Electroencephalography (EEG) provides a direct measure of neuronal activity as reflected in the scalp electrical field. Empirically, global measures of EEG topography remain stable in so-called EEG microstates for brief periods (50–100 ms) before switching to another quasi-stable state. The brains of both healthy individuals as well as neurological and psychiatric patients exhibit four dominant topographies. While neurological and psychiatric diseases rarely affect the nature of these states, they do fundamentally alter the temporal dynamics of switching between them.
Researchers have tried to model EEG-microstate sequences with hidden-Markov models, extracting transition probabilities from EEG data. Such one-step Markov construction assumes that the transition probabilities between microstates depend on the current state but are otherwise memoryless. However, EEG data shows long-range temporal correlations, which one-step Markov models cannot replicate. In a new paper, LML External Fellow Claire Postlethwaite and colleagues propose a more accurate modelling strategy based on stochastic differential equations (SDE) capturing dynamics on a “noisy network attractor” having a set of quasi-stable states. Using a single layer network of four states, they find they can reproduce the transition probabilities between microstates, but not the heavy-tailed residence time distributions. However, a more complex two layer network – having one hidden layer – allows capturing these heavy tails and associated long-range temporal correlations.
The authors suggest that analysis of the temporal dynamics of microstate sequences may provide insights into the switching between brain states required for cognition and perception in the resting state, and could also provide indicators of underlying changes characteristic of neurological disorders.
The paper is available at https://arxiv.org/abs/1903.05590