Guiding functional connectivity estimation by structural connectivity in MEG: an application to discrimination of conditions of mild cognitive impairment

José Angel Pineda-Pardo, Ricardo Bruña, Mark William Woolrich, Alberto Marcos Dolado, Anna Christina Nobre, Fernando Maestú, Diego VidaurreNeuroImage 2014 · DOI

TL;DR

This paper introduces a novel method to improve the estimation of brain functional connectivity (FC) by incorporating structural connectivity (SC) information from diffusion MRI into magnetoencephalography (MEG) data. This "structurally guided" approach leads to more accurate FC models, which in turn significantly improves the ability to distinguish between healthy individuals and those with Mild Cognitive Impairment (MCI), and even between different subtypes of MCI.

What Problem Does This Paper Address?

Diagnosing Mild Cognitive Impairment (MCI) early is crucial, as it often precedes Alzheimer's disease. MCI is characterized by a decline in cognitive function, and researchers believe this is linked to disruptions in brain network connectivity. While functional connectivity (FC), measured by brain activity, shows promise as a biomarker, traditional methods like correlation analysis can be inaccurate, picking up indirect connections. Precision matrices offer a better representation of direct connections, and methods like the graphical lasso (GL) can estimate sparse precision matrices. However, inferring these networks solely from functional data can still be challenging.

This paper addresses the challenge of accurately estimating brain connectivity and using it for MCI diagnosis. Specifically, it explores how to best integrate structural information (the physical wiring of the brain from diffusion MRI) with functional information (brain activity from MEG) to create more robust and informative network models. The authors hypothesize that by using SC to "guide" the estimation of FC, they can achieve more precise network representations and, consequently, better diagnostic accuracy for MCI.

Methods

The core of this paper lies in a novel adaptation of the graphical lasso (GL) algorithm. The standard GL estimates a sparse precision matrix (representing direct functional connections) by penalizing non-zero elements. This paper introduces two adaptive versions:

  • Functional Adaptive GL (GLa): This version uses the functional data itself to adapt the penalties, aiming to improve the estimation based on the observed functional patterns.
  • Structural Adaptive GL (GLd): This is the key innovation. Here, the penalties for estimating the functional precision matrix are guided by the structural connectivity (SC) derived from diffusion-weighted MRI (DW-MRI). Weaker anatomical connections are penalized more heavily, effectively encouraging the functional network to align with the underlying structural wiring. This is a "soft constraint," meaning the model can deviate from the strict structural pattern if the functional data strongly suggests otherwise.

The study used beamformer source reconstruction on resting-state MEG data to obtain time series for 66 cortical regions. Diffusion tensor imaging (DTI) was used to derive SC, quantified as fiber density between regions. These methods were applied to three groups: healthy controls (HC), single-domain amnestic MCI (sdMCI), and multiple-domain amnestic MCI (mdMCI). The performance of GL, GLa, and GLd was evaluated using log-likelihood (how well the model fits the data) and classification accuracy between the groups using four different machine learning classifiers (LDA, kNN, SVM, SVMrbf) with a 10-fold cross-validation approach.

Datasets

The study utilized data from 81 participants:

  • 29 Healthy Controls (HC)
  • 22 Single-Domain Amnestic Mild Cognitive Impairment (sdMCI)
  • 30 Multiple-Domain Amnestic Mild Cognitive Impairment (mdMCI)

Participants were recruited from Madrid, Spain. The data included resting-state MEG recordings and structural MRI (T1-weighted and diffusion-weighted). No new benchmarks were introduced; the focus was on applying and evaluating existing methods with multimodal data.

Key Insights and Ideas

  • SC as a "Soft Prior" for FC: The most significant insight is that structural connectivity can effectively guide the estimation of functional connectivity without rigidly constraining it. This "soft prior" approach allows for a more nuanced and accurate representation of brain networks.
  • MEG's Temporal Resolution Advantage: The paper highlights that MEG's finer temporal resolution compared to fMRI allows for the study of functional networks across a wider frequency spectrum, potentially revealing more about neural dynamics.
  • Multimodal Fusion Improves Diagnosis: The study strongly supports the idea that combining structural and functional neuroimaging data leads to better diagnostic capabilities for MCI than using either modality alone.
  • Frequency Band Specificity: Different frequency bands (especially alpha and broadband) appear to be more informative for discriminating between HC and sdMCI, while the full beta band is crucial for distinguishing HC from mdMCI.

Key Results

  • Improved Log-Likelihood: The adaptive GL methods (GLa and GLd) consistently showed higher log-likelihood values than the non-adaptive GL, indicating better data fitting. GLd, guided by SC, performed comparably to GLa, which uses functional data for adaptation, suggesting a genuine link between SC and FC.
  • Enhanced Classification Accuracy:
    • HC vs. sdMCI: The GLd approach achieved the highest accuracy of 86.27% using LDA with broadband frequency data.
  • HC vs. mdMCI: The GLd approach yielded the best accuracy of 81.36% using LDA with full beta band data.
  • sdMCI vs. mdMCI: Interestingly, the non-adaptive GL method performed best in this comparison, achieving 84.62% accuracy with LDA in the broadband.
  • Feature Importance: The classification models identified specific brain regions and connections that were most diagnostic. For HC vs. MCI distinctions, temporal and cingulate regions were consistently important. For mdMCI, frontal regions also played a significant role, aligning with the executive function deficits associated with this subtype.

Limitations and Future Work

The authors acknowledge several limitations:

  • MCI Conversion Uncertainty: The study's participants were not all confirmed converters to Alzheimer's disease, meaning the groups might have heterogeneity that affects classification accuracy. Future work should monitor these patients for conversion.
  • Limited Frequency Bands: The study focused on alpha and beta bands. Future research should explore the impact of SC priors on other frequency bands (delta, theta, gamma).
  • Stationary Connectivity: The analysis assumed stationary functional connectivity. Future work could explore dynamic connectivity models (e.g., Hidden Markov Models) with SC priors.
  • SC Quantification: DTI tractography has limitations (e.g., fiber crossing). More advanced diffusion models could improve SC estimation.
  • Brain Parcellation: The choice of brain parcellation can influence network topography. Future studies should investigate different parcellation schemes.
  • Leakage Correction: The study did not apply leakage correction to MEG data, which could potentially remove false positives but also genuine signals. Future work could explore integrating leakage correction with SC priors.
  • Sample Size: While larger than some previous studies, the sample size is still moderate.

Future work should focus on larger datasets, clinical follow-up for conversion prediction, exploring other frequency bands and dynamic connectivity models, and refining SC quantification and parcellation methods.

Summary updated

BibTeX
@article{pinedapardo2014guiding,
  author = {José Angel Pineda-Pardo and Ricardo Bruña and Mark William Woolrich and Alberto Marcos Dolado and Anna Christina Nobre and Fernando Maestú and Diego Vidaurre},
  title = {{Guiding functional connectivity estimation by structural connectivity in MEG: an application to discrimination of conditions of mild cognitive impairment}},
  journal = {NeuroImage},
  year = {2014},
  doi = {10.1016/j.neuroimage.2014.08.002},
}
Figure