Abstract: In neuroscience, functional brain connectivity describes the connectivity between brain regions that share functional properties. Neuroscientists often characterize it by a time series of covariance matrices between functional measurements of distributed neuron areas. An effective statistical model for functional connectivity and its changes over time is critical for better understanding the mechanisms of brain and various neurological diseases. To this end, we propose a matrix-log mean model with an additive heterogeneous noise for modeling random symmetric positive definite matrices that lie in a Riemannian manifold. The heterogeneity of error terms is introduced specifically to capture the curved nature of the manifold. We then propose to use the local scan statistics to detect change patterns in the functional connectivity. Theoretically, we show that our procedure can recover all change points consistently. Simulation studies and an application to the Human Connectome Project lend further support to the proposed methodology.
Introduction
In modern neuroscience, understanding how different brain regions coordinate—known as functional connectivity—is a central challenge. This coordination is typically represented by Symmetric Positive Definite (SPD) covariance matrices derived from functional measurements like fMRI. Because the brain is a dynamic system, these connectivity patterns are not static; they shift as individuals perform tasks or transition between cognitive states. Detecting these "change points" is vital for mapping brain function and diagnosing neurological disorders.
The Riemannian Challenge
A significant hurdle in modeling functional connectivity is the geometric nature of SPD matrices. These matrices do not reside in a flat Euclidean space but on a nonlinear Riemannian manifold. Applying standard statistical tools designed for flat space to these curved surfaces is akin to using a flat map to measure distances on a globe: it inevitably introduces distortion. Previous research often relied on "naive" log-transformations that failed to account for the manifold's curvature, leading to inaccuracies in change-point detection.
Methodology
The Heterogeneous Matrix-log Mean Model
To address the limitations of Euclidean models, this paper introduces a novel statistical framework: the Heterogeneous Matrix-log Mean Model. The core innovation lies in the treatment of noise. Instead of a simple additive model, the authors propose:
$$\log Y_i = \log \mu_i + \text{log}'_{\mu_i} \epsilon_i$$
Here, $\text{log}'_{\mu_i}$ represents the differential of the matrix logarithm. This term ensures that the noise ($\epsilon_i$) is properly scaled according to the "tangent space" of the mean matrix ($\mu_i$). By incorporating this location-dependent transformation, the model becomes intrinsic to the manifold, respecting its curved geometry rather than forcing it into a flat projection.
Local Scan Statistics
To identify transitions in brain states, the authors employ local scan statistics. This involves a sliding window approach that calculates a weighted average of transformed matrices. A spike in this statistic indicates a high probability of a change point. The optimal number of change points is determined through K-fold cross-validation, ensuring the model captures meaningful shifts rather than random noise.
Experimental Results
Simulation and Theoretical Consistency
The authors provide a mathematical proof (Theorem 4.1) demonstrating that their procedure consistently recovers all change points as sample sizes increase. In synthetic simulations, the proposed method achieved a higher Sure Coverage Probability (SCP) than traditional Euclidean-based methods, proving that geometric awareness leads to more accurate detection.
Application to the Human Connectome Project (HCP)
The model was applied to fMRI data from 850 healthy adults in the Human Connectome Project. Subjects watched video clips of geometric objects interacting in either "social" or "random" patterns. The model detected an average of 3.66 change points per subject. Notably, for subjects with four detected change points, the timing (approx. 40s, 72s, 107s, 138s) closely mirrored the actual task transitions in the experiment (35s, 70s, 105s, 140s).
An intriguing finding was that detected change points often slightly preceded the actual video transitions. This suggests the model is sensitive enough to capture the brain's anticipatory mechanisms as it learns to predict upcoming shifts in the visual stimuli.
Conclusion
By treating covariance matrices as points on a Riemannian manifold, this research provides a more robust framework for dynamic brain modeling. While the current model assumes independence between time points, it sets the stage for future work involving temporal dependencies. This geometric approach ensures that heterogeneity is treated not as a nuisance, but as a fundamental mathematical necessity for understanding the curved nature of brain connectivity.
Summary updated