Activation‐based association profiles differentiate network roles across cognitive loads

Abstract

Working memory (WM) is a complex and pivotal cognitive system underlying the performance of many cognitive behaviors. Although individual differences in WM performance have previously been linked to the blood oxygenation level‐dependent (BOLD) response across several large‐scale brain networks, the unique and shared contributions of each large‐scale brain network to efficient WM processes across different cognitive loads remain elusive. Using a WM paradigm and functional magnetic resonance imaging (fMRI) from the Human Connectome Project, we proposed a framework to assess the association and shared‐association strength between imaging biomarkers and behavioral scales. Association strength is the capability of individual brain regions to modulate WM performance and shared‐association strength measures how different regions share the capability of modulating performance. Under higher cognitive load (2‐back), the frontoparietal executive control network (FPN), dorsal attention network (DAN), and salience network showed significant positive activation and positive associations, whereas the default mode network (DMN) showed the opposite pattern, namely, significant deactivation and negative associations. Comparing the different cognitive loads, the DMN and FPN showed predominant associations and globally shared‐associations. When investigating the differences in association from lower to higher cognitive loads, the DAN demonstrated enhanced association strength and globally shared‐associations, which were significantly greater than those of the other networks. This study characterized how brain regions individually and collaboratively support different cognitive loads.

1. INTRODUCTION

The working memory (WM) network is a complex cognitive system in the human brain. It is involved in memory maintenance and rehearsal in the absence of external stimuli, including memory encoding, storage, and selective retrieval, and functions in the hierarchical abstraction of stimuli and conception of plans for behavioral actions (Christophel, Klink, Spitzer, Roelfsema, & Haynes, 2017). WM is also proposed to have relevance for individual capabilities, such as fluid intelligence, reasoning proficiency, and academic performance (Eriksson, Vogel, Lansner, Bergstrom, & Nyberg, 2015).

Brain regions underlying WM processes are widely distributed across the brain (Christophel et al., 2017), and include sensory‐related cortices (e.g., somatosensory and visual cortices) (Pasternak & Greenlee, 2005) as well as the frontoparietal executive control network (FPN), dorsal attention network (DAN), and salience network (Ptak, 2012). For example, memory storage of visual features not only occurs in the visual cortex but also in the parietal and prefrontal cortices and superior precentral sulcus (Eriksson et al., 2015), and in the WM process, the FPN globally interacts with other networks of the brain (Cole, Repovs, & Anticevic, 2014; D'Esposito & Postle, 2015). On the other hand, intrinsic networks, particularly the default mode network (DMN), deactivate to suppress stimulus‐unrelated physiological signals to support task execution mediated by task‐positive networks, which characterize other networks, such as the FPN and DAN, as positively activated to cope with external stimuli (Amer, Anderson, Campbell, Hasher, & Grady, 2016; Wen, Liu, Yao, & Ding, 2013; Zhou et al., 2018). This trait has been revealed from the perspective of functional activation (Pagnoni, 2012) and functional interaction (Amer et al., 2016; Zhou et al., 2018), and its action is a prominent factor predicting WM performance (Amer et al., 2016; Wen et al., 2013).

A crucial mission of system neuroscience based on neuroimaging techniques is to explore the neural mechanisms that underlie cognitive performance. Despite ample work identifying networks implicated in WM processes, the direct links between WM performance and brain network behavior, including regional activations and interactions between networks across large‐scale networks, have been less explored. Previous research has reported an association between WM performance and blood oxygenation level‐dependent (BOLD) activation and/or functional connectivity for specific networks (Salami et al., 2018). Furthermore, several studies have reported that stronger functional activation and connectivity in the FPN correspond to higher WM performance (Nagel et al., 2009; Salami et al., 2018; Ziaei, Salami, & Persson, 2017), whereas other studies have found anticorrelations between the FPN and DMN to be associated with WM performance (Xin & Lei, 2015). For associations between network reconfigurations and WM performance, functional integration and segregation have also been linked to cognitive capability (Cohen, Gallen, Jacobs, Lee, & D'Esposito, 2014; Shine et al., 2015; Shine et al., 2016). However, the individual and shared contributions of brain networks to proficient WM function have not been adequately explored. Complex cognitive processes, such as WM, may not be localized to discrete brain regions, but rather mediated by collaborative interactions among a set of brain regions. From this view, how different brain regions function cohesively and how they interact to mediate global networks to support WM task execution remain largely unknown.

Considering that activation strength (AS) usually reflects the extent to which a region is recruited in task execution (Berry, Sarter, & Lustig, 2017; Schweizer, Grahn, Hampshire, Mobbs, & Dalgleish, 2013; Thompson, Waskom, & Gabrieli, 2016), we proposed a framework to examine the relationship between network actions and behavioral performance, and particularly how interactions between networks are associated with WM performance. This was implemented by computing regional and network associations as well as shared associations with other regions to determine WM performance based on typical and semipartial correlations (Cohen, Cohen, West, & Aiken, 2003; Nimon, 2010).

Using a WM paradigm data set from the Human Connectome Project (HCP, N = 328), we first identified brain regions engaged during different cognitive loads and linked to WM performance. Following voxel‐wise analyses of the activation for the two contrasts (i.e., 2‐back vs. 0‐back and 0‐back vs. baseline), we further explored load‐dependent changes in the BOLD response in relation to WM performance. Results demonstrated that the brain regions exhibited corresponding activation and association of performance in the two contrasts. Finally, the association (ASS) and shared‐association strengths (SASS) of each brain network were used to examine their contributions and functional interactions in mediating WM performance. Thus, we developed a framework at the large‐scale network level to assess how brain regions act and interact to guide WM task executions across cognitive loads. We predicted that the degree of activation/deactivation of both the task‐positive network and the DMN would individually and collaboratively contribute to efficient WM processes, and moreover, due to its role in sustaining attention, the DAN would play a critical role at higher cognitive loads.

2. METHODS

2.1. Data sets and preprocessing

The study data were obtained from the HCP (S500 release). We used the same descriptions for data collection as used in our previous work on a partially overlapping portion of the HCP data set (Zuo, Song, Fan, Eickhoff, & Jiang, 2016; Zuo, Yang, Liu, Li, & Jiang, 2018a). Briefly, the data set was collected on a 3 T MRI Skyra scanner (Siemens, Munich, Germany) using a standard 32‐channel head coil. The magnetic field produced by the coil was modeled to provide a customized distortion correction. The primary scanning parameters were repetition time = 720 ms, echo time = 33.1 ms, flip angle = 52°, field of view = 208 × 180 mm², slice thickness = 2.0 mm, and voxel size = 2.0 mm isotropic cube (Van Essen et al., 2013).

The WM test in the HCP is an N‐back visuospatial WM test (Barch et al., 2013) (refer to the task functional magnetic resonance imaging (fMRI) paradigms on page 176) and is an effective and reliable functional localizer across subjects (Drobyshevsky, Baumann, & Schneider, 2006) and time (Caceres, Hall, Zelaya, Williams, & Mehta, 2009). There are two fMRI runs and within each fMRI run there are four long blocks separated by approximately 15 s of rest (fixation). Each block consists of two shorter blocks with either a 0‐back block (25 s each +2.5 s for cue) or 2‐back block (25 s each +2.5 s for cue) from one of the four categories (faces, places, tools, and body pictures) (Barch et al., 2013; Scott, Hellyer, Hampshire, & Leech, 2015). Although the 0‐back task was mainly related to attentional processes, which are not engaged in information maintenance or updating, it is regarded as a cornerstone of WM (Eriksson et al., 2015; Linden, 2007) and as a lower load task (compared to higher load tasks, e.g., 2‐back) for examining network reconfigurations across different cognitive demands (Scott et al., 2015; Shine et al., 2015; Shine et al., 2016; Zuo, Yang, Liu, Li, & Jiang, 2018b). The WM process has a distributed nature in recruiting functional regions across the brain (Christophel et al., 2017; Eriksson et al., 2015), for example, visuospatial perception regions (visual cortex), attentional process regions, and frontal–parietal regions for executive processes. However, it remains unclear, across different cognitive loads, how WM‐related regions act and interact to modulate WM functions and guide WM performance. For example, it is largely unknown how the attentional process sustains attention and interacts with other networks across different cognitive loads, although both the DAN and FPN exert top‐down control during WM processes (see Figure 1b). It is the issue that we addressed in the current study.

Figure 1.

Illustration for the networks involved in the WM process. In this study, we attempted to examine how different working memory (WM)‐related regions act and interact to modulate functions across different cognitive loads

The minimally preprocessed fMRI data set was processed as follows (Glasser et al., 2013): (a) gradient nonlinearity distortion; (b) 6 degree of freedom FSL/FLIRT‐based motion correction; (c) FSL/topup‐based distortion correction; (d) registration to the T1 image space; and (e) FSL/FNIRT‐based registration to MNI 2‐mm space. After receiving the data, we further scrubbed those frames with excessive head motions based on >0.5 mm criterion (percentage of frames with excessive motion was 1.5% for the WM tasks) and corrected them by interpolation (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012).

The HCP S500 data set contained records from 512 subjects. Data were excluded based on the following criteria: (a) data reported on the known issues page of the HCP website, https://wiki.humanconnectome.org/display/PublicData/HCP+Data+Release+Updates%3A+Known+Issues+and+Planned+fixes; (b) data with an insufficient number of frames (correct number of frames for each mental state is described by (Barch et al., 2013); (c) data for which the task fMRIs did not have a full explanatory variable record; and (d) data acquired without left–right or right–left phase encoding. In the end, 453 subjects (188 males, aged 29.1 ± 3.5 years) were used in the subsequent analyses. The original task data sets included seven conventional tasks, comprising WM, gambling, motor, language, social cognition, relational, and emotion processing (Barch et al., 2013); however, we only utilized the WM data set with 0‐back and 2‐back sessions plus baseline sessions to examine activation and association patterns at different cognitive demand loads.

Furthermore, to reduce the bias of ASS‐based analysis, we ensured that the subjects were actively engaged in WM execution inside the MRI scanner, so that the recorded WM scales faithfully reflected WM capability of the individuals. Thus, only subjects with >80% accuracy in the 0‐back task were retained. Here, WM task accuracy was adopted to assess individual WM capability, as reported in previous studies (Schultz & Cole, 2016; Vatansever, Menon, Manktelow, Sahakian, & Stamatakis, 2015). Additionally, subjects without 0‐back or 2‐back measures or whose measures were recognized as outliers were discarded (using MATLAB code written by Brett Shoelson, https://www.mathworks.com/matlabcentral/fileexchange/3961-deleteoutliers). In order to reduce the bias from the twin structures, only one subject from a twin structure was kept and finally 328 subjects remained, with the accuracy scale statistics shown in Figure 2. (For the results based on the cohort without manually destructing the twin structures, thus 385 subjects in total, were presented in Supporting Information.)

Figure 2.

Data used in this study. Panel a indicates correlations (Pearson correlation r = 0.41, p = 1.20e‐14) of working memory (WM) accuracy scales separately during 0‐back and 2‐back tasks. Panels b and c are histograms of 0‐back and 2‐back accuracy, respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

2.2. Generating AS maps in WM task executions

Although many reports describe the regions involved in WM task execution, we generated activation maps for the current data cohort to validate the WM task fMRI data collection rationale and for the HCP‐specific WM task configurations (Barch et al., 2013). We previously generated activation maps using the FSL/FEAT tool based on the same HCP data set (Zuo et al., 2018a), and thus followed the same methods and steps in the current study. Briefly, detection of the activation regions for the WM tasks was implemented using FSL (v5.0.9)/FEAT (v6.00) (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). The fMRI time series was filtered with a Gaussian‐weighted high‐pass filter with a cutoff of 200 s (Barch et al., 2013). The mean signal plus five principal components of white matter (WM) and cerebrospinal fluid (CSF), as well as the movement parameters and their derivatives (Movement_parameters.txt file in the HCP S500 release), were added as confounding factors to remove physiological noise. The above principal components were derived separately by decomposing the regional signal concealed by the eroded WM and CSF masks (Behzadi, Restom, Liau, & Liu, 2007). The statistical parameters were: at the subject level, a within‐subjects fixed‐effects analysis in FSL/FEAT to estimate the average effects across runs, with a cluster‐based activation threshold of Z = 1.96 and Monte Carlo‐based cluster‐level correction of p = 0.05; and, at the group level, a mixed‐effects analysis implemented in FSL/FLAME (FMRIB Local Analysis of Mixed Effects) (Beckmann, Jenkinson, & Smith, 2003) to estimate the average effects of interest with a cluster‐based threshold of Z = ±2.32 and p = 0.05.

2.3. Voxel‐wise ASS maps to WM performance in task execution

For each gray matter voxel across all subjects, the ASS to WM performance (herein measured by WM accuracy) was computed in 0‐back and 2‐back tasks, respectively. Voxel‐wise ASS can be measured as the addressed variance (AV) in regressing voxel‐wise AS (N × 1 vector, where N is the number of subjects) from the accuracy in 0‐back (or 2‐back) (N × 1 vector, where each element is a subject's accuracy), which can be readily implemented by the regression function in MATLAB. For linear correlation and regression between two vectors, A _{N × 1} and B _{N × 1}, the Pearson correlation coefficient r and AV (R ²) show the relationship R ² = r × r (Bewick, Cheek, & Ball, 2003). Therefore, for simpler computation in this study, r was used to indicate ASS. This strategy was used to examine the relationship at the group level between voxel and behavioral actions (namely, WM accuracy). A similar strategy was exploited previously to examine regional coactivations at the group level aroused by external stimuli, for example, participants watching the same movie inside a scanner (Chen et al., 2017), and has been found to be more sensitive to physiological functions while suppressing noisy signals (Simony et al., 2016).

2.4. Network‐wise ASS and SASS

We compared the ASS and SASS between networks. Here, ASS indicated the ASS (characterized by R ² in the regression) from overall nodes in a network to predict WM accuracy, whereas SASS indicated the common ASS between two groups of nodes from two networks. Figure 3 illustrates the study workflow and strategy for computing the ASS/SASS matrix. The matrix elements (Panel 4) indicated the SASS between network pairs for the 10 networks defined by Power et al. (2011).

Figure 3.

Illustration of the analysis pipeline in this study. As described in the first row, the activation maps were computed in two contrasts, that is, 2‐back versus 0‐back and 0‐back versus resting. The association maps between activation strength (AS) and working memory (WM) performance were then computed voxel by voxel. Thereafter, the network‐scale associations and shared associations within and between networks were analyzed. Panel 1 shows the parcellation of brain regions into 10 networks (plus one with undetermined nodes and functions) (Power et al., 2011). Panel 2 illustrates the shared portion (A∩B) in the Venn diagram between Variables A and B in predicting observer Y. Panel 3 indicates the yielded activation matrix across subjects with size N × K, where N is the number of participants and K is the number of nodes (namely ROIs). Finally, the SASS between the 10 networks in Panel 4 is obtained. Abbreviations for network names in Panel 1: cingulo‐opercular network (CON), default mode network (DMN), ventral attention network (VAN), frontoparietal network (FPN), and dorsal attention network (DAN) [Color figure can be viewed at http://wileyonlinelibrary.com]

Pearson correlation (or linear regression) characterizes the relationship between two Variables, A and Y, in reference to their covariance. Furthermore, the specific contribution from Predictor A in regressing Y, by removing all contributions from Predictor B, can be computed using semipartial correlation, in which the squared semipartial correlation coefficient (R ²) indicates the proportion of variance explained specifically by Predictor A (Cohen et al., 2003; Kim, 2015). This method differs from partial correlation because the relationships among independent variables are characterized while the dependent variable remains unchanged (Perrin et al., 2008), and is often used in behavioral research (Bushman et al., 2009), clinical neuroscience (Lowe et al., 2008), and imaging neuroscience (Perrin et al., 2008). As a further extension of specific variance, the shared proportion attributed to the shared variance can be measured by subtracting the sum of the independent contributions from A or B from the total amount of contributions from A and B (Perrin et al., 2008; see also Cohen et al., 2003, pp. 166–169).

In detail, semipartial correlation was used to compute the SASS between two groups of variables. Generally, for two groups of Predictors A and B (e.g., nodal activation vectors from two networks among the 10 predefined networks above) and Observer Y, we studied the AV in predicting Y. Mathematically, the vectors in A and B separately span two vector spaces. The AV by one space can be obtained by regression of the vectors in Space A (or B) from Y and can simply be implemented by multiple regression in MATLAB (regression function with constant items). Here, we denoted AV_A = reg(Y, A) and AV_B = reg(Y, B), respectively. Similarly, for the overall AV in regressing Y by both Spaces A and B, we denoted AV_AUB = reg(Y, AUB) = reg(Y, [A, B]), where AUB and [A, B] (e.g., concatenate the two spaces) denote the same space because duplicate or collinear vectors contribute no extra portions in regression. The uniquely contributed AV by Space A or B is denoted as semipartial correlation, that is, AV_A|B = reg(Y, A|B) and AV_B|A = reg(Y, B|A), where reg(Y, A|B) is the residual subspace of A after regressing out Space B to predict Observer Y. The unique AV was obtained by Space A after regressing out Space B; that is, Residual A (after B was regressed out) was perpendicular to Space B and was used to predict Observer Y.

From the Venn diagram in Figure 4 (Panel 2), the SASS between Spaces A and B toward predicting Y can be defined by the following equation (see Cohen et al., 2003, Chap. 8):

$${\mathrm{AV}}_{\mathrm{A}\cap \mathrm{B}}\triangleq {\mathrm{AV}}_{\mathrm{AUB}}-{\mathrm{AV}}_{\mathrm{A}\mid \mathrm{B}}-{\mathrm{AV}}_{\mathrm{B}\mid \mathrm{A}}\triangleq \mathrm{reg}\left(\mathrm{Y},\left[\mathrm{A},\mathrm{B}\right]\right)-\mathrm{reg}\left(\mathrm{Y},\mathrm{A}|\mathrm{B}\right)-\mathrm{reg}\left(\mathrm{Y},\mathrm{B}|\mathrm{A}\right)$$ (1)

Figure 4.

Activation strength (AS) maps for 2‐back versus 0‐back (Panel a as Contrast 1) and 0‐back versus baseline (Panel b as Contrast 2), with a cluster‐based threshold of Z = ±2.32 and p = 0.05 in the group average effects (328 participants). The regional indicators in Figure 1 were mapped to the activation maps (gold markers in Panel a). Abbreviations: caudal SFS = caudal superior frontal sulcus; DLPFC = dorsal lateral PFC; dorsal ACC = dorsal anterior cingulate cortex; FEF = frontal eye field; FFG = fusiform gyrus; IPL/IPS = intraparietal lobe/sulcus; MPFC = medial PFC; PCC = posterior cingulate cortex; SMA = somatosensory area [Color figure can be viewed at http://wileyonlinelibrary.com]

It should be noted that SASS is the same as ASS within one space, namely, reg(Y, A) = reg(Y, [A, A]), and these are the diagonal elements of the SASS matrix (Panel 4 in Figure 3).

3. RESULTS

3.1. AS maps during WM task execution

AS maps of the two contrasts, that is, 2‐back versus 0‐back (Contrast 1) and 0‐back versus Baseline (Contrast 2), were investigated and presented here with a cluster‐based threshold of Z = ±2.32 and p = 0.05 (Figure 4). Mapping from the volume image to the surface and visualization were performed using the adapted SurfStat package (http://www.math.mcgill.ca/keith/surfstat).

For the higher cognitive load (2‐back vs. 0‐back), the activated regions included the lateral middle frontal cortex, intraparietal sulcus (IPS), inferior parietal lobe (IPL), and dorsal prefrontal cortex (PFC), and the deactivated regions included the sensory cortex and medial PFC. For the lower cognitive load (0‐back vs. baseline), the activated regions included the posterior lateral frontal cortex, anterior insular cortex, superior frontal gyrus (medial), inferior frontal junction, visual cortex, and dorsal eye saccade control network (including the frontal eye field [FEF], IPS, and posterior eye field) (Pierrot‐Deseilligny, Milea, & Muri, 2004), which contain most of the FPN and DAN; the deactivated regions included the posterior cingulate cortex (PCC), anterior PFC (medial DMN), sensory cortex, posterior central gyrus, and posterior insular cortex.

3.2. Voxel‐wise ASS maps and comparisons with AS maps

First, we visually examined which brain regions predominantly contributed to predicting WM accuracy. To this end, Pearson correlation coefficients (r, threshold of |r| > 0.1 with p < 0.05) for voxels were mapped to the surface without thresholds, and colored blue‐purple for negative correlations and yellow‐red for positive correlations (Figure 5). We preserved the sign of the correlation coefficient to indicate direction; however, the absolute value of the correlation coefficient was related to the AV regardless of the sign. Results showed that, for Contrast 1 (higher load), the regions with positive ASS were located in the FPN (including the lateral middle PFC [LMPFC] and IPL), salience network (including the anterior insular cortex and dorsal MPFC), and DAN (including the FEF), whereas the regions with negative ASS were predominantly located in the DMN (including the PCC, MPFC, and superior PFC). For Contrast 2 (lower load), the regions with positive ASS included the sensory cortex, visual cortex, anterior and ventral PFC, medial temporal cortex, and left anterior inferior frontal gyrus, whereas the regions with negative ASS included the LMPFC and dorsal MPFC.

Figure 5.

Voxel‐wise association strength (ASS) maps for Contrast 1 (2‐back vs. 0‐back, Panel a) and Contrast 2 (0‐back vs. baseline, Panel b). Pearson correlation coefficients (r, threshold of |r| > 0.1 with p < 0.05) for working memory (WM) accuracy across participants are blue‐purple for negative correlations and yellow‐red for positive correlations [Color figure can be viewed at http://wileyonlinelibrary.com]

We further examined the overlaps between the AS and ASS maps. The overlaps were measured in two ways, that is, at the overall level and level after significance thresholding. The overall level was directly calculated as the overlapping region between AS and ASS in Contrasts 1 and 2, respectively. For the other level, we first thresholded the maps by their significance thresholds, |Z| > 2.32 for AS maps and p < 0.05 for ASS maps, and then calculated the overlapping regions. There were large overlapping regions between the original maps for both contrasts. Furthermore, large regions remained after significance thresholding for Contrast 1 (Figure 6b), positively located in the FPN (including the LMPFC and IPL), salience network (including the anterior insular cortex and dorsal MPFC), and DAN (including the FEF), and negatively located in the DMN (including the PCC, medial MPFC, and superior PFC). However, only sparse regions remained after the significance thresholding of Contrast 2 (Figure 6d), located mainly in the sensory cortex and medial temporal cortex. Conversely, the anterior insular cortex and lateral PFC showed significant positive activation but negative association.

Figure 6.

Overlap between activation (AS) and association strength (ASS) maps (for positive and negative pairs, respectively). Panels a and b are overlap regions for Contrast 1 (2‐back vs. 0‐back) and Panels c and d are overlap regions for Contrast 2 (0‐back vs. baseline). Two levels were used for computing the overlaps, Panels a and c are overlaps between the original AS and ASS, and Panels b and d are overlaps between AS and ASS after separate thresholding, |Z| > 2.32 for AS maps and p < 0.05 (no correction for multiple comparisons) for ASS maps (“Sig.” is significance in the legend title in Panels b and d). Blue indicates overlaps for negative regions (between negative activations and associations) and red indicates overlaps for positive regions (between positive activations and associations) [Color figure can be viewed at http://wileyonlinelibrary.com]

3.3. SASS between networks

The ASS and SASS were examined based on Equation 1. Before concentrating on each network, we first looked at the ASS of the global brain after partition sampling (Power et al., 2011) (264 nodes in total), which was estimated by multiple regression of behavioral scores with the AS of the overall 264 nodes as variables. Our results demonstrated a significant portion of the AV in the regression of both contrasts, that is, R ² = 0.78, F = 1.57, p = 0.0028 for Contrast 1 and R ² = 0.76, F = 1.41, p = 0.0164 for Contrast 2.

The SASS between network pairs was then examined. For Contrast 1, the DMN retained strong associations and comprehensive shared associations with other networks; additionally, other networks, including the visual cortex, FPN, and DAN, showed strong associations and comprehensive shared associations with other networks across the brain. For Contrast 2, the DMN showed strong associations and comprehensive shared associations with other networks. Another local community was composed of the FPN and salience network, which also showed strong ASS for the nodes within the FPN and strong shared associations between the FPN and other networks. Of note, the SASS was not normalized according to network size because, for the regression defined in Equation 1, only the noncollinear vectors (namely, heterogeneous functions) were able to independently contribute to the regression, indicating that reg(Y, A) produced an identical output as reg(Y, [A, 2 × A]), although [A, 2 × A] contained two times more vectors than [A]. In the 10 predefined networks of Power et al. (2011), the visual cortex, FPN, salience network, and DAN contained 31, 25, 18, and 11 nodes, respectively; however, for Contrast 2 (Figure 7b), the FPN showed stronger ASS than the visual cortex, whereas for Contrast 1, the DAN showed stronger ASS than the salience network (0.14 for the DAN and 0.11 for the salience network in the diagonal values in Figure 7a).

Figure 7.

Shared‐association strength (SASS, defined in Equation 1) comparisons between networks. Panels a and b are for higher load (2‐back vs. 0‐back) and lower load (0‐back vs. baseline), respectively. AV is the addressed variance in regression [Color figure can be viewed at http://wileyonlinelibrary.com]

Furthermore, we quantitatively compared the ASS and SASS between each network and the remaining nine networks and two contrasts (2‐back vs. 0‐back and 0‐back vs. Baseline). This can be achieved by the aggregation of each row (or column) of the SASS matrix separately for the two contrasts (Figure 7); however, this can result in deterioration due to the collinear vectors from different networks (as described above). Thus, we again used the semipartial correlation defined in Equation 1 to compute the within ASS and between network SASS. For ASS comparisons between networks (Panels a and c in Figure 8), traditional multiple regression was used to predict WM accuracy, whereas for SASS comparisons (Panels b and d in Figure 8), the residual of one network after regressing out all other networks was used to predict WM accuracy. The error bars in Figure 8 depict the variance of regression after 5,000 bootstraps (80% of participants were used for each sampling). Across the ASS and SASS comparisons in the two contrasts, the DMN and FPN played predominant roles. It should be noted that the DAN showed more active functions in Contrast 1 (higher load) than in Contrast 2 (lower load) for both ASS (Figure 8a) and SASS (Figure 8b, where the SASS of the DAN was significantly stronger than that of the other networks, except for the DMN and FPN (but comparable to the FPN), with p < 0.05, corrected for multiple comparisons by FDR). For a closer look at the association differences, the relative AV ratio between the two contrasts, separately for ASS and SASS, was computed by (C2–C1)/C2, where C1 is Contrast 1 (higher load) and C2 (lower load) is Contrast 2. Results showed that the DAN exhibited more significant (p < 0.05, corrected for multiple comparisons by FDR) participation enhancement for both ASS and SASS in Contrast 1 compared to that in Contrast 2 (Figure 9).

Figure 8.

Comparisons of network association strength (ASS) (Panels a and c) and shared‐ASS (SASS) (Panels b and d). Panels a and b are for Contrast 1 (2‐back vs. 0‐back) and Panels c and d are for Contrast 2 (0‐back vs. baseline). Bar height is the addressed variance (AV) for all 328 participants, and error bars indicate variance of regression with 5,000 bootstraps (80% of participants were used for each sample) [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 9.

Comparisons of the relative addressed variance (AV) ratios between two contrasts, separately for network association strength (ASS) (Panel a) and shared‐ASS (SASS) (Panel b). The ratio was computed by (C1–C2)/C2, with the error bars derived from 5,000 bootstraps in Figure 8, and where C1 is Contrast 1 (2‐back vs. 0‐back, higher load) and C2 is Contrast 2 (0‐back vs. baseline, lower load) [Color figure can be viewed at http://wileyonlinelibrary.com]

4. DISCUSSION

In this study, we proposed a framework to discriminate the behaviors of functional nodes and networks based on their ASS and SASS in predicting WM accuracy (2‐back and 0‐back). WM‐related networks, including the FPN, DAN, and salience network, showed significant activation and at the same time positively modulated cognitive performance under higher cognitive load (2‐back), but not under lower cognitive load (0‐back, also known as attentional load). Furthermore, in both contrasts, the DMN and FPN consistently showed significant ASS for intranetwork (i.e., ASS from network nodes) and internetwork (i.e., SASS), whereas the DAN showed significantly enhanced ASS and SASS under higher cognitive load (2‐back). This study provides a comprehensive characterization for how brain regions function to modulate performance across different cognitive loads.

4.1. WM‐related networks positively modulate performance under higher cognitive load

For both contrasts (2‐back vs. 0‐back and 0‐back vs. baseline) in the WM test, the medial DMN, including the PCC and MPFC, was consistently deactivated at a significant level (Z < −2.32); conversely, the lateral PFC, FEF, IPL, anterior insular cortex, and superior MPFC (anterior part of the presupplementary motor areas [e.g., SMA]) were consistently activated at a significant level (Z > 2.32) (Figure 4). The whole activation/deactivation picture coincides well with research by Barch et al. (2013), which demonstrated that the HCP‐customized task protocol efficiently recruits the visuospatial WM processing‐related brain networks.

The DMN is a coordinator of brain‐wide computing resources (Anticevic et al., 2012; Liu et al., 2017) mainly through its deactivation (Fox et al., 2005; Raichle et al., 2001). The deactivation of the DMN in WM task execution has been examined extensively (Koshino, Minamoto, Yaoi, Osaka, & Osaka, 2014; Mayer, Roebroeck, Maurer, & Linden, 2010), and our results indicated that the DMN was actively engaged in the actions of task‐positive networks (Spreng, Sepulcre, Turner, Stevens, & Schacter, 2013; Spreng, Stevens, Chamberlain, Gilmore, & Schacter, 2010). The engagement of the DMN in WM tasks has also been revealed by functional connectivity and brain network modeling (Liang, Zou, He, & Yang, 2016; Vatansever et al., 2015; Yamashita, Kawato, & Imamizu, 2015). Our results further showed that under a higher WM cognitive load (2‐back vs. 0‐back), greater deactivation of the DMN signified better WM performance (accuracy in 2‐back) (see Figure 4a where DMN is blue‐coded), consistent with the existing “less is more” perspective (Anticevic, Repovs, Shulman, & Barch, 2010). However, the correlation between DMN deactivation and performance under a lower cognitive load (0‐back task) did not exist consistently. This finding further supported that the DMN was suppressed (usually unrelated signals to the directed task) during WM execution to enhance the efficiency of the task‐positive networks (Amer et al., 2016; Wen et al., 2013; Zhou et al., 2018).

Although the lateral PFC (lateral middle frontal gyrus and inferior frontal sulcus), FEF, IPL, and anterior insular cortex were significantly positively activated in both contrasts, the AS of these task‐positive networks exhibited significant negative correlations with WM performance in 0‐back tasks but positive correlations in 2‐back tasks (Figure 5). These opposing actions confirmed that the task‐positive networks were predominantly responsible for holding information in the mind and retrieving restored memory to respond to external stimuli in the WM test (2‐back) (Diamond, 2013; Owen, McMillan, Laird, & Bullmore, 2005), whereas under control conditions (0‐back as lower cognitive load, mainly in attentional processes), the lower activations signified better performance (namely, less activation equated to better performance) (Liu et al., 2017; Zuo et al., 2018b). This finding accords with previously proposed functions of WM‐related regions, including the frontal–parietal regions, superior medial frontal gyrus (dorsal anterior cingulate cortex which also extends to the pre‐SMA), and anterior insular cortex, which cover the predominant nodes of the FPN, DAN, and salience network (Christophel et al., 2017; Cole et al., 2014; D'Esposito & Postle, 2015; Ptak, 2012).

The sensory‐related cortices, including the somatosensory cortex and visual cortex, all showed significant activations and positive associations in the 0‐back test, which were largely attenuated in the 2‐back test. In the 0‐back test, the medial part of the temporal lobe, particularly the fusiform gyrus (FFG), consistently showed significant activation and ASS for WM accuracy (see Figure 6). This is largely consistent with the perceptional process in the 0‐back task because, apart from the basic information input cortices (including the sensory and visual cortices), the FFG also exhibits versatile visual/semantic processing capability (Zhang et al., 2016), strong connectivity with multimodal sensory integration cortices (e.g., temporal pole cortex and parahippocampal gyrus (Fan et al., 2014; Zuo et al., 2016)), and particular preference for visual information processing (Binney, Parker, & Lambon Ralph, 2012).

4.2. DMN, FPN, and DAN globally modulate brain functions across cognitive loads

The ASS and SASS analyses highlighted the important roles of the DMN and FPN in both ASS to WM accuracy and active interactions with other networks (Figures 7 and 8). Based on the functional connectivity of fMRI time series, existing reports have also revealed that the DMN and FPN are the main hub regions of the WM network, and functional connectivities from these networks show significant contributions in predicting WM performance (Liang et al., 2016; Liu et al., 2017; Pagnoni, 2012). Collectively, these results confirm the predominant roles of the DMN and FPN from the perspective of both functional activation and functional connectivity. Furthermore, by decomposing the connectivity into connectional strength and connectional connectivity (namely, nodal connectivity diversity), we unveiled two different properties of the DMN and FPN under different cognitive loads (2‐back and 0‐back), that is, flexibility and strength, respectively (Zuo et al., 2018b). These results agree with our findings that the DMN and the FPN simultaneously showed strong ASS and SASS with other networks, signifying that these two networks globally modulated the functions of the brain in the WM test (Cassidy et al., 2016; Douw, Wakeman, Tanaka, Liu, & Stufflebeam, 2016; Godwin, Ji, Kandala, & Mamah, 2017).

Interestingly, the DAN displayed different behaviors in the two contrasts. Compared with Contrast 2 (0‐back vs. Baseline), the DAN exhibited both ASS and SASS with other networks in Contrast 1 (2‐back vs. 0‐back). In particular, the SASS of the DAN across the entire brain was only lower than that of the DMN and was significantly greater (p < 0.05 after 5,000 bootstraps) than that of the other networks except the FPN (Figure 8b). (For the results based on 385 subjects retaining twin structure, the SASS of the DAN was significantly greater [p = 7.16e‐22] than that of the FPN, see Figure S8, Supporting Information.) Furthermore, for the relative AV ratio, the DAN exhibited more significant (p < 0.05, corrected for multiple comparisons) enhanced participation for both ASS and SASS in Contrast 1 than that in Contrast 2 (Figure 9). Previous study has disclosed that the DAN is a goal‐directed cognitive control network that top‐down regulates responses to external stimuli (Corbetta & Shulman, 2002). Recent studies have also reported that the entire FPN is composed of three heterogeneous modules (namely, DMN aligned, DAN aligned, and dual aligned) that dissociably mediate global communication (Dixon et al., 2018; Spreng et al., 2013). This accords with the concept that the DAN globally regulates the positive networks by overseeing visuospatial encoding/planning (Spreng et al., 2010; Stawarczyk, Jeunehomme, & D'Argembeau, 2018), episodic retrieval, and WM (Luckmann, Jacobs, & Sack, 2014; Nee & Jonides, 2014). This function of the DAN has also been confirmed in studies comparing normal controls with participants exhibiting spatial neglect (Ptak & Schnider, 2010) and aging (Amer et al., 2016; Kurth et al., 2016). Thus, the current study further clarified that the mediation function of the DAN is much more remarkable under higher cognitive loads (e.g., 2‐back) than in the 0‐back task.

The frontal–parietal system, which mainly consists of the FPN and DAN, is the core system for WM, and these two networks are separately but cohesively located across the middle and dorsal frontal–parietal lobe (Nee et al., 2013; Vossel, Geng, & Fink, 2014). Although both exert top‐down control in cognitive processes, they show process‐specific roles in WM. The FPN is in charge of information representation, rehearsal, and integration for higher and conceptual level cognition, whereas the DAN cooperates with stimulus perceptional regions (e.g., the visual cortex) and the FPN to guarantee directed information sustainment with high priority (Ptak, 2012). It should be noted that the DAN is the key node along the pathway involved in saccade control (Pierrot‐Deseilligny et al., 2004). Thus, our finding that the DAN sharply increased its interaction with other networks under higher cognitive load (2‐back) compared with that under lower load (0‐back) is reasonable. It has also been reported that, in WM tests after intensive training, the DAN maintains active activation (correlated to performance) under higher load, whereas the FPN dissociates this relationship (Thompson et al., 2016).

In our study, the action of the salience network in the association profile (both ASS and SASS) was cohesively homogenous to the FPN (Figure 7), in agreement with our previous study, which demonstrated that the FPN and salience network exhibit homogenous patterns from the perspective of functional flexible reconfiguration (Zuo et al., 2018b). The salience network, as a key node of the triple network theory (along with the FPN and DMN) (Menon, 2011), is responsible for detecting salient features (i.e., those with higher priority, as proposed by Ptak (2012)) and acts as an integral hub in mediating interactions between externally oriented attention (e.g., by the DAN and FPN) and internally oriented self‐related mental processes (e.g., by the DMN) (Menon, 2011; Seeley et al., 2007). Similar to the FPN, it also positively regulates task‐positive networks but negatively regulates the DMN (Chen, Cai, Ryali, Supekar, & Menon, 2016; Zhou et al., 2018).

4.3. Future directions to extend the current study

We have characterized the action and interaction patterns between brain networks modulating performance across cognitive loads. However, questions remain regarding these action and interaction patterns. In recent years, there has been a shift in imaging study of WM from region‐based to network‐based approaches (Courtney, 2015; Eriksson et al., 2015; Ester, Sprague, & Serences, 2015). Some cognitive models, for example, those proposed by Baddeley (2012) and Shura, Hurley, and Taber (2016), emphasize that the individual functional module usually extensively recruits distributed regions in a certain stage of WM process (Baddeley, 2012; Christophel et al., 2017). To this end, the proposed shared‐association method is a process‐based strategy (Courtney, 2015; Eriksson et al., 2015; Hasson, Chen, & Honey, 2015) and can characterize how different regions coherently process information in the same dimension to collaboratively guide performance. Future study should focus on how to classify the regions according to their processing patterns and how these patterns could predict individual WM performance.

4.4. Methodological considerations

Here, the proposed semipartial correlation‐based algorithm described the pair‐wise similarity of functions between brain regions. It can be applied between fMRI time series or regional coherence across subjects, as used in the current study, or to examine other cognitive tasks. It should be noted that semipartial correlation can potentially deteriorate due to suppression effects (Nimon, 2010; Ray‐Mukherjee et al., 2014), which will, for a certain vector space, yield greater SASS than ASS (e.g., when computing SASS between A and B in predicting Y, A, and B share a subspace that is perpendicular to Y). In the current study, the results did not deteriorate by suppression as the SASS (Figure 8b,d) did not surpass the ASS (Figure 8a,c). For further discussion of our study, the MATLAB code is available from Github, https://github.com/nmzuo/shared_prediction.

CONFLICT OF INTEREST

The authors declare no potential conflict of interest.

Supporting information

Appendix S1 Supporting Information

Zuo N, Salami A, Yang Y, Yang Z, Sui J, Jiang T. Activation‐based association profiles differentiate network roles across cognitive loads. Hum Brain Mapp. 2019;40:2800–2812. 10.1002/hbm.24561