Epileptic Seizures Classification in EEG Signals Using Wavelet Based-Neural NetworksContexto UDLAP Admin
Tesis Digitales UDLAP
One of the most important organs of the human body is the brain, this consists of some billions nerve cells that not only put together thoughts and highly coordinated physical actions but regulate our unconscious body processes, such as digestion and breathing. The nerve cells of the brain are known as neurons, which make up the so-called “gray matter” of the organ. The neurons transmit and gather electrochem- ical signals that are communicated via a network of millions of nerve fibers called dendrites and axons. This interaction of physiological and chemical processes gives rise to observed neuro-electrical activity in the brain controlling the processes of our body. Electroencephalography is a medical imaging technique that reads scalp electric activity generated by the brain. An electroencephalogram (EEG) is defined as electrical activity of an alternating type, which is recorded from the scalp surface after being picked up by metal electrodes and conductive media. An EEG has been used frequently in the clinical area for the evaluation and treatment of brain diseases, such as epilepsy [SHA11], [YUE11].
Epilepsy is a common chronic neurological disorder aﬀecting around 0.1% of the world’s population [SHA11], [YUE11]. In epilepsy, the normal pattern of neuronal activity becomes disturbed, causing strange sensations, emotions, and behavior, or sometimes convulsions, muscle spasms and loss of consciousness. There are many possible causes of epilepsy. Anything that disturbs the normal pattern of neuron activity ranging from illness to brain damage to abnormal brain development can lead to seizures [SHA11]. The traditional method of EEG analysis is by visual inspection of the signals plotted on paper, and this inspection of the signals requires highly trained professionals. These professionals, although guided by the general definitions for epileptogenic sharp transient waveforms, use additional subjective criteria based on contextual information and other heuristics to reach a decision [ANU12]. In addition, some cases may require experts to visually inspect the entire length EEG recordings of up to one week, which is tedious and time-consuming [SUB05].
Modern computer analysis can extend electroencephalograph’s capabilities by supplying information not directly available from the raw data. However, visual analysis is still a widespread technique, especially for detection of transient features of signal. In most cases, the agreement of an automatic method with visual analysis is a basic criterion for its acceptance. In recent years, several models have been proposed to classify EEG signals, some of them based on wavelet analysis and ar- tificial neural networks. The combination of wavelet analysis and artificial neural networks seeks to exploit the features of analysis and decomposition of wavelet pro- cessing along with the properties of learning, adaptation and generalization of neural networks. Despite of all works recently published [YUE11], still there is a need to improve the classification accuracy obtained by the available models.
In this thesis, we look for wavelet-based neural networks architectures able to eﬃciently handle the classification of EEG signals with respect to epilepsy-related stages. A model called Multidimensional Radial Wavelons Feed-Forward Wavelet Neural Network (MRW-FFWNN) is proposed for classification of three classes of EEG related to epilepsy conditions: Ictal, Interictal and Healthy. The Discrete Wavelet Transform (DWT) and the Maximal Overlap Discrete Wavelet Transform (MODWT) are used to decompose the EEG signals into delta (δ), theta (θ), alpha (α), beta (β), and gamma (γ) sub-bands. Each EEG signal is represented by a fea- ture vector of six components, built using the mean, absolute median and variance of delta (δ) and alpha (α) sub-bands. In this regard, some researches have reported that these sub-bands are suitable for the identification of epileptic episodes [RAV12], [SUN12]. We have used binary-tree and one-vs-one (OVO) decomposition strategies [GAL11] with two aggregation methods (Voting strategy and Weighted voting strat- egy) [GAL11] obtain the predicted class of EEG signals. Cross correlation coeﬃcients were used to evaluate the degree of similarity between mother wavelets and EEG signals, and then the most appropriate mother wavelet was chosen. A database pro- vided by the University of Bonn [BON16], [AND01] (see also Table 3.1) is used to assess our proposed model and to compare it with similar works proposed by Subasi & Ercelebi [SUB05], Shaik & Srinivasa [SHA12] and Ghosh et al. [GHO07]. We used a 3-fold validation to evaluate the performance of our proposed model. The best result of classification of three classes of EEG signals was of 96.67 % of accuracy using a binary-tree strategy based on our proposed model (MRW-FFWNN).
Tesis profesional presentada por el Mtro.
Doctorado en Ciencias de la Computación
Departamento de Computación, Electrónica y Mecatrónica, UDLAP
Presidente: Dr. Edgar Guevara Codina
Secretario y Director: Dr. Vicente Alarcón Aquino
Vocal y Co-director: Dra. María del Pilar Gómez Gil
Vocal: Dr. Roberto Rosas Romero
Vocal: Dr. Juan Manuel Ramírez Cortés