Hierarchical modelling of functional brain networks in population and individuals from big fMRI data

Seyedeh-Rezvan Farahibozorg, Janine D. Bijsterbosch, Weikang Gong, Saâd Jbabdi, Stephen M. Smith, Samuel J. Harrison, Mark W. WoolrichNeuroImage 2021 · DOI

TL;DR

This paper introduces sPROFUMO, a novel hierarchical modeling framework that efficiently analyzes large-scale fMRI datasets to map individual brain networks. It demonstrates that these detailed network maps can predict cognitive abilities, offering a powerful new tool for understanding individual differences in brain function.

What Problem Does This Paper Address?

Understanding how individual brains function and how this relates to cognitive abilities is a major goal in neuroscience. Large datasets like the UK Biobank offer an unprecedented opportunity to study these individual differences. However, existing methods for analyzing functional brain networks from fMRI data struggle with the sheer scale of these datasets, leading to computational bottlenecks. Furthermore, accurately capturing the subtle variations in brain networks across individuals while maintaining a robust group-level understanding has remained a challenge. This paper aims to bridge this gap by developing a scalable and accurate method for hierarchical modeling of brain networks that can be applied to massive fMRI datasets and used to predict cognitive traits.

Methods

The core contribution is the development of stochastic PROFUMO (sPROFUMO), an advancement of the existing PROFUMO framework. The key innovation is the incorporation of stochastic variational Bayes (VB). This allows the model to process data in smaller, random batches rather than requiring the entire dataset at once, drastically reducing computational costs (by a factor of ~100) and memory requirements.

Key methodological aspects include:

  • Hierarchical Modeling: sPROFUMO maintains a bidirectional relationship between group-level network models and individual subject-level models. This allows for both learning a general population model and capturing individual deviations.
  • Stochastic Variational Inference: This is the engine for scalability. It involves updating model parameters iteratively using mini-batches of subjects, significantly speeding up computation.
  • Double-Gaussian Mixture Model (DGMM): Used for modeling spatial maps, this helps to distinguish signal from noise, leading to cleaner individual network estimates.
  • MELODIC's Incremental Group-PCA (MIGP): Used for initial data dimensionality reduction, it's an online method that scales to large datasets.
  • Elastic-Net Regression: Employed for predicting cognitive outcomes from the derived brain network features.

Datasets

  • Simulated Data: Two sets of simulations were performed using synthetic fMRI data from 500 subjects with two recording sessions each. These simulations were designed to mimic realistic fMRI settings and test the model's performance under various challenging scenarios (e.g., cross-subject variability, mode overlap, small modes).
  • UK Biobank (UKB): The primary real-world dataset used. Data from 4999 subjects were analyzed, including resting-state fMRI (one session per subject, ~6 minutes) and 68 cognitive tests.

Key Insights and Ideas

  • Scalability is Key: The stochastic approach is crucial for unlocking the potential of big fMRI datasets like UK Biobank, making complex hierarchical modeling computationally feasible.
  • Individual Networks Matter for Cognition: Detailed, subject-specific brain network maps are not just descriptive; they can predict individual cognitive abilities. This highlights the importance of moving beyond group averages.
  • Spatial Maps Hold Predictive Power: sPROFUMO's emphasis on detailed spatial configurations of brain networks proved more predictive of cognitive traits than functional connectivity alone, a finding that contrasts with ICA-DR.
  • Noise Modeling Improves Network Estimation: The DGMM approach effectively separates signal from noise, leading to cleaner and more reliable individual network maps, especially important for datasets with shorter scan times per subject.
  • High Dimensionality Reveals Nuance: Decomposing brain networks into a larger number of modes (e.g., 150) reveals more fine-grained functional subdivisions and potentially new RSNs that are missed by lower-dimensional approaches.

Key Results

  • Computational Speedup: sPROFUMO reduces RAM usage by approximately 100-fold compared to classic PROFUMO, making it feasible for datasets with tens of thousands of participants.
  • Simulation Performance: sPROFUMO demonstrated robustness to parameter choices (batch size, $\beta$) and outperformed ICA-DR in scenarios involving significant cross-subject variability and small modes, often by over 100% in accuracy.
  • UKB Network Mapping: The model successfully mapped over 100 resting-state networks (RSNs) in individual subjects from UKB, providing unprecedented detail.
  • Cognitive Prediction: sPROFUMO's multi-element model summaries achieved significant prediction accuracies for cognitive tests (average correlation ~0.3 for memory/executive functions).
  • Double Dissociation: A key finding was that sPROFUMO's spatial maps were better predictors of cognitive traits than ICA-DR's spatial maps, while ICA-DR's temporal connectivity outperformed sPROFUMO's temporal connectivity.
  • High Dimensionality Benefits: Increasing the number of modes from 100 to 150 improved prediction accuracy and revealed new cognitively-relevant fine-grained modes.

Limitations and Future Work

  • Mode Dimensionality: While 150 modes showed good results for UKB data (short recordings), the optimal dimensionality might vary. The paper suggests 100-150 modes might be a better choice than 200 for such datasets.
  • Interpretability of High Dimensionality: The addition of new, spatially overlapping modes at higher dimensions raises questions about their precise functional interpretation and how they relate to existing RSNs.
  • Future Work: The authors suggest further testing the model with other phenotypes, exploring the interplay between spatial and temporal characteristics more deeply, and investigating the generalizability of findings on high-dimensional PFMs to other studies. The sPROFUMO code and discovered PFMs will be made publicly available.

Summary updated

BibTeX
@article{farahibozorg2021hierarchical,
  author = {Seyedeh-Rezvan Farahibozorg and Janine D. Bijsterbosch and Weikang Gong and Saâd Jbabdi and Stephen M. Smith and Samuel J. Harrison and Mark W. Woolrich},
  title = {{Hierarchical modelling of functional brain networks in population and individuals from big fMRI data}},
  journal = {NeuroImage},
  year = {2021},
  doi = {10.1016/j.neuroimage.2021.118513},
}
Figure