Abstract

Individual differences in working memory relate to performance differences in general cognitive ability. The neural bases of such individual differences, however, remain poorly understood. Here, using a data-driven technique known as connectome-based predictive modeling, we built models to predict individual working memory performance from whole-brain functional connectivity patterns. Using n-back or rest data from the Human Connectome Project, connectome-based predictive models significantly predicted novel individuals' 2-back accuracy. Model predictions also correlated with measures of fluid intelligence and, with less strength, sustained attention. Separate fluid intelligence models predicted working memory score, as did sustained attention models, again with less strength. Anatomical feature analysis revealed significant overlap between working memory and fluid intelligence models, particularly in utilization of prefrontal and parietal regions, and less overlap in predictive features between working memory and sustained attention models. Furthermore, showing the generality of these models, the working memory model developed from Human Connectome Project data generalized to predict memory in an independent data set of 157 older adults (mean age = 69 years; 48 healthy, 54 amnestic mild cognitive impairment, 55 Alzheimer disease). The present results demonstrate that distributed functional connectivity patterns predict individual variation in working memory capability across the adult life span, correlating with constructs including fluid intelligence and sustained attention.

INTRODUCTION

Working memory (WM) plays a critical role in our daily lives. From reading a news article to remembering the items we need at the grocery store, our capacity to temporarily store and manipulate information facilitates our interactions with the world around us and guides our behavior. In this role, WM has been deemed the “workbench of cognition” (Klatzky, 1975), and as such, WM capability, its relation to other cognitive skills, and its variation between individuals have been of particular interest (D'Esposito & Postle, 2015; Luck & Vogel, 2013; Baddeley, 2003).

Individual differences in WM relate to performance differences in general cognitive ability (Cowan et al., 2005; Kane & Engle, 2002; Kyllonen & Christal, 1990). WM capacity, for example, accounts for 40% of variance in global fluid intelligence (gF; Fukuda, Vogel, Mayr, & Awh, 2010). This strong relationship is thought to reflect shared neural systems, particularly involving prefrontal and parietal regions, which allow us to keep a representation active in the face of interference (Engle, Tuholski, Laughlin, & Conway, 1999; Prabhakaran, Smith, Desmond, Glover, & Gabrieli, 1997). Relationships between WM and controlled attention have also been of particular interest (Chun, Golomb, & Turk-Browne, 2011), and the ability to control attention has been proposed to play a major role in complex WM tasks (Barrett, Tugade, & Engle, 2004; Kane, Bleckley, Conway, & Engle, 2001). For example, the efficiency of information flow between attentional and perceptual neural systems may impact WM precision performance, though not WM capacity (Galeano Weber, Hahn, Hilger, & Fiebach, 2017).

Deficits in WM affect a wide range of cognitive abilities and are common to a range of neurological disorders, including schizophrenia, attention-deficit disorder, and reading disabilities (Luck & Vogel, 2013). The question of what drives such individual differences in WM, however, requires more study. WM has classically been described as a limited capacity workspace (Baddeley & Hitch, 1974), wherein individual differences in storage capacity and the ability to use this capacity efficiently account for individual variation in WM capability (Luck & Vogel, 1997; Just & Carpenter, 1992). More recent work has expanded on this foundation, with leading cognitive psychological theories positing that capacity arises from a fixed number of memory slots (Luck & Vogel, 2013) or a fixed amount of attentional resources (Ma, Husain, & Bays, 2014).

Recent fMRI studies have characterized the neural representations underlying models of WM, with multivariate pattern analysis techniques showing that WM representations are sustained in regions including prefrontal, parietal, and sensory cortices (Sreenivasan, Vytlacil, & D'Esposito, 2014; Han, Berg, Oh, Samaras, & Leung, 2013; Christophel, Hebart, & Haynes, 2012). The posterior intraparietal sulcus (IPS), in particular, has been strongly linked to variation in WM capacity: IPS activity scales with WM load until capacity is reached (Xu & Chun, 2006; Todd & Marois, 2004), and functional connectivity analyses have implicated the IPS as a central hub in networks of WM-associated regions (Palva, Monto, Kulashekhar, & Palva, 2010). More generally, work has emphasized the importance of functional connections between distributed brain regions, including frontal, parietal, premotor, and occipitotemporal regions, in mediating WM (Liu et al., 2018; Yamashita et al., 2018; Liebe, Hoerzer, Logothetis, & Rainer, 2012; Fiebach, Rissman, & D'Esposito, 2006; Gazzaley, Rissman, & D'Esposito, 2004), with recent work implicating large-scale, distributed networks in WM capacity (Soreq, Leech, & Hampshire, 2019; Cohen & D'Esposito, 2016; Miller & Buschman, 2015; Luck & Vogel, 2013). In this vein, the functional networks and anatomical features giving rise to WM processes remains a topic of interest, including their relationship to WM deterioration in older age.

In addition to the neuroimaging studies that have focused on neural correlates of WM, related work has investigated individual differences in this cognitive ability (Todd & Marois, 2005). Such differences have been traditionally measured by performance on cognitive tasks, such as verbal repetition span tasks and continuous recognition tasks (Redick et al., 2012; Kane, Conway, Miura, & Colflesh, 2007), with more recent work relating performance on such tasks to individual differences in brain function (Bertolero, Yeo, Bassett, & D'Esposito, 2018; Luck & Vogel, 2013). Here, we propose a novel, brain-based measure of individual WM performance, which goes beyond former work in being applicable to novel participants and directly comparable to models of other cognitive and attentional abilities. Employing a data-driven technique known as connectome-based predictive modeling (CPM; Shen et al., 2017; Rosenberg, Zhang, et al., 2016; Finn et al., 2015), we built models predicting individual WM performance from whole-brain functional connectivity patterns observed in participants of the Human Connectome Project (HCP) as they completed a WM (n-back) task or rested. This approach complements a growing body of work predicting outcomes and behavior from large-scale functional connectivity patterns (e.g., Gong et al., 2019; Smith et al., 2015; see Woo, Chang, Lindquist, & Wager, 2017, and Dubois & Adolphs, 2016, for recent reviews).

Using 10-fold cross-validation in the HCP data set, CPMs built from task-based or resting-state fMRI data successfully predicted previously unseen individuals' 2-back accuracy. Distinct models predicted measures of attention and, replicating previous work, gF (Finn et al., 2015), and comparisons of these models revealed relationships between WM, gF, and, to a lesser degree, sustained attention. Furthermore, the same CPMs trained to predict 2-back accuracy the HCP sample generalized to predict individual differences in memory in an independent sample of older adults, including patients with amnestic mild cognitive impairment (aMCI) and with Alzheimer disease (AD), recruited from the Samsung Medical Center (SMC) in Seoul, South Korea. Thus, we provide a novel demonstration of the use of distributed patterns of functional connectivity to measure individual differences in WM, correlating with constructs such as gF and sustained attention. We also characterize commonalities and differences in the functional anatomy of brain networks that underlie WM, gF, and sustained attention.

METHODS

This study used two independent fMRI data sets.

Data Set 1: HCP S900 Release

Data were obtained from the HCP (Van Essen et al., 2013) 900-participant release of December 2015. From the available data from multiple neuroimaging modalities, we utilized fMRI data and performance on behavioral cognitive tasks. Scan parameter details can be found in Uğurbil et al. (2013). Of the 900 participants (aged 22–35 years), we identified the 515 who completed all resting and task fMRI runs with an average frame-to-frame motion of <0.10 mm and a maximum frame-to-frame motion of <0.15 mm (Greene, Gao, Scheinost, & Constable, 2018). The mean frame-to-frame motion for each condition (e.g., n-back task) was calculated by averaging the mean motion for corresponding left/right and right/left phase-encoding runs. Of these 515 participants, we analyzed the 502 (mean age = 28 ± 3.6 years, 274 women) who did not have any missing functional atlas nodes (see the Connectome-Based Predictive Modeling section) or time points. Each participant provided written consent in accordance with the institutional review boards of Washington University in Saint Louis, University of Minnesota, Oxford University, Saint Louis University, Indiana University, University d'Annunzio, Ernst Strungmann Institute, Warwick University, Radboud University Nijmegen, and Duke University.

Connectivity matrices were created as described in Finn et al. (2015). From minimally processed HCP data (Glasser et al., 2013), further preprocessing was completed using BioImage Suite and included the regression of 12 motion parameters and global, white matter, and cerebral spinal fluid signal, as well as removal of linear trend and low-pass filtering (Finn et al., 2015). Task-evoked activity was not regressed. After preprocessing, the whole brain was parcellated by applying a 268-node whole-brain gray matter atlas spanning cortical, subcortical, and cerebellar regions to define functional network nodes (Shen, Tokoglu, Papademetris, & Constable, 2013). The fMRI signal time course in each node was averaged, and the Pearson correlation between the time courses of each pair of nodes was then calculated. Pearson r values were converted to z values using the Fisher transform to produce a 268 × 268 matrix of functional connection (edge) strengths for each participant.

Functional connectivity matrices were derived from resting-state fMRI data (four runs of 15 min each) and WM task data (n-back; two runs of 5 min each). Full details of task and rest scan procedures can be found in Barch et al. (2013) and Smith et al. (2013), respectively. Although all participants completed a total of seven different task fMRI scans targeting specific cognitive functions, we selected the WM task fMRI scan for our analyses: In addition to directly targeting our construct of focus, this 10-min n-back (0-back and 2-back) task also provided more data than any of the other six tasks, which ranged from 2.5 to 4 min in length. We also chose to include resting-state fMRI data because previous work has indicated that CPM is successful even when applied to data collected in the absence of an explicit task challenge (Rosenberg, Finn, et al., 2016; Rosenberg, Zhang, et al., 2016; Finn et al., 2015). Connectome-based models were trained to predict participants' WM performance, operationalized as 2-back task accuracy, from n-back and resting-state functional connectivity patterns.

In consideration of the known mechanistic and behavioral relationships between WM and cognitive abilities such as gF and controlled attention (Barrett et al., 2004; Kane & Engle, 2002), we also assessed the degree to which our models of WM, gF, and attention reflected this overlap. To test whether models built to predict performance on the 2-back task generalized to predict other measures of cognitive task performance, we correlated model predictions from task-based and resting-state data with measures of gF and sustained attention. gF was quantified by the number of correct responses on Penn's progressive matrices (PMAT), an increasingly difficult pattern-matching task during which participants were instructed to select the missing piece of a puzzle containing visual patterns (Moore, Reise, Gur, Hakonarson, & Gur, 2015). Although gF and WM are related (Fukuda et al., 2010; Colom, Flores-Mendoza, & Rebollo, 2002), the PMAT does not explicitly recruit or assess WM capability. Sustained attention was quantified with d′ (sensitivity) on the Short Penn Continuous Performance task (SCPT), during which participants responded to displays of seven line segments whenever they formed a number or a letter (Kurtz, Ragland, Bilker, Gur, & Gur, 2001). Although attention is also required for n-back and PMAT task completion, as it is for nearly every task in daily life (Rosenberg, Zhang, et al., 2016), we utilized 2-back accuracy, PMAT score, and SCPT d′ as commonly used indications of a participant's WM, gF, or attentional ability, respectively. We report the correlations between these observed behavioral scores of gF, attention, and WM in Supplementary Table S3.1

Data Set 2: SMC

Connectome-based models of WM defined in the full HCP sample were externally validated with an independent data set collected at the SMC in Seoul, South Korea. Models applied to the SMC data set were trained using data from the full HCP sample to maximize the number of training observations. Between March 2008 and February 2009, fMRI data and neuropsychological assessments (Seoul Neuropsychological Screening Battery) were collected from 278 individuals classified into three levels of memory function: AD patients (n = 112), aMCI patients (n = 87), and healthy controls (n = 79). Full information regarding patients' medical histories and diagnostic criteria as well as neuropsychological assessment and fMRI scan procedures are detailed in Ahn et al. (2010) and Kim et al. (2015). Of the 278 patients, aged 44–90 years, we identified the 157 (mean age = 68.7 ± 9.6 years; 105 women; 55 AD, 54 aMCI, 48 healthy) who completed all neuropsychological assessments and fMRI runs with an average frame-to-frame motion of <0.2 mm, a maximum translation of <2 mm, and a maximum rotational motion of <3°. These motion thresholds were less stringent than those applied to our HCP sample in consideration of the fact that older adults, on average, tend to exhibit higher motion levels during fMRI (Fountain-Zaragoza, Samimy, Rosenberg, & Prakash, 2019; Geerligs, Tsvetanov, Cam-CAN, & Henson 2017; Van Dijk, Sabuncu, & Buckner, 2012; D'Esposito, Zarahn, Aguirre, & Rypma, 1999). Recent work utilizing older adult fMRI data has thus tended to use motion thresholds that are less stringent than those typically applied to young adult samples (Lin et al., 2018; Guo et al., 2012; Ystad, Eichele, Lundervold, & Lundervold, 2010).

Resting-state fMRI preprocessing, including motion correction and regression of nuisance covariates, was performed using SPM12 and custom scripts (MATLAB R2016b). Nuisance covariates included 24 motion-related parameters (6 translational and rotational motions, 6 derivatives of these motions, and 12 squares of these motions), 3 mean signals (global, white matter, and cerebrospinal fluid), and linear and quadratic trends. Functional connectivity matrices were created by calculating the Pearson correlation between each of 268 nodes of the whole brain (Shen et al., 2013, whole-brain gray matter atlas). Pearson r values were converted to z values using the Fisher transform to produce a 268 × 268 connectivity matrix of edge strengths for each participant. Each participant provided written informed consent in accordance with the institutional review board of the SMC.

Connectome-based Predictive Modeling

CPM was used to construct models of WM following the procedures of Finn et al. (2015) and Rosenberg, Finn, et al. (2016), with a detailed protocol and scripts available in Shen et al. (2017).

First, using the HCP data set, participants' functional connectivity and behavioral data were split into 10 groups. Throughout 10-fold cross-validation, nine of these groups constituted a training data set, and the remaining group was designated as a testing data set. As some of the 502 participants in our HCP sample were related (194 sharing the same mother, 184 sharing the same father, 123 twins), group assignments were made with regard to familial relatedness between participants such that related participants were always placed in the same groups (i.e., models were never trained on data from one family member and tested on data from another, which could compromise the independence of the training and test sets).

Using data from the nine groups of the training set, a given edge (i.e., functional connection) in the n-back or resting-state functional connectivity matrices across all training set participants was Pearson correlated with 2-back accuracy scores for those participants, and this process was repeated for every edge to identify the functional connections most strongly related to accuracy. A significance threshold of p < .01 was applied to select edges that were positively and negatively correlated with behavior across individuals. A relationship was found between individual motion and behavioral task scores (see Supplementary Table S1), and we thus controlled for motion at this edge selection step using partial Pearson correlation. Following selection of significant edges, the strength of these retained edges was then averaged for each participant to generate two summary statistics per individual: positive network strength (the mean edge value in the set of edges positively correlated with behavior) and negative network strength (the mean edge value in the set of edges negatively correlated with behavior). Linear models were then constructed to relate 2-back accuracy scores to network strength in the training sample, and these models were then applied to left-out individuals' functional connectivity data to predict their task performance. This process was repeated 1000 times with random assignment of participants into folds so as to provide a reliable statistic of average model performance. Model performance was assessed at each iteration by the strength of Pearson correlation between predicted and observed scores for each novel individual. The average of the resulting 1000 correlation coefficients was taken to represent average model performance.

To assess the significance of the observed correlation between predicted and observed scores, we performed 1000 additional iterations of 10-fold cross-validation using shuffled behavioral scores. This permutation analysis using mismatched brain and behavioral data allowed us to create a null model performance distribution. The significance of true model performance was assessed by comparing the distributions of correlation coefficients generated by the true and null models using the equation: p(r) = (1 + number of null r values ≥ observed r value)/(1 + 1000). The effect size of the difference between null and true distribution means was also measured with Cohen's d to further characterize model performance.

To compare the features and performance of WM models with those of other related cognitive abilities, we repeated the CPM procedure described above twice more, once to predict PMAT scores and once to predict SCPT performance. The predictions of gF and sustained attention CPMs from both task and resting-state data were correlated with observed 2-back accuracy values to assess their ability to predict WM task performance.

Cross-predictive power was expected and indeed observed between WM and gF models. As such overlap is thought to reflect similar, but not identical, neural underpinnings of WM and gF (Kane & Engle, 2002), we sought to isolate these cognitive abilities and their respective networks from each other to assess the degree to which they may be distinct. To this end, we applied CPM to predict 2-back task accuracy from functional connectivity matrices while controlling for gF at the edge selection step. We constructed predictive networks as described above but selected only the edges that were correlated with 2-back accuracy values and were not correlated with PMAT scores (threshold p <.01). The features and predictive power of these models were compared with those of the original WM models described above. We did not repeat this procedure to remove edges correlated with attention (i.e., SCPT task performance) because the observed correlation between WM and sustained attention performance was weaker (see Table 1).

Table 1. 
Summary of Internal and External Performance of WM, gF, and Attention CPMs
Analysis #Training Behavior, Training DataControlled CovariateTest Behavior, Test DataModel Performance Validation TypeData Type
2-Back accuracy, Task   2-Back accuracy, Task r = .36, p = 1/1001 * 10-Fold cross-validation Internal data 
    SMC composite memory, Rest r = .37, p = 1/1001 * External validation External data, different behavior 
    PMAT score, Task r = .28, p = 1/1001 * 10-Fold cross-validation Internal data, different behavior 
    SCPT d′, Task r = .09, p = 7.19E−2   10-Fold cross-validation Internal data, different behavior 
  
2-back accuracy, Task gF** 2-Back accuracy, Task r = .33, p = 1/1001 * 10-Fold cross-validation Internal data 
  
2-back accuracy, Rest   2-Back accuracy, Rest r = .20, p = 1/1001 * 10-Fold cross-validation Internal data 
    SMC composite memory, Rest r = .30, p = 1/1001 * External validation External data, different behavior 
    PMAT score, Rest r = .13, p = 2.00E−3 * 10-Fold cross-validation Internal data, different behavior 
    SCPT d′, Rest r = .06, p = 6.29E−2   10-Fold cross-validation Internal data, different behavior 
  
10 2-back accuracy, Rest gF** 2-Back accuracy, Rest r = .17, p = 2.00E−3 * 10-Fold cross-validation Internal data 
  
11 PMAT score, Task   PMAT score, Task r = .33, p = 1/1001 * 10-Fold cross-validation Internal data 
12     SMC composite memory, Rest r = .30, p = 1/1001 * External validation External data, different behavior 
13     2-Back accuracy, Task r = .34, p = 2.00E−3 * 10-Fold cross-validation Internal data, different behavior 
  
14 PMAT score, Rest   PMAT score, Rest r = .12, p = 1.70E−2   10-Fold cross-validation Internal data 
15     SMC composite memory, Rest r = .36, p = 1/1001 * External validation External data, different behavior 
16     2-Back accuracy, Rest r = .13, p = 2.70E−2   10-Fold cross-validation Internal data, different behavior 
  
17 SCPT d′, Task   SCPT d′, Task r = .14, p = 5.00E−3 * 10-Fold cross-validation Internal data 
18     2-Back accuracy, Task r = .16, p = 6.39E−2   10-Fold cross-validation Internal data, different behavior 
  
19 SCPT d′, Rest   SCPT d′, Rest r = −5.7E−3, p = 0.533   10-Fold cross-validation Internal data 
20     2-Back accuracy, Rest r = .11, p = 5.59E−2   10-Fold cross-validation Internal data, different behavior 
Analysis #Training Behavior, Training DataControlled CovariateTest Behavior, Test DataModel Performance Validation TypeData Type
2-Back accuracy, Task   2-Back accuracy, Task r = .36, p = 1/1001 * 10-Fold cross-validation Internal data 
    SMC composite memory, Rest r = .37, p = 1/1001 * External validation External data, different behavior 
    PMAT score, Task r = .28, p = 1/1001 * 10-Fold cross-validation Internal data, different behavior 
    SCPT d′, Task r = .09, p = 7.19E−2   10-Fold cross-validation Internal data, different behavior 
  
2-back accuracy, Task gF** 2-Back accuracy, Task r = .33, p = 1/1001 * 10-Fold cross-validation Internal data 
  
2-back accuracy, Rest   2-Back accuracy, Rest r = .20, p = 1/1001 * 10-Fold cross-validation Internal data 
    SMC composite memory, Rest r = .30, p = 1/1001 * External validation External data, different behavior 
    PMAT score, Rest r = .13, p = 2.00E−3 * 10-Fold cross-validation Internal data, different behavior 
    SCPT d′, Rest r = .06, p = 6.29E−2   10-Fold cross-validation Internal data, different behavior 
  
10 2-back accuracy, Rest gF** 2-Back accuracy, Rest r = .17, p = 2.00E−3 * 10-Fold cross-validation Internal data 
  
11 PMAT score, Task   PMAT score, Task r = .33, p = 1/1001 * 10-Fold cross-validation Internal data 
12     SMC composite memory, Rest r = .30, p = 1/1001 * External validation External data, different behavior 
13     2-Back accuracy, Task r = .34, p = 2.00E−3 * 10-Fold cross-validation Internal data, different behavior 
  
14 PMAT score, Rest   PMAT score, Rest r = .12, p = 1.70E−2   10-Fold cross-validation Internal data 
15     SMC composite memory, Rest r = .36, p = 1/1001 * External validation External data, different behavior 
16     2-Back accuracy, Rest r = .13, p = 2.70E−2   10-Fold cross-validation Internal data, different behavior 
  
17 SCPT d′, Task   SCPT d′, Task r = .14, p = 5.00E−3 * 10-Fold cross-validation Internal data 
18     2-Back accuracy, Task r = .16, p = 6.39E−2   10-Fold cross-validation Internal data, different behavior 
  
19 SCPT d′, Rest   SCPT d′, Rest r = −5.7E−3, p = 0.533   10-Fold cross-validation Internal data 
20     2-Back accuracy, Rest r = .11, p = 5.59E−2   10-Fold cross-validation Internal data, different behavior 

Correction for multiple comparison: Analyses 1–4, 6–9: alpha = .0125; Analyses 5 and 10: alpha = .05; Analyses 11–13, 14–16: alpha = .016; Analyses 17–18, 19–20: alpha = .025. WM CPMs strongly predict HCP 2-back accuracy, HCP PMAT, and memory performance in the SMC data set. WM CPMs built while controlling for gF at edge selection continue to strongly predict HCP 2-back accuracy. GF CPMs strongly predict HCP PMAT score, HCP 2-back accuracy, and SMC memory performance. When trained on n-back task data, but not rest data, attention CPMs predict HCP SCPT d′. p values are derived from 1000 permutations of true and null models. Task data are HCP n-back task fMRI data.

*

Significant with p < .05, Bonferroni-corrected for multiple comparisons within models defined by training behavior and data (Analyses 1–4, 6–9, 11–13, 14–16, 17–18, 19–20).

**

Additional controlling for gF at edge selection in model construction.

External Validation in an Independent, Older Age Sample with Memory Impairment

To test whether WM CPMs defined in the HCP data set were robust to participant population, scan site and parameters, and memory measure, we applied them to resting-state functional connectivity matrices from the SMC sample. We began this analysis by defining a single set of WM CPMs (i.e., network masks and linear models relating network strength to 2-back accuracy) using data from the full HCP data set rather than utilizing 10-fold cross-validation. The performance of these CPMs in the external data set was assessed by correlating predicted 2-back accuracy scores with a composite score of visual and verbal memory performance in the SMC sample. Visual memory was quantified by performance on the Rey Complex Figure Test, during which participants visually reproduced a complex line drawing by both copying and drawing from memory (Meyers & Meyers, 1995). Verbal memory was quantified by verbal recall performance on the Seoul Verbal Learning Task, during which participants verbally recalled items from a list of 12 words before and after a 20-min delay (Ahn et al., 2010). The composite score was calculated as the sum of visual and verbal memory task performance. Motion was not significantly correlated with memory score in our group of 157 participants (mean frame-to-frame displacement: r = .12, p = .12; max translation: r = .06, p = .49; max rotation: r = .05, p = .54).

Network Overlap

We characterized the overlap of the networks predicting WM, gF, and sustained attention by first defining networks that positively and negatively predicted each measure in each of 1000 rounds of 10-fold cross-validation in the HCP sample. (Note that these networks, visualized in Figures 2, 5, and 6, represent a subset of edges included in the model trained on the full HCP sample and tested on the SMC sample. We elected to visualize these subset networks because they include the edges most consistently related to behavior.) We next counted the number of edges in common between every pair of networks and assessed the statistical significance of this overlap with the hypergeometric cumulative distribution function (Rosenberg, Hsu, Scheinost, Todd Constable, & Chun, 2018). We also report the anatomical location of overlapping edges.

RESULTS

Internal Validation: Prediction of WM from Task-based and Resting-state Connectivity

CPMs defined using both task-based and resting-state data demonstrated significant correlation between observed and predicted 2-back accuracy scores of previously unseen individuals from the HCP sample (Table 1, Analyses 1 and 6; Figure 1).

Figure 1. 

Summary of WM CPM performance. Correlations between predicted 2-back accuracy and observed 2-back accuracy in the HCP sample (top) or observed composite memory score in the external SMC sample (bottom) over 1000 true model iterations (purple) or 1000 null model permutations (gray). Scatter plots depict the correlation between each individual's predicted and observed behavioral score in one model iteration whose r and p values match that of the average values reported in these plots. p values are Bonferroni-corrected for multiple comparisons. Models were trained on HCP participants' n-back task (left) or resting-state (right) functional connectivity data and tested on functional connectivity data from previously unseen individuals. Whether built from n-back task or resting-state fMRI data, our WM models significantly predict both 2-back accuracy in the HCP data set and memory performance in the independent SMC data set, including memory-impaired individuals.

Figure 1. 

Summary of WM CPM performance. Correlations between predicted 2-back accuracy and observed 2-back accuracy in the HCP sample (top) or observed composite memory score in the external SMC sample (bottom) over 1000 true model iterations (purple) or 1000 null model permutations (gray). Scatter plots depict the correlation between each individual's predicted and observed behavioral score in one model iteration whose r and p values match that of the average values reported in these plots. p values are Bonferroni-corrected for multiple comparisons. Models were trained on HCP participants' n-back task (left) or resting-state (right) functional connectivity data and tested on functional connectivity data from previously unseen individuals. Whether built from n-back task or resting-state fMRI data, our WM models significantly predict both 2-back accuracy in the HCP data set and memory performance in the independent SMC data set, including memory-impaired individuals.

External Validation: Prediction of WM Performance in a Memory-impaired Population

To assess the robustness and generalizability of WM CPMs, we applied them to a completely independent sample of 157 older adults of the SMC data set, 109 of whom were memory-impaired (AD or aMCI). Applying the n-back task-trained network models to resting-state functional connectivity from these individuals, we found a strong correlation between predicted 2-back accuracy values and observed composite memory score of visual and verbal memory task performance (Table 1, Analysis 2; Figure 1). Strong prediction of these three measures was also achieved when rest-trained models were applied (Table 1, Analysis 7; Figure 1). Model performance remained significant in the subset of 28 individuals from the SMC data set whose motion values met the stricter motion threshold that was appropriate for young adult HCP participants (average frame-to-frame motion < 0.10 mm: task-trained, r = .46, p = 7.99 × 10−3; rest-trained, r = .40, p = 2.60 × 10−2; Supplementary Figure S2).

Functional Anatomy of WM Networks

To gain a better understanding of the brain regions and systems underlying the strong performance of our WM models (defined using the full HCP sample), we examined the connections utilized between 10 macroscale regions and eight canonical brain networks. Utilization is here operationalized as the percentage of possible network–network edges included in the full set of features that were selected on every iteration. Macroscale regions were anatomically defined and covered a range of cortical and subcortical areas. Canonical networks were functionally defined using the data used for parcellation as described in Finn et al. (2015). Grouping our 268 nodes into these regions and networks, we found that a broad range of connections was utilized in the positive and negative networks of both our task- and rest-trained WM models, spanning both cortical and subcortical regions (Figure 2; anatomical, network, and surface representations). Frontoparietal, subcortical-cerebellar, and motor networks and regions were particularly heavily utilized.

Figure 2. 

Functional anatomy of positive and negative networks of WM task-trained (top) and rest-trained (bottom) WM CPMs. Left and center: In both positive (red) and negative (blue) networks of task- and rest-trained models, a broad range of connections was utilized. Line saturation indicates the proportion of edges of the particular network pair involved in prediction, with darker lines indicating a higher proportion of edges utilized. Line width indicates the total number of edges possible in the pair, with thicker lines indicating more possible edges. Canonical networks include default mode (DM), subcortical-cerebellum (SubC), frontoparietal (FP), motor (MT), medial frontal (MF), visual association (VA), VI, and VII. Edge selection threshold for visualization is p < 1 × 10−6 for clarity. Right: Positive (red) and negative (blue) edges of the task-trained (left) and rest-trained (right) models of WM described in the article. Spheres represent nodes and lines represent functional connections between them. Nodes are sized according to the total number of connections they have in both the positive and negative networks and are colored according to the network in which they have more connections. The right surface representation images were created using the publicly available BioImage Suite Connectivity Viewer Tool.

Figure 2. 

Functional anatomy of positive and negative networks of WM task-trained (top) and rest-trained (bottom) WM CPMs. Left and center: In both positive (red) and negative (blue) networks of task- and rest-trained models, a broad range of connections was utilized. Line saturation indicates the proportion of edges of the particular network pair involved in prediction, with darker lines indicating a higher proportion of edges utilized. Line width indicates the total number of edges possible in the pair, with thicker lines indicating more possible edges. Canonical networks include default mode (DM), subcortical-cerebellum (SubC), frontoparietal (FP), motor (MT), medial frontal (MF), visual association (VA), VI, and VII. Edge selection threshold for visualization is p < 1 × 10−6 for clarity. Right: Positive (red) and negative (blue) edges of the task-trained (left) and rest-trained (right) models of WM described in the article. Spheres represent nodes and lines represent functional connections between them. Nodes are sized according to the total number of connections they have in both the positive and negative networks and are colored according to the network in which they have more connections. The right surface representation images were created using the publicly available BioImage Suite Connectivity Viewer Tool.

Default mode network (DMN) connectivity was also utilized in both task- and rest-trained WM models. In task-trained models, the DMN and DMN-associated regions (including limbic, prefrontal, parietal, and temporal cortices) were utilized more heavily in networks that positively predicted WM performance than in negatively predictive networks. Both positive and negative task-trained networks also included connections in insular regions.

Modeling Fluid Intelligence and Attention in the HCP Data Set

To compare the features and performance of WM CPMs to independent CPMs computed for gF and for sustained attention, we repeated the CPM pipeline to construct models predicting PMAT scores, as a measure of gF, and SCPT d′ values, as a measure of sustained attention, using 10-fold cross-validation in the same data set of 502 HCP participants. Replicating previous work (Finn et al., 2015, 2017), gF models generated predicted PMAT scores that strongly correlated with observed PMAT scores when trained on task fMRI data (Figure 3; Table 1, Analyses 11 and 14). Sustained attention models also generated predicted SCPT d′ scores that significantly correlated with observed SCPT d′ scores, though with less internal prediction strength than that of gF models (Steiger's Z = 3.0, p = 2.4 × 10−3; Figure 3; Table 1, Analyses 17 and 19).

Figure 3. 

Summary of gF and attention CPM performance. Correlations between predicted and observed behavioral scores over 1000 model iterations (green and blue) or 1000 null model permutations (gray) utilizing gF (top) and attention (bottom) models. Scatter plots depict the correlation between each individual's predicted and observed behavioral score in one model iteration with average correlation coefficient. p values are corrected for multiple comparisons as shown in Table 1. Behavioral measures include 2-back accuracy (HCP data, left), PMAT (HCP data, middle top) or SCPT d′ (HCP data, middle bottom), and SMC composite memory score (external data, right). Models were trained on HCP participants' n-back task functional connectivity data and were tested on functional connectivity data from previously unseen individuals. gF models predict scores of HCP participant 2-back accuracy (r = .34) and PMAT (r = .33), as well as external SMC participant memory scores (r = .30). Attention models perform with significantly less strength (Steiger's Z = 3.0, p = 2.4 × 10−3), more weakly predicting HCP 2-back accuracy (r = .16) and SCPT d′ (r = .14) scores. Attention models do not predict external SMC memory scores (r = .03). The difference in correlation between predicted and observed scores of 2-back accuracy and scores of the trained behavioral measure (PMAT or SCPT d′) is not significant for gF (Steiger's Z = 0.18, p = .86) or attention (Steiger's Z = 0.32, p = .75) models.

Figure 3. 

Summary of gF and attention CPM performance. Correlations between predicted and observed behavioral scores over 1000 model iterations (green and blue) or 1000 null model permutations (gray) utilizing gF (top) and attention (bottom) models. Scatter plots depict the correlation between each individual's predicted and observed behavioral score in one model iteration with average correlation coefficient. p values are corrected for multiple comparisons as shown in Table 1. Behavioral measures include 2-back accuracy (HCP data, left), PMAT (HCP data, middle top) or SCPT d′ (HCP data, middle bottom), and SMC composite memory score (external data, right). Models were trained on HCP participants' n-back task functional connectivity data and were tested on functional connectivity data from previously unseen individuals. gF models predict scores of HCP participant 2-back accuracy (r = .34) and PMAT (r = .33), as well as external SMC participant memory scores (r = .30). Attention models perform with significantly less strength (Steiger's Z = 3.0, p = 2.4 × 10−3), more weakly predicting HCP 2-back accuracy (r = .16) and SCPT d′ (r = .14) scores. Attention models do not predict external SMC memory scores (r = .03). The difference in correlation between predicted and observed scores of 2-back accuracy and scores of the trained behavioral measure (PMAT or SCPT d′) is not significant for gF (Steiger's Z = 0.18, p = .86) or attention (Steiger's Z = 0.32, p = .75) models.

To further characterize relationships between WM, gF, and sustained attention and their respective network models, we assessed the ability of PMAT and SCPT d′ models to predict n-back task performance and vice versa. Throughout 1000 rounds of 10-fold cross-validation, these models were trained on the data of 90% of HCP participants in the specified training behavior (e.g., PMAT score) and applied to the left out 10% to predict the other behavioral measure of interest (e.g., SCPT d′). Utilizing gF networks, we found a strong correlation between observed 2-back accuracy and predicted gF performance (Figure 3; Table 1, Analyses 13 and 16; Supplementary Table S3). We observed similarly strong performance when using WM networks to predict PMAT scores (Table 1, Analyses 3 and 8; Supplementary Table S3). In comparison to that of gF and WM, the relationship between predicted attention and WM performance was significantly less strong (Steiger's Z = 3.36, p = 7.9 × 10−4): Attention network predictions correlated more weakly with observed scores of 2-back accuracy (Figure 3; Table 1, Analyses 18 and 20; Supplementary Table S3), as did predicted 2-back accuracy and observed SCPT d′ scores (Table 1, Analyses 4 and 9; Supplementary Table S3).

WM Models and Fluid Intelligence Models Predict Memory Score in an External (SMC) Data Set

To directly compare the predictive power of WM and gF models, we also applied our gF CPMs to the external SMC data set including memory-impaired individuals. Predicted gF was correlated with observed composite memory scores (Figure 3; Table 1, Analyses 12 and 15). This correlation was of similar strength to that achieved when WM model predictions were correlated with observed composite memory scores (Table 1, Analyses 2 and 12).

The current results do not speak to why we observe some cases of improved model performance when testing on untrained behavior or even an independent fMRI data set. Several possibilities include differences in participant setting during behavioral testing (in scanner or out of scanner), differences in reliability among behavioral tests used, and differences in the distribution of behavior or age ranges of HCP and SMC participants, as individual differences in memory are thought to be magnified in aging (Nagel et al., 2008).

WM CPMs Retain Predictive Power when Controlling for gF at Edge Selection

To further assess the relationship between models of WM and gF, we repeated CPM and 10-fold cross-validation procedures to create a WM model in which we selected edges correlated with 2-back accuracy while controlling for gF (PMAT score) in each fold of 10-fold cross-validation. The resulting model, consisting of edges that appeared in every round of 10-fold cross-validation, contained 1036 edges in the positive network (previously 1674) and 755 edges in the negative network (previously 1075). Despite this substantial decrease in numbers of network edges, this model retained predictive power (Figure 4,Table 1, Analyses 5 and 10), which was not significantly different (task-trained models: Steiger's Z = 0.54, p = .59; rest-trained models: Z = 0.49, p = .62) from that of our original WM model created without controlling for gF during edge selection (Figure 4; Table 1, Analyses 1 and 6).

Figure 4. 

Correlations between observed 2-back accuracy scores and scores predicted by n-back task-trained combined networks. Though exclusion of gF-correlated edges (right) resulted in models that contained 38% fewer edges in the positive network and 30% fewer edges in the negative network, these models continued to predict WM strongly, performing comparably to models created with no gF-related edge exclusion criteria (left).

Figure 4. 

Correlations between observed 2-back accuracy scores and scores predicted by n-back task-trained combined networks. Though exclusion of gF-correlated edges (right) resulted in models that contained 38% fewer edges in the positive network and 30% fewer edges in the negative network, these models continued to predict WM strongly, performing comparably to models created with no gF-related edge exclusion criteria (left).

Anatomical Features of WM and gF Networks

In a final assessment of the relationship between the models of WM and gF, we compared the functional anatomy of these networks. To do so, we first followed the procedures of Finn et al. (2015) to group our 268 nodes into 10 anatomically defined brain regions and eight functionally defined canonical networks. We then considered the relative numbers of connections between the nodes comprising each pair of regions. As depicted in Figure 5, we found that overlap between our WM and gF networks was largely limited to connections between prefrontal, parietal, and motor regions. In general, WM networks utilized a wider range of connections throughout the brain, including subcortical and cerebellar regions, whereas gF networks utilized connections primarily between prefrontal and parietal regions. These gF network connections align with previous analyses of gF model anatomical features (Finn et al., 2017).

Figure 5. 

Edge overlap between WM and gF models. WM models utilize a wide range of connections throughout the brain, including subcortical and cerebellar regions, whereas gF models largely utilize connections between prefrontal and parietal regions. Overlap is predominantly observed in connections between parietal, prefrontal, and motor regions. Top and middle: Line saturation indicates the proportion of edges of the particular network pair involved in prediction, with darker lines indicating a higher proportion of edges utilized. Line width indicates the total number of edges possible in the pair, with thicker lines indicating more possible edges. Circles represent within-region edges. Macroscale regions include prefrontal cortex (PFC), motor cortex (Mot), insula (Ins), parietal (Par), temporal (Tem), occipital (Occ), limbic (Lim; including the cingulate cortex, amygdala, and hippocampus), cerebellum (Cer), subcortical (Sub; thalamus and striatum) and brainstem (Bsm). Bottom: Positive (red) and negative (blue) edges of the task-trained WM (left) and gF (middle) models, as well as the overlap between these two models (right). Spheres represent nodes, and lines represent functional connections between them. Nodes are sized according to the total number of connections they have in both the positive and negative networks and are colored according to the network in which they have more connections. The bottom surface representation images were created using the publicly available BioImage Suite Connectivity Viewer Tool.

Figure 5. 

Edge overlap between WM and gF models. WM models utilize a wide range of connections throughout the brain, including subcortical and cerebellar regions, whereas gF models largely utilize connections between prefrontal and parietal regions. Overlap is predominantly observed in connections between parietal, prefrontal, and motor regions. Top and middle: Line saturation indicates the proportion of edges of the particular network pair involved in prediction, with darker lines indicating a higher proportion of edges utilized. Line width indicates the total number of edges possible in the pair, with thicker lines indicating more possible edges. Circles represent within-region edges. Macroscale regions include prefrontal cortex (PFC), motor cortex (Mot), insula (Ins), parietal (Par), temporal (Tem), occipital (Occ), limbic (Lim; including the cingulate cortex, amygdala, and hippocampus), cerebellum (Cer), subcortical (Sub; thalamus and striatum) and brainstem (Bsm). Bottom: Positive (red) and negative (blue) edges of the task-trained WM (left) and gF (middle) models, as well as the overlap between these two models (right). Spheres represent nodes, and lines represent functional connections between them. Nodes are sized according to the total number of connections they have in both the positive and negative networks and are colored according to the network in which they have more connections. The bottom surface representation images were created using the publicly available BioImage Suite Connectivity Viewer Tool.

Given that WM models generalized to predict gF and vice versa, we assessed the degree of overlap between networks. With regard to network overlap models trained on n-back task data, of 1674 total edges in the positively predictive WM network, 455 were shared with the positively predictive gF network in every round of 10-fold cross-validation. This overlap was determined to be significant (p < 1.0 × 10−4) using the hypergeometric cumulative distribution function (Rosenberg et al., 2018). Of 1203 total edges in the negative predictive WM network, 300 (significant, p < 1.0 × 10−4) were shared with the negatively predictive gF network. There was no overlap between positive WM networks and negative gF networks or vice versa, nor was there any overlap between any positive and any negative network. Thus, of 2877 total edges in the positive and negative networks added together, 755 overlapped in the expected direction between positive and negative WM and gF networks (significant, p < 1.0 × 10−4). Overlap was also significant but numerically less extensive between WM and sustained attention models, totaling 63 positive network edges (p < 1.0 × 10−4) and 85 negative network edges (p < 1.0 × 10−4; Figure 6).

Figure 6. 

Overlap between positive and negative networks of task-trained WM, gF, and attention models. Positive network overlap between WM and gF CPMs was significant, totaling 455 edges, as was negative network overlap between these models, totaling 300 edges (significance in network overlap analyses indicates p < 10−4). Overlap between WM and attention CPMs was also significant but numerically less, totaling 63 positive network edges and 85 negative network edges. Forty-six edges were common to the gF and attention-positive networks (significant) and 52 to their negative networks (significant). Overall, 26 positive network edges (significant) and 29 negative network edges (significant) were shared by all three CPMs.

Figure 6. 

Overlap between positive and negative networks of task-trained WM, gF, and attention models. Positive network overlap between WM and gF CPMs was significant, totaling 455 edges, as was negative network overlap between these models, totaling 300 edges (significance in network overlap analyses indicates p < 10−4). Overlap between WM and attention CPMs was also significant but numerically less, totaling 63 positive network edges and 85 negative network edges. Forty-six edges were common to the gF and attention-positive networks (significant) and 52 to their negative networks (significant). Overall, 26 positive network edges (significant) and 29 negative network edges (significant) were shared by all three CPMs.

DISCUSSION

In this study, we provide a novel demonstration that distributed patterns of functional connectivity are a broadly applicable predictor of individual differences in WM. Using CPM, we defined predictive models of 2-back task performance in 502 individuals of the HCP data set. Indicating that our models characterize individual differences in WM-related processes in general rather than 2-back accuracy in particular, these WM CPMs generalized to predict memory in the independent SMC data set of 157 older adults. Complementing previous work that has identified neural correlates of WM (Magnuson et al., 2015; McNab & Klingberg, 2008), explored individual differences (Todd & Marois, 2005), and predicted memory impairments related to psychiatric disease (Yamashita et al., 2018), our present work predicts WM function in completely novel individuals, including older adults, with relatively high accuracy (Greene et al., 2018; Hsu, Rosenberg, Scheinost, Constable, & Chun, 2018; Lin et al., 2018). Furthermore, we characterize overlap and differences in the functional brain networks underlying WM, fluid intelligence, and sustained attention.

The present results align with previous demonstrations of the strong link between WM and gF: gF models strongly predicted WM score in novel individuals, and vice versa. Though not complete, overlap between gF and WM networks was substantial (755 of 2877 total WM CPM edges, 26.2%), and this overlap was predominantly observed in connections between parietal, prefrontal, and motor regions. This supports the notion that WM and gF rely on shared neural systems, particularly those involving parietal and prefrontal regions, which allow us to keep a representation active in the face of interference (Engle et al., 1999; Prabhakaran et al., 1997).

Although our results support the strong link between WM and gF, we also find that a set of edges uniquely predicts memory task performance. WM models built while controlling for gF correlation at edge selection showed similar model performance to original WM models built without gF control. This suggests that predictive network models of WM do not rely on functional connections that support gF.

Additionally, as SMC participants were older adults, the successful external validation of WM models indicates some degree of persistence in the neural systems supporting WM into older age, even in individuals with memory impairment. Anatomically, WM models also utilized connections between a wider range of regions than did gF networks, including more subcortical and cerebellar regions. Though connections between prefrontal and parietal regions have classically been implicated in WM performance and training (Constantinidis & Klingberg, 2016; Engle et al., 1999) and are included in our networks, our findings of additional, broad WM connectivity suggest that WM processes are facilitated by coordinated activity across a wide range of cortical and subcortical regions. This notion aligns with recent propositions that communication between large-scale networks underlies WM performance (Cohen & D'Esposito, 2016; Gazzaley et al., 2004).

The present results also align with previous indications of a link between WM and attention. An individual's ability to control attention is related to performance on complex WM tasks (Barrett et al., 2004; Kane et al., 2001), and our work indeed supports this notion by demonstrating that some network components are shared between WM and attention models (148 of 2877 total WM CPM edges overlapping, 5.1%). This degree of overlap, however, was less extensive than that of WM and gF models. Furthermore, models of WM did not correlate with sustained attention performance, and models of attention did not correlate with WM performance. The fact that that our models of WM capacity and attention lack cross applicability and do not predominantly utilize the same functional connections aligns with recent propositions that WM precision, rather than WM capacity, is what is most influenced by attentional processes and information flow between attentional and perceptual neural systems (Galeano Weber et al., 2017). Furthermore, the weak performance of attention models indicates that our modeling approach reflects more than an ability to predict general cognitive performance from functional connectivity: The strong performance we observe in models of WM is specific and cannot be replicated by just any selected measure of cognitive performance. In general, our use of predictive modeling techniques to compare different cognitive abilities is novel, providing a new approach in the study of the relationships between cognitive and attentional abilities.

Additionally, the anatomical features of our task-trained and rest-trained WM models suggest that connectivity within and between regions of the DMN is highly utilized in predicting WM performance, supporting hypotheses of a role for task-persistent DMN connectivity (Hampson, Driesen, Roth, Gore, & Constable, 2010; Hampson, Driesen, Skudlarski, Gore, & Constable, 2006). Particularly in task-trained models, DMN connectivity was utilized in positive but not negative networks, and the notion that more extensive DMN connectivity may be indicative of higher WM performance is a novel finding. This aligns with previous indications that DMN connectivity is degraded in AD patients, though reduction in the integrity of several other brain networks has also been reported in these patients (Brier et al., 2012; Greicius, Srivastava, Reiss, & Menon, 2004). Furthermore, the insula was bilaterally utilized in both positive and negative task-trained networks, providing support for propositions that the insula plays a role in switching between central executive network and DMN (Sridharan, Levitin, & Menon, 2008).

In all, our present work demonstrates, for the first time, that distributed sets of task-based and resting-state functional connections predict WM performance and can help characterize relationships between memory and other cognitive abilities, including gF and sustained attention. Though not intended to be used as a diagnostic tool, our results are theoretically informative, linking brain variables to behavioral ones to enhance our understanding of the neural infrastructure that gives rise to WM processes. Further applications of modeling and analysis techniques will be of use in relating our WM models to other memory types, including long-term memory, and other cognitive abilities and clinical symptoms.

Acknowledgments

This work was supported by the Yale Faculty of Arts and Sciences MRI Program funded by the Office of the Provost and the Department of Psychology, National Science Foundation grant BCS1558497 to M. M. C., and National Institutes of Health grant MH108591 to M. M. C. Data were provided in part by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.

Conceptualization, E. W. A., M. D. R, K. Y., and M. M. C.; Methodology, E. W. A., M. D. R., and K. Y.; Validation, M. D. R.; Investigation, E. W. A.; Writing–Original Draft, E. W. A.; Writing–Review and Editing, E. W. A., M. D. R., K. Y., A. G., D. S., R. T. C., M. M. C; Supervision, M. D. R., K. Y., and M. M. C.; Funding Acquisition, M. M. C.

Reprint request should be sent to Emily W. Avery, Department of Psychology, Yale University, New Haven, CT, or via e-mail: emily.avery@yale.edu.

Note

1. 

Supplementary material for this paper can be retrieved from https://doi.org/10.6084/m9.figshare.10693736.v1.

REFERENCES

Ahn
,
H. J.
,
Chin
,
J.
,
Park
,
A.
,
Lee
,
B. H.
,
Suh
,
M. K.
,
Seo
,
S. W.
, et al
(
2010
).
Seoul neuropsychological screening battery-dementia version (SNSB-D): A useful tool for assessing and monitoring cognitive impairments in dementia patients
.
Journal of Korean Medical Science
,
25
,
1071
1076
.
Baddeley
,
A.
(
2003
).
Working memory: Looking back and looking forward
.
Nature Reviews Neuroscience
,
4
,
829
839
.
Baddeley
,
A. D.
, &
Hitch
,
G.
(
1974
).
Working memory
. In
G. H.
Bower
(Ed.),
Psychology of learning and motivation
(
Vol. 8
, pp.
47
89
).
Academic Press
.
Barch
,
D. M.
,
Burgess
,
G. C.
,
Harms
,
M. P.
,
Petersen
,
S. E.
,
Schlaggar
,
B. L.
,
Corbetta
,
M.
, et al
(
2013
).
Function in the human connectome: Task-fMRI and individual differences in behavior
.
Neuroimage
,
80
,
169
189
.
Barrett
,
L. F.
,
Tugade
,
M. M.
, &
Engle
,
R. W.
(
2004
).
Individual differences in working memory capacity and dual-process theories of the mind
.
Psychological Bulletin
,
130
,
553
573
.
Bertolero
,
M. A.
,
Yeo
,
B. T. T.
,
Bassett
,
D. S.
, &
D'Esposito
,
M.
(
2018
).
A mechanistic model of connector hubs, modularity and cognition
.
Nature Human Behavior
,
2
,
765
777
.
Brier
,
M. R.
,
Thomas
,
J. B.
,
Snyder
,
A. Z.
,
Benzinger
,
T. L.
,
Zhang
,
D.
,
Raichle
,
M. E.
, et al
(
2012
).
Loss of intranetwork and internetwork resting state functional connections with alzheimer's disease progression
.
Journal of Neuroscience
,
32
,
8890
8899
.
Christophel
,
T. B.
,
Hebart
,
M. N.
, &
Haynes
,
J. D.
(
2012
).
Decoding the contents of visual short-term memory from human visual and parietal cortex
.
Journal of Neuroscience
,
32
,
12983
12989
.
Chun
,
M. M.
,
Golomb
,
J. D.
, &
Turk-Browne
,
N. B.
(
2011
).
A taxonomy of external and internal attention
.
Annual Review of Psychology
,
62
,
73
101
.
Cohen
,
J. R.
, &
D'Esposito
,
M.
(
2016
).
The segregation and integration of distinct brain networks and their relationship to cognition
.
Journal of Neuroscience
,
36
,
12083
12094
.
Colom
,
R.
,
Flores-Mendoza
,
C.
, &
Rebollo
,
I.
(
2002
).
Working memory and intelligence
.
Personality and Individual Differences
,
34
,
33
39
.
Constantinidis
,
C.
, &
Klingberg
,
T.
(
2016
).
The neuroscience of working memory capacity and training
.
Nature Reviews Neuroscience
,
17
,
438
449
.
Cowan
,
N.
,
Elliott
,
E. M.
,
Scott Saults
,
J.
,
Morey
,
C. C.
,
Mattox
,
S.
,
Hismjatullina
,
A.
, et al
(
2005
).
On the capacity of attention: Its estimation and its role in working memory and cognitive aptitudes
.
Cognitive Psychology
,
51
,
42
100
.
D'Esposito
,
M.
, &
Postle
,
B. R.
(
2015
).
The cognitive neuroscience of working memory
.
Annual Review of Psychology
,
66
,
115
142
.
D'Esposito
,
M.
,
Zarahn
,
E.
,
Aguirre
,
G. K.
, &
Rypma
,
B.
(
1999
).
The effect of normal aging on the coupling of neural activity to the bold hemodynamic response
.
Neuroimage
,
10
,
6
14
.
Dubois
,
J.
, &
Adolphs
,
R.
(
2016
).
Building a science of individual differences from fMRI
.
Trends in Cognitive Sciences
,
20
,
425
443
.
Engle
,
R. W.
,
Tuholski
,
S. W.
,
Laughlin
,
J. E.
, &
Conway
,
A. R. A.
(
1999
).
Working memory, short-term memory, and general fluid intelligence: A latent-variable approach
.
Journal of Experimental Psychology: General
,
128
,
309
331
.
Fiebach
,
C. J.
,
Rissman
,
J.
, &
D'Esposito
,
M.
(
2006
).
Modulation of inferotemporal cortex activation during verbal working memory maintenance
.
Neuron
,
51
,
251
261
.
Finn
,
E. S.
,
Scheinost
,
D.
,
Finn
,
D. M.
,
Shen
,
X.
,
Papademetris
,
X.
, &
Constable
,
R. T.
(
2017
).
Can brain state be manipulated to emphasize individual differences in functional connectivity?
Neuroimage
,
160
,
140
151
.
Finn
,
E. S.
,
Shen
,
X.
,
Scheinost
,
D.
,
Rosenberg
,
M. D.
,
Huang
,
J.
,
Chun
,
M. M.
, et al
(
2015
).
Functional connectome fingerprinting: Identifying individuals using patterns of brain connectivity
.
Nature Neuroscience
,
18
,
1664
1671
.
Fountain-Zaragoza
,
S.
,
Samimy
,
S.
,
Rosenberg
,
M. D.
, &
Prakash
,
R. S.
(
2019
).
Connectome-based models predict attentional control in aging adults
.
Neuroimage
,
186
,
1
13
.
Fukuda
,
K.
,
Vogel
,
E.
,
Mayr
,
U.
, &
Awh
,
E.
(
2010
).
Quantity, not quality: The relationship between fluid intelligence and working memory capacity
.
Psychonomic Bulletin & Review
,
17
,
673
679
.
Galeano Weber
,
E. M.
,
Hahn
,
T.
,
Hilger
,
K.
, &
Fiebach
,
C. J.
(
2017
).
Distributed patterns of occipito-parietal functional connectivity predict the precision of visual working memory
.
Neuroimage
,
146
,
404
418
.
Gazzaley
,
A.
,
Rissman
,
J.
, &
D'Esposito
,
M.
(
2004
).
Functional connectivity during working memory maintenance
.
Cognitive, Affective, & Behavioral Neuroscience
,
4
,
580
599
.
Geerligs
,
L.
,
Tsvetanov
,
K. A.
,
Cam-CAN
, &
Henson
,
R. N.
(
2017
).
Challenges in measuring individual differences in functional connectivity using fMRI: The case of healthy aging
.
Human Brain Mapping
,
38
,
4125
4156
.
Glasser
,
M. F.
,
Sotiropoulos
,
S. N.
,
Wilson
,
J. A.
,
Coalson
,
T. S.
,
Fischl
,
B.
,
Andersson
,
J. L.
, et al
(
2013
).
The minimal preprocessing pipelines for the Human Connectome Project
.
Neuroimage
,
80
,
105
124
.
Gong
,
W.
,
Cheng
,
F.
,
Rolls
,
E. T.
,
Lo
,
C. Z.
,
Huang
,
C. C.
,
Tsai
,
S. J.
, et al
(
2019
).
A powerful and efficient multivariate approach for voxel-level connectome-wide association studies
.
Neuroimage
,
188
,
628
641
.
Greene
,
A. S.
,
Gao
,
S.
,
Scheinost
,
D.
, &
Constable
,
R. T.
(
2018
).
Task-induced brain state manipulation improves prediction of individual traits
.
Nature Communications
,
9
,
2807
.
Greicius
,
M. D.
,
Srivastava
,
G.
,
Reiss
,
A. L.
, &
Menon
,
V.
(
2004
).
Default-mode network activity distinguishes Alzheimer's disease from healthy aging: Evidence from functional MRI
.
Proceedings of the National Academy of Sciences, U.S.A.
,
101
,
4637
4642
.
Guo
,
C. C.
,
Kurth
,
F.
,
Zhou
,
J.
,
Mayer
,
E. A.
,
Eickhoff
,
S. B.
,
Kramer
,
J. H.
, et al
(
2012
).
One-year test–retest reliability of intrinsic connectivity network fMRI in older adults
.
Neuroimage
,
61
,
1471
1483
.
Hampson
,
M.
,
Driesen
,
N.
,
Roth
,
J. K.
,
Gore
,
J. C.
, &
Constable
,
R. T.
(
2010
).
Functional connectivity between task-positive and task-negative brain areas and its relation to working memory performance
.
Magnetic Resonance Imaging
,
28
,
1051
1057
.
Hampson
,
M.
,
Driesen
,
N.
,
Skudlarski
,
P.
,
Gore
,
J. C.
, &
Constable
,
R. T.
(
2006
).
Brain connectivity related to working memory performance
.
Journal of Neuroscience
,
26
,
13338
13343
.
Han
,
X.
,
Berg
,
A. C.
,
Oh
,
H.
,
Samaras
,
D.
, &
Leung
,
H. C.
(
2013
).
Multi-voxel pattern analysis of selective representation of visual working memory in ventral temporal and occipital regions
.
Neuroimage
,
73
,
8
15
.
Hsu
,
W. T.
,
Rosenberg
,
M. D.
,
Scheinost
,
D.
,
Constable
,
R. T.
, &
Chun
,
M. M.
(
2018
).
Resting-state functional connectivity predicts neuroticism and extraversion in novel individuals
.
Social Cognitive and Affective Neuroscience
,
13
,
224
232
.
Just
,
M. A.
, &
Carpenter
,
P. A.
(
1992
).
A capacity theory of comprehension: Individual differences in working memory
.
Psychological Review
,
99
,
122
149
.
Kane
,
M. J.
,
Bleckley
,
M. K.
,
Conway
,
A. R.
, &
Engle
,
R. W.
(
2001
).
A controlled-attention view of working-memory capacity
.
Journal of Experimental Psychology: General
,
130
,
169
183
.
Kane
,
M. J.
,
Conway
,
A. R. A.
,
Miura
,
T. K.
, &
Colflesh
,
G. J. H.
(
2007
).
Working memory, attention control, and the n-back task: A question of construct validity
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
33
,
615
622
.
Kane
,
M. J.
, &
Engle
,
R. W.
(
2002
).
The role of prefrontal cortex in working-memory capacity, executive attention, and general fluid intelligence: An individual-differences perspective
.
Psychonomic Bulletin & Review
,
9
,
637
671
.
Kim
,
H.
,
Yoo
,
K.
,
Na
,
D. L.
,
Seo
,
S. W.
,
Jeong
,
J.
, &
Jeong
,
Y.
(
2015
).
Non-monotonic reorganization of brain networks with Alzheimer's disease progression
.
Frontiers in Aging Neuroscience
,
7
,
111
.
Klatzky
,
R. L.
(
1975
).
Human memory: Structures and processes
.
Oxford
:
W. H. Freeman
.
Kurtz
,
M. M.
,
Ragland
,
J. D.
,
Bilker
,
W.
,
Gur
,
R. C.
, &
Gur
,
R. E.
(
2001
).
Comparison of the continuous performance test with and without working memory demands in healthy controls and patients with schizophrenia
.
Schizophrenia Research
,
48
,
307
316
.
Kyllonen
,
P. C.
, &
Christal
,
R. E.
(
1990
).
Reasoning ability is (little more than) working-memory capacity?!
Intelligence
,
14
,
389
433
.
Liebe
,
S.
,
Hoerzer
,
G. M.
,
Logothetis
,
N. K.
, &
Rainer
,
G.
(
2012
).
Theta coupling between V4 and prefrontal cortex predicts visual short-term memory performance
.
Nature Neuroscience
,
15
,
456
462
.
Lin
,
Q.
,
Rosenberg
,
M. D.
,
Yoo
,
K.
,
Hsu
,
T. W.
,
O'Connell
,
T. P.
, &
Chun
,
M. M.
(
2018
).
Resting-state functional connectivity predicts cognitive impairment related to alzheimer's disease
.
Frontiers in Aging Neuroscience
,
10
,
94
.
Liu
,
S.
,
Poh
,
J. H.
,
Koh
,
H. L.
,
Ng
,
K. K.
,
Loke
,
Y. M.
,
Lim
,
J. K. W.
, et al
(
2018
).
Carrying the past to the future: Distinct brain networks underlie individual differences in human spatial working memory capacity
.
Neuroimage
,
176
,
1
10
.
Luck
,
S. J.
, &
Vogel
,
E. K.
(
1997
).
The capacity of visual working memory for features and conjunctions
.
Nature
,
390
,
279
281
.
Luck
,
S. J.
, &
Vogel
,
E. K.
(
2013
).
Visual working memory capacity: From psychophysics and neurobiology to individual differences
.
Trends in Cognitive Sciences
,
17
,
391
400
.
Ma
,
W. J.
,
Husain
,
M.
, &
Bays
,
P. M.
(
2014
).
Changing concepts of working memory
.
Nature Neuroscience
,
17
,
347
356
.
Magnuson
,
M. E.
,
Thompson
,
G. J.
,
Schwarb
,
H.
,
Pan
,
W. J.
,
McKinley
,
A.
,
Schumacher
,
E. H.
, et al
(
2015
).
Errors on interrupter tasks presented during spatial and verbal working memory performance are linearly linked to large-scale functional network connectivity in high temporal resolution resting state fMRI
.
Brain Imaging and Behavior
,
9
,
854
867
.
McNab
,
F.
, &
Klingberg
,
T.
(
2008
).
Prefrontal cortex and basal ganglia control access to working memory
.
Nature Neuroscience
,
11
,
103
107
.
Meyers
,
J. E.
, &
Meyers
,
K. R.
(
1995
).
Rey complex figure test under four different administration procedures
.
The Clinical Neuropsychologist
,
9
,
63
67
.
Miller
,
E. K.
, &
Buschman
,
T. J.
(
2015
).
Working memory capacity: Limits on the bandwidth of cognition
.
Daedalus
,
144
,
112
122
.
Moore
,
T. M.
,
Reise
,
S. P.
,
Gur
,
R. E.
,
Hakonarson
,
H.
, &
Gur
,
R. C.
(
2015
).
Psychometric properties of the Penn computerized neurocognitive battery
.
Neuropsychology
,
29
,
235
246
.
Nagel
,
I. E.
,
Chicherio
,
C.
,
Li
,
S. C.
,
Von Oertzen
,
T.
,
Sander
,
T.
,
Villringer
,
A.
, et al
(
2008
).
Human aging magnifies genetic effects on executive functioning and working memory
.
Frontiers in Human Neuroscience
,
2
,
1
.
Palva
,
J. M.
,
Monto
,
S.
,
Kulashekhar
,
S.
, &
Palva
,
S.
(
2010
).
Neuronal synchrony reveals working memory networks and predicts individual memory capacity
.
Proceedings of the National Academy of Sciences, U.S.A.
,
107
,
7580
7585
.
Prabhakaran
,
V.
,
Smith
,
J. A.
,
Desmond
,
J. E.
,
Glover
,
G. H.
, &
Gabrieli
,
J. D.
(
1997
).
Neural substrates of fluid reasoning: An fMRI study of neocortical activation during performance of the Raven's progressive matrices test
.
Cognitive Psychology
,
33
,
43
63
.
Redick
,
T. S.
,
Broadway
,
J. M.
,
Meier
,
M. E.
,
Kuriakose
,
P. S.
,
Unsworth
,
N.
,
Kane
,
M. J.
, et al
(
2012
).
Measuring working memory capacity with automated complex span tasks
.
European Journal of Psychological Assessment
,
28
,
164
171
.
Rosenberg
,
M. D.
,
Finn
,
E. S.
,
Scheinost
,
D.
,
Papademetris
,
X.
,
Shen
,
X.
,
Constable
,
R. T.
, et al
(
2016
).
A neuromarker of sustained attention from whole-brain functional connectivity
.
Nature Neuroscience
,
19
,
165
171
.
Rosenberg
,
M. D.
,
Hsu
,
W. T.
,
Scheinost
,
D.
,
Todd Constable
,
R.
, &
Chun
,
M. M.
(
2018
).
Connectome-based models predict separable components of attention in novel individuals
.
Journal of Cognitive Neuroscience
,
30
,
160
173
.
Rosenberg
,
M. D.
,
Zhang
,
S.
,
Hsu
,
W. T.
,
Scheinost
,
D.
,
Finn
,
E. S.
,
Shen
,
X.
, et al
(
2016
).
Methylphenidate modulates functional network connectivity to enhance attention
.
Journal of Neuroscience
,
36
,
9547
9557
.
Shen
,
X.
,
Finn
,
E. S.
,
Scheinost
,
D.
,
Rosenberg
,
M. D.
,
Chun
,
M. M.
,
Papademetris
,
X.
, et al
(
2017
).
Using connectome-based predictive modeling to predict individual behavior from brain connectivity
.
Nature Protocols
,
12
,
506
518
.
Shen
,
X.
,
Tokoglu
,
F.
,
Papademetris
,
X.
, &
Constable
,
R. T.
(
2013
).
Groupwise whole-brain parcellation from resting-state fMRI data for network node identification
.
Neuroimage
,
82
,
403
415
.
Smith
,
S. M.
,
Beckmann
,
C. F.
,
Andersson
,
J.
,
Auerbach
,
E. J.
,
Bijsterbosch
,
J.
,
Douaud
,
G.
, et al
(
2013
).
Resting-state fMRI in the Human Connectome Project
.
Neuroimage
,
80
,
144
168
.
Smith
,
S.
,
Nichols
,
T.
,
Vidaurre
,
D.
,
Winkler
,
A.
,
Behrens
,
T.
,
Glasser
,
M.
, et al
(
2015
).
A positive-negative mode of population covariation links brain connectivity, demographics and behavior
.
Nature Neuroscience
,
18
,
1565
1567
.
Soreq
,
E.
,
Leech
,
R.
, &
Hampshire
,
A.
(
2019
).
Dynamic network coding of working-memory domains and working-memory processes
.
Nature Communications
,
10
,
936
.
Sreenivasan
,
K. K.
,
Vytlacil
,
J.
, &
D'Esposito
,
M.
(
2014
).
Distributed and dynamic storage of working memory stimulus information in extrastriate cortex
.
Journal of Cognitive Neuroscience
,
26
,
1141
1153
.
Sridharan
,
D.
,
Levitin
,
D. J.
, &
Menon
,
V.
(
2008
).
A critical role for the right fronto-insular cortex in switching between central-executive and default-mode networks
.
Proceedings of the National Academy of Sciences, U.S.A.
,
105
,
12569
12574
.
Todd
,
J. J.
, &
Marois
,
R.
(
2004
).
Capacity limit of visual short-term memory in human posterior parietal cortex
.
Nature
,
428
,
751
754
.
Todd
,
J. J.
, &
Marois
,
R.
(
2005
).
Posterior parietal cortex activity predicts individual differences in visual short-term memory capacity
.
Cognitive, Affective, & Behavioral Neuroscience
,
5
,
144
155
.
Uğurbil
,
K.
,
Xu
,
J.
,
Auerbach
,
E. J.
,
Moeller
,
S.
,
Vu
,
A. T.
,
Duarte-Carvajalino
,
J. M.
, et al
(
2013
).
Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project
.
Neuroimage
,
80
,
80
104
.
Van Dijk
,
K. R.
,
Sabuncu
,
M. R.
, &
Buckner
,
R. L.
(
2012
).
The influence of head motion on intrinsic functional connectivity MRI
.
Neuroimage
,
59
,
431
438
.
Van Essen
,
D. C.
,
Smith
,
S. M.
,
Barch
,
D. M.
,
Behrens
,
T. E.
,
Yacoub
,
E.
, &
Ugurbil
,
K.
(
2013
).
The WU-Minn Human Connectome Project: An overview
.
Neuroimage
,
80
,
62
79
.
Woo
,
C. W.
,
Chang
,
L. J.
,
Lindquist
,
M. A.
, &
Wager
,
T. D.
(
2017
).
Building better biomarkers: Brain models in translational neuroimaging
.
Nature Neuroscience
,
20
,
365
377
.
Xu
,
Y.
, &
Chun
,
M. M.
(
2006
).
Dissociable neural mechanisms supporting visual short-term memory for objects
.
Nature
,
440
,
91
95
.
Yamashita
,
M.
,
Yoshihara
,
Y.
,
Hashimoto
,
R.
,
Yahata
,
N.
,
Ichikawa
,
N.
,
Sakai
,
Y.
, et al
(
2018
).
A prediction model of working memory across health and psychiatric disease using whole-brain functional connectivity
.
eLife
,
7
,
e38844
.
Ystad
,
M.
,
Eichele
,
T.
,
Lundervold
,
A. J.
, &
Lundervold
,
A.
(
2010
).
Subcortical functional connectivity and verbal episodic memory in healthy elderly—A resting state fMRI study
.
Neuroimage
,
52
,
379
388
.