Research On Nonlinear Dynamics, Causality Brain Network And Clustering Stability In The Application Of Eeg Signals Analysis

As an important part of the project,“Construction and abnormal featureresearch analysis of Depression EEG functional brain network”(No.61472270)and “Analysis and application research of Multi-mode functional brain complexnetwork”(NO.61373101) supported by National Natural Science Foundation ofChina, this thesis aims at exploring nonlinear dynamics patterns, dynamicchanges of characteristic causality brain networks and pattern recognition ofclustering stability theory in EEG signal for the specific brain cognitiveactivities. As we know, human brain is a nonlinear, network-like andnon-stationary complex system. All kinds of cognitive functional activities aredominated and processed by the human brain. By means of EEG imagingtechnology, nonlinear dynamics, brain network theory and clustering stabilitytheory are applied to EEG signal processing technology in this thesis, in order toreveal the correlation between emotional EEG signals and nonliear complexity,causality change law of brain network in error related negativity (ERN) andclustering stability analysis in motor imagery EEG signals. Overall, thecontributions of this thesis lie in the following three aspects.(1) This thesis presents a pointwise Lempel-Ziv complexity (pLZC)algorithm based on nonlinear dynamics, and then analyzes the emotional EEGsignals with the proposed pLZC algorithm. By pointwise difference in this way,the proposed method will obtain nonlinear complexity mode with correlation of emotionl cognitive activities in micro-scale and new coarse-graining method,and extract nonlinear features through computation of Lempel-Ziv complexity.In the view of statistic analysis, according to compare with classical andmulti-scale LZC algorithms, the experiment results show that the proposedpLZC algorithm dynamically characterizes the nonlinear features and accuratelydescribes the more complex strcture and component of emotional EEG signals.The proposed pLZC algorithm will be effective for the analysis and recognitionof emotional EEG signals, and it conducts a new perspective for nonlineardynamics application in EEG signals analysis.(2) This thesis puts forward a construction method of casuality brainnetwork, and applies it into the analysis for error related negativity EEG signals.The proposed method is based on indepent component analysis (ICA),multivariate autoregressive (MVAR) model and granger causality analysis(GCA), and jointly analyzes multi-region, multi-components generates from thecorresponding brain cognitive activities. The observed EEG signals aredecomposed by ICA, then we can obtain independent component sources (ICs)associated with the brain cognitive activities by source localization method.Furthermore, the effective connectivity among the ICs is establised by GCA, andthe casuality brain network is constructed by MVAR model in the brain sourcespace. Finally, the error related negativity EEG signals are analyzed with theabove casulity brain network. The experiment result proves that the proposedmethod of casulity brain network can obtain information flow among ICs fromthe perspective of effective connectivity, and detect the casuality effect amongthe ICs. Therefore, it has important significance for the construction of casualitybrain network to the attributive mechanism of ICs and the instantaneousinformation flow model of brain nerve system in the brain source space.(3) This thesis proposes a new affinity propagation cluster algorithm,termed stability-based affinity propagation (SAP), which is based on clusteringstability, and it is utilized for pattern recognition and automatically clustering of motor imagery EEG signals. This algorithm generates a set of candidatepreferences based on similarity between data points. According to eachcandidate preferences, the next step is computing clustering stability throughcomparing normalizaed mutual information (NMI) between different twoclusterings. Finally, according to normalized stability of the candidatepreferences, the optimal shared exemplar preference is selected to be the onethat generates the maximum normalized stability. We compare the proposed SAPalgorithm with the classical exemplar-based clustering method, namely k-centersand k-means performed on the motor imagery EEG signals. The experimentresult demonstrates that the SAP algorithm obtains more accurate clusteringperformance, and based on the clustering stability, the best shared exemplarpreference that generates the most stable clustering results is obtained. The studyalso shows that the proposed SAP algorithm is effective and robust to dataclustering of nonstationary EEG signals, and indicates a new direction for theclustering analysis on EEG signals.To sum up, this thesis focus on nonlinear complexity features in emotionalEEG signals, construction and analysis of casulity brain network in error relatednegativity EEG signals and clustering pattern recognition in motor imagery EEGsignals. This research provides new evidence and perspective for nonlineardynamics and brain network analysis, and extends the application of clusteringstability in EEG signals. This is a new research archievement of multi-domainand multi-discipline study.