EEG Microstate Sequences From Different Clustering Algorithms Are Information-Theoretically Invariant

Abstract

We analyse statistical and information-theoretical properties of EEG microstate sequences, as seen through the lens of five different clustering algorithms. Microstate sequences are computed for n = 20 resting state EEG recordings during wakeful rest. The input for all clustering algorithms is the set of EEG topographic maps obtained at local maxima of the spatial variance. This data set is processed by two classical microstate clustering algorithms (1) atomize and agglomerate hierarchical clustering (AAHC) and (2) a modified K-means algorithm, as well as by (3) K-medoids, (4) principal component analysis (PCA) and (5) fast independent component analysis (Fast-ICA). Using this technique, EEG topographies can be substituted with microstate labels by competitive fitting based on spatial correlation, resulting in a symbolic, non-metric time series, the microstate sequence. Microstate topographies and symbolic time series are further analyzed statistically, including static and dynamic properties. Static properties, which do not contain information about temporal dependencies of the microstate sequence include the maximum similarity of microstate maps within and between the tested clustering algorithms, the global explained variance and the Shannon entropy of the microstate sequences. Dynamic properties are sensitive to temporal correlations between the symbols and include the mixing time of the microstate transition matrix, the entropy rate of the microstate sequences and the location of the first local maximum of the autoinformation function. We also test the Markov property of microstate sequences, the time stationarity of the transition matrix and detect periodicities by means of time-lagged mutual information. Finally, possible long-range correlations of microstate sequences are assessed via Hurst exponent estimation. We find that while static properties partially reflect properties of the clustering algorithms, information-theoretical quantities are largely invariant with respect to the clustering method used. As each clustering algorithm has its own profile of computational speed, ease of implementation, determinism vs. stochasticity and theoretical underpinnings, our results convey a positive message concerning the free choice of method and the comparability of results obtained from different algorithms. The invariance of these quantities implies that the tested properties are algorithm-independent, inherent features of resting state EEG derived microstate sequences.

Determination of the Time Window of Event-Related Potential Using Multiple-Set Consensus Clustering

“…From the cluster analysis view of point, the various clustering strategies such as using the single clustering method on the different types of datasets; repeated clustering with a single clustering method and combining the results; and the multipleclustering methods applied to the individual dataset potentially affect the clustering quality (Abu-Jamous et al, 2013, 2015bLiu et al, 2015;von Wegner et al, 2018). To investigate this issue and reliably feeding consensus clustering, two datadriven based mechanisms were appropriated before multilevel cluster analysis.…”

Decoding Brain Dynamics in Speech Perception Based on EEG Microstates Decomposed by Multivariate Gaussian Hidden Markov Model

“…EEGLAB microstate plugin is a free, graphical user interface designed toolkit and developed by Koenig et al [48] for the identification and quantifications of EEG microstates in time-series EEG data. In this work, we used the two classical methods that appear in most of the current literature [49], i.e., modified K-means clustering and Atomize and Agglomerative Hierarchical Clustering (AAHC) [14,39], to extract a microstate sequence representing for the broadband EEG data. The EEG microstate analysis approach is based on the observation that the Global Field Potential (GFP) periodically achieves local maxima and that EEG microstates are defined at these maxima [49].…”

“…In this work, we used the two classical methods that appear in most of the current literature [49], i.e., modified K-means clustering and Atomize and Agglomerative Hierarchical Clustering (AAHC) [14,39], to extract a microstate sequence representing for the broadband EEG data. The EEG microstate analysis approach is based on the observation that the Global Field Potential (GFP) periodically achieves local maxima and that EEG microstates are defined at these maxima [49]. The optimal number of EEG microstate clusters is determined through a cross-validation strategy with Global Explained Variance (GEV) criterion.…”

“…Using these algorithms, EEG topographies can be assigned with microstate labels by competitive fitting based on spatial correlation, leading to a microstate sequence to represent the EEG data. Further details of AAHC, modified K-means algorithms, and experiment setups can be further provided in [14,39,49] and at www.thomaskoenig.ch, respectively. The temporal windows for each microstate FC are set to be aligned with the time intervals, where the existence of the estimated EEG microstate is expected.…”

EEG microstate architecture does not change during passive whole-body accelerations

“…Segments contaminated with artifacts were removed. The data were clustered within subjects using the widely used 'atomize and agglomerate hierarchical clustering' (AAHC) algorithm [14] and by ignoring polarity. Averages across subjects were calculated for any of the four movement conditions (Fig.…”