https://orcid.org/0000-0003-4175-918X
This scientific commentary refers to ‘Dysfunctional brain dynamics and their origin in Lewy body dementia’, by Schumacher et al. (doi:).
The prevailing functional imaging modalities that non-invasively probe the brain’s dynamics mandate a trade-off between superior temporal resolution (in the case of EEG), versus spatial resolution (for functional MRI). Consequently, EEG techniques have flourished in the search for neural correlates of rapid cognitive and perceptual processes, and have acquired clinical roles in epilepsy and anaesthesia. Conversely, functional MRI occupies a unique niche in the mapping of cortical and subcortical brain function over longer time frames. The combination of these techniques has particular promise in the study of Lewy body dementia (LBD), a neurodegenerative illness characterized by mental slowness, fluctuating cognition, visual hallucinations, and parkinsonian motor symptoms. In this issue of Brain, Schumacher and co-workers report a novel analysis of LBD using EEG microstate analysis to characterize features of cerebral information processing, while also using functional MRI to extend our understanding of subcortical dysfunction in the disease (Schumacher et al., 2019).
EEG microstates are specific topographies of electrical field potential across the scalp, which remain quasi-stable for time frames on the order of ∼100 ms before transitioning to another microstate (Fig. 1). EEG microstates have been touted as fundamental ‘atoms of thought’, supported by research showing that a specific microstate distinguishes visual depth from contour perception (Michel et al., 1992) and that reduced microstate duration in schizophrenia may abruptly terminate information processing and manifest as disordered thinking (Strelets et al., 2003). More recently, it has been shown that specific EEG microstates are mapped onto specific functional MRI resting state networks (Britz et al., 2010; Musso et al., 2010). This unexpected finding (given the very different time frames involved) can be reconciled because microstates have a complex, fractal-like temporal architecture, which manifests as self-similar sequence likelihoods across a breadth of timescales (Britz et al., 2010)—hence from milliseconds (EEG) to seconds (functional MRI). The formulation of EEG microstates as the quanta of cognition makes them a particularly interesting probe for the mental slowness and fluctuating cognition of LBD.
Microstates are quasi-stable topographies of the EEG scalp potential. Functional brain states dwell in the local neighbourhood of each local minima for a short period before stochastically switching to another microstate. Subjects with LBD exhibited greater mean microstate durations, suggestive of impaired state switching. This could be due to deeper ‘wells’ or weaker noise. Figure adapted from Cocchi et al. (2017).
Schumacher et al. (2019) assessed 46 patients with LBD, 32 patients with Alzheimer’s disease, and 18 healthy control subjects. They first acquired high density resting state eyes-closed EEG and used a hierarchical clustering algorithm to map the scalp electrical potential at each time point to one of five microstate classes A–E. They observed a widespread increase in mean microstate duration across all microstate classes, an outcome that points intuitively to the clinical finding of mental slowness in LBD. Notably, microstate duration was longer in LBD than in both Alzheimer’s disease and controls. There was also a ‘dose-response relationship’ between increases in mean microstate duration and more severe cognitive fluctuations expressed by the LBD participants, suggestive of causation rather than merely correlation.
The authors also acquired resting state functional MRI on a subset of subjects and computed time-varying functional connectivity between different canonical resting state networks. Given prior evidence of dysfunction of the basal ganglia and thalamus in LBD, they focused a priori on connectivity between these two subcortical regions and resting state cortical networks. Intriguingly, another dose-response relationship was evident; LBD subjects with longer mean EEG microstate durations had weaker dynamic functional connectivity between the two subcortical and the cortical networks. These changes were reflected in reduced dynamic functional connectivity between the basal ganglia and motor, visual and the default mode networks, as well as between the thalamus and insular, sensorimotor, occipital and cerebellar networks (although the medial visual network was the only network whose dynamic functional connectivity with the basal ganglia survived correction for multiple comparisons). It is unfortunate that the authors did not report on correlations with their neuropsychiatric inventory for visual hallucinations.
Current consensus criteria for diagnosis of LBD have low sensitivity due to crossover of symptoms between the major neurocognitive disorders. Diagnostic biomarkers currently in use include SPECT evidence of dopamine transporter loss in the striatum, PET hypometabolism in the occipital cortex, and a relatively preserved medial temporal lobe in comparison to Alzheimer’s disease. Microstate analyses are low-cost and non-invasive. The findings of Schumacher et al. (2019)—showing sensitivity to LBD, but also specificity with respect to Alzheimer’s disease—could therefore aid clinical diagnosis. To achieve this, however, validation is required in larger and independent datasets, as well as comparison with other neurodegenerative diseases such as frontotemporal and vascular dementia.
Taken together, the results of Schumacher et al. (2019) speak to an intuitively compelling understanding of LBD. Primary dysfunction originating in the basal ganglia and thalamus may propagate to distant cortical sites and impair the flexibility of subcortical-cortical dynamics, manifesting as reduced dynamic functional connectivity. The resultant ‘functional rigidity’, suppressing transitions between EEG microstates, then contributes to the ‘mental slowness’ that characterizes the LBD phenotype. The functional nature of this impairment in LBD permits a fluctuating cognition, in contrast to the fixed structural temporal lobe defects in Alzheimer’s disease, manifesting as stable memory deficits.
This intuitive reading of Schumacher et al. (2019) must be interrogated by the weight of evidence. Individual EEG microstates last only ∼100 ms, but as mentioned earlier, there are more prolonged patterns in their mean duration or frequency of occurrence. In this study on LBD, whether EEG microstates are the origin or consequence of resting state dynamics remains unclear. On one hand, it is more plausible that fast EEG dynamics, low-pass filtered by the haemodynamic response function, produce slower BOLD dynamics. Conversely, scalp EEG signals represent mainly superficial cortical sources and hence cannot mechanistically explain subcortical functional connectivity. Overall, it appears more likely that LBD primarily slows the fast electrophysiological signalling between subcortical structures and cortex, manifesting secondarily as (i) reduced variability in resting state connectivity between subcortex and cortex; and (ii) increased microstate duration in scalp EEG.
Computational modelling as well as source-space analyses may help confirm or refute this hypothesis. Hidden Markov models are a recently proposed alternative to microstate analysis, which can incorporate source space signals. The underlying source-based states in these models lead to the observed multivariate distribution of electrophysiological signals in the lead field through volume conduction. The states themselves are thus ‘hidden’ and cannot be observed directly, but are inferred from the time series of observed signals. Transition probabilities governing state switching are also inferred during the procedure. Hidden Markov models have been applied to MEG source space (Baker et al., 2014) and EEG sensor space data (Hunyadi et al., 2019), finding states lasting ∼100 ms reminiscent of EEG microstates. This type of analysis has significant promise for deepening our insights into the causality structure of breakdowns in fast EEG dynamics and slow functional MRI dynamics in LBD. Unfortunately, potential implementation of such methods in the present study is constrained by the poor resolution of source-localized EEG in the absence of head space digitization. The non-concurrence of EEG and functional MRI acquisition in this study also limits any conclusions about causation.
A more fundamental issue is the assumption of the EEG microstate representing an ‘atom of thought’, which is prerequisite to the idea that increased microstate duration causes mental slowness in LBD. This idea was initially based on correlations between the brain’s microstate at a precise moment, and the type of thought that subjects reported (Lehmann et al., 1998). This formulation seems oversimplified. If the stream of consciousness occupies the entire scalp EEG topography, where are the unconscious cognitions? Why is thought discretized into exactly five (or, in other formulations, four) microstates? It appears more plausible that the apparent discretization of microstates and hidden Markov model states into ∼100 ms blocks, is a complex manifestation of more fundamental conscious and unconscious mental processes. Perhaps another manifestation of these underlying processes is the finding that visual detection task performance fluctuates with EEG theta/alpha phase, reaching a maximum every ∼120 ms (Busch et al., 2009). Such (continuous) fluctuations could appear discrete only through the imposition of a (discrete) clustering solution. The implication for Schumacher et al. is that both increased microstate duration and mental slowness could be symptoms of a more fundamental alteration in the dynamical landscape of the brain in LBD.
Where to next? Multimodal data fusion with spatially sensitive techniques like optical imaging or even single unit recordings in mouse models of LBD may be fruitful in extending this work. Computational modelling of neuronal dynamics may be able to map the phenomenologically defined ‘microstates’ onto the mathematical notion of ‘metastability’, which is grounded in the more principled (mathematical) language of manifolds, flows and attractors (Roberts et al., 2019). Extension to other neuropsychiatric disorders, as noted above, but also to disorders where cognition ‘speeds up’, such as bipolar disorder, could help test the specificity and face validity of these findings. Future consolidation of these techniques promises a mechanistic understanding, from function at the spatial and temporal level of the neuron’s action potential, to the movements of large-scale brain dynamics, and from there, to an understanding of the mechanisms behind dysfunctions of the brain and how to treat them.
Glossary
Dynamic functional connectivity: Temporal variability of functional connectivity between two brain regions or two different resting state networks. Dynamic functional connectivity is typically calculated in a short time period determined by a sliding window, and hence can itself fluctuate over longer time frames.
EEG microstate: Briefly stable topographies of patterns of the scalp electric potential derived from EEG. The EEG signal is segmented into non-overlapping microstates each lasting ∼80–120 ms. Over time, the pattern jumps from one microstate to another.
Fractal: Complex structures that exhibit self-similarity and appear approximately the same across a broad range of observation scales. Typical examples include Mandelbrot sets, fern leaves, and the DNA sequence, but one-dimensional time series can also exhibit fractal features.
Hidden Markov model: A mathematical model comprising a set of unobserved or ‘hidden’ states that evolve according to the Markov rule: namely, that future states depend only on the current state and not on states that occurred before it. Observations in a hidden Markov model are sampled at each time step from a probability distribution (over possible observations), contingent on the current state.
Resting state network: A set of brain regions whose functional MRI BOLD signal activity shows a high degree of positive correlation in subjects resting in the scanner (i.e. with no explicit tasks).
Competing interests
The authors report no competing interests.
