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 198

Fig. 9.13 Representation of
the energy of signal X m
obtained using a discrete
dyadic wavelet decomposition
as a function of frequency. To
calculate the slope criterion, a
simple regression is
performed ( dotted line ) on the
energies calculated for 4, 8,
and 16 Hz ( circles )

x 10 7

6

5

4

3

2

1

0

1/64 1/16 1/4 1 4 16 64 128

Fréquency (Hz)

Figure 9.14 gives a representation of the data matrix after a dimension reduction.
On the left, the data obtained after a discrete wavelet transform. There are still three
dimensions: one for the participants, one for the 15 frequencies, and one for the
electrodes. Compared to Fig. 9.9 , we switched between time (46,000 points) and
frequencies (15 scales). On the right, after the calculus of the slope coef cient, only

two dimensions are remaining: one for the participants and one for the electrodes.
To construct a model able to predict the alertness state, some usual classi

cation

tools (classi cation and regression trees or k nearest neighbors for example) will be

applied on this matrix in 2 dimensions.

Fig. 9.14 Representation of
the data matrix after a
dimension reduction. On the
left , the data obtained after a
discrete wavelet transform.
There are still three
dimensions: one for the
participants, one for the 15
frequencies, and one for the
electrodes. On the right , after
the calculus of the slope
coef cient, only two
dimensions are remaining:
one for the participants, one
for the electrodes