| Epilepsy is a noninfectious chronic neurological disease accompanied by abnormal discharge of brain neurons.The clinician completes the detection and diagnosis of epilepsy by visually detecting the patient's 24-hour electroencephalogram(EEG),which contains rich physiological and pathological information.Due to the shortcomings such as time-consuming,energy-consuming and strong subjective factors in visual detection as well as the characteristic that EEG signals can generally reflect the electrophysiological activities of brain nerve cells,the epilepsy detection technology combined with signal processing and pattern recognition has become a research hotspot.In order to reduce the damage caused by epileptic seizures,on the premise and basis of the classification and detection of epileptic events by EEG signal after epileptic seizures,applying the long-term EEG records of epileptic patients to detect the seizure aura characteristics that cannot be observed visually can complete the forecast task of epilepsy.The suddenness and high risk of epilepsy seriously affect the physical and mental health of patients,increase the workload of medical staff,and increase social risk indicators.The early prediction of epilepsy is the bottleneck in the treatment of epilepsy.Studies have shown that the brain pattern of the epileptic has changed before the onset of epilepsy,so the detection of abnormal changes in neural activity within a specific time interval before the seizure can effectively achieve the epileptic forecast.Therefore,some measures,such as predicting seizure events before the epileptic seizures,and performing interventional control such as drug delivery or electrical stimulation by the medical staff timely after the alarm signal is detected,can reduce the number of seizures and improve the quality of life,which is of great research significance and clinical value for the early intervention and treatment of epilepsy diseases.Currently,the research of epilepsy prediction algorithm based on EEG signals is in its infancy.In actual clinical application,there are many problems,such as high false prediction rate and poor universality.In addition,the complex and diverse EEG signals also bring challenges to epilepsy research.In this paper,based on the detection of EEG signals in various periods of epilepsy,the prediction analysis on epilepsy can be realized through the identification of precursor features from epileptic seizures in a specific time interval.Researches on the stability,complexity and universality of the existing clinical diagnosis-oriented EEG signal detection algorithms are respectively carried out,and the corresponding models are established to explore the effectiveness of the methods.The main research work and innovative findings of this paper are as follows:(1)In order to solve the problem that the insufficient representation ability for epileptic EEG signals leads to massive differences in the recognition results of analysis algorithms for multiple classification tasks,the detection algorithm of the epileptic EEG signals based on scattering transform is proposed.Specifically,the scattering transform is combined with the analysis characteristic of wavelet transform and complex field,so the signal features with time-shift invariance and local stability are obtained by cascading complex wavelet decomposition and the local weighted average method.The iterative decomposition of multiple scattering paths in different directions and scales contributes to improving the stability of the characterization ability.The fuzzy entropy and log energy entropy features of scattering transform domain are applied to obtain complementary representations of multiple epileptic EEG signals in different periods,and the effective dynamic characteristics that can distinguish the epileptic seizure signal from other periods are fully excavated.On the Bonn University EEG dataset,the extreme learning machine classifier is utilized to complete the eight different "seizure-other" classification tasks,achieving the following evaluation indexes: not less than 99.56% sensitivity,99.50% specificity,99.50% accuracy,and 0.99 Matthews correlation coefficient.The stable recognition results indicate that the proposed algorithm can effectively characterize epileptic EEG signals in different periods,and the distinguish ability of epileptic seizure signals has been effectively improved.(2)Aiming at the high complexity and the need for artificial experience to select features of the existing detection models,this paper proposes the detection algorithm for epileptic EEG signals based on symplectic geometry.Through the symplectic similarity transformation in symplectic space,the self-adaptive feature extraction of different types of epileptic EEG signals is accomplished directly,avoiding the defect of the artificial feature design.As a regular transformation in Hamiltonian system,the symplectic similarity transformation can maintain the measurability and basic characteristics of the original EEG signal.The obtained feature vectors are orthogonal to each other.Moreover,the essence of nonlinear transformation of symplectic geometry algorithm is more suitable for dynamic analysis on epileptic EEG signals,which greatly reduces the complexity of the model while improving the representation ability.The symplectic feature vector is sent to the k-nearest neighbor classifier to complete the classification tasks.In the ten clinical multi-classification tasks of the epilepsy EEG dataset from Bonn University,the evaluation indicators of sensitivity,specificity,accuracy and Matthews correlation coefficient are not less than 99.17%,99.17%,99% and 0.96,respectively.In the " epileptic seizures/non-epileptic seizures" task of the multi-lead scalp EEG database from 23 subjects in CHB-MIT,the average performance of the above indexes are 97.17%,99.72%,99.62% and 0.92,respectively.The experimental results obtained in the long-range and short-range data sets respectively verify the higher classification accuracy and lower complexity of the proposed detection model,which lays a foundation for the development of an auxiliary diagnosis system for epileptic seizures.(3)For the purpose of completing the warning to patients before the onset of seizures,on the basis of realizing the stability and low-complexity of detection algorithm of epileptic EEG signals in different periods in the preliminary work,the research on the seizure prediction based on the aura characteristics of epileptic seizure within a specific time interval is carried out.In order to solve the problem of poor universality in multiple subjects,the epileptic seizure prediction algorithm based on synchroextracting chirplet transform is proposed.By means of combining the advantages of short-time Fourier transform reversibility and ideal time-frequency representation sparsity,a relatively ideal and high-resolution time-frequency representation of epilepsy EEG signals is obtained.The time-frequency ridge with high energy concentration is obtained by introducing the chirp rate.Only the time-frequency information most relevant to the time-varying characteristics of the original signal is retained by discarding the time-frequency coefficients of the energy diffusion area.Finally,the symplectic geometric decomposition algorithm is applied to obtain the effective precursor features of the seizures onset,and the seizure prediction is completed by the support vector machine.Under the conditions of 1min seizure prediction horizon and 30 min seizure occurrence period,an average sensitivity of 90.92% and an false prediction rate of 0.14/h are achieved on 83 seizure events from 17 subjects in the scalp epilepsy EEG database of CHB-MIT.The average sensitivity and false prediction rate obtained on the i EEG data of all subjects in the Kaggle epilepsy prediction competition dataset are 91.5% and 0.16/h,respectively.The universality of the proposed seizure prediction algorithm is verified on various types of EEG signals of multiple subjects in different epilepsy prediction datasets,providing a new solution for the clinically-oriented epilepsy seizure prediction algorithm.In summary,based on the research on the detection of posterior epileptic seizures of epilepsy EEG signals in different periods,this paper proposes an epilepsy EEG signal analysis algorithm model for early prediction tasks according to actual clinical application requirements based on epilepsy EEG signal research,realizing the accurate and reliable predictions to patients before epileptic seizures.The work of this paper lays the theoretical foundation for the early prediction algorithm of epilepsy based on EEG signals,and provides a solution for the development of the clinical early predicting treatment system for epilepsy in the next step. |
PDF Full Text Request