Statistical parametric network analysis of functional connectivity dynamics during a working memory task

Abstract

Network analysis has become a tool of choice for the study of functional and structural Magnetic Resonance Imaging (MRI) data. Little research, however, has investigated connectivity dynamics in relation to varying cognitive load. In fMRI, correlations among slow (< 0.1 Hz) fluctuations of blood oxygen level dependent (BOLD) signal can be used to construct functional connectivity networks. Using an anatomical parcellation scheme, we produced undirected weighted graphs linking 90 regions of the brain representing major cortical gyri and subcortical nuclei, in a population of healthy adults (n = 43). Topological changes in these networks were investigated under different conditions of a classical working memory task — the N-back paradigm. A mass-univariate approach was adopted to construct statistical parametric networks (SPNs) that reflect significant modifications in functional connectivity between N-back conditions. Our proposed method allowed the extraction of ‘lost’ and ‘gained’ functional networks, providing concise graphical summaries of whole-brain network topological changes. Robust estimates of functional networks are obtained by pooling information about edges and vertices over subjects. Graph thresholding is therefore here supplanted by inference. The analysis proceeds by firstly considering changes in weighted cost (i.e. mean between-region correlation) over the different N-back conditions and secondly comparing small-world topological measures integrated over network cost, thereby controlling for differences in mean correlation between conditions. The results are threefold: (i) functional networks in the four conditions were all found to satisfy the small-world property and cost-integrated global and local efficiency levels were approximately preserved across the different experimental conditions; (ii) weighted cost considerably decreased as working memory load increased; and (iii) subject-specific weighted costs significantly predicted behavioral performances on the N-back task (Wald F = 13.39, df₁ = 1, df₂ = 83, p < 0.001), and therefore conferred predictive validity to functional connectivity strength, as measured by weighted cost. The results were found to be highly sensitive to the frequency band used for the computation of the between-region correlations, with the relationship between weighted cost and behavioral performance being most salient at very low frequencies (0.01–0.03 Hz). These findings are discussed in relation to the integration/specialization functional dichotomy. The pruning of functional networks under increasing cognitive load may permit greater modular specialization, thereby enhancing performance.

Introduction

Network analysis has become a promising framework within which systems neuroscience can address a range of challenging questions (Bassett and Bullmore, 2006, Bullmore and Sporns, 2009). Some of this interest has been motivated by the fact that several topological metrics capturing core aspects of cortical networks have been found to be heritable (see Bassett and Bullmore (2009), for a review). Brain functional and structural networks rely on the assumption that correlation in BOLD signal or correlation in gray matter (GM) density are indicative of neuronal connectivity. These networks have been found to have predictive validity as network architecture appears to be related to medication dosage and to psychometric scores such as the Positive and Negative Syndrome Scale (PANSS) (Rubinov et al., 2009). In addition, whole-brain network topological metrics have also been used as biomarkers for specific neurodegenerative diseases such as Alzheimer's disease (AD) with high levels of specificity and sensitivity (Supekar et al., 2008). Research on the human connectome seems especially important from a clinical perspective, as several psychopathological disorders can be specifically described as ‘disconnection syndromes’ (Bassett and Bullmore, 2009, Catani and Mesulam, 2008).

One particular network architecture, which is commonly emphasized in this emerging field, is small-worldness (Milgram, 1967, Watts and Strogatz, 1998). The small-world property is a qualitative description of a network characterized by high levels of local clustering and short path lengths linking all nodes of the network. This constitutes a particularly attractive model of brain network organization, since it can account for the combination of both specialized and distributed information processing, as well as minimizing wiring cost in brain circuitry (Achard et al., 2006). Small-worldness of functional networks in humans has been validated by several teams of researchers working with different modalities, such as functional MRI (Achard et al., 2006), magnetoencephalography (Bassett et al., 2006) and electroencephalography (Micheloyannis et al., 2009). Although networks can be directed or undirected, weighted or unweighted, most of the research in systems neuroscience thus far has concentrated on the simplest case: undirected and unweighted graphs using a thresholding function. Following this trend, we will similarly focus on unweighted, undirected networks based on functional MRI data.

Only a small subset of human behavior has been investigated from a network analytic perspective. Resting-state functional networks have received most research attention (Achard et al., 2006, Salvador et al., 2005). These investigations have shed light on the intrinsic architecture of spontaneous fluctuations in BOLD signal when subjects are at rest. In addition, the functional connectivity networks underlying finger tapping and listening to music have been summarized by Eguiluz et al. (2005). For these two activities, the brain also appears to display a small-world architecture. Intriguingly, Eguiluz et al. (2005) reported that listening to music was associated with functional networks characterized by higher mean degrees than those subtending finger tapping. That is, listening to music seems to induce, on average, a greater amount of functional connectivity than finger tapping. Other authors have investigated functional connectivity patterns linked with movement execution (De Vico Fallani et al., 2008, Cecchi et al., 2007, Astolfi et al., 2009). By contrast, there have been very few studies exploring the relationship between human cognition and functional networks' topological metrics (Sporns et al., 2004, Bullmore and Sporns, 2009). The exploration of such changes in topology will aid systems neuroscience in probing the dynamics of functional networks.

Several studies have investigated changes in functional network topologies across different cognitive tasks. Some research in this area has been based on the N-back task and has used either EEG (Pachou et al., 2008), MEG (Bassett et al., 2009) or fMRI data (Salvador et al., 2008). The N-back paradigm is a popular cognitive task, which allows evaluation of changes in working memory load under several conditions of increasing difficulty (Gevins and Cutillo, 1993). The task requires on-line monitoring of a series of stimuli, as well as updating and manipulating remembered information. It is generally thought to place great demands on several key processes involved in memory (Honey et al., 2002, Owen et al., 2005). Some authors, however, have recently argued that N-back performance is not a pure measure of working memory, but may be better described as a measure of processing speed (Miller et al., 2009). Research in this area generally seeks to unveil the neural basis of the central executive system of working memory (Baddeley, 1998, D'Esposito et al., 1995, Kane et al., 2007). The N-back task has extensively been used in fMRI to identify pathological responses in clinical populations (Harvey et al., 2005, Caseras et al., 2006, Wishart et al., 2006, Van Snellenberg et al., 2006). A recent meta-analysis of 24 primary studies using the N-back task demonstrated that the cortical regions that tend to be robustly activated under this paradigm are the lateral premotor cortex, the dorsal cingulate and medial premotor cortices, the dorsolateral and centrolateral prefrontal cortices, the frontal poles, and the medial and lateral posterior parietal cortices (Owen et al., 2005). The pattern of interactions between these different regions, however, is still one of the key questions in the study of working memory. Network analysis may be particularly useful in this respect, by enabling researchers to map the functional connectome subtending different working memory loads. Our first objective in this article is therefore to explore the dynamics of functional connectivity patterns under different levels of cognitive load.

The second contribution of this paper is methodological. We introduce a concise way of summarizing inference on networks, by using a mass-univariate approach. This strategy is similar in spirit to the one adopted in the statistical parametric mapping (SPM) framework (Friston, 1994), and will therefore be referred to as a statistical parametric network (SPN). SPM revolutionized the analysis of PET and later MRI data, by providing summary maps, which facilitate the identification of a set of regions of interest, through the application of specific levels of significance (Friston, 1994). We argue that network analysis urgently needs to adopt a similar approach when combining networks over a population of subject-specific correlation matrices. Currently, networks constructed on the basis of structural and functional data are treated in a different manner. Whereas in fMRI studies, researchers have generally computed inter-regional correlations over time (Achard et al., 2006, Achard and Bullmore, 2007), studies based on structural MRI data have utilized inter-regional correlations over subjects (e.g. Bassett et al., 2008). In this paper, we will focus on the specific problem posed by the study of functional MRI cortical networks, where correlation matrices are subject-specific with regional covariances computed with respect to time. It will therefore be assumed throughout this paper that a population of subject-specific networks is available for the network analysis.

Our proposed SPN framework pursues some of the ideas put forward by Achard et al., 2006, He et al., 2007, He et al., 2009b and generalizes them in order to produce summary networks constructed over the different conditions of an experiment, or similarly for distinct populations of interest. Achard et al., 2006, He et al., 2009b have presented summary networks, where an edge is only included if a test statistic for that edge computed over all the subjects in the population of interest is significant. We will term these networks mean or summary SPNs. In addition, we will also construct what we term differential or difference SPNs, which provide information about which edges have been significantly ‘lost’ and which edges have been significantly ‘gained’ from one experimental condition to another, or from one subject population to another. Similar approaches to functional network comparison have been adopted by Zalesky et al., 2010, Richiardi et al., in press using network-based statistics and machine learning methods, respectively, for the comparison of two populations of subjects and controls. Our approach is slightly more general since it accommodates complex experimental designs where information is pooled over several (more than two) experimental conditions. The SPN framework shares with SPM a concise visualization of the data, and is therefore particularly useful for the presentation and reporting of results. Specific to network analysis, differential SPNs have the additional advantage of replacing some of the concerns with the choice of a threshold value to compute the adjacency matrix with the selection of a specific p-value. In this context, thresholding is supplanted by inference (Achard et al., 2006, He et al., 2009b).

A recent review of the literature on complex brain networks has emphasized the paucity of research papers linking network topology with cognitive and behavioral functions (Bullmore and Sporns, 2009). The present investigation should therefore be regarded as exploratory in nature, as only few previous results can be used for comparison. As little research has been done on the dynamics of functional networks under different levels of cognitive demand, no prior hypothesis was formulated on the nature of the potential relationship between the global and local efficiencies of these functional networks under different levels of working memory load.