Research On Epilepsy Detection And Representation Algorithm Based On Deep Learning

Epilepsy is one of the most common life-threatening and challenging neurological diseases.At present,there are still many challenges in seizure detection methods based on electroencephalogram data: EEG signals have severer imbalance with more non-epileptic data and less epileptic data;EEG signals have the characteristics of unsteady state and diversity;The manual detection of EEG signals is time-consuming and labor-consuming with a high misjudgment rate,and the results detected by different physicians are different;Clinical raw EEG data often contain a variety of noise and physiological artifacts,which interfere with seizure signals.To solve the above existing problems,it requires to explore an unsupervised algorithm model to get low-dimensional feature representation for linear classification,which can effectively distinguish seizure signals.In addition,we further study efficient and reliable automatic seizure detection technology to prepare for developing online seizure detection technology and neural intervention technology.Based on the analysis of the published Bonn data and the clinical raw EEG data collected from 301 Hospital of China,this thesis proposes two seizure detection methods:one is the low-dimensional representation algorithm based on Deep Boltzmann Machine unsupervised learning,which can linearly classify seizure data and non-seizure data;The other is to propose a model with continuous double-layer convolution structure based on one-dimensional convolutional neural network,which can carry out supervised learning,and automatically detecting seizure signals efficiently and reliably.In the unsupervised learning algorithm,we firstly use discrete wavelet transform to extract the spectral-temporal information of EEG data in the epileptic frequency band,and the statistics in time domain and frequency domain are calculated.Then DBM is used to train the statistics to reduce the feature dimension.Compared to principal component analysis,the representation performance of the two unsupervised training models were evaluated by employing Fisher discriminant function as well as support vector machine.DBM is trained to two states: transient state（DBM＿transient）and converged state（DBM＿converged）.For REC detection,the DBM＿transient on Bonn data is 9.15% and7.2% higher than PCA and DBM＿converged respectively,while the DBM＿transient on C301 data is 4.39% and 45.1% higher than PCA and DBM＿converged respectively;The evaluation results show that on the calculated Fisher discriminant function,the DBM＿transient of Bonn data is 0.4859 and 0.0453 higher than that of PCA and DBM＿converged respectively,while the DBM＿transient of C301 data is 0.00894 and0.014029 higher than that of PCA and DBM＿converged respectively.Therefore,our results demonstrate that DBM＿transient shows great low-dimensional representation of seizure EEG signals,rendering the following-up linear classification available,and thus facilitating reliable and robust automatic seizure detection.In the supervised learning training,we propose a model with continuous double-layer convolution structure based on one-dimensional convolutional neural network,which can detect seizure signals from C301 data efficiently and stably.The sensitivity,specificity,accuracy and F1-score are 99.7%,95.4%,99.6% and 96.1%respectively,and the running time of the training using GPU is 2 to 3 times faster than those of comparison models.The results show that the one-dimensional convolution neural network model proposed in this thesis is superior to the existing methods and is efficient and stable in the performance of seizure detection.