Combining electro- and magnetoencephalography data using directional archetypal analysis

Anders S. Olsen, Rasmus M. T. Høegh, Jesper Løve Hinrich, Kristoffer H. Madsen, Morten MørupFrontiers in Neuroscience 2022 · DOI

Abstract

Metastable microstates in electro- and magnetoencephalographic (EEG and MEG) measurements are usually determined using modified k-means accounting for polarity invariant states. However, hard state assignment approaches assume that the brain traverses microstates in a discrete rather than continuous fashion. We present multimodal, multisubject directional archetypal analysis as a scale and polarity invariant extension to archetypal analysis using a loss function based on the Watson distribution. With this method, EEG/MEG microstates are modeled using subject- and modality-specific archetypes that are representative, distinct topographic maps between which the brain continuously traverses. Archetypes are specified as convex combinations of unit norm input data based on a shared generator matrix, thus assuming that the timing of neural responses to stimuli is consistent across subjects and modalities. The input data is reconstructed as convex combinations of archetypes using a subject- and modality-specific continuous archetypal mixing matrix. We showcase the model on synthetic data and an openly available face perception event-related potential data set with concurrently recorded EEG and MEG. We find that our approach successfully models scale and polarity invariant data, such as microstates, accounting for intersubject and intermodal variability.

Introduction

In cognitive neuroscience, the study of "microstates"—brief, stable patterns of electrical activity—has traditionally relied on clustering methods like k-means. While effective, these "hard assignment" models assume the brain snaps instantly between states. In reality, neural activity is a fluid, dynamical process. Furthermore, integrating data from Electroencephalography (EEG) and Magnetoencephalography (MEG) is technically challenging due to differing scales, sensitivities, and the inherent polarity invariance of neural signals (where a signal and its inverse often represent the same physical source).

Methodology: Directional Archetypal Analysis (DAA)

To address these challenges, we propose Directional Archetypal Analysis (DAA). Unlike standard Archetypal Analysis (AA) which uses Euclidean distance, DAA employs a loss function based on the Watson distribution. This tool from directional statistics is specifically designed for axial data, making the model naturally invariant to both the scale (amplitude) and polarity (sign) of the signals.

Geometric representation of archetypes on a sphere, illustrating how DAA identifies extreme points (archetypes) that define the convex hull of the data.

Key features of the DAA framework include:

  • Archetypes as Extremes: Instead of finding "centers" or averages, DAA identifies archetypes—the most extreme, representative patterns that define the boundaries of the data space.
  • Continuous Mixing: Every time point is described as a weighted blend of these archetypes, allowing for smooth trajectories through state-space rather than discrete jumps.
  • Multimodal Integration: By sharing a "generator matrix," the model ensures consistent timing of brain responses across EEG and MEG while allowing for modality-specific spatial maps.

Experimental Validation

The model was validated using synthetic 3D datasets and a high-quality face perception dataset (Wakeman and Henson, 2015) featuring concurrent EEG and MEG recordings from 16 subjects. Participants viewed images of famous, unfamiliar, and scrambled faces, providing a classic Event-Related Potential (ERP) challenge.

Key Results and Insights

DAA demonstrated superior fit in synthetic tests, converging to true underlying archetypes where Euclidean methods failed, particularly when handling polarity-inverted data. In the face perception study, the model successfully identified archetypes corresponding to the N170, a well-known signature for face processing.

Model consistency (Normalized Mutual Information) across different numbers of components (K), comparing the DAA solution to directional clustering and least squares.

Analysis of the continuous mixing weights revealed significant condition sensitivity; "scrambled" faces followed different trajectories than "famous" faces, failing to recruit specific face-related archetypes. Despite high inter-subject variability, the multisubject DAA found a common set of archetypal "rules" applicable across the entire group, suggesting that brain activity lives on a high-dimensional manifold where cognitive processes are paths across its surface.

Discussion and Future Work

While DAA provides a more biologically plausible model of brain state transitions, some limitations remain. The model projects data onto a hypersphere, which can struggle to represent states of zero activity (the origin). Additionally, the current framework assumes zero latency between subjects, an assumption that could be relaxed in future iterations by incorporating shift-invariant components to account for individual differences in processing speed. Nonetheless, DAA represents a significant step forward in the unified analysis of multimodal neural data.

Summary updated

Abstract

Metastable microstates in electro- and magnetoencephalographic (EEG and MEG) measurements are usually determined using modified <i>k</i>-means accounting for polarity invariant states. However, hard state assignment approaches assume that the brain traverses microstates in a discrete rather than continuous fashion. We present multimodal, multisubject directional archetypal analysis as a scale and polarity invariant extension to archetypal analysis using a loss function based on the Watson distribution. With this method, EEG/MEG microstates are modeled using subject- and modality-specific <i>archetypes</i> that are representative, distinct topographic maps between which the brain continuously traverses. Archetypes are specified as convex combinations of unit norm input data based on a shared generator matrix, thus assuming that the timing of neural responses to stimuli is consistent across subjects and modalities. The input data is reconstructed as convex combinations of archetypes using a subject- and modality-specific continuous archetypal mixing matrix. We showcase the model on synthetic data and an openly available face perception event-related potential data set with concurrently recorded EEG and MEG. In synthetic and unimodal experiments, we compare our model to conventional Euclidean multisubject archetypal analysis. We also contrast our model to a directional clustering model with discrete state assignments to highlight the advantages of modeling state trajectories rather than hard assignments. We find that our approach successfully models scale and polarity invariant data, such as microstates, accounting for intersubject and intermodal variability. The model is readily extendable to other modalities ensuring component correspondence while elucidating spatiotemporal signal variability.
BibTeX
@inproceedings{olsen2022combining,
  author = {Anders S. Olsen and Rasmus M. T. Høegh and Jesper Løve Hinrich and Kristoffer H. Madsen and Morten Mørup},
  title = {{Combining electro- and magnetoencephalography data using directional archetypal analysis}},
  booktitle = {Frontiers in Neuroscience},
  year = {2022},
  doi = {10.3389/fnins.2022.911034},
}
Figure