On null models for temporal small-worldness in brain dynamics

Aurora Rossi, Samuel Deslauriers‐Gauthier, Emanuele NataleNetwork 2024 · DOI

TL;DR

This paper introduces a new statistical null model, the Random Temporal Hyperbolic (RTH) graph, designed to better capture the "small-worldness" properties of dynamic brain networks derived from fMRI data. The RTH model proves superior to existing null models in replicating the temporal small-worldness observed in resting-state brain activity, offering a more robust tool for hypothesis testing in neuroscience.

What Problem Does This Paper Address?

Understanding how the brain functions involves analyzing its complex network dynamics over time. Functional magnetic resonance imaging (fMRI) allows us to capture these dynamics by representing brain activity as temporal networks, where brain regions are nodes and their interactions are edges that change over time. To rigorously test hypotheses about these brain networks, it's crucial to compare them against a "null model" – a statistical construct that shares certain properties with the real data but is otherwise random. The challenge is that existing null models for temporal brain networks don't adequately capture the specific characteristics of real brain activity, particularly its "temporal small-worldness." This property, related to efficient information exchange, is vital for brain function, and its accurate representation in null models is essential for reliable scientific inference.

Methods

The authors propose and evaluate the Random Temporal Hyperbolic (RTH) graph model. This model is an extension of the static Random Hyperbolic (RH) graph, known for its ability to reproduce properties of real-world networks. The RTH model is a synthetic model, meaning it generates networks based on mathematical principles rather than directly randomizing empirical data.

Key methodological aspects include:

  • Temporal Network Extraction: They first process fMRI data to extract temporal brain networks. This involves linear regression and filtering of BOLD signals, followed by a sliding window approach to capture time-varying correlations between brain regions.
  • Temporal Small-Worldness Definition: A new definition of temporal small-worldness ($S = C/L$) is proposed, using the temporal clustering coefficient ($C$) and temporal path length ($L$). This differs from previous definitions by using the temporal clustering coefficient directly, rather than a temporal correlation coefficient.
  • Model Comparison: They compare the RTH model against two data-based null models (Random Temporal Permuted Times - RTPT, and Random Temporal Edges - RTE) and two other synthetic geometric models (Random Temporal Square - RTS, and Random Temporal Torus - RTT).
  • Parameter Optimization: Two optimization methods are used to find the best parameters for the synthetic models to match empirical data properties like average degree and temporal small-worldness.

Datasets

The primary dataset used in this study is from the Human Connectome Project (WU-Minn HCP 1200 Subjects Data Release). Specifically, they utilized resting-state fMRI (rfMRI) data from 1,047 subjects. The data was preprocessed and minimally preprocessed according to established protocols. The authors also make their code and extracted temporal brain networks publicly available.

Key Insights and Ideas

  • RTH Model's Superiority: The RTH graph model, with its hyperbolic geometry, is shown to be a significantly better null model for temporal brain networks than existing methods, particularly in capturing temporal small-worldness.
  • Importance of Geometric Structure: The success of the RTH model suggests that the underlying geometric properties of the network space (hyperbolic in this case) are crucial for accurately modeling brain dynamics, beyond simple randomization of edges or time.
  • New Temporal Small-Worldness Definition: The proposed definition of temporal small-worldness ($S = C/L$) is shown to be more consistent with the static definition and better reflects the temporal dynamics of brain networks compared to alternative definitions.
  • Parameter Efficiency: The RTH model achieves good fits with relatively few parameters, making it a parsimonious yet powerful tool.

Key Results

  • The RTH graph model, particularly with specific parameter settings (e.g., $\alpha=0.65, v=0.6$ for $S$ and $\alpha=0.8, v=0.9$ for $SSB$), consistently demonstrated the closest temporal small-worldness values to empirical resting-state fMRI data across various average degrees.
  • In quantitative comparisons (Tables 2 and 3), the RTH model showed the smallest distances (area, minimum, maximum, mean absolute, and mean Euclidean) between its median temporal small-worldness curves and those of the empirical data, for both definitions of small-worldness.
  • The data-based models (RTPT and RTE) performed worse than the synthetic geometric models (RTS, RTT, and RTH), likely because they lack tunable parameters to match specific network properties.

Limitations and Future Work

  • Focus on Resting-State: The study exclusively focuses on resting-state fMRI data from healthy individuals. The authors suggest future work could extend this to patients with neurological or psychiatric conditions (e.g., Alzheimer's, Parkinson's, ADHD, depression) to see if the RTH model remains optimal.
  • Task-Based Networks: Investigating task-based fMRI data, where cognitive processes are engaged, is another avenue for future research.
  • Scale-Freeness: While the paper focuses on small-worldness, the authors note that extending the analysis to temporal scale-freeness would be a valuable next step, and it would be interesting to see if the RTH model also excels in this regard.
  • Alternative Null Models: Future work could also explore comparing the RTH model with null models based on phase randomization and autoregressive randomization of brain dynamics.

Summary updated

Abstract

Brain dynamics can be modeled as a temporal brain network starting from the activity of different brain regions in functional magnetic resonance imaging (fMRI) signals. When validating hypotheses about temporal networks, it is important to use an appropriate statistical null model that shares some features with the treated empirical data. The purpose of this work is to contribute to the theory of temporal null models for brain networks by introducing the random temporal hyperbolic (RTH) graph model, an extension of the random hyperbolic (RH) graph, known in the study of complex networks for its ability to reproduce crucial properties of real-world networks. We focus on temporal small-worldness which, in the static case, has been extensively studied in real-world complex networks and has been linked to the ability of brain networks to efficiently exchange information. We compare the RTH graph model with standard null models for temporal networks and show it is the null model that best reproduces the small-worldness of resting brain activity. This ability to reproduce fundamental features of real brain networks, while adding only a single parameter compared with classical models, suggests that the RTH graph model is a promising tool for validating hypotheses about temporal brain networks.
BibTeX
@inproceedings{rossi2024null,
  author = {Aurora Rossi and Samuel Deslauriers‐Gauthier and Emanuele Natale},
  title = {{On null models for temporal small-worldness in brain dynamics}},
  booktitle = {Network},
  year = {2024},
  doi = {10.1162/netn\_a\_00357},
}
Figure