Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 136

New approaches in anticipating seizures involve the so-called active algorithms in
contrast to passive algorithms described above. The active probe may be intermittent
photic stimulation. It was found [Kalitzin et al., 2002, Parra et al., 2003] that in EEG
and MEG signals, before the seizure onset, the phenomenon of phase clustering of
harmonically related frequency components occurs. It doesn't mean that the stimulus
provokes the seizure in any way. Based on this observation phase clustering index
(PCI), which reflects the ability of neuronal systems to generate epileptic discharges,
was introduced [Kalitzin et al., 2002]. PCI is zero if the phases are random, and tends
toward high values if the phases are strongly clustered. The development of seizures
was correlated with the value of PCI in the gamma band, therefore rPCI, PCI in
gamma band relative to PCIs in the other frequency bands, was introduced. Based on
rPCI the time horizon of the occurrence of seizures may be estimated [Kalitzin et al.,
2005], which opens possibilities of counteracting the seizures. Using computational
model it has been shown that active paradigms yield more information regarding the
transition to a seizure than a passive analysis [Suffczynski et al., 2008]. The method
seems to be promising, however it requires further experimental and model studies.
The evolution of the approach to seizure prediction leads to the conclusion that
before addressing the question whether a developed method might be sufficient for
clinical applications, it has to be tested if the performance of a new method is bet-
ter than chance. To this aim methods for statistical validation are needed. They may
rely on comparison with analytical results (derived from random or periodic predic-
tion schemes) [Winterhalder et al., 2003] or on simulations by Monte Carlo method,
e.g., [Andrzejak et al., 2003b, Kreuz et al., 2004, Jerger et al., 2005]. The statis-
tical significance of results may be tested, e.g., using the concept of seizure time
surrogates [Andrzejak et al., 2003a]. To this aim the seizure-onset times of the orig-
inal EEG recordings are replaced by the seizure onset times generated by random
shuffling of the original values. If a measure's predictive performance for the orig-
inal data is higher than the one for a certain quantile of the shuffled ones, then the
performance of this measure can be considered as significantly better than the ran-
dom prediction. In the study of Mormann et al. [Mormann et al., 2005] comprising
linear and non-linear methods the concept of surrogates introduced in [Andrzejak
et al., 2003a] was used. The authors found significant predictive performance for the
estimators of spatial synchronization, contrary to the univariate methods such as cor-
relation dimension, Lyapunov exponent, and the signal energy which were unable
to discriminate between pre-ictal and ictal periods. It was also found that non-linear
measures didn't have higher predictive performance than linear ones.
To assure the methodological quality of future studies on seizure prediction the
following guidelines were formulated [Mormann et al., 2007]:

• Prediction should be tested on unselected continuous long-term recordings

• Studies should assess both sensitivity and specificity

• Results should be tested using statistical validation methods based on Monte

Carlo simulations. If prediction algorithms are optimized using training data,
they should be tested on independent testing data.