Age-dependent changes in task-based modular organization of the human brain

Highlights

• Dynamics of task-active brain communities are quantified in adults of varying ages.

• Community number and node flexibility are positively associated with subject age.

• Dynamics of intrinsic functional networks are differentially modulated by age.

• No correspondence is found between age or brain dynamics and task performance.

Abstract

As humans age, cognition and behavior change significantly, along with associated brain function and organization. Aging has been shown to decrease variability in functional magnetic resonance imaging (fMRI) signals, and to affect the modular organization of human brain function. In this work, we use complex network analysis to investigate the dynamic community structure of large-scale brain function, asking how evolving communities interact with known brain systems, and how the dynamics of communities and brain systems are affected by age. We analyze dynamic networks derived from fMRI scans of 104 human subjects performing a word memory task, and determine the time-evolving modular structure of these networks by maximizing the multislice modularity, thereby identifying distinct communities, or sets of brain regions with strong intra-set functional coherence. To understand how community structure changes over time, we examine the number of communities as well as the flexibility, or the likelihood that brain regions will switch between communities. We find a significant positive correlation between age and both these measures: younger subjects tend to have less fragmented and more coherent communities, and their brain regions tend to change communities less often during the memory task. We characterize the relationship of community structure to known brain systems by the recruitment coefficient, or the probability of a brain region being grouped in the same community as other regions in the same system. We find that regions associated with cingulo-opercular, somatosensory, ventral attention, and subcortical circuits have a significantly higher recruitment coefficient in younger subjects. This indicates that the within-system functional coherence of these specific systems during the memory task declines with age. Such a correspondence does not exist for other systems (e.g. visual and default mode), whose recruitment coefficients remain relatively uniform across ages. These results confirm that the dynamics of functional community structure vary with age, and demonstrate methods for investigating how aging differentially impacts the functional organization of different brain systems.

Introduction

Humans experience notable changes in cognitive ability and behavior as they age, often in situations involving memory encoding, memory retrieval, and executive control functions (Balota et al., 2000, Grady and Craik, 2000, Cepeda et al., 2001, West et al., 2002, Treitz et al., 2007). Over the past few decades, advances in brain imaging have made it possible to observe and quantify neural changes associated with advanced age. One of the most widely-reported phenomena associated with aging is the loss of segregation between neural systems: many networks become less internally coherent, while at the same time they become more similar to other networks. This result has been reported using a number of methodological approaches, including whole-brain ICA (Onoda et al., 2012), whole-brain parcel-based functional connectivity methods (Betzel et al., 2014, Chan et al., 2014, Song et al., 2014, Ferreira et al., 2015, Geerligs et al., 2015) as well as similar analyses confined to a subset of systems (Grady et al., 2016, Ng et al., 2016), whole brain voxel-wise analyses (Tomasi and Volkow, 2012), and seed-based methods (Zhang et al., 2014) (for reviews, see Dennis and Thompson, 2014, Contreras et al., 2015, Sala-Llonch et al., 2015). Moreover, these changes have been tracked longitudinally within participants (Ng et al., 2016), have been shown to affect various properties theoretically associated with the efficiency and efficacy of information processing in the brain (Sala-Llonch et al., 2014, Gomez-Ramirez et al., 2015), and have been associated with behavioral effects (Ng et al., 2016, Sala-Llonch et al., 2014).

Although the dominant change associated with aging is one of decreased intra-network connectivity and increased inter-network connectivity, this pattern varies across networks. The loss of intra-network connectivity is found most consistently in the default mode network (DMN), even among those studies that consider brainwide connectivity (Onoda et al., 2012, Tomasi and Volkow, 2012, Wang et al., 2012, Song et al., 2014, Ferreira et al., 2015, Geerligs et al., 2015, Ng et al., 2016). Some studies also report similar decreases in networks associated with higher cognitive functions (Onoda et al., 2012, Wang et al., 2012, Chan et al., 2014, Geerligs et al., 2015, Ng et al., 2016). However, other networks consistently show no change, or even an increase in intra-network connectivity, especially those associated with sensory functions (Tomasi and Volkow, 2012, Song et al., 2014, Geerligs et al., 2015). Similarly, connectivity between the DMN and other networks tends to increase (or, equivalently, the uniqueness of the networks decreases) (Ferreira et al., 2015, Ng et al., 2016).

In parallel with this line of research on how the brain's functional architecture changes with age, a largely separate effort has sought to extend connectivity methods by accounting for the fact that the brain is not static (for a review, see Calhoun et al. (2014)). To the contrary, this work has demonstrated that patterns of connectivity are quite variable (Gonzalez-Castillo et al., 2014), which can be characterized as constituting a series of transitions between fairly well-defined brain states (Hansen et al., 2015). It has been proposed that the greatest variability occurs in regions that serve to connect fairly well-segregated systems (Zalesky et al., 2014), and that a small set of networks may modulate the organization across a large number of others (Di and Biswal, 2015). The time-resolved approach adds yet another dimension for investigating age-related effects; for example, Qin et al. (2015) report increased variability in connectivity across networks including DMN and cerebellum, and decreased variability between those two and within the cingulo-opercular network, as a function of age.

Having established these aging-related changes in functional connectivity—along with some general principles of dynamic connectivity—in the resting state, an obvious next question is how the results differ during task performance. “Task-free” paradigms dominate studies of functional connectivity. Incorporating a task could affect connectivity, including its relationship with age and its dynamics, in a number of ways. For instance, compensatory strategies employed by older—but not younger—adults could drive the connectivity profiles of the two groups even further apart; alternatively, the presence of an extrinsic input could impose structure on the systems that have become homogenized in older adults. Indeed, Dubois (2016) demonstrated widespread changes in the relationship between age and connectivity across resting and task scans, with the largest effects being a weakening in the age–connectivity relationship during tasks compared with rest. Likewise, connectivity between and within networks could change as participants learn, change strategies, or even simply become fatigued.

For the present study, we used a memory task that incorporated a strong element of cognitive control. In particular, after studying a list of items, participants were presented with the studied items, along with novel (unstudied) items, and instructed to indicate whether each item was studied or not. Items occurred in one of two contexts: a “liberal” context indicating that each item in that context was likely to have been studied (70% of items were studied items) or a “conservative” context indicating that each item was unlikely to have been studied (30% of items studied). In the face of imperfect memory evidence, participants must exert cognitive control—adjusting the criterion they use to endorse an item as studied—in order to perform well on this task. Given that the domains of memory and cognitive control are fundamental in human cognition, and are associated with changes over the lifespan (Jacoby et al., 2005), this task is an appealing choice for studying how the brain's architecture changes with age when not at rest. Previous results with this task revealed wide individual differences in adaptability (Aminoff et al., 2012), and implicated a network of regions including lateral prefrontal and lateral posterior parietal cortex in performing this task (Aminoff et al., 2015).

Although the brain regions associated with the performance of this task are well documented, these results are derived from the standard mass-univariate GLM analysis of BOLD data, and therefore give little basis for predictions in terms of network-level dynamics. In fact, by definition, these existing results assume stationarity and consider each voxel as independent. Even results derived from methods that explicitly model the spatiotemporal nature of brain activity (e.g., ICA) would require a theoretical framework in order to define regions of interest in the context of how network dynamics relate to other factors, such as age. Thus, there remains a gap in understanding of the neural processes related to performance of this task on the level of dynamic interactions between large-scale brain regions and networks. Our current understanding of these processes, based on existing theories and results, is specified on a very different level from the target of our current investigation. Our goal in this work is to apply a data-driven analysis method to investigate the dynamics of these regions and networks, which allows us to uncover age-related changes at scales at which it is difficult to make specific hypotheses based upon existing literature.

We apply a dynamic community detection method to quantify several higher-order aspects of task-based functional connectivity and their dependence on age. This method and other network science approaches have proved successful in distilling the information in fMRI data into intuitive, descriptive, and predictive network characteristics (Bassett et al., 2012, Davison et al., 2015, Siebenhühner et al., 2013, Bullmore and Sporns, 2012, Bassett et al., 2009, Bassett et al., 2011, Bassett et al., 2015). While previous results suggest that static community structure will meaningfully differ on a group level between older and younger adults at rest (Meunier et al., 2008), we ask whether the dynamic changes in these communities are affected by age during task-based cognition, and how such effects vary across individual participants. We quantify the size and number of functional brain communities, the degree to which brain regions flexibly switch between communities, and the association of the community structure with known intrinsic functional connectivity networks or systems, in order to determine whether these systems are differentially involved in age-related changes.