Imbalanced Classification For Epileptic EEG Signals Based On Deep Neural Network

Epilepsy is a disorder syndrome of the brain nervous system caused by abnormal synchronous discharge of nerve cells in the brain.Electroencephalogram is the most widely used method among many auxiliary tools for epilepsy.Not only is traditional visual detection of electroencephalogram time-consuming,but also susceptible to the subjective factors of the doctors.Therefore,developing an accurate automatic epileptic seizure detection system is of great significant.Based on the assumption that the distribution of dataset is balanced,the traditional epilepsy detection method makes classification judgment of the data of seizures and non-seizures.However,in the actual electroencephalogram signals of epilepsy patients,the time of seizure period is far less than the time of non-seizure period.Using the traditional classifier to classify and recognize the epileptic electroencephalogram signals,the decision boundary of classifier will be biased to the seizure periods,which leads to the expansion of epileptic electroencephalogram signals further aggravate the impact of two unbalanced types of electroencephalogram data.In this paper,an imbalanced classification method for epileptic EEG signals based on deep neural network is proposed.First of all,Borderline-SMOTE algorithm is used to synthesize new samples for the boundary data of seizure and non-seizure.The classifier can learn the boundary features of the two types of data as accurately as possible,solving the problem of decision boundary deviation effectively.At the same time,the electroencephalogram training data with different proportions were processed,respectively,to avoid blurring classification boundary caused by the synthesis of too many boundary data.So as to obtain the training sets with different proportions of distribution in the seizure and non-seizure periods.Then,the pyramidal one-dimensional deep convolutional neural network is designed.Compared with the common two-dimensional convolutional neural network,the pyramidal one-dimensional convolutional neural network reduces the training parameters,improves the training rate,and avoids the overfitting effectively caused by the small number of training samples.The proposed epilepsy detection method is evaluated on a long-scalp electroencephalogram database.The comparative experiment shows that the pre-balance treatment can significantly improve the classification performance of the deep network.Compared with other epilepsy detection methods and long-short-term neural network,the combination of Borderline-SMOTE applying to one-third training set and pyramidal one-dimensional convolutional neural network achieves better seizure detection performance.