<title>Classification of epileptic EEG using neural network and wavelet transform</title>

Abstract

One of the major conthbutions of electroencephalography has been its application in the diagnosis and clinical evaluation of epilepsy [1]. The interpretation of the EEG is achieved through visual inspection by a trained electroencephalographer. However, descriptions of rules used during the visual analysis of data are often subjective and can vary from one reader to another. Computerized methods are a means to standardize this process. In recent years, much effort has been made to develop such methods that can characterize different interictal, ictal (seizure), and postictal stages. The main issue of whether there exists a preictal phenomenon remains unresolved. In the present study we address this issue making use of specifically designed and trained recurrent neural networks in conjunction with signal wavelet decomposition technique. The purpose of this combined consideration was to demonstrate the potential for seizure prediction by up to several minutes prior to its onset.Epilepsy is a serious disease affecting tens of millions of people around the world. It is known, that in many epileptic patients the occurrence of seizures can be effectively suppressed by accurately planned medical treatment.Consequently, an important and challenging task is the creation of efficient computerized systems to differentiate between various interictal, ictal (seizure), and postictal stages in BEG recordings. Over the past decade most theoretical and experimental studies focused on developing computerized methods for quantitative characterization of underlying nonlinear dynamical systems. These included "chaos theory" measures of correlation dimension, Kolinogorov entropy, and Lyapunov exponents [2][3][4]. As this work progressed, the limitations of available computational procedures became increasingly apparent. Those limitations of the related mathematical algorithms are based on the requirements of long-term stationarity of the time series (which is often not the case in epileptic BEG considerations) and an extremely high sensitivity to the noise, both electrical and physiologic in its nature. Our efforts were therefore directed to constructing alternative measures, which could reflect short-term signal "textural" complexity conditions which may then result in seizures [5][6][7]. In [6] we showed the feasibility of using signal local texture features in conjunction with a wavelet transform. The purpose of the present study is to explore the ability of specifically designed and trained recurrent neural networks (RNN) combined with wavelet preprocessing in segmentation of regions with varying texture information. Our goal was to demonstrate the existence of several minute long preictal stages rather than to develop an on-line detection tool. With the application of novel effective methods of training recurrent neural nets we have been able to show that for a particular single patient data set these preictal stages are in fact apparent . The RNN were chosen because they can implement extremely nonlinear decision boundaries and ...

Protocol Design Challenges in the Detection of Awareness in Aware Subjects Using EEG Signals

“…Examples of good results with this method in the EEG context can be found, for example, in some other studies. 17,[23][24][25][26] In our experiment, a wavelets-based extraction was implemented by using the signals coming from the 63 electrodes and sampled at 100 Hz. For each channel, the wavelets expansion in terms of the Morlet family of the total sample length (5.5 seconds) were computed and the coefficients corresponding to a frequencies interval of 1 to 30 Hz with a frequency step of 0.25 Hz and a time step of 0.01 seconds were kept.…”

Comparative study of stock trend prediction using time delay, recurrent and probabilistic neural networks

“…This may be a potential bottleneck of the training procedure, particularly when both and are large. While applying the multistream EKF procedure to a variety of real-world tasks, we have, however, experienced no problems [3], [28], [29]. We have found singular-value decomposition best for inverting up to 100 100 elements [9].…”

“…For large data sets, this advantage increases. One-stream training fails for sets of several thousand vectors [28], [29]. We have used multistream EKF training of RNN for predicting trends of several stocks.…”

“…From the scalar time series, we form points of embedding dimensions (27) 2) For every point, we count the number of points within a hypersphere of radius centered at that point. Assuming random distribution, the probability that a point lies within of the chosen point will be (28) Over points, we get the average probability (29) where if otherwise.…”

Patient-Specific Early Seizure Detection From Scalp Electroencephalogram

“…In order to overcome these difficulties, alternative measures reflecting short-term signal “textural complexity” changes that precede seizures have been employed by a number of investigators within a linear system framework. For example, Petrosian et al (1996, 1997, 2000) have explored the ability of specifically designed and trained recurrent neural networks (RNN) combined with wavelet decomposition preprocessing to predict seizures. The RNN approach is particularly useful for the task of patient specific EEG prediction because it can be trained in a straightforward manner, is able to implement extremely nonlinear decision boundaries and possesses memory of previous states.…”