Research On Automatic Seizure Detection Algorithm Based On Tensor Analysis

Epilepsy is a brain disease caused by abnormal discharge of neurons in the brain,which is one of the most common neurological diseases.Seizures onset is usually characterized by clenching teeth,limb convulsions which also could cause urinary incontinence,loss of consciousness,and even life-threatening in severe cases.Electroencephalogram(EEG)is the main tool for diagnosing epilepsy containing a lot of physiological and pathological information,which can provide diagnosis guidance for doctors by analyzing the EEG signals.Currently,the clinical diagnosis of epilepsy is mostly based on the doctor's own knowledge and experience,meanwhile the onset and end of seizures are judged by visually observing the original EEG recordings.However,EEG recording is usually uninterrupted more than 24 hours resulting in the amount of data is too large.It is not only difficult and time-consuming for doctors to judge epileptic seizures by naked eye observation,but also puts forward high requirements on the medical level of doctors.Therefore,it is of great significance to construct an automatic seizure detection system for computer-assisted diagnosis and treatment.This paper proposes a new method for automatic seizure detection,which is based on tensor distance and Bayesian linear discriminant analysis(Bayesian LDA).EEG tensor constituted by multi-channels EEG signals fully considers the spatiotemporal information of multiple channels,which can effectively improve and promote the performance of seizure detection.Meanwhile,Tucker decomposition is performed to extract principal components of the EEG tensor and reduced the complexity of algorithm effectively.The tensor distance feature calculated based on the EEG core tensor conforms to the spatial characteristics of the multi-channels EEG signals and reflects the spatial information of EEG tensor which can effectively characterize the difference between seizure and non-seizure EEG signals.Firstly,discrete wavelet decomposition is operated to get the time-frequency characteristics of EEG signals in this method and the tensor representation of EEG signals is constructed.Then,the principal components of the EEG tensor are obtained by Tucker decomposition and the tensor distance between different types of EEG tensors is calculated.Finally,the tensor distance is input into the Bayesian LDA classifier for training and classification,and the output results of Bayesian LDA are processed by post-processing technologies to get the final detection results.This algorithm is evaluated on the long-term EEG dataset from the Epilepsy Center of the University Hospital of Freiburg,and the sensitivity of 95.12%,specificity of 97.60%,recognition accuracy of 97.60%,as well as the false detection rate of 0.76/h were obtained,which showed good detection effect and its potential clinical application.The research work of this article expands the application of tensor analysis in EEG processing and seizure detection with the high-dimensional data form of EEG tensor and effectively promotes the research and development of automatic seizure detection system.Due to the limitation of experimental data,the automatic seizure detection algorithm proposed in this paper needs to be further verified the performance by more clinical EEG data.