Research On Algorithms Of Automatic Seizure Detection

Epilepsy is a chronic neurological disease with recurrent and sudden epileptic seizures, which is caused by intermittently abnormal neuronal electrical discharges in a group of brain cells. It clinically produces the disturbance of movement, sensation, or mental function, such as loss of consciousness and whole-body convulsion, which greatly threatens the life of person. It is reported that more than 1% of the world’s population suffers from this disease. The etiology of seizure is complex and the human knowledge is still insufficient to understand the pathogenetic mechanism of epilepsy. As an essential tool for investigating the epilepsy, electroencephalogram (EEG) records the electrical activities of nerve cells in the brain and contians a mass of physiological and pathological information, which plays an important role in the diagnosis, epileptogenic zone location and treatment of epilepsy.At present, the analysis of EEG usually depends on the visual inspection of neurologists according to clinical experience. But, massive amounts of EEG recordings make this task very tedious and time-consuming, and the subjectivity of neurologistsmay will influence the judgement of seizures. Hence, the computer-assisted automatic EEG analysis and seizure detection system is necessary and urgent, which can relieve the workload of neurologists and assist the diagnosis of epilepsy. Moreover, it has great significance for treating the epilepsy, improving the lives of patients and revealing the pathogenesis of epilepsy.This thesis conducts a comprehensive research for the field of automatic seizure detection, carries out some studied in the aspects of EEG dissimilarity distance feature extraction, EEG classification based on sparse representation and collaborative representation, and several algorithms are proposed for automatic seizure detection. The works and main contributions of the thesis are summarized as follows:(1) The distance metric is introduced to EEG analysis and a novel automatic seizure detection algorithm is proposed based on diffusion distance and Bayesian Linear Discriminant Analysis (BLDA). In this method, the wavelet transform is performed to the EEG signals for time-frequency analysis and three detail sub-signals are selected to form EEG distribution. The diffusion distances between seizure and non-seizure EEG distributions are computed to quantitatively measure the differences between seizure and non-seizure signals. Based on the property that the dissimilarity between the same class of EEG signals is lower than that between different class of EEG signals, the diffusion distances are fed into BLDA classifier to detect seizures. Compared with the Earth Mover’s Distance (EMD), the diffusion distance has not only powerful ability to distinguish seizure and non-seizure signals but also stronger anti-noise performance and lower computational complexity. The BLDA applies regularization to prevent overfitting for possibly noisy datasets, which lead to better classification effect. The experimental results on the Freiburg long-term EEG database manifests the validity of the diffusion distance as EEG feature and indicates that the proposed seizure detection method has promising classification ability.(2) On the basis of the theory of sparse representation, an automatic seizure detection algorithm based on multilayer kernel collaborative representation is developed. This algorithm applies wavelet transform to decompose EEG into multilayer sub-signals and the collaborative representation is combined with kernel method to classify EEG sub-signals on selected layers, then a multi-decision rule is presented to fuse the results of three channels and three layers, which constructs a multilayer kernel collaborative representation based classification system. In the scheme of kernel collaborative representation based classification, the sparse vectors of testing samples over the training dictionary are obtained by solving l2-minimization problem and the residuals between raw EEG signals and their reconstructions with each class are compared to classification, which avoids the choice of EEG features and the design of classifier in conventional detection methods. The kernel function enhances the separability between seizure and non-seizure EEG signals, which makes the classification more efficacious. In addition, it reduces computational burden greatly as well as retains the performance of classification via applying l2-minimization instead of l1-minimization. The design of multilayer system and decision rule make an effective combination of EEG signals in the frequency domain and spatial domain, which further improves the detection accuracy. The proposed method is evaluated on the Freiburg EEG dataset and shows its notable performance with high sensitivity and low false detection rate. The satisfying results shows that this algorithm is efficient and high real-time for seizure detection.(3) The sparse representation of symmetric positive definite (SPD) matrix is studied and a novel algorithm based on log-Euclidean Gaussian kernel sparse representation is proposed to detect seizures. The covariance descriptor is used to model and represent all channels of EEG data and the covariance matrix is modified to become the SPD matrix. Since the space of SDP matrix is a non-linear Riemannian manifold, the log-Euclidean Gaussian kernel function is applied to embed it into a linear reproducing kernel Hilbert space (RKHS) for performing sparse representation, and the classification of testing samples is achieved by computing the reconstructed residuals. In this work, the covariance descriptor combines the statistical properties of EEG signals in time domain, frequency domain and spatial domain, as well as suppresses noise effectively. Unlike the traditional sparse representation of vector data in Euclidean space, the sparse representation of SPD matrix is reasonable and valid by the reason that the log-Euclidean Gaussian kernel function takes the geometric manifold structure into consideration. In addition, the traditional sparse representation based detection method need the iterative analysis for processing every channel of EEG signals, which has the the drawback of repeat operations and complex system design. This algorithm can handle multile channels of EEG signals synchronously by the sparse representation of SPD matrix, which overcomes the defects of traditional method and reduces the computational burden greatly. The experimental results on the long-term EEG dataset indicate that the proposed method has better detection performance and stronger robustness, as well as faster speed, which satisfies the clinical demands of real-time detection system.The researches in this thesis are conducive to promote the study of EEG analysis and automatic seizure detection in theory and clinical application, which make a contribution in the development of automatic seizure detection technique. Due to the limitation of the EEG database, the effectiveness and robustness of the proposed seizure detection algorithms remain to be further verified and improved.