Research On Epileptic Seizure Prediction Based On EEG Signal Feature Processing And Deep Learning

Epilepsy is one of the most common neurological disorders in the world,in which seizures are caused by pathological neuronal activity in the brain that leads to disruption of normal brain function.Seizures can lead to abnormal behavior in patients.Seizure prediction is an important goal in clinical management and treatment,and accurate seizure prediction has the potential to transform clinical epilepsy care and create new treatment strategies for individuals in clinical decision support systems.Currently,the diagnosis and monitoring of epilepsy disorders is still primarily based on clinicians’ visual analysis of electroencephalography(EEG)recordings,which is highly subjective and wastes a lot of labor and time.EEG signal feature processing and deep learning techniques have made many important advances in the study of various types of EEG signals.In order to solve the problem of unclear seizure timing,which is the most important problem in actual clinical practice of epilepsy prediction,this thesis explores the multiple states in the pre-ictal period.In view of the actual clinical signal processing and analysis needs,this thesis investigates three aspects of EEG signal pre-processing,feature processing and classification algorithms based on EEG signal processing and deep learning methods.An end-toend patient-specific seizure prediction model is built around the pre-seizure period and the effectiveness of the method is verified.The research content and innovations of this thesis include the following three aspects:(1)To achieve effective automated and standardized end-to-end preprocessing for clinical applications of epilepsy prediction,an improved empirical mode decomposition denoising method based on detrended fluctuation analysis and an automated end-to-end standardized EEG signal preprocessing pipeline are proposed in this thesis.The pipeline can enhance the discriminability of EEG features in the preprocessing stage and save a lot of time and effort for data set preparation and clinical use.The proposed method and process were validated on raw EEG signals acquired in the laboratory and showed significant noise removal effect.(2)To address the problem of high feature dimensionality in feature processing of epileptic EEG signals,this thesis proposes a feature channel fusion algorithm based on t-distributed stochastic neighbor embedding.The algorithm processes the feature information of multi-channel highdimensional epileptic EEG signals,and reduces the dimensionality of the high-dimensional features,which improves the expression of feature information and provides a good support for subsequent classification.The proposed algorithm showed a maximum sensitivity improvement of 2.61%,a maximum false alarm rate reduction of 0.25 times/h,and a maximum improvement of 0.05 under the working feature surface of the subjects in the controlled experiment of epilepsy prediction tasks.(3)To achieve high accuracy in multi-state classification and recognition in the pre-seizure period,an epilepsy prediction network based on deep residual shrinkage network and gated recurrent units is proposed in this thesis,which incorporates soft thresholding signal noise reduction mechanism and attention mechanism to achieve accurate prediction of epilepsy state using the redundant information in EEG signals.The network is combined with an epilepsy signal feature channel fusion algorithm to build an end-to-end patient-specific epilepsy prediction model.The designed model achieved satisfying results in the epilepsy prediction task using the open-source epilepsy EEG dataset collected at Boston Children’s Hospital.Specifically,the designed model obtained an average sensitivity of91.36%,an average false alarm rate of 0.135 times/h,and an average area under the subject’s working feature surface of 0.9.