Automatic Seizure Detection Using Convolutional Neural Networks And Random Forest

Epilepsy is a common neurological disease,which is characterized by anomalous discharge of the neuron in the brain.Electroencephalogram(EEG),which can record these abnormal electrical activity,becomes an important tool to explore epileptic characteristics and is widely used in epileptic research,including diagnosis,focus location,qualitative research,seizure prediction and seizure control.At present,the diagnosis of epilepsy is mainly done by the experienced medical staff using visual inspection of the EEG records of patients.As long time EEG recording can produce a large amount of EEG data,manual detection is time-consuming and inefficient.The judgment of analysts is influenced by the long hours of labor,leading to an increase in the rate of misdiagnosis.As a result,automated epileptic detection becomes more and more important and automatic seizure detection and recognition of epileptic EEG signal by computer has become an important auxiliary detection method.Automatic detection of epilepsy is essentially the classification of seizure and non-seizure EEG signals.The purpose of automatic seizure detection system is to distinguish these two kinds of signals according to the different characteristics of seizure EEG signals and normal EEG signals.Generally,there are limitations in traditional methods for automatic detection.In most automatic seizure detection algorithms,the feature extractor by artificial design has a great impact on the performance of the algorithm.And in many studies,a separate model for each patient is constructed because of the complexity of EEG signal and the difference between different people,which means a new model must be created when working on the data form a new patient.Furthermore,sometimes samples form the patient are not enough to build a new model.In order to solve these problem,an automatic seizure detection method combined with multiple transfer convolutional neural networks(CNNs)and random forest(RF)is proposed for classification of between seizure and non-seizure EEG signals in this study.CNN has the ability of automatically learning model parameters from training data without the need to manually design feature extractor.Multiple CNNs are built in our system,each of which is trained by the data of one patient form the training set.For the patient to be tested,model parameters of each CNN are adjusted through a segment of the data of testing patient based on transfer learning method.Each adjusted CNN discriminates test samples and outputs of the CNNs form the verdict vectors.Then a random forest is created based on the verdict vectors that are generated by the data for transfer learning.Corresponding decision vectors of testing data are classified by random forest to get the final category.This system achieves an average sensitivity of 90.22%,specificity of 98.76%and accuracy of 98.75%in a dataset consisting of 543 hours of long-term EEG signals.The experimental results show that the method proposed in this paper can effectively detect seizures.It does not need to re-establish the model and just takes a short time when dealing with data from new patient,which has certain research value.