EEG Epilepsy Prediction And Source Localization Based On Neural Networks

Epilepsy is a common mental illness,and its evolution process has obvious manifestations in Electroencephalogram(EEG)signals.Therefore,the analysis of epilepsy EEG is of great significance for the classification,prediction and source localization of epilepsy.Nowadays,in epilepsy treatment,a large amount of non-labeled EEG signal can be easily accessed.Therefore,learning to extract EEG features from unlabeled data is very important for epileptic EEG classification.Meanwhile,learning to extract proper EEG features and construct the spatio-temporal model for EEG signals with machine learning methods is crucial for seizure prediction and source localization.The application of these methods gives a great help in objective quantitative analyzing and assisting doctors in diagnosis.Therefore,in order to solve the problems in it,we studied the feature clustering,prediction and source localization of epilepsy EEG with machine learning methods.The main research contents and contributions include:For unsupervised clustering problem,a novel framework for clustering based on exemplar discriminative information is proposed.Making full use of the discriminative information of exemplar samples in the clustering,the proposed algorithm establishes a new objective function.Moreover,the objective function is optimized by updating the exemplar data and partition hyperplane alternately.The final clustering results are obtained when stopping rules are satisfied.In addition,we analyze the performance of the proposed algorithms on artificial datasets and UCI datasets.The results are also compared with various related algorithms.Experiments show that the proposed algorithm has good clustering performance and achieves high purity result.Moreover,the proposed method is applied to the clustering of epileptic EEG signals,and the visualization analysis is carried out.These studies lay the foundation for in-depth research on epilepsy EEG analysisInspired by bag-of-word feature extraction in natural language processing,a novel framework for seizure prediction is proposed by learning synchronization patterns with bag-of-wave features in EEG.The proposed feature extraction method represents the spatio-temporal information in different scales.Firstly,the clustering model is used to quantify the expression of local signals,and construct the EEG feature dictionary.Then,the sliding time window is used to extract feature of the multi-lead signal over a period with feature dictionary.The statistical histogram in a time window is expressed as the spatio-temporal feature of the segment signal.Thus,each dimension of the histogram represents the frequency at which the local feature words occur synchronously within the time window.We further integrate the feature in different time windows,and the temporal pattern of epilepsy is mined on a larger time scale.Finally,combined with the ELM algorithm to classify the features and give early warning of the seizure onset.The experimental results show that the proposed feature extraction method gives a promising performance in epilepsy prediction.In EEG source localization problem,the traditional modeling method relies on the prior knowledge of the head model and the conduction model.Usually,the head model and the conduction model are difficult to modeling accurately,which affects the accuracy of the solution of EEG inverse problem.Therefore,a new source localization method using spatio-temporal neural networks is proposed.Firstly,based on the Bayesian model,we derive that the temporal of scalp signal and source distribution need to be modeled in solving source localization problem.According to this,the sequence-to-sequence long-term memory network(LSTM)is used to learn the relationship between scalp EEG and source location.By training on the analog signal,the spatio-temporal relationship between the scalp signal and the source signal is optimized,and the relationship is implicit in the weights of the neural network.The experiment results show that the proposed approach can solve the EEG source localization problem effectively and robust with noise.Moreover,the proposed approach is applied in real epileptic EEG signals to show the effectiveness.