Research On The Feature Extraction And Identification From The Spatiotemporal Patterns Of Brain Oscillations

Brain science and brain-inspired intelligence aims to develop brainmachine intelligence and diagnose brain diseases based on the cognition mechanism of brain,which is one of the main frontier fields of science and technology in the Long-Range Objectives through the Year 2035 in China.Brain oscillations directly reflect the regular discharges from neuronal ensembles.Their diverse spatiotemporal patterns are of significant theoretical value and practical importance for advancing research in brainmachine intelligence and assisting in the diagnosis and treatment of brain diseases.Scalp electroencephalography(EEG)is widely used in physiological state identification.Although the features of scalp EEG have been used to assess physiological states and diagnose brain diseases,the inter-subject variability in EEG signal patterns limits the generalization capability of artificial intelligence models.Intracranial EEG(iEEG)is mainly used in epileptic tissue localization for refractory patients.At present,the localization primarily relies on the evolving patterns during seizure onset and the localization potential from large amount of interictal data remains unexplored.High-frequency oscillations(HFOs)act as promising biomarkers closely related to the epileptic tissue.However,particular challenges in spatiotemporal pattern analysis of HFOs are posed due to their short transience with various waveforms and high randomness.In addition,with the epileptic network hypothesis in ictogenesis widely confirmed,highly active nodes in the network exhibit rich spatiotemporal activities within and outside the epileptic tissue,which can provide crucial information on the epileptic tissue and postoperative outcomes.In order to address the above-mentioned challenges,this thesis conducts research on the spatiotemporal patterns of brain oscillations.First,it achieves model generalization for physiological state identification through covariance matrix optimization and efficient classification of highdimensional spatiotemporal features.Subsequently,combined with the characteristics of brain oscillations in high-frequency band,it comprehensively characterizes the activities of the nodes in the epileptic network from spatial,temporal,and spatiotemporal aspects,which enables a deep understanding of the dynamics of the epileptic network and enhancing the reliability of epileptic tissue localization.The research encompasses four main aspects:1.With respect to the generalization problem in physiological state identification,multi-source signal alignment and tensor network(MSSATN)framework is proposed in Chapter 2.Domain adaptation algorithms and deep neural networks are currently the two main approaches to improving model generalization.However,the feature extraction process in domain adaptation often leads to the loss of spatiotemporal information in EEG signals,and deep neural networks involve tremendous parameters,making the training and optimization processes complex.This thesis proposes MSSA-TN framework:first,the covariance matrices are minimized to reduce the signal distribution discrepancy among subjects.Then high-dimensional features of EEG signals are efficiently classified by optimizing the parameters of the designed tensor network.In the generalization experiments conducted on the simulated driving EEG dataset released by the University of Technology Sydney in 2019,compared with existing algorithms,the accuracy is improved by at least 3.71%.Furthermore,when coupled with convolutional modules in neural networks,the framework enables automatic feature extraction and optimization,resulting in a 2.73%increase in accuracy on the mental workload dataset released by the French Institute for Research in Computer Science and Automation in 2021.2.With respect to the problem of spatial pattern characterization and separation in epileptic tissue localization,spatial pattern clustering of high frequency activity(SFC-HFA)is proposed in Chapter 3.The dynamic activities of nodes in the epileptic network are closely linked to the spatial patterns of high-frequency activities.Current research lacks the quantitative measurement and separation of such spatial patterns utilized in improving epileptic tissue localization.This thesis proposes SPC-HFA algorithm:First,the intensity of pathological HFOs is quantified with skewness to construct the spatiotemporal feature matrix.Next,k-means clustering is employed to assess and separate the intrinsic spatial patterns in column vectors.Finally,the most active cluster centroid is selected for epileptic tissue localization.In the localization experiment conducted on the epileptic patient dataset released by University Hospital Zurich in Switzerland in 2017,compared to other algorithms,SPC-HFA demonstrates improvement in terms of area under the curve(effect size>0.2)and achieves the best localization results in 10 patients.Furthermore,SPC-HFA can be effectively extended to HFO detection algorithms and enhance their localization performance(effect size≥ 0.48).3.With respect to the problem of the functional connectivity(FC)analysis in high-frequency band in epileptic tissue localization,the skewness-based functional connectivity(SFC)analysis is proposed in Chapter 4.FC analysis plays a crucial role in understanding the mechanisms of ictogenesis and network characteristics.However,existing methods are suitable for low-frequency band(<80 Hz).To enable effective FC analysis in high-frequency band,this thesis proposes SFC algorithm:Based on the feature matrix in Chapter 3,SFC characterizes the temporal correlations between nodes using rank correlation to construct the FC network.Subsequently,it employs the node connectivity strength for localization and postsurgical outcome evaluation.In the experiments conducted on the datasets released by institutions such as University Hospital Zurich in Switzerland in 2017 and National Institutes of Health in the United States in 2021,significant difference(p<0.001)is found in node connectivity strength within and outside the epileptic tissue.SFC achieves better results in both overall and individual localization as well as outcome evaluation compared to low-frequency band.4.Based on Chapter 3 and 4,this thesis proposes the algorithm for dynamic spatiotemporal pattern analysis in epileptic tissue localization.The algorithm uses non-negative matrix factorization to capture the dynamic spatiotemporal patterns simultaneously and evaluates the effectiveness in epileptic tissue localization。In the experiment conducted on the dataset released by Hospital of the University of Pennsylvania in the United States in 2022,when the number of patterns is set to 5,the algorithm achieves the best localization performance in general.The average areas under the curve are 0.79、0.82、0.83 and 0.78 in interictal,preictal,ictal and postictal phase,respectively.Compared to the algorithm presented in Chapter 3,this approach further enhances the reliability of localization except for the ictal phase(effect size>0.5).The research in this thesis is laid on the foundation of prior knowledge on brain oscillations.It aims to improve the identification and localization ability through novel signal processing and artificial intelligence algorithms.Although the effectiveness of the proposed algorithms has been validated on open-source datasets,integration with clinical medicine and neuroscience is still needed to further refine the algorithms.Within the trend of open-source EEG acquisition systems and datasets as well as standardization of data formats and processing pipelines,future research will fully leverage the potential of big data analysis and deep learning,integrate multiple disciplines,and further advance the translation and application of brain science and brain-inspired intelligence.