Research On Epileptic Seizure Prediction Based On Machine Learning

With the implementation of brain project in many countries and the needs of human development,more and more attention has been paid to the research of nervous system diseases.Epilepsy,as one of the most common neurological diseases,seriously affects the safety and life of patients and their families.Epileptic seizure prediction can remind patients before seizure onset,so as to take timely measures to prevent patients from further injury.However,due to the high similarity between pre-ictal and inter-ictal,the high complexity and the great difference between different individuals and seizure types in EEG signal,the current epileptic seizure prediction algorithm can not achieve the desired results.On the basis of understanding the current situation of epileptic seizure prediction,this paper proposes the corresponding framework and algorithm based on EEG signal to solve these problems,so as to improve the performance of epileptic seizure prediction.Firstly,we made detailed analysis of the commonly used features in time domain,frequency domain and time-frequency domain of EEG signal research.Several features were extracted from these domains for epileptic seizure prediction.By comparing the impact of these three types of features on prediction performance,it provides a basis for further feature design and analysis in epileptic seizure prediction.Then,we implement an automated framework for epileptic seizure prediction based on sequential forward feature selection method.In this framework,67 features from 18 channels of EEG signal are extracted to form a feature set.Then the minimal-redundancy-maximal-relevance criterion is used to evaluate all features.Finally,the sequential forward selection method is used to select the optimal feature subset for each patient according to the classification accuracy of support vector machine.Among them,the optimal parameters of support vector machine are obtained by grid search method,and the classification accuracy of patient model is evaluated by leave-one-out cross validation method based on the onset.The framework was evaluated using 23 cases in CHB-MIT database.Good results are obtained in 7 metrics,with an overall average sensitivity of 0.902 and a false positive rate of 0.096/h.Our innovation in this part is to propose an automated framework,which can be very convenient for each patient to design the optimal feature subset for seizure prediction,and provide the basis for clinical treatment for patients,and obtain good predictive performance on one of the datasets.However,due to the high time complexity of the sequence forward feature selection method in the training phase,we introduce the data gravitation model and improve it for the prediction of epileptic seizures.In this algorithm,we use the same data preprocessing method and features with seizure prediction framework based on sequential forward selection method.We introduce the feature-weighted data gravitation to evaluate the features.According to the data used,we use the stochastic gradient descent method to optimize the feature weight.In the classification stage,we combine gravitation and K-Nearest Neighbor algorithm to predict the signal segments.Finally,23 cases in CHB-MIT database were evaluated using leave-one-out cross validation method.The overall average sensitivity of 0.949 and the false positive rate of0.071/h were obtained.The average training time was 13.71 seconds.In this part,our innovation is to improve the data gravity model and apply it to epileptic seizure prediction.To a certain extent,we can analyze the features of patients and obtain good prediction performance.In this paper,aiming at the problems in the research of epileptic seizure prediction,we analyzed the features comprehensively,and design a framework and a algorithm from the perspective of feature design,and applied them to a popular database,which can provide a reference for researchers using this database.The framework and algorithm can also provide technical reference for tasks using other databases.