Subject identification using edge-centric functional connectivity

Abstract

Group-level studies do not capture individual differences in network organization, an important prerequisite for understanding neural substrates shaping behavior and for developing interventions in clinical conditions. Recent studies have employed "fingerprinting" analyses on functional connectivity to identify subjects' idiosyncratic features. Here, we develop a complementary approach based on an edge-centric model of functional connectivity, which focuses on the co-fluctuations of edges. We first show whole-brain edge functional connectivity (eFC) to be a robust substrate that improves identifiability over nodal FC (nFC) across different datasets and parcellations. Next, we characterized subjects' identifiability at different spatial scales, from single nodes to the level of functional systems and clusters using k-means clustering. Across spatial scales, we find that heteromodal brain regions exhibit consistently greater identifiability than unimodal, sensorimotor, and limbic regions. Lastly, we show that identifiability can be further improved by reconstructing eFC using specific subsets of its principal components. In summary, our results highlight the utility of the edge-centric network model for capturing meaningful subject-specific features and sets the stage for future investigations into individual differences using edge-centric models.

Individualized event structure drives individual differences in whole-brain functional connectivity

“…High-amplitude cofluctuations make proportionally bigger contributions to time-averaged FC than low-amplitude cofluctuations. This statement is non-controversial; edge time series are a mathematically precise "temporal unwrapping" of the Pearson correlation into its framewise contributions, the average of which is simply FC [19,[23][24][25][26]. For this reason, it makes sense to focus on frames where many edges simultaneously make big contributions -those same frames necessarily will, on average, make bigger contributions to FC than, say, frames where only a few edges exhibit high-amplitude edge time series.…”

“…Here, we address these questions directly. Our approach leverages a recently-proposed method for decomposing FC into its framewise contributions, detecting events, and assessing the impact of events on timeaveraged FC [19,[23][24][25][26]. We apply this framework to two independently acquired datasets: the Midnight Scan Club [11,12] and the MyConnectome project [10,27].…”

An edge-centric model for harmonizing multi-relational network datasets

“…For instance, they could represent connectivity estimates made using different imaging modalities, while subjects complete different tasks, or different measurements of connection weight made on the same network dataset. They could even reflect connectivity estimates at different points in time, in which case edge covariance is nearly identical to edge functional connectivity, which we analyzed in previous studies [28][29][30].…”

“…Other recent papers have deliberately adopted edge-centric approaches, usually citing [33,34] as inspiration. In addition to the edge time series and edge functional connectivity models discussed earlier [28][29][30][31][32], others have generated edgeedge networks by embedding edges in a some metric space and using inter-edge distances to quantify the connection weight between edges [35].…”

“…Both group-and subject-level morphometric networks are defined at the regional level, meaning that network nodes represent brain regions. Recently, we proposed an alternative, edge-centric method for constructing brain networks, emphasizing a network's edges as its irreducible units and generating a higher-order network [28][29][30][31][32]. Edge-level networks have a number of useful properties not shared by node-level networks, e.g.…”