Multi-dynamic modelling reveals strongly time-varying resting fMRI correlations

Data and model considerations for estimating time-varying functional connectivity in fMRI

“…For instance, if temporal modulations in first-order statistics (the average pattern of activity within a state —i.e. the mean of the Gaussian distribution) were temporally independent from modulations in time-varying FC, this would violate the assumptions of the HMM and could potentially affect model stasis; in this case, modeling the mean as a separate temporal process would likely improve the estimation of time-varying FC ( Pervaiz et al., 2022 ). Furthermore, we have shown in the supplementary results that our measure of model stasis must be regarded as a summary measure and that the distribution of FC can vary above and beyond that measure within the dataset.…”

“…It should also be noted that we here only focussed on model stasis, because it is among the most fundamental measures of performance of a time-varying FC model. However, other evaluative measures, such as the ability to predict individual traits and behavior may be of interest when evaluating time-varying FC model performance, as shown in Pervaiz et al., 2020 , Vidaurre, 2021 , Pervaiz et al., 2022 and many other works. It is likely that some of the variables we here showed to reduce model stasis, such as higher similarity between subjects and fewer free parameters per state (as obtained, e.g., through a coarser parcellation), would indeed be disadvantageous when considering other evaluative measures or when conducting a time-averaged FC study.…”

Temporal segmentation of EEG based on functional connectivity network structure

“…In this simulation, the number of communities changed from 2 to 3 and back. In the interval [1,20] seconds, the functional connectivity network had two community structures. Community 1 consisted of nodes 1-10 and Community 2 consisted of nodes 15-32.…”

“…To add random diversity, 22 edges and 76 edges were selected from Community 1 and Community 2, respectively, and random weights (N(0.2, 0.1)) were added to the original edge weights (the R-MAT algorithm 39 , one of the most commonly-used network generation models; it models graph structure by a degree distribution that follows a power-law distribution, approximating the properties of real-world networks; R-MAT has been previously applied in simulating dynamic brain networks 14 .). The network in the interval [40,60] s is the same as that in the interval [1,20] s.…”

“…In the interval [20,40] s, the functional connectivity network had three community structures. Community 1 and Community 2 were the same as in the interval [1,20] s. Community 3 consisted of nodes 11-14.…”