Application Research Of Deeply Ensemble Classification Model In Epileptic EEG Prediction

As a brain disease,epilepsy requires neurologists to monitor the electroencephalogram status of patients’ brains in real-time to predict seizures.The manual method is not only time-consuming,but also limited by the professional level of the doctor.Therefore,it is very important to use computeraided diagnosis to achieve automatic seizure detection.However,the data used by the traditional method is limited to the temporal and frequency features of the EEG signal itself and ignores the implicit spatial characteristics.At the same time,the unbalanced EEG data set adversely affects on the performance of the crosspatient seizure prediction model.In this thesis,a deep ensemble network is used to automatically extract various features of EEG signals,and a Generative Adversarial Networks is used to reasonably amplify EEG data.The following work has been carried out around the construction of epilepsy patient-specific prediction models and cross-patient prediction models:(1)This thesis proposes a multi-feature deep ensemble networks classification model to achieve automatic extraction of multiple features of EEG signals.First,the one-dimensional EEG signals are converted into image data through spatial projection,which achieves the explicit writing of implicit spatial information into the data,and provides conditions for using the spatial information in the EEG data analysis.Secondly,two vertically ensembled feature extractors are used to extract EEG signals.That is,the convolution feature extractor with parallel convolution and anti-aliasing filter is used to extract the spatial and frequency domain features in the EEG signal;the time domain features in the EEG signal segment are mined by a bidirectional long-short-term memory network.Finally,the fully connected layer was use to classify the extracted comprehensive features.Through comparative experiments on the CHBMIT dataset,the multi-feature deep ensembled classification model achieved excellent performance in patient-specific seizure detection,indicating that the model can automatically analyze and extract a variety of features of EEG data to achieve effective seizures prediction.(2)This paper proposes a generative multi-feature deep ensemble model to solve the problem that it is difficult to fit a cross-patient epilepsy prediction model on an imbalanced data set.The EEG signals of the same state are different in different patient,which easily leads to the problem that the generator has too large loss fluctuation during training and it is difficult to fit the real distribution.By using the improved Wasserstein distance instead of the KL divergence to train the generator,the generator parameter update speed is automatically matched with the rate of change of the model loss,and the reliability of the generator is increased while improving network convergence.Experimental results show that the proposed algorithm achieves excellent performance in cross-patient epilepsy prediction classification model.It shows that the generative multi-feature deep ensemble model can alleviate the influence of the imbalance of EEG data samples on the performance of the classification model.