Information Technology Reference
In-Depth Information
developing ef
cient and accurate recognition systems that can automatically
interpret the mental state of a person through EEG measurements.
9.1.1 Electroencephalographic Signals and Previous Works
The electrical activity of the brain is divided into different oscillatory rhythms
characterized by their frequency bands. The main rhythms in ascending order of
frequency are delta (1
3.5 Hz),
theta (4
8 Hz), alpha (8
12 Hz), and beta
-
-
-
(12
30 Hz). Alpha waves are characteristic of a diffuse awake state for healthy
subjects and can be used to discern the normal awake and relaxed states, which is
the topic of this experimental study. The oscillatory alpha rhythm appears as
visually observable puffs on the electroencephalogram, especially over the occipital
brain areas at the back of the skull, but also under certain conditions in more frontal
recordings sites. The distribution of cortical electrical activity is taken into account
in the characterization of an oscillatory rhythm. This distribution can be compared
between studies reported in the literature through the use of a conventional elec-
trode placement; the international system de
-
ned by Jasper (
1958
) and shown in
Fig.
9.1
.
Furthermore, the brain electrical activity is non-stationary, as speci
ed in Subasi
et al. (
2005
); i.e., the frequency content of EEG signals is time varying. EEG
signals are almost always pre-treated before any analysis is performed. In most
cases, the Fourier transform or discrete wavelet decomposition (DWT) are used (see
Sect.
9.4.1
). In Subasi et al. (
2005
), authors use a DWT to pick out the wavelet sub-
band frequencies (alpha, delta, theta, and beta) and use it as an input to a neural
networks classi
cients of a DWT are used as
features to describe the EEG signal. These features are given as an input to an
arti
er. In Hazarika et al. (
1997
), coef
cial neural network.
In Ben Khalifa et al. (
2005
), the EEG signal is decomposed in 23 bands of 1 Hz
(from 1 to 23 Hz) and a short term fast Fourier transformation (STFFT) is used to
calculate the percentage of the power spectrum of each band. In Cecotti and Graeser
(
2008
), a Fourier transform is used between hidden layers of a convolutional neural
network to switch from the time domain to the frequency domain analysis in the
network.
To predict the state of alertness, the most common method is neural networks
[see for example Subasi et al. (
2005
) or Vuckovic et al. (
2002
)]. However, the
disadvantage of this approach is that it requires having a large set of test subjects
relative to the number of predictive variables. To avoid this problem, the authors of
Subasi et al. (
2005
) and Vuckovic et al. (
2002
) split their signal into several
segments of a few seconds, called
“
epochs.
”
Other approaches use different sta-
tistical methods. For example, Yeo et al. (
2009
) uses support vector machine,
Anderson and Sijercic (
1996
) uses autoregressive models (AR), and Obermaier
et al. (
2001
) use hidden Markov chains.
