriminant information through the spatial and frequency domains is a challenging problem in brain computer interfaces (BCI), especially when the experimental paradigm is not as well-known as motor imagery. We propose a new classification algorithm that extracts relevant oscillatory activities from ongoing electroencephalographic (EEG) signals without prior knowledge of the discriminant frequency bands. In this approach, a spatial filtering matrix transforms EEG signals into several components. In order to minimize the generalization error of a Bayesian classifier, the spatial filter is tuned using spectral density matrices. The algorithmic implementation of this method uses explicitly the cross-spectral properties of EEG data. Results on both realistic simulations of EEG data and real asynchronous BCI experiments show the interest of the method and its superiority over the common spatial pattern (CSP) algorithm.
[1]
Antoine Souloumiac,et al.
Jacobi Angles for Simultaneous Diagonalization
,
1996,
SIAM J. Matrix Anal. Appl..
[2]
G. Pfurtscheller,et al.
Designing optimal spatial filters for single-trial EEG classification in a movement task
,
1999,
Clinical Neurophysiology.
[3]
J. J. Ermer,et al.
Rapidly recomputable EEG forward models for realistic head shapes.
,
2001,
Physics in medicine and biology.
[4]
M. Rosenblatt.
Stationary sequences and random fields
,
1985
.
[5]
Shigeo Abe DrEng.
Pattern Classification
,
2001,
Springer London.