Quantitative EEG-Based Brain-Computer Interface - Quantitative EEG Analysis Methods and Clinical Applications - page 213

The features extracted for classification were bandpass power of mu rhythms
on left and right primary motor areas (C3 and C4 electrodes). LDA was used to clas-
sify the bandpass power features on C3/C4 electrodes referenced to FCz [9]. A linear
classifier was defined by a normal vector w and an offset b as

( )

T

y = sign wx + b (8.18)

where x was the feature vector. The values of w and b were determined by Fisher
discriminant analysis (FDA). The three-class classification was solved by combining
three binary LDA discriminant functions:

[ ]

T

() () ()

x t = Pt Pt

C 3 C 4

(8.19)

( )

() ()

yt = sgn wx

T

t + b i , = − 13

i i i

where P C3 ( t ) and P C4 ( t ) are values of the average power in the nearest 1-second time
window on C3 and C4, respectively. Each LDA was trained to discriminate two dif-
ferent motor imagery states. The decision rules are listed in Table 8.1, in which six
combinations were designated to the three motor imagery states, respectively, with
two combinations not classified.
An adaptive approach was used to update the LDA classifiers trial by trial. The
initial normal vectors w i T of the classifiers were selected as [

1
0] (corresponding to the three LDA classifiers in Table 8.1) based on the ERD distri-
butions. They were expected to recognize the imagery states through extracting the
power changes of mu rhythms caused by contralateral distribution of ERD during
left- and right-hand imagery, but bilateral power equilibrium during foot imagery
over M1 areas [47, 48]. The initial b was set to zero.
When the number of samples reached five trials per class, the adaptive training
began. Three LDA classifiers were updated trial by trial, gradually improving the
generalization ability of the classifiers along with the increase of the training sam-
ples. This kind of gradual updating of classifiers provided a chance for initial user
brain training and system calibration in an online BCI.
Figure 8.11 shows the probability that three progress bars won during an online
feedback session. In each motor imagery task, the progress bar that has the maxi-

+ 1 − 1], [0 − 1], and [ −

Table 8.1 Decision Rules for Classifying the Three
Motor Imagery States Through Combining the Three
LDA Classifiers
Left Versus
Right

Left Versus
Foot

Right Versus
Foot

Decision

+ 1 + 1 − 1 Left

Left

+ 1 + 1 + 1

Right

− 1 + 1 + 1

Right

− 1 − 1 + 1

Foot

+ 1 − 1 − 1

− 1 − 1 − 1 Foot

+ 1 − 1 + 1 None