Feature Extraction and Classification of EEG Signals. The Use of a Genetic Algorithm for an Application on Alertness Prediction - Guide to Brain-Computer Music Interfacing - page 207

9.6.3 Results

The algorithm, programmed using Matlab, is run 100 times for each evaluation
method with 300 parents and 150 children. The training and test sets are different
for two different runs. Figure 9.21 gives CCR values for each run of the GA with
CART (stars) and SVC (circles).
For each run, the algorithm is launched two times (one time with CART and the
other time with SVC). During a run, CART, and SVC use the same training and test sets
in order to obtain comparable results. The correct classi cation rate obtained by CART

(mean of 86.68 % and standard deviation of 1.87 %) exceed signi cantly (Mann -

-

p-value ¼ 5 : 57 10 14 ) those obtained by SVC (mean of
83.49 % and standard deviation of 2.37%), as mentioned in Table 9.3 . At the end of the
algorithm, some of the best genomes have the same evaluation (due to the low number
of individuals and the evaluation method). It is therefore necessary to choose a genome
(BEST) among those who have the same score. Thus, the best genomes at the end of
each run of the algorithm are stored. The genome that appears most often is considered
as the BEST for the evaluation method considered. The two BEST (for CART and
SVC) get a correct classi

Whitney paired test with a

cation rate equal to 89.33 %. For CART, the BEST is
obtained by performing regression between 1/8, 1/4, 2, 4, and 64 Hz on electrode F4
(right frontal area on Fig. 9.1 ). For SVC, the BEST is obtained from electrode F2 (right
frontal area) and the regression between 1/32, 1/16, 2, 4, 8, 64, and 128 Hz (see
Table 9.4 ). Frequencies chosen for these genomes aremore extensive than those used in
the preliminary study.

Fig. 9.21 Correct
classi cation rates calculated
with CART ( stars ) and SVC
( circles ) for each run of the
GA with 300 parents and 150
children

90

88

86

84

82

80

CART
SVC

78

76

0 10 20 30 40 50 60 70 80 90 100

Launch number

Table 9.3 CCR for the two
evaluation methods

Evaluation
methods

CCR
Mean Standard deviation

CART

86.68 1.87

SVC

83.49 2.37