Working memory‐related changes in functional connectivity persist beyond task disengagement

Abstract

We examined whether altered connectivity in functional networks during working memory performance persists following conclusion of that performance, into a subsequent resting state. We conducted functional magnetic resonance imaging (fMRI) in 50 young adults during an initial resting state, followed by an N‐back working memory task and a subsequent resting state, in order to examine changes in functional connectivity within and between the default‐mode network (DMN) and the task‐positive network (TPN) across the three states. We found that alterations in connectivity observed during the N‐back task persisted into the subsequent resting state within the TPN and between the DMN and TPN, but not within the DMN. Further, both speed of working memory performance and TPN connectivity strength during the N‐back task predicted connectivity strength in the subsequent resting state. Finally, DMN connectivity measured before and during the N‐back task predicted individual differences in self‐reported inattentiveness, but this association was not found during the post‐task resting state. Together, these findings have important implications for models of how the brain recovers following effortful cognition, as well as for experimental designs using resting and task scans. Hum Brain Mapp 35:1004–1017, 2014. © 2012 Wiley Periodicals, Inc.

INTRODUCTION

The functional architecture of the human brain is composed of discrete, large‐scale networks with widely spaced nodal regions, which demonstrate highly correlated activity across time. Such correlations, termed “functional connectivity,” are evident in low frequency spontaneous BOLD signal fluctuations observed while subjects are resting, a state defined by the absence of directed cognitive engagement. Connectivity measured during rest is said to be “intrinsic” connectivity. However, functional connectivity networks are known to be sensitive to cognitive state, as the strength of connectivity is altered during task engagement relative to the resting state. Specifically, regions engaged by auditory [Arfanakis et al., 2000], visual [Arfanakis et al., 2000; Hampson et al., 2004; Nir et al., 2006], motor [Arfanakis et al., 2000; Jiang et al., 2004], and cognitive [Newton et al., 2011] tasks show altered connectivity during performance of those tasks. These findings indicate that task‐evoked cognitive states induce selective changes in the strength of brain networks relative to the resting state.

Two lines of research indicate that task‐driven changes in functional connectivity persist beyond the duration of the task and into a subsequent resting state. First, task content affects the strength of connectivity during a subsequent resting state. Resting state connectivity between visual and frontal regions immediately following performance of a task varied depending on whether the task involved classifying faces or scenes [Stevens et al., 2010]. Similarly, precuneus connectivity after a language comprehension task varied depending on the content of the language in that task [Hasson et al., 2009]. Second, direct comparisons of resting state functional connectivity before and after task performance reveal marked differences. Such task‐driven differences in resting connectivity have been observed after finger tapping [Duff et al., 2008], motor [Albert et al., 2009; Vahdat et al., 2011] and visual [Lewis et al., 2009] learning, language generation [Waites et al., 2005], memory encoding [Tambini et al., 2010], and semantic classification [Grigg and Grady, 2010]. These persistent alterations in functional connectivity were observed not only within task‐relevant networks [Albert et al., 2009; Tambini et al., 2010; Waites et al., 2005], but also between networks [Duff et al., 2008; Grigg and Grady, 2010; Lewis et al., 2009]. Notably, the direction of change in resting‐state connectivity was not consistent across these studies, with connectivity after task performance becoming more positive [Albert et al., 2009; Duff et al., 2008; Grigg and Grady, 2010; Tambini et al., 2010; Vahdat et al., 2011; Waites et al., 2005], more negative [Lewis et al., 2009; Vahdat et al., 2011], and less negative [Lewis et al., 2009]. Taken together, these results indicate that not only is the network structure of the brain altered during task performance, but also that such alterations persist after the task ends, likely in ways specific to the task and connections being tested.

Several characteristics of these persistent connectivity changes remain unknown. First, how effortful cognition affects subsequent resting‐state connectivity is not known. An investigation of the resting brain using non‐connectivity measures [Barnes et al., 2009] showed that the brain's oscillatory dynamics were perturbed after engagement of working memory (WM). WM is defined as the temporary maintenance and manipulation of task‐relevant information, and it contributes critically to higher order cognitive functioning [Engle et al., 1999]. Further, recovery from this perturbation took longer after greater WM loads, suggesting that effortful cognition requires a subsequent post‐task recovery period, which is affected by the degree of that effort. While it is known that engagement of WM alters connectivity within and across the default mode network (DMN) and the task‐positive network (TPN) [Bluhm et al., 2011; Fransson, 2006; Fransson and Marrelec, 2008; Gordon et al., 2012a], whether these changes persist after WM disengagement remains to be examined. Second, while connectivity after a task may be altered compared to connectivity before the task, it is not known whether these pre/post task differences are predicted by the strength of connectivity during the task. Third, while the strength of resting state connectivity predicts individual differences in cognitive performance [e.g., Hampson et al., 2010], the stability of those connectivity‐cognition relationships is not known. Specifically, it is unknown whether resting state connectivity still predicts individual differences in cognition if the relevant networks have been altered by a preceding task.

This study addressed these gaps by examining whether altered functional connectivity patterns during WM performance persist following conclusion of that performance. We examined functional connectivity within the DMN and TPN, as well as between the DMN and TPN, during three different scans: a baseline resting state session, an N‐back working memory task, and a subsequent resting state session immediately following the task. Based on previous findings that task performance generally alters functional connectivity in a persistent fashion, and that working memory in particular affects connectivity within and between the DMN and TPN, we first hypothesized that these network relationships would be altered during the N‐back task relative to baseline, and that alteration would persist during the subsequent resting state scan. Further, the magnitude of the changes in resting state connectivity from baseline to post‐task scans should be predicted by the strength of connectivity during the N‐back task. Such a finding would strongly suggest that any persistent changes in resting state functional connectivity are specifically due to the performance of the task. Notably, as the reported direction of change in functional connectivity during N‐back tasks has varied widely across previous studies [Bluhm et al., 2011; Fransson, 2006; Fransson and Marrelec, 2008; Gordon et al., 2012a], we have no clear basis for specifying the direction in which connectivity will be persistently altered by the task. Second, while it is known that within‐DMN and DMN‐TPN resting state connectivity predict task‐irrelevant thought [Buckner et al., 2008], attention lapses [Weissman et al., 2006], and mind‐wandering [Mason et al., 2007], it is unknown whether such relationships exist if the connectivity has been persistently altered by a working memory task. Thus, we hypothesized that resting state DMN connectivity would correlate with a behavioral measure of inattentiveness, and we examined whether this correlation remains stable across the three states: the initial, “baseline” resting state session, the N‐back task, and the subsequent resting state session.

METHODS

Subjects

Fifty Georgetown University undergraduates (35 female; 48 right handed) ages 18 to 22 years (M ± SD = 20.44 ± 0.90) participated in the study for payment. Informed consent procedures were carried out according to Georgetown University's Institutional Review Board guidelines. Exclusion criteria included (1) self‐reported use of psychotropic medication (e.g., stimulants, anti‐anxiety/depression); (2) self‐reported history of neurological injury or disease, seizure disorder, psychiatric diagnosis; and (3) contraindications for MRI—e.g., metal implants in the body, dental work involving metal, pregnancy.

Behavioral Testing

Subjects completed the Adult ADHD Self‐Report Scale v1.0 [Kessler et al., 2005], which reported trait measures of Inattention and Hyperactivity‐Impulsivity. This scale consists of 18 questions about subjects' inattentive and impulsive behaviors consistent with DSM‐IV criteria for ADHD. It is scored on a five‐point Likert scale; scores can range from 0 to 36 for both inattention and for hyperactivity‐impulsivity. The questionnaire was administered at the time of subject recruitment. Analysis focused on the inattention score, as the hyperactivity‐impulsivity measure did not relate to connectivity in past work [Gordon et al., 2012a]. Across subjects, mean Inattention scores were 13.58 ± 3.64; these scores were approximately normally distributed across the sample (D'Agostino‐Pearson omnibus test statistic = 1.27, P = 0.47).

Scanning Procedure

Each scanning session included five scans over the course of 30 minutes. The first scan was a resting state scan (Rest1) lasting 5:04 min in which subjects were told to relax with eyes closed and to not think of anything in particular. This was followed by two scans during a simple shape discrimination task (not described here) lasting 5:46 min each, and then by acquisition of a high‐resolution anatomical image lasting 4:18 min. Subjects then performed an N‐back working memory task lasting for 6:26 min (task described below) followed immediately by a second resting state scan (Rest2) identical to the Rest1 scan. Note that as the focus of this study is on post‐task modification of resting state connectivity, we consider the impact of only the N‐back task on the Rest2 scan, as it immediately preceded the Rest2 scan; effects of the shape discrimination task cannot be directly linked with the Rest2 scan as it was separated in time from the Rest2 scan by the intervening N‐back task and by the anatomical image acquisition.

The N‐back task consisted of nine 30s N‐back blocks (three blocks each at 1‐, 2‐, and 3‐back) alternating with eight 15s blocks of fixation. Each N‐back block consisted of nine serially presented consonants appearing for 500 ms, with an inter‐trial‐interval of 2,500 ms. The N‐back load condition (1‐, 2‐, or 3‐back) varied between task blocks, with condition order pseudorandomized using a modified Latin Square. Each block was preceded by a 3,000 ms screen informing the subject of the N‐back condition. Subjects were instructed to press a hand‐held button with their right hand when the current letter matched the letter n trials ago (e.g., for the two‐back condition, subjects see: R V N W N—button‐press for N). Targets were present on 19% of trials; each block contained between one and three targets with target frequency balanced across conditions. No condition contained sequences of stimuli that were targets in any other condition. Stimuli were presented using E‐Prime (Psychology Software Tools Inc., Pittsburg, PA).

N‐back accuracy was calculated as (% target hits—% false alarms). N‐back accuracy was very high (see Supporting Information), and therefore was not examined in any further analyses. Reaction time (RT) was calculated as the mean time to respond to targets correctly hit.

fMRI Data Acquisition

Imaging was performed on a Siemens Trio 3T scanner (Erlangen, Germany). For the resting runs, 152 whole‐brain images were acquired using a gradient echo pulse sequence (37 slices, TR = 2,000 ms, TE = 30 ms, 192 × 192 mm FOV, 90 degree flip angle, voxel dimensions 3 mm isotropic). For the N‐back run, 197 whole‐brain images were acquired using a gradient echo pulse sequence (34 slices, TR = 2,000 ms, TE = 30 ms, 256 × 256 mm FOV, 90 degree flip angle, voxel dimensions 4 × 4 × 4.2 mm). The first four images of each functional run were discarded to allow for signal stabilization. A high resolution T1‐weighted structural scan (MPRAGE) was acquired with the parameters: TR/TE = 2,300/2.94 ms, TI = 900 ms, 90 degree flip angle, 1 slab, 160 sagittal slices with a 1.0 mm thickness, FOV = 256 × 256 mm², matrix=256 × 256, resulting in an effective resolution of 1.03 mm isotropic voxels.

Image Preprocessing

Using SPM8 (Wellcome Department of Cognitive Neurology, London, UK) implemented in MATLAB (Version 7.10 Mathworks, Inc., Sherborn, MA), images were corrected for translational and rotational motion by realigning to the first image of each scanning run (Rest1, N‐back, and Rest2). Across all runs, all subjects demonstrated less than 2.0 mm of translational motion from the first image in any one direction (max translation = 1.25 mm) and less than 1.0° of rotation from the first image around any one axis (max rotation = .54°). Root mean squared (RMS) motion was further calculated as a summary measure of motion for each subject in each run. RMS motion was 0.076 ± 0.23 in the Rest1 run, 0.085 ± 0.39 in the Nback run, and 0.083 ± 0.31 in the Rest2 run, and paired t‐tests indicated that it did not differ between runs (all pairwise ts < 1.8, all ps > 0.05). Further analyses indicated that the reported results were not caused by effects of motion (see Supporting Information).

Images were then slice‐time corrected, normalized to an EPI template, and smoothed using a Gaussian kernel with full‐width at half‐maximum (FWHM) of 8 mm. Finally, a band‐pass filter was applied to the data to restrict signal variation to frequencies between 0.01 Hz and 0.1 Hz, corresponding to the frequency range established in the literature for fluctuations in resting state data [Biswal et al., 1995].

Functional Connectivity Calculation

Network node creation

Regions of interest (ROIs) representing nodes of the DMN and TPN networks were created based on a previously published Independent Components Analysis (ICA) of resting state data conducted in an independent sample (N = 44) of subjects demographically similar to those in the present study [Gordon et al., 2012b]. Two ICA components were identified which together visually matched the canonical DMN, one comprising the anterior aspect of the DMN and one comprising the posterior aspect, and one component (described as the “Set Maintenance” component in Gordon et al. 2006) was identified which visually matched the canonical anticorrelated TPN. ROIs were created as the large clusters (k > 50 voxels) of the identified components after thresholding the Z‐map at Z > 7.0. This resulted in five DMN nodes located in ventral and anterior dorsal medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), precuneus (Prec), and bilateral angular gyrus (lAG, rAG), as well as six TPN nodes located in dorsal anterior cingulate cortex (dACC), bilateral dorsolateral prefrontal cortex (ldlPFC, rdlPFC), bilateral anterior insula (laIns, raIns), and right supramarginal gyrus (rSMG). These nodes well‐matched regions strongly correlated and anticorrelated with an independent PCC seed in the present data (see Supporting Information Fig. S1), and also had good overlap with regions activated and deactivated by the present N‐back task (see Supporting Information Fig. S2).

Nuisance signal identification

To identify the effects of motion and physiological noise (such as respiration and heart rate) that would be common to all nodes, timecourses approximating these signals were calculated for each scanning run. Physiological noise timecourses were approximated by obtaining signal timecourses from white matter and CSF segmentations of the MPRAGE image [Van Dijk et al., 2010]. Motion timecourses were obtained as the six realignment parameter timecourses from the motion correction preprocessing step, expressed as absolute differences from the first timepoint in each of the three translation and rotation directions. Notably, the global signal was not included as a nuisance signal, as regression of the global signal is known to strongly affect observed negative correlations [Chang and Glover, 2009; Fox et al., 2009; Murphy et al., 2009], and recent work suggests that the procedure may reduce the accuracy of connectivity estimates [Saad et al., 2012].

Functional connectivity calculations

For each network node in each run, the average node timecourse was calculated across all voxels. Connectivity between two nodes in a given run was calculated by conducting partial correlations between the timecourses of those nodes, while partialling out the motion and physiological noise timecourses. The resulting r values were converted to Z‐scores using Fisher's transformation in order to increase normality of the distribution, allowing further statistical analysis of correlation strengths. In the N‐back run, analysis was restricted to task performance by excluding the 15 s in each fixation block plus 6 subsequent seconds (to allow for hemodynamic response stabilization). The partial correlations were thus assessed on 148 timepoints in each Rest run and 108 timepoints in the N‐back run. Results obtained without removal of these fixation blocks are presented in Supporting Information; they did not differ from those obtained with removal of the fixation blocks. To further avoid any artificial inflation of the N‐back correlations that could be driven by the block structure of the task load conditions, effect of load was partialled out. Load‐effect timecourses were obtained by convolving three boxcar timecourses (one for each load condition) with a canonical hemodynamic response, and were partialled out of the N‐back connectivity correlations (see Gordon et al. 2006, and Jones et al. 2009 for discussions of the rationale for removing task‐load effects).

Functional connectivities calculated for each run, in each network

Using this method, for each subject in each run (Rest1, N‐back, Rest2), within‐DMN connectivity calculations were conducted between the PCC node and each other DMN node. Cross‐network DMN‐TPN connectivity calculations were conducted between the PCC node and each TPN node. Finally, we calculated the strength of anticorrelation between each node and the PCC node in the independent dataset. This calculation revealed that the rSMG node was most negatively correlated with the PCC, followed by the raIns node. As the SMG is not usually considered a key TPN node, we selected the raIns as the central TPN node. Therefore, within‐TPN connectivity calculations were conducted between the raIns node and each other TPN node.

Effects of Performing a Working Memory Task on Connectivity

Changes in functional connectivity across runs

For the within‐DMN connections calculated above, mean connectivity values were entered into a 3 (State: Rest1, N‐back, Rest2) × 4 (connections from PCC to other DMN nodes) repeated measures Analysis of Variance (ANOVA). A similar 3 (State) × 6 (connections from PCC to TPN nodes) repeated measures ANOVA was conducted for the cross‐network PCC‐TPN connections, and a 3 (State) × 5 (connections from raIns to other TPN nodes) repeated measures ANOVA was conducted for the within‐TPN connections. In the three ANOVAs, main effects of State tested whether overall connectivity differed across Rest1, N‐back, and Rest2 states, and State X connection interactions tested whether state‐related differences were present more strongly in some connections than in others. Significant effects were followed up with paired t‐tests.

Effect of task connectivity and performance on subsequent resting state connectivity

For within‐DMN, DMN‐TPN, and within‐TPN connections, we tested whether connectivity in these connections during the N‐back task predicted the subsequent Rest2 connectivity. To reduce multiple comparisons, we calculated average connectivity from the PCC to all other DMN nodes as overall within‐DMN connectivity, average connectivity from the PCC to all TPN nodes as overall DMN‐TPN connectivity, and average connectivity from the raIns to all other TPN nodes as overall within‐TPN connectivity. We then correlated the strength of these overall connectivities during the N‐back task to their strength during Rest2. Significant findings were followed up with post‐hoc connection‐wise analyses to identify connections with the most robust N‐back vs. Rest2 correlations. Finally, we tested whether the strength of Rest2 connectivity was also predicted by N‐back task performance in these nodes by correlating average RT against the strength of Rest2 connectivity.

Correlations Between Within‐DMN Connectivity and Trait‐Level Inattention Measures

For each run (Rest1, N‐back, Rest2), we tested whether the overall strength of within‐DMN connectivity (calculated above) was correlated with trait‐level inattention; significance values were corrected for the three runs tested using Bonferroni correction. We then tested whether the correlations with Inattention differed in the different runs using Meng's t‐test. Significant correlations were followed up with post‐hoc nodewise analyses to determine for which nodes the correlations were most robust. Though our a priori interest was in associations between inattention and within‐DMN connectivity, we also tested within‐TPN and DMN‐TPN connections for completeness.

RESULTS

Effect of Performing the N‐Back Task on Connectivity

Within‐DMN

The State X Node ANOVA revealed a significant main effect of State [F(2,48)=5.12, P = 0.01], indicating that connectivity between PCC and other DMN nodes differed between Rest1, N‐back, and Rest2. Effects are graphed in Figure 1. Overall PCC connectivity with DMN nodes became significantly less positive from the Rest1 run to the N‐back run [t(49) = 3.18, P = 0.003], and then became more positive again from the N‐back run to the Rest2 run [t(49) = 2.35, P = 0.023], with no differences observed between the Rest1 and Rest2 runs [t(49) = 0.45, P = 0.65]. Further, a significant interaction indicated that the effect of State varied by node [F(6,44) = 8.03, P < 0.001] such that connectivity with the angular gyrus nodes became significantly less positive from the Rest1 run to the N‐back run [rAG: t(49) = 7.29, P < 0.001; lAG: t(49) = 3.55, P = 0.001], and then became significantly less positive again from the N‐back run to the Rest2 run in the rAG node [t(49) = 5.94, P < 0.001] but not the lAG node [t(49) = 1.90, P = 0.064], with no differences observed between the Rest1 and Rest2 runs in either node [ts(49)<1.5, ps > 0.30]. In the Prec and mPFC nodes, no differences by state were observed [all ts(49) < 2.0, all ps > 0.05]. These findings indicate that the N‐back task altered within‐DMN functional connectivity in the AG nodes, but that connectivity returned to baseline after the task was complete.

Figure 1.

Default mode network nodes (brain slices) and within‐DMN connectivity strength in those nodes (small graphs) and across the whole network (larger bottom graph) measured during the baseline Rest1 scan, the N‐back working memory scan, and the post‐task Rest2 scan. The seed region for connectivity calculations was the PCC (in green). Unless otherwise indicated, * indicates significant differences from Rest1 at P < 0.05. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

DMN‐TPN

The State X Node ANOVA revealed a significant main effect of State [F(2,48)=15.82, P < 0.001]. Effects are graphed in Figure 2. Overall PCC connectivity with TPN nodes became significantly less negative from the Rest1 run to the N‐back run [t(49) = 4.48, P < 0.001] and remained less negative in the Rest2 run than it was in the Rest1 run [t(49)=4.52, P < 0.001], with no difference observed between the N‐back and Rest2 runs [t(49) = 0.35, P = 0.72]. Further, a significant interaction indicated that the effect of State varied by node [F(10,40) = 7.00, P <0. 001] as follows: Connectivity of the PCC with the dACC, raIns, and laIns nodes became less negative from the Rest1 run to the N‐back run [ts(49) > 3.50, ps < 0.001] and remained less negative in the Rest2 run than it was in the Rest1 run [ts(49) > 2.9, ps < 0.005], with no differences observed between the N‐back and Rest2 runs [ts(49) < 1.7, ps > 0.10]. Thus, in these nodes, connectivity became less negative during the N‐back run and remained so during the Rest2 run. A similar pattern was observed for the rSMG node, with connectivity becoming less negative from the Rest1 run to the N‐back run [t(49) = 8.51, P < 0.001], though connectivity did become more negative again from the N‐back run to the Rest2 run [t(49) = 3.78, P < 0.001] but still remained less negative in the Rest2 run than it was in the Rest1 run [t(49) = 4.73, P < 0.001]. Thus, in this node connectivity became less negative during the N‐back run and only returned partially back to its original baseline in the Rest2 run. By contrast, in the rdlPFC and ldlPFC nodes, connectivity with the PCC did not change from the Rest1 run to the N‐back run [ts(49) < 0.80, ps > 0.45], but became less negative from the N‐back run to the Rest2 run [significantly for the ldlPFC: t(49) = 2.02, P = 0.049; and at trend level for the rdlPFC: t(49) = 1.70, P = 0.095) and, critically, remained less negative in the Rest2 run than in the Rest1 run [ts(49) > 2.60, ps < 0.015]. These findings indicate that the N‐back task induced PCC connectivity with dACC, aIns, and rSMG TPN nodes to become less negative, and, critically, PCC connectivity with all TPN nodes remained less negative during the post‐task resting state relative to the pretask resting state.

Figure 2.

Task‐positive network nodes (brain slices) and cross‐network DMN‐TPN connectivity strength in those nodes (small graphs) and across the whole network (larger bottom graph) measured during the baseline Rest1 scan, the N‐back working memory scan, and the post‐task Rest2 scan. The seed region for connectivity calculations was the PCC (in green). Unless otherwise indicated, * indicates significant differences from Rest1 at P < 0.05. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Within‐TPN

The State X Node ANOVA revealed a significant main effect of State [F(2,48)=6.00, P = 0.005]. Effects are graphed in Figure 3. Overall raIns connectivity with TPN nodes became significantly less positive from the Rest1 run to the N‐back run [t(49)=3.15, P =0.003] and remained less positive in the Rest2 run than it was in the Rest1 run [t(49)=2.97, P =0.005], with no difference observed between the N‐back and Rest2 runs [t(49)=1.29, P =0.20]. Further, a significant interaction indicated that the effect of State varied by node [F(8,42)=4.32, P =0.001] as follows: In the laIns and ldlPFC nodes, connectivity with the raIns followed the pattern of the overall network, becoming significantly less positive from the Rest1 run to the N‐back run [ts(49) > 2.00, ps < 0.05] and remaining less positive in the Rest2 run than it was in the Rest1 run [ts(49) > 2.40, ps < 0.02], with no differences observed between the N‐back and Rest2 runs [ts(49) < 1.0, ps > 0.35]. By contrast, in the rSMG node, connectivity with the raIns also became less positive from the Rest1 run to the N‐back run [t(49) = 4.98, P < 0.001], but then became more positive again from the N‐back run to the Rest2 run [t(49) = 4.71, P < 0.001], returning all the way back to baseline such that no difference was observed between the Rest1 and Rest2 runs [t(49) = 1.66, P = 0.10]. Finally, State did not appear to affect connectivities of the raIns with the dACC or rdlPFC nodes, as no significant pairwise connectivity comparisons emerged [ts(49) < 2.0, ps > 0.05]. These findings indicate that the N‐back task induced connectivity between the raIns and the laIns, ldlPFC, and rSMG nodes to become less positive, and that connectivities remained less positive in the laIns and ldlPFC nodes during the subsequent resting state than they were in the initial resting state.

Figure 3.

Task‐positive network nodes (brain slices) and within‐TPN connectivity strength in those nodes (small graphs) and across the whole network (larger bottom graph) measured during the baseline Rest1 scan, the N‐back working memory scan, and the post‐task Rest2 scan. The seed region for connectivity calculations was the right anterior insula (in green). Unless otherwise indicated, * indicates significant differences from Rest1 at P < 0.05. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Effect of Task Connectivity and Performance on Subsequent Resting State Connectivity

We examined whether the strength of within‐DMN, DMN‐TPN, or within‐TPN connectivity observed during the N‐back task predicted the subsequent Rest2 connectivity. We found that N‐back connectivity did not predict Rest2 connectivity in within‐DMN [r(48) = 0.07, P = 0.61] or DMN‐TPN [r(48) = 0.20, P = 0.16] connections, but it did predict Rest2 connectivity in within‐TPN connections [r(48) = 0.31, P = 0.031]; see Figure 4A. As the previous analyses indicated that within the TPN, persistent post‐N‐back changes in connectivity were localized to raIns‐laIns and raIns‐ldlPFC connections, we conducted post‐hoc nodewise correlations between N‐back and Rest2 connectivity in these two connections. We found that N‐back connectivity predicted Rest2 connectivity in the raIns‐laIns connection [r(48) = 0.34, P = 0.017; Fig. 4B] but not the raIns‐ldlPFC [r(48) = 0.14, P = 0.34]. We further investigated whether reaction time on the N‐back task also predicted subsequent Rest2 connectivity in these two connections. We found that RT predicted Rest2 connectivity in the raIns‐laIns connection [r(48) = −0.28, P = 0.05; Fig. 4C], but not in the raIns‐ldlPFC connection [r(48) = 0.11, P = 0.45]. These findings suggest that the reduction in positive connectivity that persists after the N‐back task within the interhemispheric aIns connection is predicted by the strength of the connectivity during the task, as well as by the speed of performance on the task, such that less positive N‐back connectivity and slower N‐back performance predict less positive resting‐state connectivity afterwards.

Figure 4.

A: Connectivity during the N‐back task predicts subsequent resting‐state connectivity within the TPN but not within the DMN or in DMN‐TPN connections. B: Post‐hoc nodewise analyses show that relationships between N‐back connectivity and Rest2 connectivity are strongest in the connection between right and left anterior insula. C: Post‐task Rest2 connectivity between right and left anterior insula was predicted by the speed of N‐back performance.

Associations Between Within‐DMN Connectivity and Trait‐Level Inattentiveness

We examined whether trait‐level Inattentiveness assessed before scanning was associated with within‐DMN connectivity measured during the initial resting‐state scan, and whether it continued to be associated with within‐DMN connectivity measured during and after the N‐back task. After correcting for the three correlations conducted, we found that overall connectivity within the DMN did predict Inattentiveness during the Rest1 run [r(48) = −0.36, P = 0.01] and the N‐back task [r(48) = −0.36, P = 0.009] but not during the Rest2 run [r(48) = 0.065, P = 0.65]; see Figure 5A. Meng's t‐tests indicated that this correlation was higher in both the Rest1 [t(47) = 2.90, P < 0.01] and N‐back runs [t(47) = 2.33, P < 0.05] than in the Rest2 run, with no differences observed between the Rest1 and N‐back runs [t(47) = 0.016, P > 0.9]. These findings indicate that individuals who were more attentive demonstrated more positive connectivity within the DMN during the Rest1 and N‐back runs, but that the degree of Inattentiveness had no association with resting connectivity after the N‐back task. Post‐hoc correlations between Inattentive scores and each separate within‐DMN connection indicated that reduced Inattention scores were most strongly predicted by a stronger PCC‐lAG connection during the Rest1 [r(48) = −0.31, P = 0.028] and N‐back [r(48) = −0.29, P = 0.039] runs, but the relationship emerged only at trend level, and, notably, in the opposite direction during the Rest2 run [r(48) = 0.26, P = 0.072]; see Figure 5B. The PCC‐Prec and PCC‐mPFC connections also demonstrated negative correlations with Inattention at trend level during the Rest1 and N‐back runs [rs(48) < −0.24, ps < 0.10], but no significant correlation during the Rest2 run [|rs(48)| < 0.15, ps > 0.45]. By contrast, no significant associations emerged between Inattention scores and within‐TPN or DMN‐TPN connections.

Figure 5.

A: Correlations between trait‐level Inattentiveness and within‐DMN connectivity emerged during the initial resting state and during the N‐back task, but not during the post‐task resting state. B: Similarly, post‐hoc nodewise correlations revealed that trait‐level Inattentiveness correlated significantly with connectivity between posterior cingulate and left angular gyrus during the initial resting state and during the N‐back task, but not during the post‐task resting state.

DISCUSSION

This study examining the status of functional connectivity networks during working memory (WM) engagement and after its disengagement led to two principal novel findings. First, connectivity with the task‐positive network (TPN), both within the network and in cross‐network connections to the default mode network (DMN), was altered during N‐back task performance relative to the resting state preceding the task. Critically, this altered connectivity persisted following task completion and into the subsequent resting state. Additionally, individuals with a less coordinated TPN during the N‐back task demonstrated a less coordinated TPN after the task. Second, individuals who had higher scores on a self‐reported attention scale demonstrated a more coordinated DMN during the pretask resting state and during the N‐back task, but this association was eliminated during the post‐task resting state. These findings suggest that brain network function is altered during the period immediately following WM performance, with properties that are associated with the preceding cognitive engagement.

Persistence of Working Memory‐Induced Functional Connectivity

This study is the first to demonstrate that alterations in functional connectivity during WM task performance can persist beyond the task's conclusion and into the subsequent resting state. For both within‐network connections between the raIns and lateral frontal, parietal, and insular TPN nodes, as well as for cross‐network connections between the PCC and TPN nodes, connectivity during the WM task was altered compared to the baseline resting state. These changes largely did not fade away after the task concluded, as connectivity during the immediately following resting state remained altered compared to the baseline resting state. Persistently altered connectivity within the task‐relevant network has been previously observed after sensorimotor (finger tapping [Duff et al., 2008], motor learning [Albert et al., 2009; Vahdat et al., 2011], and visual learning [Lewis et al., 2009]), language (language generation [Waites et al., 2005] and semantic classification [Grigg and Grady, 2010]), and episodic memory (face‐object pair encoding [Tambini et al., 2010]) tasks. Similar to several of those previous studies [Duff et al., 2008; Grigg and Grady, 2010; Lewis et al., 2009], we also observed pre/post‐task connectivity changes between the task‐relevant network and the DMN, indicating that both within‐ and between‐network relationships are persistently affected by the performance of a task. The present findings extend this phenomena to WM, a key ability subserving effortful cognitive functions [Engle et al., 1999].

This study is also the first to demonstrate that the strength of resting state functional connectivity observed after completion of a task is predicted by the strength of connectivity during the task itself. Less positive overall connectivity of TPN nodes with the right anterior insula during the N‐back task, most prominently with the homologous left anterior insula, predicted less positive connectivity in the same connections after the task, indicating that individuals who had low TPN coordination during the task, compared to other individuals, also had low TPN coordination during the period after the task. By contrast, post‐task connectivity in cross‐network DMN‐TPN connections was not predicted by connectivity during the N‐back task, suggesting that effects of the N‐back task on cross‐network connections may be indirect. The extent to which this relationship is stable will depend upon replication in future studies.

We offer three possible interpretations of the functional significance of the observed findings that we cannot confirm with the present data but which could be tested in future work. First, the persistence of WM‐related connectivity changes may reflect a cognitive aftereffect, such that the brain continues to perform the N‐back task even after the task stimulation has ceased. However, our results showing selectivity in the persistence of connectivity alterations at the network level (e.g., within‐TPN and DMN‐TPN but not within‐DMN) and at the connection level (e.g., raIns with laIns and ldlPFC but not raIns with dACC and rdPFC) go against this interpretation, as all N‐back task‐related changes in connectivity ought to have persisted if the brain continued to be similarly engaged in the post‐task period. Instead, the specific connections that were persistent—and particularly the persistent connection that was most predictive of individual variation in post‐task connectivity, the connection between bilateral anterior insula—has been previously associated with interoceptive processing [Critchley et al., 2004]. We speculate that persistent connectivity changes in this connection may reflect persistence of the subjective aspects of an effortful cognitive experience. This speculation can be tested empirically in the future. Second, the persistence of WM‐related connectivity changes may reflect consolidation of the N‐back experience. Indeed, studies of resting periods following skill learning [Albert et al., 2009; Duff et al., 2008; Lewis et al., 2009] and memory encoding [Tambini et al., 2010] suggest that post‐task connectivity changes reflect consolidation of those experiences in the service of learning. Although subjects were not instructed to learn anything in the present study, any experience has the potential to be a memory and will be consolidated following its conclusion.

Third, persistence of WM‐related connectivity may reflect a post‐task recovery period following effortful cognition that reflects persistent depletion of cognitive resources. This interpretation is based on the direction of connectivity changes we observed (e.g., less positive connectivity within the TPN and less negative DMN‐TPN connectivity during and after the WM task), which have been associated with worse cognitive performance. Less positive connectivity within the TPN predicted reduced executive control abilities [Gordon et al., 2011], and less negative DMN‐TPN connectivity was associated with increased inattentiveness and impulsivity [Gordon et al., 2012a], increased trial‐to‐trial behavioral variability [Kelly et al., 2008] and reduced working memory performance [Hampson et al., 2010; Sala‐Llonch et al., 2012]. This interpretation also agrees with Barnes et al. 2009, who posited that such a recovery period is required following effortful N‐back task performance and that it scales with the effort involved, as they observed that post‐task oscillatory alterations returned to baseline about 8 min after a 1‐back task and about 15 min after a 2‐back task. In the present task, although accuracy was very high (see Supporting Information), verbal reports by subjects indicated that they found the task to be highly effortful. Indeed, we found that subjects who were slower on the N‐back task (indicating that more cognitive effort was required for performance) demonstrated less positive within‐TPN connectivities after the completion of the task than subjects who were faster. While this correlation would not have survived a stricter statistical threshold controlling for multiple comparisons, it suggests that it may be worthwhile testing whether the persistent suboptimal network configurations we observed may correspond to the state of depleted cognitive resources and mental fatigue experienced during [Kato et al., 2009; Lorist, 2008] and after [Holtzer et al., 2010; van der Linden et al., 2003] periods of demanding cognition. Unfortunately, our post‐task resting state period did not extend long enough to examine whether the timecourse of network recovery to baseline that was observed by Barnes et al. 2009 may generalize to connectivity strength.

The present results may help disambiguate mixed findings of how within‐ and between‐network functional connectivity changes from the resting state to WM task performance. Compared to a resting state, within‐TPN connectivity during an N‐back task became both more positive [Gordon et al., 2012a; Newton et al., 2011] and less positive [Fransson, 2006; Repovš and Barch, 2012], as in the present study. Within‐DMN connectivity both remained unchanged in primary DMN nodes [Bluhm et al., 2011; Gordon et al., 2012a], became less positive [Fransson, 2006; Newton et al., 2011], as in the present study, and became more positive at low loads but less positive at high loads [Repovš and Barch, 2012]. Further, cross‐network DMN‐TPN connectivity became more negative [Fransson, 2006; Repovš and Barch, 2012], more positive [Bluhm et al., 2011], similar to the present study, and changed inconsistently depending on the specific TPN node [Gordon et al., 2012a]. One possible explanation for these varied results is that the studies differed in when the resting state session was performed, reflecting a mixture of pre and post‐task connectivity changes. Gordon et al. 2004 and Newton et al. 2012 conducted resting scans after the WM task, while Fransson 2012a counterbalanced scan order across subjects, Bluhm et al. 2011 conducted resting scans interspersed with the task, and Repovš and Barch 2011 did not report the order of scanning. Thus, in all five studies, the resting state scan may have included some remnants of connectivity changes from the preceding WM task for at least some subjects. The present findings thus draw attention to the fact that the order in which resting and task scans are collected may have a substantial impact on measurements of functional connectivity.

Association of DMN Connectivity with Individual Variation in Inattentiveness

This study is the first to demonstrate that healthy adults with lower self‐reported inattentiveness have more positive functional connectivity within the DMN during a baseline resting state session. Previous work has established that DMN engagement (either strong within‐network resting state functional connectivity or strong task‐related deactivation) predicts improved performance across a variety of cognitive domains, including self‐referential processing and rumination [Zhu et al., 2012], executive function and processing speed [Andrews‐Hanna et al., 2007], cognitive control [Eichele et al., 2008; Li et al., 2007], working memory performance [Hampson et al., 2006; Sala‐Llonch et al., 2012; Sambataro et al., 2010], and behavioral symptoms of inattentiveness [Brown et al., 2011] and distractibility [Fassbender et al., 2009] in children and adults with Attention Deficit Hyperactivity Disorder (ADHD). The present results show that the association with behavioral inattentiveness extends to healthy young adults who do not meet criteria for ADHD.

Notably, this association with trait‐level inattentiveness was observed both in the resting state preceding working memory engagement and during working memory, but not in the resting state following the task. This lack of connectivity‐behavior correlations during the post‐task session could be explained by the substantial individual variation in post‐task resting state network connectivity strengths [as shown by the wide range of Z(r) values in Fig. 4A]. If post‐task changes in resting state connectivity are due to a recovery period following demanding cognitive effort, as posited above, individual differences in the timing of recovery may function as a source of unexplained variance that would weaken the ability to detect a correlation between behavior and connectivity strength. From a methodological perspective, the fact that resting state connectivity can be altered enough by a previous task to eliminate correlations between connectivity and behavior means that scanning protocols should be designed with care so that resting state connectivity is not contaminated by the effects of previous tasks.

Caveats

The following factors ought to be noted, as they may influence comparison of functional connectivity measures during the resting‐state and the working memory task. First, the working memory run has one more nuisance regressor (load) than the resting run. Regressing out load is necessary in order to directly compare the resting and working memory runs, as it limits the possibility of connectivity being driven by load‐related activation differences, as discussed in the Methods section and in previous work [Gordon et al., 2012a]. Second, the acquisition voxel size differed between the resting and working memory runs. However, previous work has indicated that differences in voxel size do not affect measures of functional connectivity [Van Dijk et al., 2010]. Finally, the design of the present study cannot account for other possible changes in a subject's cognitive state over the course of the scanning session, which may include increasing fatigue and decreasing alertness. Thus, it is possible that subjects were more drowsy or tired in the second rest than in the initial rest session, which may have driven some portion of the connectivity alterations (though the fact that within‐DMN connections were not persistently altered mitigates this concern to some extent). A study design that would have controlled for this (e.g., a control group scanned during Rest1‐Rest2‐Rest3 or counterbalanced task and rest runs) is needed in order to definitively determine a causal effect of the WM task on post‐task resting state.

CONCLUSIONS

Functional connectivity within the task‐positive network, as well as between the default mode and task‐positive networks, was altered during working memory performance compared to a baseline resting state. Most importantly, these alterations in connectivity persisted after disengagement from working memory, during the subsequent resting state. We characterized two properties of this persistently altered connectivity. First, both slower working memory performance and weaker connectivity between two key task‐positive nodes during working memory predicted weaker connectivity in that connection after working memory disengagement, suggesting that the putative post‐task recovery period was associated with cognitive effort during task performance. Second, stronger default mode connectivity measured during the baseline resting state and during the working memory task predicted lower inattentiveness, but this brain‐behavior association disappeared during the immediately following resting state. This suggests that post‐task recovery from cognitive effort may obscure connectivity‐behavior relationships that are observable before and during a task. These findings have important implications not only for models of how the brain recovers following transient effortful cognition, but also for the design of future studies investigating the resting state and rest‐task relationships.

Supporting information

Supporting Information Figure 1.

Supporting Information Figure 2.

Supporting Information