Research On Epilepsy Detection Based On Adaptive Decomposition And Deep Fores

Epilepsy is a chronic disease of the brain characterized by uncontrolled neuronal hyperexcitability that spreads from the lesion to the surrounding brain tissue,causing a variety of motor,sensory,cognitive,and psychological abnormalities.In addition,severe seizures can induce lasting changes in brain structure and function or even endanger the life of person.Thus,making early diagnosis and effective control of epileptic disorders is critical to maintaining long-term health.Electroencephalography(EEG)is an important tool in the diagnosis of epilepsy,which can reveal much information about the disease pathology.In clinical practice,detecting and assessing seizures by visual observation is time-consuming and labor-intensive.The use of computers for assisted detection and identification of seizures could facilitate early diagnosis and treatment,which helps patients improving the quality of life timely and effectively.In this paper,based on the topic of automatic seizure detection,we analyze the adaptive decomposition method of EEG signal and deep forest algorithm,and propose a series of automatic seizure detection algorithms.The main research includes the following three aspects:(1)Analyzing and comparing the performance of three time-frequency decomposition methods in classifying EEG signals.This study applies Discrete wavelet transform(DWT),Empirical mode decomposition(EMD),and Variational mode decomposition(VMD)to decompose the EEG signals into sub-signals of different frequency bands,and extracts five effective EEG features: mean,line length,skewness,average absolute deviation,and sample entropy.Then,they are combined into feature vectors and sent to three classifiers,Bayesian linear discriminant analysis(BLDA),Random forest(RF)and Support vector machine(SVM),to conduct EEG signal classification.The algorithm was evaluated on the Bonn database and the Bern-Barcelona database.The results indicate that VMD has better classification performance compared to the other two time-frequency decomposition methods.(2)To further explore the relationship between seizures and non-seizures in EEG signals,this paper proposes an automatic detection algorithm for epileptic seizures based on Variational mode decomposition(VMD)and Deep forest(DF).Firstly,the VMD is performed on the original EEG signal,and the first three Variational modal functions(VMFs)are selected to construct the EEG time-frequency distribution.Then,the Log-Euclidean covariance matrix(LECM)is calculated to represent the characteristics of EEG and form the EEG features.Finally,the deep forest model is applied to complete the classification of EEG signals.In addition,to improve the accuracy,the discriminant results are generated by post-processing with moving average filtering and adaptive collar technique extension operations.The algorithm achieved good performance on both the Bonn database and the Freiburg long-term EEG database.(3)To address the issues of high computational complexity in Gaussian Mixture Model and the excessive memory usage and loss of original feature information in traditional Deep Forest methods,this study proposes an automatic detection algorithm for epileptic seizures based on LogEuclidean Gaussian Mixture Models(LE-GMMs)and the Multiple Pooling and error Screening Forest(MPSForest)algorithm.In this algorithm,VMD is also applied to decompose the EEG signal into five layers,and the first three layers are selected for constructing the EEG timefrequency distribution.Then its Gaussian mixture model is estimated and LE-GMMs are constructed to extract the effective EEG features.These features are fed into the MPSForest model to achieve classification recognition of seizure and non-seizure data.After that,the output results are post-processed to obtain the final detection results.Experimental results have demonstrated that the automatic epileptic seizure detection algorithm proposed in this study has achieved excellent performance on both the UPenn and Mayo Clinic database and the Freiburg long-term EEG database.The research results in this thesis extend the application of adaptive decomposition and deep forest algorithm in epilepsy EEG signal detection,which effectively contribute to the progress and development of automatic seizure detection system.However,due to the limitation of experimental data,our proposed automatic seizure detection algorithm still needs further optimization and improvement in actual clinical research.