A Methodological Study On Seizure Automatic Detection

Epilepsy is one of the most common diseases of nervous. The pathological mechanism of epilepsy is abnormally synchronization discharge of neurons which results in brain dysfunction. Electroencephalogram(EEG) and electrocorticogram (ECoG) are representation of the neural electric activities, and they are commonly used in seizure detection. Seizure automatic detection based on EEG and ECoG is important for the localization and classification of epileptic seizures.In order to solve the problem that the accumulated energy is unstable, we defined the feature combined time-frequency analysis with the energy. In order to solve another problem brought by large amount of data and high feature space in EEG for quickly and accurately detecting the seizures, a method that based on maximum-relevance and minimum-redundancy(mRMR) criteria and extreme learning machine with parameter(pELM) was proposed. Firstly, the time-frequency measures by short-time Fourier transform(STFT) and empirical mode decomposition(EMD) were extracted as the energy cells, and we got the features combined with the spatial location. Secondly, using the sequential forward selection method based on mRMR criteria to generate the feature subsets, then using the wrapper feature selection to select the optimal feature subsets by taking every feature subset as evaluation unit. Finally, the states were classified using the extreme learning machine with parameters, support vector machine based on particle swarm optimization(psoSVM) and back propagation algorithm based on particle swarm optimization(psoBP).The results are as follOws:1)With the ECoG signals,the time-frequency energy features based on stet got the classification accuracy, up to0.97,which was better than the classification accuracy based on empirical mode decomposition.2)Among the psoSVM,psoBP and pELM,the classification accuracy of the EEG Was appoximately0.85, the classification accuracy between seizure and subclincal seizure recorded with ECoG was approximately0.82.3)Between the seizure and subclinical seizure in ECoG recordings,the classification accuracy is0.98for psoSVM,0.97for pELM,and the trainibng time was28.1seconds for psoSVM and0.8s for pELM respectively.However the classification accuracy based on psoBP was significantly changing,which resulted in unstable performance. The performance of pELM is better than psoSVM and psoBP algorithms in terms of computation time and classification accuracy.4)Use the pELM to detect seizure In three ECoG segments, the false detection rate is1false deteclions per24hour period,and the seizure onset delay is0.1.The results show that the features definition based on STFT is efficient,and this approach proposed in this paper call defect epileptic seizures accurately in real-time.