| Epilepsy,as a sudden and life-threatening nervous system disease,seriously affected around 50 million people in the world,with more than 9 million patients in China.However,about 10%?50% epilepsy patients are impossible to be cured via medication or surgical treatment.For these patients,if the preictal duration before the seizure onset can be accurately identified,it will allow patients to take precautions in advance to reduce the secondary injury caused by the sudden loss of consciousness.The scalp electroencephalogram(EEG)signals record rich information of the electrical activities in the brain,which plays an enormous role in the field of epileptic seizure detection.Up to now,existing research are mainly focus on seizure/non-seizure detection while few attention has been paid to the preictal states classification.The accurate preictal classification can offer more protection so that the damages caused during a seizure could be prevented or reduced.In addition,traditional epileptic detection algorithms generally define preictal state as an hour before seizure.For practical clinical application,this classification method is not detailed enough to eliminate the threat and anxiety caused by seizure.Thus,a timely and accurate seizure detection algorithm has great practical significance.The main research contents and results of this paper are as follows:1.A method of epilepsy detection based on deep transfer learning feature is proposed.Firstly,a detailed division mode of preictal state is designed,and then 5deep neural network models(Alex Net,VGG19,Inception-v3,Res Net152 and InceptionRes Net-v2)pre-trained by Image Net database are used to extract the feature of subband mean amplitude spectrum(MAS)map of multi-channel EEGs.Finally,the extracted deep feature vector is fed into a 2-layer network(the fully connected layer and the Softmax layer)for epilepsy EEG classification.This method obtain the overall accuracy of 89.81% in the 5 epileptic states(the preictal state is equally divided into 3non-overlapping periods of 20 minutes)classification task on the CHB-MIT database.2.A method of epilepsy detection based on deep neural network(DNNs)fusion transfer learning features is presented.Firstly,we concatenate the transfer features of MAS map captured by 5 pre-trained models,and then the merged features are combined with the 7-layer neural network composed of fully connected layers,dropout layers,a softmax layer to identify epileptic EEG signals.In order to bring greater help to patients in clinical,the preictal state is further divided to form a 8-category task in this paper.To evaluate the generalization performance of the proposed algorithm,the clinical i Neuro database EEG signals are also used for verification in this paper.We conduct extensive experiments on 2 benchmark epileptic EEG databases,and the proposed algorithm gets the best results of classification.On the CHB-MIT database with 20 and 10 minutes non-overlapped equilong segments of the preictal state,the proposed method achieves the highest overall accuracies of 96.97% and 92.28%,and the highest overall recognition accuracies of 91.21% and 87.87% with 60 and 10 minutes duration respectively.3.A method of epilepsy detection based on Hand-Crafted features and deep transfer features fusion is proposed.The MAS can reflect the frequency domain information of EEG signals but ignore the time domain information.In order to obtain more recognizable Hand-Crafted features,we fuse the frequency domain features of MAS and MPSD with the time-frequency domain feature extracted from wavelet packet decomposition,and then combines the proposed algorithm for epilepsy detection.The multiple experiments results show that the recognition effect based on Hand-Crafted features fusion algorithm is significantly improved compared with single Hand-Crafted feature.On the 5-state epileptic classification problem,the proposed algorithm achieves the overall accuracies of 98.97% and 92.04% with CHB-MIT and i Neuro databases respectively. |
PDF Full Text Request