An Enhanced Wavelet Neural Network Model For Epileptic Seizure Detection And Prediction

Loading...
Thumbnail Image
Date
2016-03
Authors
Lai, Kee Huong
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Epilepsy is a very common and much-feared neurological disorder. Much research has been done in developing better automated classi ers with higher accuracy that can help clinicians identify the di erent segments of electroencephalography (EEG) signals. In this research work, an enhanced wavelet neural network (WNN) model is proposed for the purpose of epileptic seizure detection and prediction. The architecture and con guration of WNNs can be further enhanced using metaheuristic strategies. Speci cally, the harmony search (HS) algorithm is employed and incorporated in the learning of WNNs. The contribution of this thesis is threefold. Firstly, the HS algorithm is used in the feature selection stage. The HS algorithm, which is originally used for optimization problems involving real numbers, is modi ed and employed in the task of feature selection, which involves binary values. Apart from simplifying the network architecture, the reduction in the number of features also reduces computational cost. Secondly, the HS algorithm is employed to nd the translation vectors of the hidden nodes of WNNs. A good set of translation vectors will indirectly increase the e ciency of the learning process of WNNs. To achieve this goal, the HS algorithm is hybridized with the type-2 fuzzy c-means clustering algorithm. Thirdly, the HS algorithm is incorporated in the learning algorithm of WNNs. In particular, the HS algorithm is used to determine the synaptic weights and bias terms of WNNs. Novel harmony memory initialization and improvisation strategies are incorporated in the proposed HS-based learning algorithm. The e ectiveness of the three aforementioned improved methods are rst tested using ten sets of UCI machine learning data sets. The preliminary simulations report that the hybridized methods give superior performance than the conventional stand-alone algorithms. Also, WNNs models that are trained using the HS algorithm and other metaheuristic approaches report comparable results. The WNNs models with enhancements in three di erent aspects are then tested using two real world applications, namely in the tasks of epileptic seizure detection and prediction. The discrete wavelet transform (DWT) method is used to pre-process the EEG signals to yield di erent groups of wavelet coe cients, which correspond to di erent frequency sub-bands. Simulation results show that the enhanced WNN model outperforms most of the other machine learning methods reported in the literature. This suggests the potential usage and implementation of the developed classi ers in the eld of epileptology.
Description
Keywords
Epilepsy
Citation