Research On Epileptic Features Extraction Using Time-frequency-nonlinear Analysis Joint Brain Network

Epilepsy is a neurological disease caused by the hypersynchrony of neurons in the brain.It has the characteristics of uncertainty,transient and repeatability,which brings huge hidden dangers to the life safety of patients.However,epilepsy early warning is still a huge challenge at present,due to the complex spatiotemporal propagation mechanism of epileptic seizures and the lack of comprehensive analysis of the spatiotemporal characteristics of epilepsy EEG.This paper deeply excavates epilepsy information from multiple dimensions such as time-frequency,nonlinearity,and brain network,constructs a spatiotemporal feature extraction and screening algorithm for epilepsy EEG,and verifies the feasibility of the algorithm on scalp EEG and intracranial EEG.The specific research contents of this paper are as follows:1.Time-frequency-nonlinear joint feature extraction and analysis.Combined with the characteristics of time-frequency and nonlinear two dimensions,fuzzy entropy was selected to measure the complexity of epileptic EEG signals.Power spectral density was used to measure the frequency domain energy change of epileptic EEG.Validation of two datasets found that the distribution of time-frequency and nonlinear joint features in the preictal and interictal states were different,and there were individual differences.But multiple epileptic seizures in the same patient conformed to the same trend.2.Feature extraction and analysis of brain networks.The spatial network characteristics of epilepsy EEG were analyzed from the perspective of the brain network.Phase-locked value was selected to measure the phase synchronization of epileptic EEG signals,and the functional brain network of epilepsy was constructed by adjacency matrix to reveal the brain network variety between interictal and preictal states.Through the validation for the two datasets,it can be found that the phase synchrony of multiple channels EEG signals in preictal states was higher than that of interictal.The small-world properties,global efficiency,average degree,clustering coefficient and eigenvector centrality of the preictal brain network were greater than those in the interictal period,but the characteristic path length was smaller than that in the interictal period.The results showed that the brain network had changed before the seizure,which became more compact,and the ability to communicate epileptic information between different brain regions was stronger.3.An epilepsy spatiotemporal feature screening algorithm and a classification performance evaluation model were established.Through evaluating from the two aspects of feature independence and informativeness,the features with high independence and rich information content were selected and combined into the spatiotemporal features of epilepsy,which were input into the support vector machine for training to identify preictal states.In the kaggle dataset,98.01% accuracy,0.96 AUC,98.3% F-Score and 3.83% false positive rate were obtained,and in the CHB-MIT dataset,95.93% accuracy,0.92 AUC,94.97% F-score and 4.73% false positive rate were obtained.Compared with the ablation experiments of separate time-frequency,nonlinear joint features and brain network features,the classification performance of spatiotemporal features was more excellent.The experimental results showed that the method proposed in this study can effectively extract the spatiotemporal information of epilepsy and identify the preictal states data.