Spectrally resolved fast transient brain states in electrophysiological data

Diego Vidaurre, Andrew J. Quinn, Adam P. Baker, David Dupret, Álvaro Tejero-Cantero, Mark W. WoolrichNeuroImage 2016 · DOI

TL;DR

This paper introduces a novel method, HMM-MAR, to identify rapidly changing, multi-region brain states in electrophysiological data. By combining Hidden Markov Models with Multivariate Autoregressive models and a non-parametric spectral estimation technique, it can characterize these transient brain states in terms of their temporal, spectral, and spatial properties without needing pre-defined task timings. This allows for a more nuanced understanding of brain dynamics than traditional methods.

What Problem Does This Paper Address?

Understanding how the brain coordinates activity across different regions at very fast timescales (sub-second) is a major challenge. Traditional methods often rely on sliding time windows, which pre-define the temporal resolution and can be inefficient if the same brain states repeat. While some approaches can pool data across repeated trials, they often assume stationarity within those trials. This paper tackles the problem of identifying and characterizing these rapid, dynamic brain states, including their spectral and connectivity properties, in an unsupervised manner, even when they are very short-lived and occur at unpredictable times.

Methods

The core of the proposed method is the HMM-MAR (Hidden Markov Model - Multivariate Autoregressive Model).

  • MAR Model: This component models the relationships between different brain regions over time, capturing spectral properties and inter-regional interactions (like coherence).
  • HMM: This component models the brain's activity as a sequence of discrete "states." Each state has its own unique MAR model, meaning each state represents a distinct pattern of multi-region spectral properties. The HMM learns the transitions between these states.
  • Non-parametric Multitaper Spectral Estimation: To get spectral information for each identified state, the authors use a modified multitaper method. This approach leverages the inferred state time courses as temporal windows to estimate state-specific power spectra and coherence. This is crucial for capturing spectral details even for very short state durations.
  • Bayesian Hierarchy and Variational Inference: The entire model is framed within a Bayesian framework, allowing for principled model selection and parameter estimation using variational Bayes. This approach helps manage model complexity and infer parameters even with noisy data.
  • Sign Ambiguity Handling: A specific method is proposed to address the sign ambiguity inherent in MEG source reconstruction, which is crucial for reliable multi-subject analyses.

Datasets

The authors evaluate their method on two types of data:

  • Synthetic Data: A simulated dataset with known ground-truth brain states and spectral properties was generated to rigorously test the method's ability to recover these known dynamics.
  • Real MEG Data: Data from a volitional motor task (button press) collected from ten right-handed volunteers was used to demonstrate the method's application to real-world neuroscience problems.

Key Insights and Ideas

  • Unsupervised Detection of Transient States: The HMM-MAR can identify distinct brain states that are dynamically changing over time, without prior knowledge of when these states occur or what task is being performed.
  • Spectral and Connectivity Signatures of States: Each identified state has a unique spectral profile (e.g., power in specific frequency bands) and connectivity pattern (e.g., coherence between regions), providing rich characterizations of brain activity.
  • Short-Lived Dynamics are Recoverable: The method is effective even for brain states that last only around 100 milliseconds, which are often missed by traditional sliding window approaches.
  • Task-Relevant States Emerge Naturally: In the motor task data, the HMM-MAR identified states that corresponded to known event-related desynchronization (ERD) and synchronization (ERS) patterns, even though the model was not explicitly told about the button press timing.
  • Bridging Parametric and Non-parametric Strengths: The approach combines the interpretability of MAR models with the robustness of non-parametric spectral estimation, offering a powerful hybrid.

Key Results

  • The HMM-MAR successfully recovered simulated brain states and their spectral properties, demonstrating the validity of the approach.
  • In the MEG motor task data, the method identified three distinct states, which were labeled as ERD, ERS, and a baseline state based on their temporal occurrence and spectral characteristics.
  • The identified states showed statistically significant differences in power spectral density (PSD), coherence, and partial directed coherence (PDC) between different frequency bands and across different brain regions.
  • The HMM-MAR regularized time-frequency (T-F) analyses provided a more robust and detailed view of brain activity changes around the button press compared to traditional sliding window T-F analyses, particularly for coherence and PDC.
  • The authors showed that functional connectivity (cross-channel coefficients) alone can drive state transitions, as demonstrated by the HMM-crossMAR model.

Limitations and Future Work

  • Computational Cost: While the method is powerful, it can be computationally intensive, especially for a large number of brain regions. The authors suggest fixing some connections to reduce complexity.
  • Matching States Across Subjects: Similar to other unsupervised methods like ICA, matching identified states across different subjects can be non-trivial. They suggest prediction-based measures for this.
  • Scale-Free Dynamics: The current first-order Markov assumption in HMMs doesn't explicitly model long-term dependencies characteristic of scale-free systems. Future work could explore hierarchical models or investigate if inferred state time courses exhibit scale-free properties.
  • Continuous Stimulus Levels: The current state-switching paradigm might not be ideal for tasks with continuously varying stimulus levels. Future work could explore models that modulate state "strength" rather than just switching states.
  • Application to Resting-State and fMRI: The authors suggest applying the HMM-MAR to study whole-brain resting-state dynamics and adapting it for fMRI data by incorporating the hemodynamic response function.

Summary updated

BibTeX
@article{vidaurre2016spectrally,
  author = {Diego Vidaurre and Andrew J. Quinn and Adam P. Baker and David Dupret and Álvaro Tejero-Cantero and Mark W. Woolrich},
  title = {{Spectrally resolved fast transient brain states in electrophysiological data}},
  journal = {NeuroImage},
  year = {2016},
  doi = {10.1016/j.neuroimage.2015.11.047},
}
Figure