Time-varying functional connectivity as Wishart processes

Onno P. Kampman, Joe Ziminski, Soroosh Afyouni, Mark van der Wilk, Zoe KourtziImaging 2024 · DOI

TL;DR

This paper introduces Wishart Processes (WPs) as a powerful new tool for analyzing how brain connectivity changes over time in fMRI data. By developing a comprehensive benchmarking framework, the authors demonstrate that WPs are competitive with existing methods and even outperform them in certain key tasks, offering a more robust and flexible approach to understanding dynamic brain function.

What Problem Does This Paper Address?

Understanding the human brain requires grasping its dynamic nature – how different regions coordinate and communicate over time. Functional connectivity (FC) measures this coordination by looking at correlated activity between brain regions in fMRI data. However, traditional FC analysis often assumes this connectivity is static, which is a simplification. Time-varying functional connectivity (TVFC) aims to capture these dynamic changes, but estimating it accurately is challenging. Existing methods have limitations, such as sensitivity to noise, difficulty in capturing rapid changes, or requiring careful manual tuning. This paper addresses the need for a robust, flexible, and principled method for TVFC estimation that can handle the complexities of real fMRI data.

Methods

The core contribution is the application and evaluation of Wishart Processes (WPs) for TVFC estimation. WPs are stochastic processes that model dynamic covariance matrices, making them a natural fit for TVFC. The authors leverage recent advances in approximate inference and open-source libraries to make WPs computationally viable.

Key methodological aspects include:

  • Wishart Process (WP) Model: This model treats the covariance matrix (representing connectivity) as a dynamic entity evolving over time. It's built upon Gaussian Processes (GPs) and constructs a Wishart-distributed random matrix at each time point.
  • Variational Inference: To handle the computational complexity of fitting the WP model, the authors employ variational inference, a technique that approximates the true posterior distribution.
  • Sparse GPs: To further improve computational efficiency, they use sparse Gaussian Processes, which employ "inducing points" to reduce the number of parameters.
  • Benchmarking Framework: A crucial aspect is the development of a comprehensive framework to evaluate TVFC methods. This includes:
  • Simulations: Generating synthetic data with known ground-truth covariance structures (null, constant, periodic, stepwise, state transition, boxcar).
  • Subject Phenotype Prediction: Using TVFC estimates to predict subject characteristics (e.g., age, cognitive scores).
  • Test-Retest Reliability: Assessing how consistent TVFC estimates are across multiple scans of the same subject.
  • Brain State Analysis: Identifying recurring patterns of connectivity.
  • External Stimulus Prediction: Using TVFC to predict the timing of an external stimulus in task-based fMRI.
  • Imputation Benchmark: A novel data-driven approach where methods predict unseen time points using observed data, serving as a proxy for ground truth.
  • Baseline Methods: The WP is compared against established methods: static FC (sFC), sliding window (SW) with cross-validated window length (SW-CV), and dynamic conditional correlation (DCC-MGARCH).

Datasets

  • Human Connectome Project (HCP) Resting-State fMRI Data: Used for subject phenotype prediction, test-retest reliability studies, and the imputation benchmark. This dataset consists of 812 subjects with four scans each, processed using improved fMRI reconstruction.
  • Rockland Dataset (Task-Based fMRI): Used for the external stimulus prediction benchmark. This dataset includes fMRI scans from subjects performing a visual task, with data preprocessed using FSL and denoised with ICA-AROMA.

The paper also introduces a novel data-driven imputation benchmark that can be applied to any dataset, serving as a valuable tool for future TVFC research.

Key Insights and Ideas

  • WPs Offer a Principled Approach to TVFC: The WP model provides a statistically grounded way to model dynamic covariance, moving beyond heuristic methods like sliding windows.
  • Benchmarking is Crucial and Challenging: The lack of ground truth in real fMRI data makes evaluating TVFC methods difficult. The proposed comprehensive benchmarking framework, especially the imputation benchmark, offers a more reliable way to compare methods.
  • Static FC is Insufficient for Dynamic Data: The study consistently shows that methods outperforming sFC are likely capturing genuine time-varying dynamics, suggesting that sFC is a good null model for detecting dynamic components.
  • Smoothness vs. Abrupt Changes: WPs excel at capturing smoothly changing connectivity, but struggle with sudden, sharp transitions. Conversely, methods like SW-CV can be overly sensitive to noise or miss gradual changes.
  • Imputation Benchmark as a Proxy for Ground Truth: The imputation benchmark shows strong correspondence with results from simulated data where ground truth is known, suggesting its utility for real-world scenarios.

Key Results

  • Competitive Performance: WPs performed competitively across most benchmarks, often matching or exceeding established methods.
  • External Stimulus Prediction: WPs outperformed SW-CV and DCC on the external stimulus prediction task, capturing the task-related covariance changes more effectively.
  • Reduced False Positives: WPs were less prone to spurious covariance estimates in TVFC null models (static cases) compared to SW-CV and DCC.
  • Simulation Performance: WPs showed good performance on smoothly changing covariance structures but struggled with abrupt changes (stepwise, state transition).
  • Resting-State fMRI: WP estimates showed good predictive power for subject measures (age, variance) and comparable test-retest reliability to other methods.
  • Imputation Benchmark: WPs and SW-CV generally outperformed DCC-J and sFC on the imputation benchmark, especially for dynamic data.

Limitations and Future Work

  • Computational Cost: WPs can be computationally intensive, especially for large datasets (many time points or regions). The authors suggest exploring sparse GP innovations and low-rank implementations to mitigate this.
  • Handling Abrupt Changes: WPs, due to their smooth nature, do not effectively model covariance structures with sudden change points (e.g., state transitions). Future work could incorporate change point kernels or explore state-based models.
  • Test-Retest Reliability Interpretation: The utility of test-retest studies for TVFC is questioned, as it's unclear whether methods should capture stable individual traits or session-specific states.
  • Kernel Choice: The choice of kernel function in GPs introduces inductive biases. Future work could explore more expressive kernels to incorporate domain knowledge (e.g., periodic kernels for task data).
  • Multiverse Analysis: The authors suggest using a multiverse analysis framework to increase the robustness of conclusions by testing various methodological choices in parallel.
  • Further Benchmarking: Including tb-fMRI benchmarks with richer task structures is recommended.

Summary updated

Abstract

We investigate the utility of Wishart processes (WPs) for estimating time-varying functional connectivity (TVFC), which is a measure of changes in functional coupling as the correlation between brain region activity in functional magnetic resonance imaging (fMRI). The WP is a stochastic process on covariance matrices that can model dynamic covariances between time series, which makes it a natural fit to this task. Recent advances in scalable approximate inference techniques and the availability of robust open-source libraries have rendered the WP practically viable for fMRI applications. We introduce a comprehensive benchmarking framework to assess WP performance compared with a selection of established TVFC estimation methods. The framework comprises simulations with specified ground-truth covariance structures, a subject phenotype prediction task, a test-retest study, a brain state analysis, an external stimulus prediction task, and a novel data-driven imputation benchmark. The WP performed competitively across all the benchmarks. It outperformed a sliding window (SW) approach with adaptive cross-validated window lengths and a dynamic conditional correlation (DCC)-multivariate generalized autoregressive conditional heteroskedasticity (MGARCH) baseline on the external stimulus prediction task, while being less prone to false positives in the TVFC null models.
BibTeX
@inproceedings{kampman2024time,
  author = {Onno P. Kampman and Joe Ziminski and Soroosh Afyouni and Mark van der Wilk and Zoe Kourtzi},
  title = {{Time-varying functional connectivity as Wishart processes}},
  booktitle = {Imaging},
  year = {2024},
  doi = {10.1162/imag\_a\_00184},
}
Figure