Brain–Computer Interfaces - Index of

The Graz Brain-Computer Interface 87
selection algorithm frequently used in feature selection tasks in the Graz BCI is the
so-called distinction sensitive learning vector quantization (DSLVQ) [54], an extension
of the learning vector quantization method [19]. In contrast to many classical
feature selection methods such as SFFS or genetic algorithms, DSLVQ does not
evaluate the fitness of many candidate feature subsets.
5 Frequency Band and Electrode Selection
One aim of the Graz group is to create small, robust and inexpensive systems. The
crucial issue in this context is the number of electrodes. Generally, wet EEG electrodes
are used in BCI research. Wet sensors, however, require scalp preparation and
the use of electrolytic gels, which slows down the sensor placement process. There
is also the fact that EEG sensors need to be re-applied frequently. For practical issues
it is consequently advantageous to minimize the number of EEG sensors.
We studied the correlation between the number of EEG sensors and the
achievable classification accuracies. Thirty mastoid-referenced EEG channels were
recorded from 10 naive volunteers during cue-based left hand (LMI), right hand
(RMI) and foot MI (FMI). A running classifier with a sliding 1-s window was used
to calculate the classification accuracies [30]. An LDA classifier was used each
time to discriminate between two out of the three MI tasks. For reference, classification
accuracies were computed by applying the common spatial pattern (CSP)
method [13, 29, 57]. For comparison, LDAs were trained by individual band power
estimates extracted from single spatially filtered EEG channels. Bipolar derivations
were computed by subtracting the signals of two electrodes, and we derived orthogonal
source (Laplacian) derivations by subtracting the averaged signal of the four
nearest-neighboring electrodes from the electrode of interest [14]. Band power features
were calculated over the 1-s segment. Finally, the DSLVQ method was used
to identify the most important individual spectral components. For a more detailed
description see [58].
The results are summarized in Table 1. As expected, CSP achieved the best overall
performance because the computed values are based on the 30-channel EEG data.
Laplacian filters performed second best and bipolar derivations performed worst.
Of interest was that LMI versus RMI performed slightly worse than LMI versus
FMI and RMI versus FMI. Although the statistical analyses showed no significant
Table 1 Mean (median)±SD (standard deviation) LDA classification accuracies. The values for
Laplacian and bipolar derivations are based on single most discriminative band power feature; CSP
on the 2 most important spatial patterns (4 features) (modified from [58])
Spatial filter LMI vs. RMI LMI vs. FMI RMI vs. FMI
Bipolar 68.4 (67.8) ± 6.6 % 73.6 (74.6) ± 9.2 % 73.5 (74.9) ± 10.4 %
Laplacian 72.3 (73.5) ± 11.7 % 80.4 (83.2) ± 9.7 % 81.4 (82.8) ± 8.7 %
CSP 82.6 (82.6) ± 10.4 % 87.0 (87.1) ± 7.6 % 88.8 (87.7) ± 5.5 %