Biomedical Engineering Reference
In-Depth Information
segments of nonepileptic seizures. An NN with a multilayer perceptron was
proposed for the automatic analysis of the seizure-related EEG waveforms.
Fuzzy clustering techniques have been applied to the detection of epilep-
tic events. Geva and Kerem (1998) used rats to trial an EEG-based brain-
state identification technique, which allowed neural behavior prediction to
estimate the occurrence of epileptic seizures. The rats were exposed to epilep-
tic conditions from which the features leading up to and during the seizure
were analyzed using the wavelet transform method. These extracted fea-
tures were then used as input to the unsupervised optimal fuzzy clustering
(UOFC) algorithm from which a classification of the data was undertaken.
The classification was successful in determining the unique behavioral states
as well as the features associated with the seizure. Figure 6.6 illustrates a
flow diagram of the steps followed by Geva and Kerem (1998). Although the
results hold some uncertainty due to the fluctuation in the data, the applica-
tions of this technique have promise for recognizing seizures.
A number of researchers have applied hybrid approaches to improve epilepsy
detection. Harikumar and Narayanan (2003) applied fuzzy logic techniques
through genetic algorithms (GA) to optimize the EEG output signals for the
classification of patient risk associated with epilepsy. The risk factor was deter-
mined by inserting data regarding energy, variance, peaks, sharp and spike
waves, duration, events, and covariance measurements into a binary GA and
continuous GA. The performance index measures and quality value measures
were also evaluated. The authors claimed over 90% accuracy in detecting the
epileptic seizures using their technique, and from the results it appears that
the continuous GA may provide the most accurate risk assessment.
A hybrid CI approach based on genetic search techniques has been applied
by D'Alessandro et al. (2003) to identify patient-specific features and electrode
sites for optimum epileptic seizure prediction. The genetic search algorithm
was trained using features extracted from a subset of baseline and preictal
data, the trained algorithms were then validated on the remaining data sets.
A three-level feature-extraction process was adopted by the authors as illus-
trated in Figure 6.7. Feature selection was applied to identify the most impor-
tant features contributing to the prediction task. Synchronized video and EEG
signals were recorded from four patients at 200 Hz, which included information
regarding 46 preictal states and 160 h of baseline data. A 60-Hz notch filter
was applied to remove the power line frequency noise. The training data con-
stitutes 70% of these data with testing carried out on the remaining data sets;
a block diagram of the steps in the prediction of epileptic seizures is illustrated
in Figure 6.7. EEG signals were collected from multiple implanted intracranial
electrodes and quantitative features derived from these signals. Features rep-
resenting information in the time domain, frequency domain, and nonlinear
dynamics were used in the study. By using an intelligent genetic search pro-
cess and a three-step approach (Figure 6.7), important features were selected
for use with the classifiers. A probabilistic NN classifier was used to clas-
sify “preseizure” and “no preseizure” classes; and the predicted performance,
based on four patients implanted with intracranial electrodes, was reported to
Search WWH ::
Custom Search