patterns is the decrease/increase of the power spectrum with respect to a baseline period [1]-[3] Additionally, our findings suggest that mental activity can be reliably detected by preprocessing the EEG signal with a SL estimator, even when using only nine electrodes. Though the SL estimator used is not absolutely accurate due to the reduced number of electrodes, the linear combination of channels that implements SL estimation is likely to apply a favorable transformation in the features'

more »

... . This transformation would simplify the distribution of patterns representing the EEG power distribution corresponding to different mental tasks. These findings are in line with those already obtained in a previous study [4]. In this study, we demonstrated that for the purpose of controlling by EEG the movement of a cursor on a computer screen, an interelectrode distance of 6 cm returned SL estimations more efficient than those obtained with an interelectrode distance of 3 cm. The difference in performance between the SL9 and SL methods (presented in Table II ) presumably indicates that the former was better matched to the topographical extent of the EEG control signal. This study suggests that reduction of the number of scalp electrodes is possible for BCI devices. We observed that in a rather large group of normal subjects (13), EEG activity related to the imagination of movements detected by two electrodes placed over C3 and C4 scalp positions can be discriminated by a simple quadratic classifier. The characteristic property of the best performing classifier tested in this study is its ability to take into account the covariance of features' vectors. The evidence of higher recognition scores of this classifier may suggest that the spectral features used are actually correlated. This correlation should be taken into account in the classifier design to increase substantially the recognition scores. The present result enlarges to a higher number of subject the observations made by Pfurtscheller et al. [2], also extending their results to normal subjects, whose cortical representation of the hand area, sampled from the two electrodes, is still intact and engaged in the network that allows the subject to control his muscles. REFERENCES [1] J.