Discovering dynamic brain networks from big data in rest and task

Diego Vidaurre, Romesh G. Abeysuriya, Robert Becker, Andrew J. Quinn, Fidel Alfaro-Almagro, Stephen M. Smith, Mark W. WoolrichNeuroImage 2018 · DOI

TL;DR

This paper introduces a computationally efficient method, stochastic variational inference, to analyze dynamic brain networks from massive neuroimaging datasets. This advancement allows for more reproducible and interpretable insights into brain function during both rest and task states, overcoming limitations of previous methods.

What Problem Does This Paper Address?

Understanding how brain activity changes over time, known as dynamic brain networks, is crucial for comprehending cognition and neurological disorders. Traditional methods like sliding windows struggle with the trade-off between capturing fast dynamics and having enough data for reliable estimation. Furthermore, the increasing availability of large-scale datasets (e.g., Human Connectome Project, UK Biobank) has outpaced the computational capacity of existing analysis techniques, particularly for complex models like Hidden Markov Models (HMMs) that can capture rich temporal and spectral information. This paper addresses the challenge of scaling HMM-based dynamic brain network analysis to these massive datasets, enabling more robust and reproducible findings.

Methods

The core contribution is the development and application of a stochastic variational inference (SVI) approach for Hidden Markov Models (HMMs).

  • Hidden Markov Models (HMMs): The paper builds upon HMMs, which model time series data as a sequence of discrete "states." Each state represents a distinct pattern of brain activity (e.g., functional connectivity, spectral properties). The model infers these states and the probabilities of transitioning between them.
  • Standard Variational Inference: This is a common technique for approximating the complex probability distributions in HMMs. However, it requires loading the entire dataset into memory and can be computationally intensive, especially for large datasets.
  • Stochastic Variational Inference (SVI): The authors adapt SVI, a method rooted in stochastic optimization, to the HMM framework. Instead of using the full dataset at each iteration, SVI uses random subsets (batches) of subjects. This significantly reduces memory requirements and computational time, making it feasible to analyze datasets with thousands of subjects.
  • Observation Models: The paper demonstrates the SVI approach with different HMM observation models:
  • Gaussian: Models each state's activity using a multivariate Gaussian distribution, capturing mean activation and functional connectivity.
  • MAR (Multivariate Autoregressive): Models each state's spectral properties and cross-channel interactions, suitable for higher temporal resolution data like MEG.

Datasets

The proposed method was evaluated on several large-scale datasets:

  • Human Connectome Project (HCP) Resting-State fMRI: Data from 820 healthy adults.
  • UK Biobank Resting-State fMRI: Data from 5248 healthy adults.
  • Human Connectome Project (HCP) Task MEG: Data from 52 subjects performing a motor task.
  • Simulated Data: Both Gaussian and MAR models were used to generate synthetic data for validation.

The paper does not introduce new benchmarks but rather demonstrates the scalability of HMM analysis on existing large datasets.

Key Insights and Ideas

  • Scalability is Key: The primary insight is that computational efficiency is the bottleneck for analyzing large-scale neuroimaging data with advanced models like HMMs. SVI provides a practical solution.
  • Reproducibility Across Runs: The SVI approach demonstrated high consistency across multiple runs with different random seeds, suggesting robustness and reliability even with the stochastic nature of the algorithm.
  • Robustness to Data Splits: Splitting large datasets in half and running the HMM separately yielded highly consistent results, further validating the reliability of the inferred brain states.
  • Task Dynamics Revealed: Even when blind to task timing, the HMM with MAR models could identify states that reflected task-specific dynamics in MEG data, highlighting the power of spectral analysis in capturing temporal patterns.
  • Beyond Sliding Windows: The paper reinforces that HMMs offer a more principled way to capture dynamic brain activity without being constrained by arbitrary window sizes, allowing for the discovery of faster and more complex temporal patterns.

Key Results

  • Computational Speedup: The SVI approach significantly reduces the computational burden, allowing HMMs to be run on datasets with thousands of subjects in a matter of hours, a feat impossible with standard variational inference. For example, the HCP resting-state fMRI data (820 subjects) took approximately 221 minutes per run.
  • High Correlation with Standard Inference: For simulated data, the SVI approach achieved high correlations (often > 0.75) with the standard, non-stochastic HMM inference, demonstrating its accuracy.
  • Robust State Inference: The method successfully identified consistent brain states across different runs and data splits for both resting-state fMRI and task-based MEG data. For instance, in the HCP resting-state data, 12 states were reliably inferred.
  • Discovery of Task-Relevant States: In the MEG task data, the HMM identified states that clearly differentiated between task-related activity and baseline states, capturing rhythmic patterns related to button presses.

Limitations and Future Work

  • Overfitting with Many Regions: While SVI addresses computational scaling with large datasets, it does not inherently solve the problem of overfitting when dealing with a very large number of brain regions (high dimensionality). The authors suggest that collecting more data can help mitigate this.
  • Data Harmonization: Combining data from different modalities or studies requires sufficiently uniform preprocessing pipelines and a common brain space, which is an ongoing challenge.
  • Combining Models: The authors propose future work on combining multiple HMM models (ensemble methods) to achieve even richer descriptions of brain dynamics, leveraging the computational efficiency of SVI.
  • Exploring Joint Models: Estimating a joint model across different datasets or modalities in a single inference process is another promising avenue for future research.

Summary updated

BibTeX
@article{vidaurre2018discovering,
  author = {Diego Vidaurre and Romesh G. Abeysuriya and Robert Becker and Andrew J. Quinn and Fidel Alfaro-Almagro and Stephen M. Smith and Mark W. Woolrich},
  title = {{Discovering dynamic brain networks from big data in rest and task}},
  journal = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.neuroimage.2017.06.077},
}
Figure