Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 134

early idea was to use the rate of occurrence of interictal epileptic spikes as a mea-
sure indicating the time of seizure onset. There were some positive reports in this
respect [Lange et al., 1983], but in the systematic study based on large experimen-
tal material it was demonstrated that spiking rates do not change markedly before
seizures, but may increase after them [Gotman and Marciani, 1985, Katz et al., 1991].
Other attempts involved application of autoregressive modeling which allowed for
detecting pre-ictal changes within up to 6 s before seizure onset [Rogowski et al.,
1981, Salant et al., 1998].
In the 1980s, the non-linear methods based on chaos theory were introduced
to epilepsy research. They involved calculation of attractor dimension, e.g., [El-
ger and Lehnertz, 1998], Lyapunov exponent [Iasemidis et al., 2003], correlation
density [Martinerie et al., 1998]. However, more recently the optimistic results ob-
tained by non-linear measures have been questioned. Starting from early 2000 a
number of papers based on rigorous methodology and large experimental material
showed that the results reported earlier were difficult to reproduce and too optimistic
[Aschenbrenner-Scheibe et al., 2003, Lai et al., 2003, Harrison et al., 2005, Mc-
Sharry et al., 2003]. In a study by [Wendling et al., 2009], concerning coupling of
EEG signals in epileptic brain, nonlinear regression, phase synchronization, gener-
alized synchronization were compared with linear regression. The results showed
that regression methods outperformed the other tested measures. The re-evaluation
of the earlier works led to the conclusion that the presence of non-linearity in the
signal does not justify the use of non-linear complicated measures, which might be
outperformed by simpler linear ones.
The detection of seizures is especially important in case of patients with frequent
seizure episodes and for neonates. The methods applied for this purpose usually con-
sist of two stages. In the first stage signal features such as spike occurrence, spike
shape parameters, increase of amplitude or time-frequency characteristics found by
means of wavelet transform were extracted. In the second step artificial neural net-
works (ANN) were used for identification of seizure onset. The review of the works
based on this approach may be found in [Sanei and Chambers, 2009].
Saab and Gotman [Saab and Gotman, 2005] proposed to use for patient non-
specific seizure detection a wavelet transform (Daubechies 4 wavelet). From WT
the characteristic measures were derived, namely: relative average amplitude (the ra-
tio of the mean peak-to-peak amplitudes in the given epoch to the mean of amplitude
of the background signal), relative scale energy (the ratio of the energy of the coef-
ficients in the given scale to the energy of the coefficients in all scales), coefficients
of the variation of the amplitude (square of the ratio of the standard deviation to
the mean of the peak-to-peak amplitude). The classification was performed based on
Bayes conditional formula which allows to describe behavior of the system based on
how it behaved in the past. The a priori probabilities were obtained using the training
data and the a posteriori probabilities served as the indicator of the seizure activity.
The reported sensitivity was 76% and median value of delay in detecting seizure
onset was 10 s.
The patient specific method for the detection of epileptic seizure onset was pre-
sented by Shoeb at al. [Shoeb et al., 2004]. They applied wavelet decomposition