Research On Classification And Recognition Of Epileptic EEG Based On Sparse Representation And Feature Extraction

Epilepsy is a neurological comprehensive disease that attacks the brain abnormal discharge of neurons to disrupt normal biological electrical signals in the nervous system, and the external form of convulsions, language barriers, trances and terrible confusions, which has a serious influence on life and work of patients. The EEG(Electroencephalogram) contains important information about physiological and pathological, and it can be obtained conveniently and has no traumatic injuries for patients, which is currently the most valuable auxiliary examination in the epilepsy diagnosis. In the clinical diagnosis of epilepsy, it usually relies on visual inspection. But because of the uncertainty of the seizures, it often needs long-term real-time monitor of the patient's brain waves. So the testing time is longer and the efficiency is low, and according to the doctor's clinical experience and subjective judgment, it is easy to appear inconsistent inspection situation. Therefore, the identification of seizures with higher precision to realize automatic detection of ictal EEG is worth effort.Combined with the complexity of the nervous system and nonlinear and non-stationary characteristics of EEG signals, this paper studied EEG signals classification problems from two aspects: nonlinear feature extraction method and sparse representation based classification method. The nonlinear dynamic method mainly studied the different functional status of the brain, which can reveal the seizure and abnormal discharge mechanism in the cerebral cortex. The study of complex behavior and the inherent law of epileptic EEG signals provide a new thought and method for classification. Recurrence quantification analysis, as a kind of nonlinear recursive analysis method, describes quantitatively the system periodically embedded process and change rule of certainty. When the training sample space is large enough, the objects can be approximation to represented by the same samples subspace. Sparse representation is from the compression perception, breakthrough the limitation of the Nyquist sampling theorem, by using the l1 norm minimization constraints to compare reconstruction error of test sample with different categories of training samples, which to realize the high precision EEG signal classification. The research content and innovation points are as follows:(1) This paper proposed epileptic EEG signals feature extracted method based on recurrence quantification analysis(RQA). We did phase space reconstruction for the sample data firstly and then calculated the recursive figure of epileptic EEG signals to extract RQA parameter values as nonlinear characteristics, selected the appropriate classification threshold directly classified ictal and interictal EEG signals. In order to further improve the accuracy of epileptic EEG automatic detection, we adopted the method of combining the nonlinear and linear features, combining RQA quantitative values with the variation coefficient and fluctuation index as the input of SVM, and realize the automatic detection of epileptic EEG. Experimental results showed that the extracted RQA features can well reflect two different nonlinear dynamic characteristics of ictal EEG signals and interictal EEG signals.(2) The classification and recognition of epileptic EEG based on sparse representation and the wavelet transform was proposed. Sparse representation pursued the minimum number of dictionary elements to represent the test sample by building a complete dictionary and solve the l1 minimize problem. Firstly, epileptic EEG signals were preprocessed. Secondly, we used wavelet transform to obtain different frequency subbands under multi-scale decomposition of epileptic EEG signal samples. Based on the time-frequency analysis characteristics of wavelet frequency subbands to construct the dictionary, the SRC model with wavelet subband were expressed as a linear combination of dictionary atom, which help to recognize epileptic EEG features of different frequency ranges. Finally, the reconstruction errors with ictal and interictal EEG signals were calculated separately, the test sample was determined by minimal error.(3) Epileptic EEG signal classification based on K-SVD dictionary learning and sparse representation was proposed. Because of the EEG signal was smooth, first of all, we used the EMD to decompose epileptic EEG signal into a finite number of intrinsic mode function(IMF), each IMF component contained the original signals' local characteristics of different time scales. Extracted the IMF component signal features such as variation coefficient and fluctuation index to construct the dictionary. Then, orthogonal matching pursuit algorithm was used to test samples in the current dictionary of sparse representation, and the K-SVD algorithm was used to learn feature dictionary and coefficient of sparse representation. Then, the test sample was reconstructed by dictionary characteristics and corresponding sparse coefficient vector of ictal EEG signals and interictal EEG signals respectively. Finally, the redundancy error to category of the given test samples were judged to realize the automatic classification of epileptic EEG. The dictionary based on the frequency characteristics of the IMF components greatly reduced the data dimension and complex computation, which improved the performance of epileptic EEG signals detection.