Automatic Detection Of Epileptic EEG Signals Based On Semi-supervised Learning

Epilepsy is a common psychiatric disorder with approximately 70 million patients worldwide.The epileptic EEG signal is an important tool for determining seizures.The epileptic EEG signal requires a priori knowledge of neurologists for manual labeling,and this labeling method is time-consuming and laborious.The automatic detection of epileptic EEG signals is important for the treatment of epileptic patients.In recent years,deep learning in epileptic EEG signal detection has become a research hotspot and achieved very excellent detection results,the most prominent of which is a hybrid model combining convolutional neural network CNN and long short-term memory network LSTM,which can achieve 100% detection accuracy in the Bonn University dataset.However,the model was not tested with new clinical patient data in these studies,and the dataset has limitations when it comes to building the model.In this paper,we implement a hybrid model combining CNN and LSTM for classification of epileptic EEG signals,and the accuracy of this model decreases dramatically to nearly50% when tested with new patient clinical data.In actual clinical diagnosis,our model often has to face new patients,and it is important to improve the generalization ability of the model so that the model can maintain a high correct detection rate even in the face of new patients.To address the above problems,in order to improve the stability and generalization ability of the model,this paper proposes a multilayer weighted integrated self-learning algorithm weighted by different classifiers and a confidence evaluation method based on a hybrid Gaussian model.The main works are as follows:(1)A multilayer weighted integrated self-learning algorithm for weighting different classifiers is proposed,which first weights the different classifiers to vote on the diagnosis results,and then the voting results are weighted again to produce the final diagnosis results.The algorithm improves the problem that traditional self-learning algorithms are influenced by data noise,and exhibits strong stability in the detection of EEG signals from different data sets and clinical epilepsy,which greatly reduces the workload of neurologists and provides support and assistance for the diagnosis and treatment of epilepsy.The experimental results show that the algorithm can improve the stability and reliability of automatic diagnosis of epileptic EEG signals,and the accuracy and AUC area can reach 0.80-0.95 in two different public datasets and clinical data classification.(2)On the basis of multi-layer weighted ensemble self-learning algorithm,a confidence evaluation method based on Gaussian mixture model is proposed.After the multi-layer weighted classifier marks a large number of unlabeled epilepsy EEG data,the method will The labeled data uses multiple Gaussian mixture models to fit the data distribution,and then selects the high-confidence data in each Gaussian model,and uses the high-confidence data selected by the evaluation method to increase deep learning The number of training data in the middle,improve the generalization ability of the model.This method makes full use of unlabeled data.After a large amount of training data is added for training,the hybrid model of CNN and LSTM is used to test the new patient clinical data after training,and the accuracy rate is increased by 16-48%.