Measuring temporal, spectral and spatial changes in electrophysiological brain network connectivity

Matthew J. Brookes, George C. O'Neill, Emma L. Hall, Mark William Woolrich, Adam P. Baker, Sofia Palazzo Corner, Siân E. Robson, Peter G. Morris, Gareth R. BarnesNeuroImage 2014 · DOI

TL;DR

This paper introduces a novel method combining beamforming and Canonical Correlation Analysis (CCA) to analyze the dynamic, frequency-specific, and spatially distributed nature of brain network connectivity using MEG data. By accounting for temporal non-stationarity and spatial inhomogeneity, this approach offers a more nuanced understanding of brain function than traditional static connectivity measures.

What Problem Does This Paper Address?

Traditional neuroimaging studies often analyze functional connectivity by averaging signals over several minutes, assuming that connectivity is static over time. However, growing evidence suggests that brain connectivity is dynamic, changing over time and varying across different frequency bands. Furthermore, large-scale brain networks might not be uniform but rather composed of transiently synchronized sub-networks with distinct spatial patterns. This means that static, time-averaged connectivity measures may miss crucial aspects of brain function. The authors aim to address this by developing a method that can capture these temporal, spectral, and spatial variations in brain network connectivity.

Methods

The core of the proposed method involves a multi-stage process:

  • Beamforming: This technique is used to project MEG sensor data into source space, estimating the electrical activity within specific brain regions (voxels). Beamforming offers good spatial resolution and helps reduce artifacts.
  • Multivariate Leakage Reduction: A significant challenge in MEG source analysis is "signal leakage," where activity from one brain region can erroneously appear in the estimated activity of another. The authors extend previous work by using a multivariate linear regression approach to suppress these zero-phase-lag interactions between seed and test clusters, aiming to isolate genuine connectivity.
  • Windowed Canonical Correlation Analysis (CCA): To capture temporal dynamics, the analysis is performed within sliding time windows. CCA is then applied to the band-limited amplitude envelopes of the signals from seed and test clusters. CCA identifies the linear combinations of signals (canonical variates) that maximize temporal correlation between the two regions within each window. This allows for the computation of a canonical correlation coefficient that reflects the strength of connectivity.
  • Phase Randomization for Statistical Testing: To ensure that observed correlations are not spurious, the authors employ a phase randomization technique. This generates surrogate data with the same spectral properties but no genuine correlation, creating a null distribution against which the observed connectivity is tested. This allows for the generation of dynamic statistical thresholds that adapt to the temporal structure of the data.

Datasets

The study utilizes two main types of data:

  • Simulated MEG Data: The authors generated simulated MEG data to test the performance of their method, particularly in assessing leakage reduction and the validity of statistical testing. This involved simulating six spatially separate sources with controlled coupling.
  • Real Resting-State MEG Data: The developed method was applied to resting-state MEG data acquired from human subjects.

Key Insights and Ideas

  • Connectivity is Dynamic and Frequency-Specific: The results strongly suggest that functional connectivity is not static but varies significantly over time and across different frequency bands, particularly in the alpha and beta bands.
  • Spatial Inhomogeneity within Networks: The method reveals that connectivity patterns are not uniform but exhibit spatial inhomogeneity, suggesting the presence of distinct "sub-networks" that can be identified within larger brain regions.
  • MEG's Advantage in Temporal Resolution: The excellent temporal resolution of MEG is highlighted as a key advantage for capturing these dynamic changes in connectivity, which might be missed by slower neuroimaging techniques like fMRI.
  • Importance of Leakage Reduction: The simulations clearly demonstrate that without proper leakage reduction, spurious correlations can be detected, leading to inaccurate conclusions about functional connectivity.

Key Results

  • The proposed method successfully generated time-frequency connectivity plots, visualizing the temporal and spectral evolution of coupling between brain regions.
  • Simulations showed that the leakage reduction technique effectively controlled false positive rates, especially for Gaussian distributed data.
  • Analysis of real MEG data revealed significant temporal and spectral variations in connectivity between the left and right sensorimotor cortices.
  • The method identified distinct spatial patterns (spatial modes) of connectivity, with the first mode resembling known sensorimotor networks and a second mode potentially reflecting secondary somatosensory regions.

Limitations and Future Work

  • Gaussianity Assumption: The leakage reduction method's effectiveness is dependent on the assumption of Gaussian distributed data. Deviations from Gaussianity in real MEG data can lead to an overestimation of false positives, meaning the algorithm suppresses rather than perfectly corrects leakage.
  • Loss of Genuine Zero-Phase-Lag Interactions: The leakage suppression technique comes at the cost of potentially losing genuine zero-phase-lag interactions, which could be problematic in specific scenarios.
  • Constant Leakage Assumption: The assumption of constant leakage across all frequencies and time might be invalidated in certain cases.
  • Spatial Resolution Limitations: While beamforming improves spatial resolution, MEG's inherent spatial resolution limitations due to the ill-posed inverse problem remain a challenge.
  • Future Work: The authors suggest extending the method to task-based studies, exploring other brain networks (e.g., the default mode network), and investigating multiple current orientations within voxels. They also propose exploring alternative techniques like Hidden Markov Models (HMMs) for temporally adaptive analysis and developing principled methods for identifying the number of spatial and temporal modes of correlation.

Summary updated

BibTeX
@article{brookes2014measuring,
  author = {Matthew J. Brookes and George C. O'Neill and Emma L. Hall and Mark William Woolrich and Adam P. Baker and Sofia Palazzo Corner and Siân E. Robson and Peter G. Morris and Gareth R. Barnes},
  title = {{Measuring temporal, spectral and spatial changes in electrophysiological brain network connectivity}},
  journal = {NeuroImage},
  year = {2014},
  doi = {10.1016/j.neuroimage.2013.12.066},
}
Figure