TL;DR
This study compares two methods, microstates and Hidden Markov Models (HMMs), for analyzing spontaneous brain activity. They found that these methods capture different aspects of brain dynamics: microstates reflect rapid, localized events, while HMMs capture slower, network-level activity. This suggests they are complementary tools for understanding brain function.
What Problem Does This Paper Address?
The human brain exhibits complex, dynamic activity even at rest. Researchers use electroencephalography (EEG) and magnetoencephalography (MEG) to study this, looking for recurring patterns of brain activity. Two popular methods for this are EEG microstate analysis and Hidden Markov Modeling (HMM) of MEG power envelopes. While both methods seem to identify similar timescales of brain states, their underlying assumptions and how they process data are quite different. This raises a key question: do these methods actually capture the same underlying neural processes, or are they sensitive to different aspects of brain dynamics? This paper aims to disentangle these differences by directly comparing microstates and HMMs using simultaneous MEG/EEG recordings.
Methods
The researchers used simultaneous MEG and EEG recordings from participants at rest. They applied two main analytical approaches:
- EEG Microstate Analysis: This method identifies recurring, quasi-stable patterns in EEG scalp topography. They used an Atomize-Agglomerate Hierarchical Clustering (AAHC) algorithm, typically applied to time points with high Global Field Power (GFP), to define microstate topographies. They also explored adapting this to MEG data.
- Hidden Markov Modeling (HMM) of Power Envelopes: This approach models brain activity as a sequence of discrete "hidden" states. They used HMMs to analyze the power envelopes of MEG (and later EEG) signals. This method captures transitions between states based on the statistical properties of the signal over time.
A key aspect of their methodology was to perform direct comparisons by feeding similar data to both algorithms and by adapting each method to work with both MEG and EEG data, allowing them to isolate the effects of the clustering algorithm versus the recording modality.
Datasets
The study used resting-state MEG/EEG data from 42 healthy young adults. Some of these participants were previously included in another study by the same group. No new datasets were introduced.
Key Insights and Ideas
- Microstates vs. HMMs: Different Scales, Different Dynamics: The most striking finding is that microstates and HMM states are not interchangeable. Microstates seem to capture very brief, localized events (around 37 ms without temporal smoothing), while HMM states capture longer-lived, network-level activity (around 150 ms for MEG sensor-level HMM).
- Microstates are Modality-Specific: EEG microstates did not reliably correspond to microstates derived from MEG data. This suggests that microstate analysis is highly sensitive to the specific recording modality (EEG vs. MEG) and the way signals are processed.
- HMMs are More Cross-Modal: In contrast, HMMs applied to both EEG and MEG data revealed similar functional networks and some degree of temporal correspondence, suggesting they capture more fundamental, cross-modal brain dynamics.
- Microstates are Driven by Peaks, HMMs by Transitions: Microstates are biased towards time points of high signal power (GFP), essentially capturing "snapshots" of brain activity. HMMs, by analyzing power envelopes, are better at capturing the transitions between low and high power states, thus reflecting the dynamics of bursting activity.
- Temporal Smoothing Matters for Microstates: The "classical" long lifetimes reported for microstates (e.g., 120 ms) are often achieved through temporal smoothing (interpolation). When analyzing raw microstate activation, lifetimes are much shorter, highlighting the transient nature of the underlying events.
Key Results
- Spatial and Temporal Dissimilarity: Direct comparison of EEG microstates and MEG HMM states showed poor spatial and temporal correlation.
- MEG Microstates Differ from EEG Microstates: Microstates derived from MEG data did not match the canonical EEG microstates.
- EEG and MEG HMM States Show Overlap: HMM states inferred from EEG and MEG data showed significant spatial and temporal correlations, indicating they capture similar network-level dynamics.
- Microstate Lifetimes are Short (Raw): Raw microstate activation lifetimes were around 37 ms for EEG and 32-47 ms for MEG, significantly shorter than typical reported values which often involve temporal smoothing.
- HMM Lifetimes are Longer: HMM states had mean lifetimes around 150 ms for sensor-level data, consistent with capturing bursting activity.
- Sensor-Level HMM is Viable: The study demonstrated that HMM analysis can be effectively performed directly on sensor-level MEG/EEG data, simplifying the process and reducing reliance on source reconstruction.
Limitations and Future Work
- Microstate Interpretation: The study suggests that MEG microstates might be a "split" version of HMM states, capturing more localized aspects of network activity. Further work could explore this relationship more deeply.
- HMM for Phase Dynamics: The current HMM approach focuses on power envelopes, ignoring phase information. Future research could use time-embedded HMMs that incorporate phase dynamics to better understand synchronization and network interactions.
- Microstate Reference Dependence: For EEG, the choice of reference can influence microstate analysis. While the average reference was used, exploring other reference schemes or standardized techniques could be beneficial.
- Source Reconstruction Biases: While the study attempted to control for some biases, source reconstruction can still introduce differences, particularly in frontal regions. Future work could investigate how different source reconstruction methods impact HMM state inference.
- Complementary Approaches: The authors conclude that microstates and HMMs are complementary. Future research should focus on integrating findings from both methods to gain a more comprehensive understanding of brain dynamics.
Summary updated