Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 115

frequency disturbances, power supply interference, abrupt slopes, electrode pop ar-
tifacts. The eye movements and eye blink artifacts identification was based on the
correlation between EOG electrodes and/or certain EEG derivations. For muscle and
low frequency artifacts criteria based on spectral content of the signals were used.
The thresholds for each kind of artifact were set separately. Sensitivity of the system
to each type of artifact can be a priori adjusted or can be set by means of an au-
tomatic procedure for thresholds optimization based on minimalization of the cost
function. The agreement with the visual scoring was on the level of inter-expert
agreement. The MATLAB version of the above described software is available from
http://eeg.pl/Members/hubert/artefakt . It includes a complete source code
for all the computations, the user interface as well as input/otput routines for some
formats.
4.1.6 Analysis of continuous EEG signals
Since the first recording of EEG from the human scalp by Hans Berger in 1929,
clinicians and scientists have investigated EEG patterns by visual inspection of sig-
nals recorded on paper charts and, up to the present time, visual analysis is still in
use, but the signals are displayed on a computer screen. The complicated character
of EEG consisting of rhythmical components and transient structures has promoted
early attempts for an automatic EEG analysis. Berger assisted by [Dietch, 1932] ap-
plied Fourier analysis to EEG series. The first digital analysis of EEG was performed
in the time domain [Brazier and Casby, 1956]. The development of theoretical back-
ground of spectral analysis by [Blackman and Tukey, 1958] promoted the devel-
opment of signal analysis in frequency domain and in 1963 the first paper on digital
spectral analysis of EEG appeared [Walter and Adey, 1963]. With the development of
computer techniques and implementation of fast Fourier transform, spectral analysis
became a basic digital tool for EEG analysis, since the contribution of characteristic
rhythms in brain activity has an important impact on clinical diagnosis.
Due to its complexity, the EEG time series can be treated as a realization of a
stochastic process, and its statistical properties can be evaluated by typical meth-
ods based on the theory of stochastic signals. These methods include: probability
distributions and their moments (means, variances, higher order moments), corre-
lation functions, and spectra. Estimation of these observables is usually based on
the assumption of stationarity. While the EEG signals are ever changing, they can
be subdivided into quasi-stationary epochs when recorded under constant behav-
ioral conditions. On the basis of empirical observations and statistical analysis per-
formed by several authors, quasi-stationarity can be assumed for EEG epochs of 10 s
length approximately, measured under constant behavioral conditions [Niedermayer
and Lopes da Silva, 2004].
EEG signal can be analyzed in the time or frequency domain, and one or several
channels can be analyzed at a time. The applied methods involve spectral analysis
by Fourier transform (FT), autoregressive (AR) or MVAR parametric models, time-
frequency and time-scale methods (Wigner distributions, wavelets, matching pur-
suit). The most common methods used for post-processing include: cluster analysis,