Modelling subject variability in the spatial and temporal characteristics of functional modes

Samuel J. Harrison, Janine D. Bijsterbosch, Andrew R. Segerdahl, Sean P. Fitzgibbon, Seyedeh-Rezvan Farahibozorg, Eugene P. Duff, Stephen M. Smith, Mark W. WoolrichNeuroImage 2020 · DOI

TL;DR

This paper introduces PROFUMO, a novel probabilistic framework for analyzing fMRI data that explicitly models and accounts for individual differences in brain activity patterns. By jointly inferring group and subject-level functional brain networks, PROFUMO offers a more accurate and interpretable understanding of how brain connectivity varies across individuals, moving beyond traditional methods that often conflate spatial and temporal variations.

What Problem Does This Paper Address?

Analyzing functional Magnetic Resonance Imaging (fMRI) data, especially from large-scale studies, presents a significant challenge due to substantial variability in brain activity across individuals. Traditional methods often struggle to disentangle true biological differences from noise or artifacts, and can conflate variations in the spatial layout of brain regions with differences in their temporal activity and connectivity. This makes it difficult to accurately interpret findings, particularly when relating them to behavior or clinical conditions. The authors aim to address this by developing a more sophisticated model that can simultaneously capture and differentiate these sources of variability, leading to more robust and interpretable results.

Methods

The core of this work is the Probabilistic Functional Modes (PFMs) framework, extended and implemented as the PROFUMO software package. Key methodological advancements include:

  • Joint Group and Subject-Level Inference: Unlike traditional methods that estimate group-level patterns first and then infer subject-specific details, PROFUMO models these jointly. This allows for a more integrated and less biased estimation of both group and individual variations.
  • Hierarchical Probabilistic Model: The model explicitly parameterizes subject variability in:
  • Spatial Maps: Capturing differences in the size, shape, and location of functional brain regions.
  • Functional Coupling: Modeling how different brain networks interact over time.
  • Mode Amplitudes: Accounting for variations in the overall strength of activity within networks.
  • Advanced Inference: A computationally efficient variational Bayesian approach is used for inference, allowing the model to scale to large datasets.
  • Flexible Noise Modeling: The model incorporates specific noise components for spatial maps and time courses, improving the stability and accuracy of the decomposition.
  • HRF-Based Temporal Model: The temporal dynamics are modeled using a canonical hemodynamic response function (HRF), providing a physiologically grounded approach.

Datasets

The authors evaluated PROFUMO using three types of data:

  • Simulated Data: Specifically designed to be challenging, testing the model's ability to handle spatial and temporal correlations, spatial variability, amplitude variations, and structured noise.
  • Human Connectome Project (HCP) Data: A large-scale dataset from over 1000 subjects, providing a robust testbed for population-level analyses.
  • Active-State fMRI Data: A smaller dataset (14 subjects) collected during various cognitive tasks, used to examine how PROFUMO captures state-dependent changes in functional connectivity.

Key Insights and Ideas

  • Spatial Variability is Key: The study strongly suggests that individual differences in the spatial organization of brain networks are a primary driver of observed functional connectivity variations, and that failing to account for this can lead to biased interpretations of temporal connectivity.
  • Temporal Connectivity Reflects Cognitive State: Once spatial variability is properly modeled, the functional coupling between brain networks appears to be more reflective of an individual's current cognitive state rather than stable, trait-like characteristics.
  • PROFUMO's Advantage in Interpretability: By disentangling spatial and temporal variations, PROFUMO offers a clearer and more interpretable view of brain function, allowing for more direct links between neuroimaging findings and behavior or cognitive states.
  • Limitations of ICA-DR: The paper highlights that standard Independent Component Analysis with Dual Regression (ICA-DR) often conflates spatial and temporal information, leading to biased estimates and reduced interpretability, especially when dealing with significant individual differences in spatial organization.

Key Results

  • Improved Spatial Map Recovery: PROFUMO demonstrated superior accuracy in recovering subject-specific spatial maps compared to ICA-DR in simulations.
  • Increased Heritability of Spatial Maps: Analyses on HCP data showed that PROFUMO's spatial maps captured significantly more heritable variance, suggesting a stronger link to genetic factors.
  • More Consistent and Specific Fingerprints: PROFUMO's subject-specific spatial maps were more consistent across different scanning sessions and provided more specific individual "fingerprints."
  • Better Classification of Cognitive States: In active-state data, PROFUMO's network matrices allowed for significantly better classification of different cognitive conditions compared to ICA-DR, indicating a more accurate capture of state-dependent connectivity changes.
  • Reduced Bias in Functional Coupling: PROFUMO's approach led to functional coupling estimates that were less biased and more sensitive to cognitive state changes compared to ICA-DR.

Limitations and Future Work

  • Computational Demands: While PROFUMO is designed to be computationally efficient, analyzing large datasets still requires significant computational resources.
  • Model Assumptions: Like all models, PROFUMO relies on simplifying assumptions about fMRI data and brain function. The authors acknowledge that the model is not perfect and that further refinements are possible.
  • Future Directions: The authors suggest that future work could explore more complex models for spatial organization, incorporate multimodal data (e.g., structural MRI) more deeply, and investigate finer-scale functional parcellations. They also highlight the potential for PROFUMO to be extended to dynamic functional connectivity analyses.

Summary updated

BibTeX
@article{harrison2020modelling,
  author = {Samuel J. Harrison and Janine D. Bijsterbosch and Andrew R. Segerdahl and Sean P. Fitzgibbon and Seyedeh-Rezvan Farahibozorg and Eugene P. Duff and Stephen M. Smith and Mark W. Woolrich},
  title = {{Modelling subject variability in the spatial and temporal characteristics of functional modes}},
  journal = {NeuroImage},
  year = {2020},
  doi = {10.1016/j.neuroimage.2020.117226},
}
Figure