Inter-subject and inter-parcellation variability of resting-state whole-brain dynamical modeling

Abstract

Modern approaches to investigate complex brain dynamics suggest to represent the brain as a functional network of brain regions defined by a brain atlas, while edges represent the structural or functional connectivity among them. This approach is also utilized for mathematical modeling of the resting-state brain dynamics, where the applied brain parcellation plays an essential role in deriving the model network and governing the modeling results. There is however no consensus and empirical evidence on how a given brain atlas affects the model outcome, and the choice of parcellation is still rather arbitrary. Accordingly, we explore the impact of brain parcellation on inter-subject and inter-parcellation variability of model fitting to empirical data. Our objective is to provide a comprehensive empirical evidence of potential influences of parcellation choice on resting-state whole-brain dynamical modeling. We show that brain atlases strongly influence the quality of model validation and propose several variables calculated from empirical data to account for the observed variability. A few classes of such data variables can be distinguished depending on their inter-subject and inter-parcellation explanatory power.

Towards an Efficient Validation of Dynamical Whole-brain Models

“…The global weights for coupling 𝐶 ∈ [0, 0.945] and delay 𝜏 ∈ [0, 94] were tuned along a set of 64 × 48 = 3072 discrete, equidistant points. The selected parameter range for the considered model was suspected to include the optimal model parameters of the best correspondence between simulated and empirical data, which was confirmed by postsimulation analyses 22 . Furthermore, the considered grid granularity was selected as a trade-off between the computational costs and a favorable density that allowed for a close approximation of the goodness-of-fit values as confirmed in this study by comparing the fitting results with those of other optimization methods.…”

“…All methods were performed in accordance with the relevant guidelines and regulations. The given datasets had been utilized for the extraction of structural and resting-state functional connectivity (SC and FC, respectively) in related works [20][21][22] . We used the obtained empirical connectomes for the model derivation and validation.…”

“…For a broad exploratory investigation of the noise impact, its intensity was varied between 0 (noisefree scenario) and a maximal amplitude of 2, which was inspired by Deco et al 14 and indicates a very strong perturbation of the system dynamics. In view of the rising computational requirements, the highest value of delay 𝜏 was reduced to 48, which however did not distort the results, because the parameters maximizing the model fit did not suggest very large delays 22 .…”

“…In other words, they enable researchers to investigate the relationship between brain structure and function with a particular focus on the question whether and how the latter emerges from the former and how they correlate [13][14][15][16][17][18][19] . Additionally, models offer a quick in silico experimental way to study and compare different brain parcellations, network configurations and parameters of datapreprocessing, which again contributes to a deeper understanding of the interaction between the brain architecture and dynamics [20][21][22] . Another advantage of the discussed modeling approach is that it allows for a meaningful description of complex natural phenomena by a set of interpretable parameters.…”

Searching for Consistent Brain Network Topologies Across the Garden of (Shortest) Forking Paths

“…However, perhaps the most common approach in human network neuroscience is the use of parcellations: pre-defined assignments of neighbouring voxels into regions-of-interest (ROIs). A wide variety of parcellations exist (Eickhoff et al, 2018), and recent work reported how the choice of parcellation scheme can affect aspects such as structure-function similarity estimation (Messé, 2020) but also the intra-subject and inter-subject variability of the whole-brain resting-state modeling (Popovych et al, 2021). Parcellation schemes vary on two main dimensions: the criterion based on which clusters are identified (e.g.…”