Mixtures of large-scale dynamic functional brain network modes

Chetan Gohil, Evan Roberts, Ryan C. Timms, Alex Skates, Cameron Higgins, Andrew Quinn, Usama Pervaiz, Joost van Amersfoort, Pascal Notin, Yarin Gal, Stanislaw Adaszewski, Mark W. Woolrich

TL;DR

This paper introduces DyNeMo, a novel generative model for neuroimaging data that captures dynamic brain network activity as a flexible mixture of underlying "modes." Unlike previous methods that assume mutually exclusive brain states, DyNeMo's ability to model overlapping modes and long-range temporal dependencies leads to more accurate and interpretable representations of brain dynamics, particularly in MEG data.

What Problem Does This Paper Address?

Understanding how the brain's functional networks evolve over time is crucial for deciphering cognition. Traditional methods for analyzing dynamic functional connectivity (FC), like sliding window approaches or Hidden Markov Models (HMMs), often assume that only one brain network (or "state") is active at any given moment. This "mutual exclusivity" constraint, while simplifying analysis, can be an oversimplification of the brain's complex, fluid dynamics. Researchers are increasingly interested in capturing these dynamics at millisecond timescales, but existing models struggle to represent situations where multiple network patterns might co-exist or interact. This paper addresses this limitation by proposing a model that allows for flexible, non-exclusive mixtures of brain network "modes."

Methods

DyNeMo employs a novel generative model that represents time-varying functional connectivity as a linear mixture of static spatial "modes." Each mode is characterized by a mean and covariance, representing a specific pattern of brain activity. The temporal evolution of these mixtures is governed by a recurrent neural network (RNN), allowing the model to capture rich temporal structures.

Key methodological innovations include:

  • Amortized Variational Inference: This Bayesian approach, adapted from variational autoencoders, uses an inference RNN to efficiently learn model parameters from data, enabling scalability to large datasets.
  • Reparameterization Trick: This technique allows for sampling from the variational posterior distribution, enabling the use of sophisticated, non-linear generative models.
  • Linear Mixture of Modes: Unlike HMMs, DyNeMo does not enforce mutual exclusivity, allowing modes to co-activate and mix dynamically.
  • Model RNN for Temporal Dynamics: A unidirectional RNN models the temporal evolution of the mixing coefficients, capturing long-range temporal dependencies.

Datasets

The authors used the following datasets:

  • Simulated Data:
    • Simulation 1 (Long-Range Dependencies): Generated using a Hidden Semi-Markov Model (HSMM) to specifically test the model's ability to learn long-range temporal dependencies.
  • Simulation 2 (Linear Mode Mixing): Designed to evaluate DyNeMo's capacity to infer a linear mixture of co-activating modes.
  • Real MEG Data:
    • Resting-State MEG Data: Collected from 55 healthy participants.
  • Task MEG Data (Visuomotor Task): Collected from 51 healthy participants.

Key Insights and Ideas

  • Brain dynamics are more fluid than mutually exclusive states: The success of DyNeMo in modeling MEG data suggests that brain networks often co-activate and mix, rather than strictly switching between discrete states. This challenges the long-standing assumption of mutual exclusivity in many neuroimaging analyses.
  • DyNeMo offers a more interpretable decomposition: By learning "modes" that can be mixed, DyNeMo provides a richer and more nuanced understanding of brain activity compared to the discrete states of an HMM.
  • Long-range temporal dependencies are crucial: The model's ability to capture these dependencies, particularly in resting-state MEG, highlights their importance in understanding brain function and suggests they are not solely an artifact of analysis methods.
  • Task-evoked responses can be learned without explicit task timing: DyNeMo's ability to identify task-related activations in mixing coefficients, even when unaware of the task structure, demonstrates its power in uncovering meaningful neural dynamics.

Key Results

  • Improved performance over HMM: DyNeMo outperformed HMMs in simulations where mutual exclusivity was violated, particularly in reconstructing time-varying covariance.
  • Identification of plausible resting-state networks: DyNeMo inferred 10 modes from resting-state MEG data that corresponded to known functional networks (e.g., visual, sensorimotor, default mode networks).
  • Task-dependent evoked responses: In task MEG data, DyNeMo identified modes showing clear activation patterns time-locked to visual stimuli and motor responses, without prior knowledge of task timings.
  • Accurate spectral property modeling: DyNeMo provided a more accurate model of dynamic spectral properties compared to HMMs, with a statistically significant reduction in reconstruction error.
  • Short-lived mode activations: DyNeMo revealed mode activations with lifetimes of approximately 100-150 ms in resting-state MEG, consistent with fast transient brain states.

Limitations and Future Work

  • Hyperparameter sensitivity: DyNeMo, like many deep learning models, has a large number of hyperparameters that need careful tuning. The number of modes (J) is particularly important for interpretability.
  • Computational cost: While DyNeMo is designed for scalability, training can still be computationally intensive.
  • Future Work: The authors suggest several avenues for future research, including:
  • Applying DyNeMo to other neuroimaging modalities (fMRI, EEG, ECOG).
  • Using the inferred latent representation for decoding tasks (e.g., predicting behavior).
  • Developing semi-supervised learning extensions.
  • Exploring transfer learning capabilities by fine-tuning models trained on large datasets.
  • Investigating spectral pre-whitening techniques to improve high-frequency spectral content modeling.
  • Extending the model to handle unprepared sensor-level data for more direct source reconstruction.

Summary updated

BibTeX
@article{gohil2022mixtures,
  author = {Chetan Gohil and Evan Roberts and Ryan C. Timms and Alex Skates and Cameron Higgins and Andrew Quinn and Usama Pervaiz and Joost van Amersfoort and Pascal Notin and Yarin Gal and Stanislaw Adaszewski and Mark W. Woolrich},
  title = {{Mixtures of large-scale dynamic functional brain network modes}},
  journal = {NeuroImage},
  year = {2022},
  doi = {10.1016/j.neuroimage.2022.119595},
}
Figure