Research On Fracial Analysis Of Epiieptic EEG And Automatic Seizure Detection Methods

Epilepsy is a common chronic neurological disorder affecting about 1% populations of the world. The recurrent and sudden epileptic seizures bring huge suffering and injury for epilepsy patients. Epileptic seizure is a result of abnormal excessive and hypersynchronous neuronal discharges in the brain and 80% of epileptic patients have abnormal waves in EEG. Thus the EEG examination and analysis are important tools for epilepsy diagnosis, focus localization and the judgment of seizure types. Meanwhile, the study on computer-aided automatic methods of epileptic EEG analysis and seizure detection is significant both in improving the diagnosis efficiency and developing closed-loop electric simulator.As a total reflection of the brain neurons’ electric activities, EEG signals recorded on scalp surface or in cerebral cortex have complex non-linear characteristics. Although the non-linear analysis of EEG has been received the attentions from epilepsy researchers, the study on the fractal characteristics of EEG signals is still fewer. Fractal theory is an important branch of non-linear science, and the fractal analysis of EEG signals can further clear the nature of chaos dynamic activities of the epileptic brain. In addition, because the mechanism of epileptic seizure is very complex and the seizure types and processes are various, the existing automatic detection technologies of epileptic seizures do not meet the demand of clinical application in aspect of accuracy, real-time performance and robustness. To resolve the problems in EEG analysis and seizure detection, this work systematically studies the fractal and multi-fractal characteristics of epileptic EEG, and introduces the advanced learning methods and classification models to research new seizure detection methods with high accuracy and good real-time performance. The research contents include the following aspects.Firstly, the evolution law of EEG signals’Higuch fractal dimensions in the preictal stage is studied and explained from the epileptic seizure mechanism. The variation of EEG signals’Higuch fractal dimensions can be used as an indication of the impending seizure and a seizure prediction algorithm is proposed based on the dimension variation and Bayesian linear discrimination analysis (BLDA). The prediction algorithm obtained a high sensitivity and a low false prediction rate when it was tested on Freiburg EEG database. In addition, the seizure prediction algorithm has a low compute cost.Secondly, the K-nearest-neighbor fractal dimensions of EEG signals are found having significant statistic difference between ictal and inter-ictal stages. Based on the dimension difference, a seizure detection algorithm was proposed through introducing a gradient boosting method. The detection algorithm was tested on Freiburg EEG database and obtained a high sensitivity and a low false detection rate. In addition, the average detection time delay of the seizures’onset was only 2.46 seconds.Thirdly, the study on the fractal characteristics of EEG signals based on a fractal dimension is further extended to the multi-fractal analysis through a multi-fractal spectrum, which can describe the local singularity and the inhomogeneity of fractal EEG. After proving that EEG signal has multi-fractal characteristics, the physical significances of the spectrum parameters (a0、αmin、αmax、Δα、f(αmin)、f(αmax)、Δf R) are explained. The significant statistic difference between ictal and inter-ictal EEG can be found on all the multi-fractal spectrum parameters by a comparative study.Finally, the multi-fractal spectrum features of epileptic EEG signals are combined with Relevance Vector Machine (RVM) to propose a seizure detection system integrating the multi-channel discrimination results. In the post-processing stage of the seizure detection system, the ictal probabilities of EEG segments outputted by RVM are integrated on multi-channels, which conforms the diagnosis process of clinicians and can make the detection more accurate. The proposed seizure detection system obtained satisfactory results on the sensitivity and the recognition rate when it was tested on Freiburg EEG dataset. Further more, the system’s computational complexity is low enough that the seizure detections created by the system is real-time. The running time of the system on one-hour three-channel EEG is only about 1.2 minutes.In this thesis, the Higuchi algorithm and the K-nearest-neighbor algorithm used in the computation of EEG fractal dimension both are not need of reconstructing phase space, thus have a low computational complexity. In addition, the Moment method used for multi-fractal analysis also has the advantage of computing simply compared with the multi-fractal detrended fluctuation analysis method (MF-DFA) commonly used in other research fields. Therefore, the time needed for features extraction in the epileptic seizure detection algorithms proposed in this paper is greatly reduced, which promotes that the seizure detection is real-time. On the other hand, BLDA, gradient boosting method and RVM are introduced into the seizure detection algorithms in this work. Since these learning method and classification models are advanced and have good classification accuracy, the detection algorithms proposed in this work can classify the epileptic EEG patterns more accurately. All in all, the study made in this work further carry forward the non-linear research of epileptic EEG, provide new ideas for the research of seizure detection methods with high accuracy and good real-time performance. In the follow-up work, the seizure detection algorithms proposed in this work will be further tested and perfected based on a large number of EEG data recorded from epileptic patients in clinic.