Dynamic Brain Coherence Networks Uncovered using Directional Statistics
Poster No:
1811
Submission Type:
Abstract Submission
Authors:
Anders S. Olsen1, Anders Brammer1, Morten Mørup2
Institutions:
1Technical University of Denmark, Kgs. Lyngby, Denmark, 2Technical University of Denmark, Kgs. Lungby, Denmark
First Author:
Co-Author(s):
Introduction:
Most fMRI functional connectivity (FC) studies evaluate interregional correlation, either across the entire scan (static) or in sliding windows (time-varying). However, fMRI data are usually filtered within a narrow frequency band, which enables tracking the oscillatory nature of regional signals. This approach may be favourable for investigating FC, since it disregards signal amplitude information, which generally encompasses region-specific activity and transient noise such as motion [1].
Here we characterize interregional, instantaneous phase coherence in the resting human brain. For each volume, we produce a connectivity vector, the pool of which is clustered using mixture models. We uncover time-varying resting-state networks in a cohort of 1003 healthy young adults from the human connectome project (HCP) at fsLR resolution (p=91.282 voxels). To our knowledge, we present the first atlas of instantaneous FC in a large cohort. We provide three versions with varying degrees of mixture model complexity, which may serve as alternatives to existing FC network atlases by relying on instantaneous connectivity rather than time-averages.
Here we characterize interregional, instantaneous phase coherence in the resting human brain. For each volume, we produce a connectivity vector, the pool of which is clustered using mixture models. We uncover time-varying resting-state networks in a cohort of 1003 healthy young adults from the human connectome project (HCP) at fsLR resolution (p=91.282 voxels). To our knowledge, we present the first atlas of instantaneous FC in a large cohort. We provide three versions with varying degrees of mixture model complexity, which may serve as alternatives to existing FC network atlases by relying on instantaneous connectivity rather than time-averages.
Methods:
Preprocessed and denoised HCP-fMRI data were decomposed into phase by applying a Hilbert transform, followed by establishing a -dimensional frame-wise phase coherence matrix A_t=cos(\theta_it-\theta_jt) for each region i and j and time point t (see Figure 1). Due to the angle difference identity, A_t has rank 2 [2]. Thus, the eigendecomposition allows us to capture the majority of the variance in the phase coherence matrix using a px2-dimensional basis. Orthogonal bases are distributed on the Grassmann manifold, while each eigenvector (e.g., the leading eigenvector as in LEiDA [3]) is distributed on the sign-symmetric unit hypersphere [4]. Directional statistics offers suitable statistical distributions for these manifolds, namely the Watson and Angular Central Gaussian (ACG) distributions for the sign-symmetric hypersphere, and matrix ACG (MACG) for the Grassmann manifold [5-7]. While the Watson distribution is parameterized by a mean direction vector and a scalar precision, the ACG and MACG are parameterized by a full or rank-reduced covariance matrix.
Results:
We split the data into a train and test set and evaluated a range of mixture sizes, K={2,...,20}. For Watson mixtures, the test-log-likelihood did not saturate within the evaluated range of K. For ACG/MACG mixtures, K=7 networks were found to be suitable based on an elbow in the test-log-likelihood curve. For simplicity, we present seven Watson networks trained on the full data set (Figure 2). The inferred networks include high-level networks such as the frontoparietal and default-mode networks and some more diffuse background states.
Conclusions:
We introduce time-varying high-resolution functional brain networks estimated without assuming any signal stationarity through, e.g., windowing, and by assessing only interregional phase coherence, disregarding potentially spurious amplitude information. For Watson mixtures, we observed states involving brain areas responsible for higher-order cognitive function, while somatosensory regions were less clear. Collectively, the current observations suggest an opportunity to gain complementary information about brain connectomics via time-varying functional coherence based analyses.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 1
Methods Development
Task-Independent and Resting-State Analysis 2
Neuroinformatics and Data Sharing:
Brain Atlases
Keywords:
Computational Neuroscience
Data analysis
FUNCTIONAL MRI
Machine Learning
Other - Directional statistics
1|2Indicates the priority used for review
·Figure 1: Methodological framework for leading eigenvector dynamics analysis and Watson mixture models
·Figure 2: Networks for a K=7 solution for Watson mixture models, including centroids, prior \pi, and precision \kappa.
Provide references using author date format
[2] Olsen A.S. et al. (2022), 'Psilocybin Modulation of Time-Varying Functional Connectivity is Associated with Plasma Psilocin and Subjective Effects', NeuroImage.
[3] Cabral J. et al. (2017), 'Cognitive performance in healthy older adults relates to spontaneous switching between states of functional connectivity during rest', Nature Scientific Reports
[4] Olsen A.S. et al. (2023), 'Angular Central Gaussian and Watson Mixture Models for Assessing Dynamic Functional Brain Connectivity During a Motor Task', ICASSP2023.
[5] Watson G.S. (1965), 'Equatorial Distributions on a Sphere', Biometrika
[6] Tyler D.E. (1987), 'Statistical Analysis for the Angular Central Gaussian Distribution on the Sphere', Biometrika.
[7] Chikuse Y. (1990), 'The Matrix Angular Central Gaussian Distribution', Journal of Multivariate Analysis.