Multivariate synchrony modules identified through multiple subject community detection in functional brain networks

The functional connectivity of the human brain may be described by modeling interactions among its neural assemblies as a graph composed of vertices and edges. It has recently been shown that functional brain networks belong to a class of scale-free complex networks for which graphs have helped define an association between function and topology. These networks have been shown to possess a heterogenous structure composed of clusters, dense regions of strongly associated nodes, which represent multivariate relationships among nodes. Network clustering algorithms classify the nodes based on a similarity measure representing the bivariate relationships and similar to unsupervised learning is performed without a priori information. In this paper, we propose a method for partitioning a set of networks representing different subjects and reveal a community structure common to multiple subjects. We apply this community identifying algorithm to functional brain networks during a cognitive control task, in particular the error-related negativity (ERN), to evaluate how the brain organizes itself during error-monitoring.

I. INTRODUCTION

Functional connectivity is defined as the temporal correlations between spatially remote neurophysiological events [1] and describes the neural processes required for cognitive and motor tasks. Functional connectivity has been quantified by applying coherence or nonlinear synchronization measures to various neuroimaging data. In previous work, the bivariate relationships between neuronal populations have been represented as graphs composed of vertices and edges. These graphs were then analyzed using various measures from graph theory including small-world measures and centrality measures for hub classification. One remaining issue is to identify the functional modules in these neural networks through community detection. A community structure is defined as the natural tendency of a network's vertices to divide into modules containing a dense number of intra-connecting edges within each module and a sparse number of interconnecting edges between modules. Community detection methods can be categorized as divisive, agglomerative, or optimal where a particular objective function is maximized. These categories can include methods pertaining to spectral analysis [2], random walks [3], and min-cut problems [4].