EEG microstates have distinct scalp potential topographies, often captured by four microstate maps explaining 64-84% of brain dynamics variance using clustering methods. Recent research suggests 5-15 cluster maps are needed to account for 80% of the variance. The clustering algorithms employed are typically chosen arbitrarily, with tools like Cartool and EEGLAB only supporting a limited range of shallow clustering methods. Such methods can be suboptimal for complex, high-dimensional data. While deep clustering has shown promise in fields like computer vision and NLP, its potential for EEG microstates remains underexplored. This study examines the eficacy of deep clustering on EEG microstates, proposing a unified framework that combines representation learning and clustering for improved microstate analysis.
Microstates are pivotal in revealing the intricate temporal and spatial nuances of multichannel,
lfuctuating EEG signals [
Microstates ofer distinct scalp potential mappings that encapsulate the spatial and temporal
dimensions of EEG brain signals[
The assumption of microstate theory is the existence of a finite number of non-overlapping, quasi-stable brain activation states leading to microstates. However, the problem within microstate theory is the lack of stability in the number of cluster maps learnt from shallow clustering (modified k-means, hierarchical clustering) and the resulting EEG microstates, which can explain up to 80% of the variance from the input EEG signal. To the aforementioned problem statement, the Research Question (RQ) is formulated as follows:
(RQ)- To what extent do deep clustering methods improve the identification and stability of cluster maps and enhance the eficiency of microstate sequences across multiple resting and cognitive states over shallow clustering?
The remaining sections of the report are devoted to articulating the precise research hypothesis and the experimental details necessary to address the previously defined research question.
A secondary experiment is designed using quantitative research methods with deductive reasoning to test the following research hypothesis.
Hypothesis - IF deep clustering methods (autoencoder-based) are used to form cluster maps across person and activity-specific and agnostic experimental configurations for highdimensional EEG signals acquired with multiple resting and cognitive conditions, THEN the stability of these cluster maps and the eficiency of microstates will be significantly higher than those cluster maps and microstates obtained by applying shallow clustering methods (modified k-means clustering).
A comprehensive set of general and specific research objectives is meticulously crafted to test the research hypothesis and realise the research aim.
1. To select a distinct EEG dataset. 2. To prepare and pre-process EEG data • To identify an EEG dataset with multiple resting and cognitive state activities [21]. • To pre-process EEG signals from the selected dataset in accordance with Makoto’s pre-processing pipeline. 3. To implement microstate generation pipeline using pre-processed EEG signals involving multiple resting and cognitive conditions [22].
• To implement person-specific and activity-specific microstate generation pipeline. • To implement person-agnostic and activity-agnostic microstate generation pipeline. • To implement person-specific and activity-agnostic microstate generation pipeline. • To execute person-agnostic and activity-agnostic microstate generation pipeline. 4. To formulate microstate cluster maps.
• To generate cluster maps using shallow clustering methods (modified k-means and agglomerative hierarchical clustering).
• To generate cluster maps using deep clustering methods (autoencoder-based). 5. To perform hyperparameter tuning • To optimise hyperparameters, including the number of layers and kernels for convolutional autoencoders, optimisation, regularisation and dropout rates by performing tuning. 6. To measure the microstate’s eficiency • To evaluate the microstate’s eficiency using microstate parameters: Global Explained Variance(GEV), mean duration, and time coverage. 7. To evaluate the outcome of distribution comparison and test the research hypothesis • To run Friedman Fr test between clustering techniques.
The eficiency of the generated microstate formed using multiple shallow and deep clustering methods can be evaluated from the microstate’s temporal parameters 1. Global Explained Variance (GEV): is the global variance accounted for by each microstate, which is the ratio of the summation of variances to the GFPs for all time points. 2. Duration (ms): provides the stable microstate period in milliseconds, which is the mean time when a label is present without interruption. 3. Coverage (%): is the contribution of a microstate expressed in percentage, representing the fraction of recording time that a label is present.
EEG microstates provide stable configurations of scalp electric fields and are thought to be the
atoms of thought of human information processing [
1. 2D Topographic Maps as Inputs: Traditional methods using 1D vector representations of EEG data for clustering abstracted the spatial intricacies of brain activity. Using 2D topographic maps, we preserve the spatial domain knowledge crucial for understanding neural source activations and their topological distributions. This spatial resolution becomes a source of explainability: a) Spatial Activation Patterns: by analysing which regions of the 2D map the autoencoder focuses on, we can infer the most salient spatial features that define specific microstates [23]. b) Comparison with Known Electrode Activations: by juxtaposing our activation patterns with traditionally known electrode activations for specific cognitive tasks, we can validate and interpret the spatial significance of our findings [24]. 2. Deep Clustering vs. K-means: Transitioning from traditional k-means clustering to deep clustering introduces additional layers of complexity. However, the hierarchical nature of deep neural networks allows us to understand EEG data at multiple levels of abstraction: a) Hierarchical Feature Learning: As the data progresses through the autoencoder layers, the network learns increasingly abstract features. By examining activations at diferent layers, we can understand granular and high-level patterns that define microstates [25]. b) Cluster Activation Analysis: Post clustering, analysing which nodes or features activate for specific clusters can provide insights into the defining characteristics of each microstate cluster. This is a more nuanced approach than the often binary partitioning seen in k-means [26].
Microstate analysis provides a finite number of whole-brain representative maps with distinct topographies that capture the spatiotemporal characteristics of EEG brain signals with varying degrees of explained variance. However, there are uncertainties with the number of dominant cluster maps representing the coverage of the global explained variance. Moreover, the choice of algorithms to compute cluster maps is limited, primarily using shallow clustering methods. Deep clustering has produced substantially better results than shallow clustering for image data with self-supervision. The research conducted in the context of this doctoral project is one of the earlier studies to assess the efectiveness of deep clustering techniques on EEG microstates without supervision. Deep clustering methods are incorporated to enhance the stability of microstates for intra and inter-person and task-conditioned resting and cognitive data. The eficiency of the proposed method is compared against the existing pipeline using microstate parameters. Friedman Fr test is performed to evaluate the significance of the proposed method.
The primary contribution anticipated from this doctoral research is the demonstration of the suitability of deep clustering methods incorporated into the microstate generation pipeline, which performs representation learning and clustering interactively and iteratively. The proposed pipeline with deep clustering is expected to improve the stability of microstates, which are demonstrated across four distinct cases, namely person-specific activity-specific, personagnostic activity-specific, person-specific activity-agnostic, person-agnostic activity-agnostic using multiple resting and cognitive conditions. Moreover, testing the pipeline with data from multiple resting and cognitive conditions is expected to validate the robustness of the microstate generation pipeline across varied natural and dynamic conditions. In addition, the proposed pipeline might be ofered to the microstate community for performing microstate analysis with alternative input modalities, such as 2D topographic input maps, as opposed to the trivial 1D vector, which fails to capture the spatial characteristics of the input EEG data.
This work was conducted with the financial support of the Science Foundation Ireland Centre for Research Training in Digitally-Enhanced Reality (d-real) under Grant No. 18/CRT/6224. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. www.sciencedirect.com/science/article/pii/S105381191000220X. doi:https://doi.org/ 10.1016/j.neuroimage.2010.02.052. [15] L. D. Hernandez, K. Rieger, A. Baenninger, D. Brandeis, T. Koenig, Towards using microstate-neurofeedback for the treatment of psychotic symptoms in schizophrenia. a feasibility study in healthy participants, Brain Topography 29 (2015) 308–321. [16] V. Férat, M. Arns, M.-P. Deiber, R. Hasler, N. Perroud, C. M. Michel, T. Ros, Electroencephalographic microstates as novel functional biomarkers for adult attentiondeficit/hyperactivity disorder, Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 7 (2022) 814–823. URL: https://www.sciencedirect.com/science/article/pii/ S2451902221003190. doi:https://doi.org/10.1016/j.bpsc.2021.11.006. [17] E. Min, X. Guo, Q. Liu, G. Zhang, J. Cui, J. Long, A survey of clustering with deep learning: From the perspective of network architecture, IEEE Access 6 (2018) 39501–39514. doi:10.1109/ACCESS.2018.2855437. [18] S. Zhou, H. Xu, Z. Zheng, J. Chen, Z. Li, J. Bu, J. Wu, X. Wang, W. Zhu, M. Ester, A comprehensive survey on deep clustering: Taxonomy, challenges, and future directions, ArXiv abs/2206.07579 (2022). [19] A. V. Chikkankod, L. Longo, On the dimensionality and utility of convolutional autoencoder’s latent space trained with topology-preserving spectral eeg head-maps, Machine Learning and Knowledge Extraction 4 (2022) 1042–1064. URL: https://www.mdpi.com/ 2504-4990/4/4/53. doi:10.3390/make4040053. [20] S. Chambon, M. N. Galtier, P. J. Arnal, G. Wainrib, A. Gramfort, A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series, IEEE Transactions on Neural Systems and Rehabilitation Engineering 26 (2018) 758–769. doi:10.1109/TNSRE.2018.2813138. [21] A. Damborská, C. Piguet, J.-M. Aubry, A. G. Dayer, C. M. Michel, C. Berchio, Altered electroencephalographic resting-state large-scale brain network dynamics in euthymic bipolar disorder patients, Front. Psychiatry 10 (2019) 826. [22] P. Croce, A. Quercia, S. Costa, F. Zappasodi, Eeg microstates associated with intra- and inter-subject alpha variability, Scientific Reports 10 (2020). doi: 10.1038/ s41598-020-58787-w. [23] B. H. van der Velden, H. J. Kuijf, K. G. Gilhuijs, M. A. Viergever, Explainable artificial intelligence (xai) in deep learning-based medical image analysis, Medical Image Analysis 79 (2022) 102470. doi:https://doi.org/10.1016/j.media.2022.102470. [24] G. Vilone, L. Rizzo, L. Longo, A comparative analysis of rule-based, model-agnostic methods for explainable artificial intelligence, in: Irish Conference on Artificial Intelligence and Cognitive Science, 2020. URL: https://api.semanticscholar.org/CorpusID:229345223. [25] G. Vilone, L. Longo, Notions of explainability and evaluation approaches for explainable artificial intelligence, Information Fusion 76 (2021) 89–106. URL: https:// www.sciencedirect.com/science/article/pii/S1566253521001093. doi:https://doi.org/ 10.1016/j.inffus.2021.05.009. [26] J. Kaufmann, M. Esders, L. Ruf, G. Montavon, K.-R. Samek, Wojciech and, From clustering to cluster explanations via neural networks, IEEE Transactions on Neural Networks and Learning Systems (2022) 1–15. doi:10.1109/TNNLS.2022.3185901.