Recursive dynamic functional connectivity reveals a characteristic correlation structure in human scalp EEG

Abstract

Time-varying neurophysiological activity has been classically explored using correlation based sliding window analysis. However, this method employs only lower order statistics to track dynamic functional connectivity of the brain. We introduce recursive dynamic functional connectivity (rdFC) that incorporates higher order statistics to generate a multi-order connectivity pattern by analyzing neurophysiological data at multiple time scales. The technique builds a hierarchical graph between various temporal scales as opposed to traditional approaches that analyze each scale independently. We examined more than a million rdFC patterns obtained from morphologically diverse EEGs of 2378 subjects of varied age and neurological health. Spatiotemporal evaluation of these patterns revealed three dominant connectivity patterns that represent a universal underlying correlation structure seen across subjects and scalp locations. The three patterns are both mathematically equivalent and observed with equal prevalence in the data. The patterns were observed across a range of distances on the scalp indicating that they represent a spatially scale-invariant correlation structure. Moreover, the number of patterns representing the correlation structure has been shown to be linked with the number of nodes used to generate them. We also show evidence that temporal changes in the rdFC patterns are linked with seizure dynamics.

Multiplex dynamic networks in the newborn brain disclose latent links with neurobehavioral phenotypes

“…First, it extracts latent representations from group‐level mdFCNs to effectively suppress noise and/or intersubject variations. This is essential to uncover the group‐level structure (Mehrkanoon et al, 2021 ; Zhu, Liu, Mathiak, et al, 2020 ), to support the network decomposition task because high variability leads to superfluous components, and to yield reliable/repeatable findings (Panwar et al, 2021 ; Shellhaas et al, 2022 ; Westende et al, 2020 ). The latter is paramount because phase‐based functional connectivity measures are challenged with relatively lower test–retest reliability when compared to amplitude‐based connectivity metrics (Colclough et al, 2016 ).…”

“…First, the 3 min time‐courses of cortical signals were segmented by a 2 s sliding window with 50% overlap, resulting in 179 temporal segments for each cortical signal. The 2 second window duration was taken to capture at least one cycle of the lowest frequency (low‐δ oscillations), and to be short enough to reflect the rapid temporal dynamics in FCNs (Bassett & Sporns, 2017 ; Du et al, 2018 ; Haartsen et al, 2020 ; Panwar et al, 2021 ). A segment wise cross‐spectrum was computed between all pairs of cortical parcel signals using a discrete Fourier transform, which yielded estimates of phase–phase synchronization via the debiased weighted phase lag index (dwPLI) that is known for robustness against volume conduction effects (Yu, 2020 ).…”

“…Studies on the functional connectivity are usually focused on measuring time‐averaged neuronal coupling at specific oscillatory frequencies; however, neuronal activity is known to exhibit multiplex network dynamics (Bassett & Sporns, 2017 ; Zhu, Liu, Mathiak, et al, 2020 ; Zhu, Liu, Ye, et al, 2020 ) characterized by rapid changes at sub‐second time scales (Bassett & Sporns, 2017 ; Haartsen et al, 2020 ; Panwar et al, 2021 ), as well as coincident interactions at a wide range of oscillatory frequencies (Bassett & Sporns, 2017 ; Haartsen et al, 2020 ; Westende et al, 2020 ). Recently, several systematic approaches of multiplex network analysis were proposed to accommodate coexisting networks operating at different oscillatory frequencies (Kivelä et al, 2014 ), and/or to measure the network dynamics seen as time‐varying changes in the network properties (Du et al, 2018 ; Haartsen et al, 2020 ).…”

Dynamic Connectivity Analysis Using Adaptive Window Size

“…The 75 s window size was chosen because the most relationships were found with it. The higher-order dynamic functional connectivity time series were computed in [ 31 ] using the constant sliding window analysis method. In [ 32 ], the wPLI was computed based on predefined 2 s epochs, and constant sliding window analysis was proposed to extract information about connectivity patterns in the future, focusing on determining the tradeoff between temporal resolution and estimation error, i.e., determining the optimal window size.…”