Since the discovery of resting state connectivity there has been an explosion of interest in relating differences in functional networks measured at rest with behavioral or cognitive measures in both healthy and patient populations. However, to date, the majority of studies doing so have employed univariate techniques. In univariate analyses, the behavioral measure of interest is tested against (e.g., correlated with) a resting state functional connectivity measure separately at the voxel or region of interest level (ROI), with the major assumption being that each voxel, or ROI, varies independently. However, this is certainly not the case, and more powerful multivariate methods need to be established in order to find relationships between behavior and resting state functional connectivity measures.
Here, we use para-ICA to identify relationships between fluctuations in RSNs and behavior. This method has several advantages over traditional univariate methods. First and foremost, para-ICA is a multivariate method, which considers all data point and voxels within each modality simultaneously, not independently as in univariate analyses. This allows for identification of complex relationships that exist between modalities that would be missed using univariate analyses. Second, this method is completely data-driven, and no hypothesis regarding the relationship between modalities is required.
Using simulation data we first demonstrate that para-ICA results in realistic correlations between data sets with large differences in dimensionality. The validity of para-ICA on datasets with more balanced dimensions has been previously established (Liu et al., 2008
). Based on our simulated data, we conclude that para-ICA performs well in the null scenario, where there is no relationship between modalities, as well as in cases with varying levels of correlations between modalities even when applied to datasets with unbalanced dimensions. This is because, under the Infomax-ICA framework, the performance of para-ICA depends on the dimensionality of the samples (e.g., subjects), and not on the dimensionality of the data modalities (Liu et al., 2008
). This finding is reassuring given the data-driven nature of the method, and demonstrates that para-ICA will not induce a relationship between modalities if no real relationship exists.
In line with our primary goal, we demonstrate that common RSNs can be spatially decomposed into sub-components that covary with specific behavioral profiles. Of the nine networks we performed para-ICA on, four did not result in significant correlations between the spatial components and behavior components. As hypothesized, there were no sub-components of the auditory network that significantly covaried with any of the behavioral indices collected. This was expected, as the behavioral indices were mostly cognitive in nature, and there were no tasks with auditory information. In addition, no significantly correlated pairs of components were observed for the two lateralized attention networks and the fronto-polar network. It is interesting that the para-ICA did not identify any components between lateralized attention networks and some of the lateralized behavioral tasks. This could be related to various statistical thresholds applied in the current research. Another possible explanation is that the ICA on the behavioral measures was driven by factors other than task lateralization. However, as hypothesized, other RSNS including the default mode networks, the bilateral frontal network, and an attention network did have sub-components that covaried with components consisting primarily of specific cognitive measures. As this method is exploratory in nature, we provide limited interpretation of the results when possible.
Temporo-parietal attention network
The para-ICA on the temporo-parietal attention network identified a spatial sub-component located in the precuneus that differentially covaried with response time in congruent and incongruent Stroop task trials. Less connectivity of this area in the precuneus to temporo-parietal attention RSN negatively covaried with better performance (faster response time) in incongruent trials, but worse performance (slower response time) in congruent trials.
The precuneus has previously been implicated in visuo-spatial processes, including attention (Beauchamp et al., 2001
; Simon et al., 2002
; Cavanna and Trimble, 2006
). In addition, the precuneus has been shown to be preferentially activated in a color-word Stroop task with a hypothesized role of selecting task relevant information together with the dorsolateral PFC (Banich et al., 2000
). Here, we show that the level of involvement of a cluster of the precuneus to the temporo-parietal attention networks has different effects on incongruent and congruent trials in the Stroop task, a task often used to measure selective attention.
Posterior–superior default network
For the posterior–superior default network, greater connectivity from a sub-component in the cuneus was found to negatively covary with two separate working memory tasks that require the retention of spatial information. Subjects with more cuneus involvement in this RSN had slower response times in a spatial working memory task, and a combined spatial and verbal working memory task. This area of the primary visual cortex is essential for visuo-spatial processing, including working memory (for review, see Kravitz et al., 2011
). In addition, previous studies have shown that the ability to suppress areas of the default mode network, including the cuneus, during working memory tasks is correlated with performance on those tasks (Hampson et al., 2006
; Anticevic et al., 2010
). Here we found a sub-component of the default mode network at rest that negatively covaries specifically with two tasks involving spatial working memory.
Posterior–inferior default network
There were two sub-components of the posterior–inferior default network that covaried with specific behavioral components. The first sub-component consisted of a small cluster in the right middle occipital gyrus (BA 19) that differentially covaried with congruent Stroop task trials and a combined verbal and spatial working memory task. Greater connectivity of this area to the posterior–inferior default network positively covaried with faster response time in the working memory task and slower response time in the Stroop congruent task trials. Therefore, the connectivity of this region to this RSN has different effects on selective attention and working memory tasks.
The second sub-component of the posterior–inferior default network included a region in the vermis of the cerebellum that negatively covaried with response time to all Stroop trial types, as well as response time to incongruent Stroop trial types. Less connectivity to this RSN by the cerebellum, which has previous been implicated in Stoop task performance (Egner and Hirsch, 2005
), was inversely related to slower response times in these Stroop trial types.
The identification of a sub-component of the default mode network that covaried differently to an attention task and a working memory task, and the fact that a second sub-component specifically covaried with just a selective attention task, is not without precedent. Although regions that deactivate during cognitive tasks are generally similar across task type, subtle task-dependent deactivations have been reported (Tomasi et al., 2006
). Interestingly, a recent study by Mayer and colleagues found that task-induced deactivations of the default mode network differed between an attention task and a working memory task (2010). Our results are consistent with these, and suggest that the roles of specific sub-regions of the posterior–inferior default network are dependent on task-type.
Medial frontal default network
The para-ICA on the medial frontal default network identified a spatial component consisting of the brainstem and mid-cingulate cortex that differentially covaried with two working memory tasks. The greater connectivity of the mid-cingulate and less connectivity of the brainstem to this default network was inversely related to faster response time in the verbal working memory tasks, but slower response time in a combined verbal and spatial working memory task. For the medial frontal default network, we again see regions of the default mode network that differentially respond to different tasks. However, in contrast to our findings for the posterior–inferior default network, here we identify regions of the medial frontal default network that responds differently to two working memory tasks, one involving just spatial information and one involving both verbal and spatial information. Therefore, it appears that the relationship between variations of the default mode network and cognition are not only task-dependent, but also depend on the specific division of the default network being investigated.
Bilateral frontal network
The para-ICA performed on the bilateral frontal network resulted in a spatial sub-component that split the left inferior frontal gyrus into three separate clusters that were negatively covaried with response time in a combined spatial and verbal working memory task (bound). These three clusters included the pars opercularis (BA 44), pars triangularis (BA 45), and pars orbitalis (BA 47). Greater connectivity of pars opercularis and pars triangularis, which together are generally thought to make up Broca's Area (Keller et al., 2009
), and less connectivity of pars orbitalis, was inversely related to faster response time in the combined spatial and verbal working memory task. This is especially interesting in light of the hypothesis put forward by Badre and colleagues that a dual system of cognitive control exists in the left inferior frontal gyrus (2005). According to this model, for any cognitive task requiring retrieval and selection of mnemonic information, such as a working memory task, the anterior portion of the left ventrolateral prefrontal cortex (pars orbitalis; BA 47) is responsible for the controlled retrieval of information while the mid-ventrolateral prefrontal cortex is responsible for post-retrieval selection of task-relevant representations (Badre and Wagner, 2007
). Here, we show a similar dissociation between the left anterior ventrolateral prefrontal cortex (pars orbitalis) and the left mid-ventrolateral prefrontal cortex (pars opercularis and pars triangularis), as the involvement of these areas to the bilateral frontal RSN covaried differently with performance in a dual spatial and verbal working memory task.
This study has several limitations. One common issue with ICA analyses, including para-ICA analyses, is the number of components that are estimated. However, we believe that the 40 components estimated in the group ICA is appropriate due to the high degree of spatial correlation between our components and the common RSN templates. For the para-ICA, we based our decision by identifying the number of components that explained approximately 90% of the variance for both the behavioral data and each resting state component. Finally, this study is limited by a relatively small sample size, a larger sample size would provide more detection power, and by the fact that nine separate para-ICAs were performed. Several steps were taken in attempts to control for this issue as several different levels of the analyses (Figure ). In addition, this study is intended as an exploratory study to show the feasibility of indentifying sub-components of RSNs that covary with specific behaviors. Accordingly, we only provide cautious interpretation of results when possible, as these findings need to be verified with other means, such as more hypothesis driven approaches, before strong interpretations can be made.