18-12-2012, 03:03 PM
Automated Detection of Epileptic Seizures Using Wavelet Entropy Feature with Recurrent Neural Network Classifier
Automated Detection of Epileptic Seizures.pdf (Size: 169.01 KB / Downloads: 41)
INTRODUCTION
The brain is a highly complex system and the EEGs
originate from it reflect the electrical activity of the cerebral
cortex. Epilepsy is a common neurological disorder that is
characterized by recurrent unprovoked seizures [1-3]. These
seizures are transient signs and symptoms due to abnormal,
excessive or synchronous neuronal activity in the brain. About
50 million people worldwide have epilepsy at any one time [2].
The correct classification of each person's epilepsy by seizure
is important for proper treatment. It involves an expert’s efforts
in analyzing the entire length of the EEG recordings to detect
traces of epilepsy. As the traditional methods of analysis are
tedious and time-consuming, many automated epileptic EEG
detection systems have been developed in recent years [4-10].
With the advent of technology, processing of the EEG data has
been done digitally. Due to this facility, any long duration EEG
recordings can be fed to an automated seizure detection
system in order to detect the presence of seizures. In recent
years different features and classification approaches have been
developed for the automated diagnostic systems for epilepsy
[4-10]. Similarly different artificial neural network (ANN)
based classification systems for epileptic diagnoses have been
proposed by several researchers. The method proposed
includes the features, namely, average EEG amplitude, average
EEG duration, coefficient of variation, dominant frequency,
relative spike amplitude, spike rhythmicity, average power
spectrum, periodogram, spectral entropies, autoregressive (AR)
features, as inputs to different types of ANN like, adaptive
structured neural network, learning vector quantization (LVQ)
network, LAMSTAR network, back propagation neural
network (BPNN), Elman network (EN) and probabilistic neural
network (PNN) [4-12].
RECURRENT NEURAL NETWORK CLASSIFIERS
Neural network models are found to be the potential
candidate for pattern classification problems due to special
characteristics such as massive parallelism, self organization,
adaptive learning capability and robustness [18-19].
Recurrent neural network model, Elman neural network is
used in this work. Being a feedback backpropagation network,
it has the ability to store information. The delay introduced in
the connection between the layers stores values from the
previous time step, which can be used in the current time step.
The simple architecture used for classification of the EEG
signal is shown in Fig.7 [20]. For training the network,
gradient descent bakpropogation algorithm is used. The
activation functions used are: tan sigmoid for the hidden layer
and log sigmoid for the output layer.
CONCLUDING REMARKS
This paper discusses automated detection of epileptic
seizures using wavelet entropy with recurrent neural network
classifiers. EEG recordings obtained from extra cranial and
intracranial electrodes were used and the performances were
evaluated in terms of specificity, sensitivity and overall
accuracy. It has been observed from experimental results that
an overall classification accuracy of 99.75 was achieved to
detect normal and epileptic seizures and 96.3 to discriminate
interinctal and ictal seizures. The proposed wavelet based
feature provides promising results compared to spectral
entropy and will be suitable for real-time seizure detections.