Classification Of Epileptic EEG Signals Based On Machine Learning

Epilepsy is one of the most common diseases in the nervous system and can cause physical and mental harm to the human body severely,so it is a difficult medical problem that needs to be solved urgently.The automatic and accurate recognition of the starting and ending points of epileptic seizure contributes to the timely treatment of patients.In this paper,the automatic recognition algorithm of epileptic seizure is studied by machine learning method,and the algorithm is validated based on the data of 129 epileptic seizure cases of 24 patients in the CHB-MIT Database taking 2 seconds as a unit.Firstly,epileptic EEG signals are classified by means of the support vector machine(SVM).Using the wavelet decomposition,four characteristic frequency bands of EEG signals are extracted.The features of time-domain signal variance,frequency domain energy and wavelet energy are obtained respectively from each frequency band to form twelve-dimension feature vectors.These feature vectors are classified by the SVM and the optimal patient-specific threshold classification method is used to combine the data from different channels.For all patients,the test results show that the average accuracy is 80.53%,the average True Positive Rate(TPR)is 76.06%,the average False Positive Rate(FPR)is 11%,and the average recognition error of the starting and ending points of the seizure is 11.5 seconds.Secondly,epileptic EEG signals are recognized using the convolutional neural network(CNN).The time-frequency analysis of EEG is carried out,and the one-dimensional EEG signals are transformed into two-dimensional time-frequency images within the frequency range of 32 Hz by Short Time Fourier Transform(STFT).These images are classified by the CNN with LeNet-5 structure,and the multichannel joint method is used to combine the data from different channels.For all patients,the test results based on the time-frequency images with the size of 32×32 show that the average accuracy is 90.13%,the average TPR is 96.05%,the average FPR is 7%,and the average recognition error of the starting and ending points of the seizure is 5.3 seconds.Also,in terms of the precision and speed of the recognition algorithm,the sensibility of the accuracy of the algorithm on different size of time-frequency images such as 32×32,64×64,128×128 and 256×256 is compared and it is concluded that the time-frequency image of the size of 32×32 is the most suitable.Finally,the two classification methods presented in this paper respectively based on the SVM and the CNN are compared from four aspects of feature extraction,training parameters,classifiers and experimental results.The comparison results by all kinds of evaluation indexes show that the CNN-based classification method discussed in the paper has better performance than the SVM-based classification method discussed in the paper.After that,the influence of different training parameters,such as the iteration times,the initial learning rate,the momentum and the weight attenuation rate,on the CNN model is analyzed,as well as the influence of different sizes of time-frequency images on this model is also discussed.Meanwhile,the fine-tuning method of network parameters,which uses the previous network parameters as the initial parameters for the present training process,is proposed to reducing training time.The results of this paper provide a theoretical basis for automatic recognition of EEG signals in epileptic status,contributing to the study of clinical treatments for epileptic seizure.