Age-related Change and the Predictive Value of the ‘Resting State’: A Commentary on Campbell and Schacter (2016)

Abstract

This is a commentary on Campbell and Schacter (2016), ‘Aging and the Resting State: Is Cognition Obsolete?’

‘Resting state functional connectivity’ (RSFC) is gaining ascendency as a primary tool for investigating the functional organization of the human brain. The overall topography of the brain’s large-scale functional networks has been shown to be highly reproducible, purportedly transcending a multiplicity of task contexts and changing in systematic ways with development, aging, and disease [see Stevens & Spreng (2014) and Matthews & Hampshire (2016) for recent reviews]. The advent of RSFC research has coincided with methodological innovations, such as the integration into brain research of graph theoretical network models and big data analytics. In their current article, Campbell & Schacter (C&S) acknowledge the proliferation of studies using RSFC measures of cognition and challenge its status as the preeminent predictor of cognitive function and dynamics. They advocate for the essential role of task-based designs in testing specific hypotheses about cognition and aging. In brief, C&S argue that RSFC does not provide a privileged window into the functional architecture of the human brain, and that it is prone to variability and influences by age differences in non-cognitive factors, such as motion and vascular health. Critically, RSFC is of limited use for testing hypotheses about cognition because its measures are always remote from the cognitive phenomena that the researchers are trying to relate them to.

While we agree with C&S’s position that carefully designed experimental manipulations are essential for testing specific mechanistic hypotheses about cognition and its neural correlates, we will argue that, under specific circumstances, RSFC measures may provide certain advantages over task-evoked activity, and that a genuine and increasingly thorough integration of RSFC and task-related approaches is essential for advancing our understanding of brain function and aging. In our commentary to the C&S article, we first highlight some specific advantages of the RSFC approach, especially to study neurocognitive aging, then we review key evidence supporting the predictive value of the brain’s ‘resting state’, and finally we discuss the implications of a perspective emphasizing the brain’s ‘intrinsic’ activity.

As C&S state, the major goal of neuroscientific research on aging is to understand how age-related brain changes impact cognitive function. Still, the majority of studies to date have employed cross-sectional designs and thus have informed only about age differences. While this work has stimulated important new hypotheses, reliance on cross-sectional studies with relatively small, elite and non-representative samples has limited the conclusions we can draw, and our understanding of neurocognitive aging (e.g., Falk et al., 2013). By contrast, longitudinal designs are necessary for evaluating change over time and for isolating contributions of brain maintenance and compensatory processes from pre-existing individual differences in neurocognitive characteristics (Nyberg et al., 2010; Raz & Lindenberger, 2013; Reuter-Lorenz & Park, 2014). Assessing performance and neural activity changes on the same set of measures over time is the defining feature of longitudinal studies of neurocognitive aging, but also one that introduces confounding effects of practice even when new task versions are used for each measurement point [see also Lövdén, Bäckman, Lindenberger, Schaefer, & Schmiedek (2010) and Morcom & Johnson (2015)]. Because the RSFC approach is not directly influenced by practice effects, it may provide a valuable tool for assessing decline, along with mitigating processes such as functional compensation and/or plasticity over time. Furthermore, by linking RSFC measures with behavioral performance in relevant tasks, it may also help determine whether this neural reorganization benefits cognitive performance.

Additional advantages of the RSFC approach stem from the simplicity of its protocol. First, because it has minimal cognitive demands, RSFC is not impacted by degraded cognitive task performance, and offers a practical tool for functional assessment of participants with cognitive impairment. Thus, RSFC can boost the power of neurocognitive studies by allowing the investigation of participants spanning a larger range of abilities. Furthermore, its simpler experimental protocol compared to task-based fMRI, permits the aggregation of unprecedented amounts of RSFC data for data mining and hypothesis testing [e.g., Power et al., 2011; Yeo et al., 2011; see also Laumann et al. (2015) for an example of high sampling in an individual human]. Thus, RSFC may provide an advantage in studies of functional organization over task-based approaches which are often underpowered and have discrepant designs that preclude easy aggregations of the data (Falk et al., 2013). On the other hand, the trend towards acquiring large datasets featuring both RSFC and task-related data using consistent protocols, such as in the Human Connectome Project (HCP; Essen et al., 2013), may soon compensate for this advantage.

Importantly, recent studies suggest that RSFC is sufficiently consistent and shares enough variance with task-evoked activity that it can be used to identify trait-like features of brain function and predict a broad spectrum of task-related activations (Finn et al., 2015; Laumann et al., 2015; Tavor et al., 2016). RSFC demonstrates high within-subject stability across multiple scanning sessions in younger (Shehzad et al., 2009) and older adults (Song et al., 2012). Furthermore, RSFC patterns are sufficiently consistent to allow successful identification of individuals across sessions (Finn et al., 2015). An individual’s RSFC profile may thus serve as an intrinsic functional ‘fingerprint’ that captures trait-like features of brain organization that are relevant for neurocognitive function and may be sensitive to trajectories of change over time. Although C&S question the value of RSFC for testing mechanistic hypotheses about cognitive function, it is worth noting that task-evoked brain activations have been effectively linked to performance on different cognitive tasks outside of the scanner to address mechanisms underlying individual differences in successful aging (Cabeza, Anderson, Locantore, & McIntosh, 2002; Pudas et al., 2013). Likewise, by relating RSFC to cognitive measures obtained outside the scanner environment, cognitive mechanisms can also have a place in RSFC research.

The potential value of using RSFC to complement task-based measures in predicting age-related changes in neurocognitive performance and function is further supported by recent evidence showing high consistency of functional networks topography across rest and a multitude of tasks (Cole, Bassett, Power, Braver, & Petersen, 2014; Laumann et al., 2015; Smith et al., 2009), as well as reciprocal influences between performance, task-evoked activity, and RSFC (Stevens & Spreng, 2014). Recently, Tavor et al. (2016) showed that RSFC measures can predict individual differences in task-evoked activity over several task domains. Using HCP data, Tavor and colleagues trained a model to predict qualitative differences (shape, position, size, and topography) in task-evoked activity over multiple behavioral domains, based on RSFC and structural MRI input. They avoided circularity by using a leave-one-out approach such that the predictive model for a particular subject did not see the task-evoked activity map for the same subject during training, but it did for all the other subjects. The model’s predictions based on the subjects’ RSFC and morphology were found to be more similar to the task-evoked activity maps it did not see than to the rest of the subjects’ maps. Furthermore, removing structural features from the model did not affect its predictive ability, which suggests that RSFC may be sufficient to predict activations in subjects in the absence of task-related data.

What information in the RSFC signal drives the predictions reported in the Tavor et al. study? The features selected by the predictive models can be considered intrinsic because the models are blind to the subjects’ task data (Tavor et al., 2016). Less clear is whether these features relate more to ‘architectural’ aspects of the brain’s functional organization or to ‘functional’ aspects related to the subject’s cognitive state during RSFC acquisition. Consistent with the ‘architectural’ perspective is the view that RSFC reflects the history of co-activation among distributed brain regions, governed by Hebbian learning. More specifically, repeated synchronized activations drive selective increases and decreases in functional connectivity, which in turn bias future couplings between regions, and facilitate cognitive performance (Guerra-Carrillo, Mackey, & Bunge, 2014; Stevens & Spreng, 2014). An emerging stronger position suggests that the functional specialization of particular brain regions is largely determined by their patterns of connectivity. In other words, what a region ‘does’ depends more on ‘with whom’ it is connected rather than ‘where’ it is located [see Saygin et al. (2016) and Baldassarre et al. (2012) for recent developmental and training evidence, respectively].

On the other hand, consistent with the ‘functional’ perspective is the idea that the variance intrinsic in the RSFC signal, driven by the absence of a particular ‘task set’, might render it better suited for a multiple-states-to-multiple-contexts mapping, and thus a better predictor for a larger variety of behavioral outcomes, compared to specific tasks. The opposing view, also expressed by C&S, states that RSFC is just a task state, and thus it does not provide better predictions than any other task-evoked activity, no matter the domain [see Buckner, Krienen, & Yeo (2013) for a related discussion].

The advent of RSFC research has focused attention on a more ‘intrinsic’ perspective of brain operation, which emphasizes the brain’s ‘background’ activity of ongoing accumulation and integration of information, rather than its momentary exchanges with the environment. According to this ‘intrinsic’ view, experimental task manipulations reveal only a small fraction of the brain’s true activity (Raichle, 2010, 2015). In a strong sense, the ‘intrinsic’ view raises the intriguing possibility that variations in the blood-oxygen-level-dependent (BOLD) signal recorded during fMRI task performance may reflect changes in the slow components of the brain’s intrinsic activity rather than delayed responses to rapid neural events (Raichle, 2010, 2015). Given the current limited understanding of the coupling between neural and BOLD signals and the way this relation is influenced by age-related physiological changes (e.g., Fabiani et al., 2014), the value of this claim remains to be determined. In a weaker sense, however, this view is compatible with the idea that the brain’s intrinsic activity may be predictive of a wide range of behavioral and cognitive outcomes, including the brain’s task-evoked activity.

Nevertheless, even if RSFC can reveal certain intrinsic features of the brain’s overall functional organization and activity, we recognize that much caution is needed when attempting to link these with individual differences in cognition and behavior. As C&S argue, a correlation clearly does not provide, by itself, proof of a mechanism. However, a model that can successfully predict specific cognitive outcomes based on RSFC measures alone may ultimately have bearing for a mechanistic account. One ongoing challenge, though, is the proper translation of the abstract features captured by multivariate analyses of functional connectivity into concepts with psychological relevance. (Tavor et al., 2016). On the other hand, task-based approaches are also prone to over-interpretation (i.e., ‘reverse-inference’) (Poldrack, 2006, 2011; cf. Hutzler, 2014), and we should bear in mind that experimental cognitive tasks are simply analytic models for testing putative brain functions and hypotheses about the neurocognitive architecture. The ultimate goal is to understand human behavior in vivo, not merely inside the scanner or any other test environment (Falk et al., 2013).

In summary, we agree with C&S’s argument for the preeminence of task-based manipulations in testing specific mechanistic hypotheses about cognition, but we also value the advantages of RSFC over task-based approaches under particular circumstances, especially for investigating neurocognitive aging. These advantages pertain to the assessment of change and functional plasticity in longitudinal designs while minimizing practice effects, enabling the investigation of individuals spanning a larger range of abilities, unconfounded by variations in performance, and facilitating the aggregation of large datasets that permit more robust conclusions. A fundamental relationship between RSFC and task-evoked activity is supported by recent evidence that the former can reliably predict the latter and that an individual’s functional ‘fingerprint’ robustly manifests across multiple functional states. Over recent years, it has become increasingly clear that evoked neural responses cannot be understood in isolation from the larger context of the brain’s ongoing activity, chiefly because the brain’s functional architecture is organized at multiple temporal and spatial scales (Sadaghiani, Hesselmann, & Friston, 2010). Thus, a practical understanding of neurocognitive function and age-related change, reflective of the brain’s true complexity, requires a better integration between RSFC and task-related approaches. This means not only using task-evoked activity to test hypotheses generated by RSFC, as C&S rightfully argue, but also adapting cognitive task design, analysis, and interpretation to truly capitalize on results from RSFC studies [see also Buckner et al. (2013) and Cole, Smith, & Beckmann (2010)].