EEG Signals Classification Based On Entropy Measures And SVM

Electroencephalography (EEG) is the electrical activity of nerve cells in the cerebral cortex or overall reaction on the surface of the scalp, which contains important informations about the conditions and functions of the brain. EEG is fundamental for diagnosing nervous diseases. The electrical activities of the brain usually exhibit complex behavior with nonlinear dynamic properties. Epilepsy is a common chronic disease of brain, causing great damages to patients’mental and physical health. Epileptic seizure led to central nervous system dysfunction caused by transient abnormal excitement and excessive synchronization discharge. For scalp EEG, the epileptic EEG detection methods are usually based on the feature extraction methods and the classification algorithms.There is growing evidences that the brain is a complex nonlinear dynamical system, and nonlinear dynamical methods have been widely applied to analyze the EEG signals. Based on the study of nonlinear methods, this paper proposes two nonlinear analysis methods:the feature extraction of epileptic EEG signals based on approximate entropy and modified approximate entropy; the feature extraction of epileptic EEG signals based on sample entropy and modified sample entropy. Firstly, calculating the entropy values of the ictal EEG and interictal EEG, such as approximate entropy, modified approximate entropy, sample entropy and modified sample entropy. Secondly, the four entropies are applied as features to analyze the EEG signals of epileptic parents. Entropic measurements are nonlinear time analysis methods based on complexity. Entropies don’t depend on the record length to estimate the stability of the value, and have the ability of anti-interference and resisting noise. Modified entropies methods of the fuzzy function based Heaviside function have solved the transient problem from0to1, to a great extent improving the efficiency of distinguishing the time series.Due to the different of complexity between epilepsy patients and healthy person, two methods are employed to classify these two kinds of signals. The one is to choose an appropriate threshold value as the discrimination criteria; the other is based on the four entropies and SVM to discriminate epileptic seizure signals from epileptic EEG automatically.Statistical learning theory based SVM (Support vector machine) is a new machine learning tool. With the VC dimension theory and structural risk minimization principle, SVM has a good performance on the analysis of high dimension nonlinear system, which is widely applied in the analysis of epileptic EEG signals. For nonlinear classification problems, a new higher dimensional feature space is reconstructed form the original signals where the linear separability of projected samples is enhanced. The optimal hyperplane could separate the samples without error and maximize the distance from either class. To improve the accuracy of epileptic EEG automatic detection, SVM is used as a classifier to distinguish epileptic seizure signals from epileptic EEG. The classifier performance depends on the features, such as approximate entropy, modified approximate entropy, sample entropy and modified sample entropy.The experimental results show that two kinds of nonlinear feature extraction methods could effectively distinguish between the ictal EEG signals and interictal EEG signals. The classification accuracies of feature extraction of epileptic EEG signals based on the approximate entropy are increasing from0.15times the standard deviation to two standard deviations, with the highest classification accuracy of87.25%. The classification accuracies of nonlinear classification using support vector machine than linear classification using the nonlinear characteristics are improved. The classification accuracies of combining the linear with nonlinear features than only the nonlinear characteristics are also improved. Combined four entropies, the accuracy are up to96.00%for interictal and ictal EEG signals. In General, the combination of linear and nonlinear characteristics has the higher accuracy of epileptic EEG.