Content uploaded by Monica D. Rosenberg
Author content
All content in this area was uploaded by Monica D. Rosenberg on Jan 03, 2018
Content may be subject to copyright.
Connectome-based Models Predict Separable Components
of Attention in Novel Individuals
Monica D. Rosenberg
1
, Wei-Ting Hsu
1
, Dustin Scheinost
2
,
R. Todd Constable
1,2
, and Marvin M. Chun
1,2
Abstract
■Although we typically talk about attention as a single process,
it comprises multiple independent components. But what are
these components, and how are they represented in the functional
organization of the brain? To investigate whether long-studied
components of attention are reflected in the brain’s intrinsic func-
tional organization, here we apply connectome-based predictive
modeling (CPM) to predict the components of Posner and
Petersen’s influential model of attention: alerting (preparing
and maintaining alertness and vigilance), orienting (directing
attention to a stimulus), and executive control (detecting and
resolving cognitive conflict) [Posner, M. I., & Petersen, S. E.
The attention system of the human brain. Annual Review of
Neuroscience,13,25–42, 1990]. Participants performed the
Attention Network Task (ANT), which measures these three fac-
tors, and rested during fMRI scanning. CPMs tested with leave-
one-subject-out cross-validation successfully predicted novel
individual’s overall ANT accuracy, RT variability, and executive
control scores from functional connectivity observed during
ANT performance. CPMs also generalized to predict participants’
alerting scores from their resting-state functional connectivity
alone, demonstrating that connectivity patterns observed in
the absence of an explicit task contain a signature of the ability
to prepare for an upcoming stimulus. Suggesting that significant
variance in ANT performance is also explained by an overall sus-
tained attention factor, the sustained attention CPM, a model de-
fined in prior work to predict sustained attentional abilities,
predicted accuracy, RT variability, and executive control from
task-based data and predicted RT variability from resting-state
data. Our results suggest that, whereas executive control may
be closely related to sustained attention, the infrastructure that
supports alerting is distinct and can be measured at rest. In the
future, CPM may be applied to elucidate additional independent
components of attention and relationships between the func-
tional brain networks that predict them. ■
INTRODUCTION
The ability to pay attention differs substantially across
people. Although these individual differences are most
pronounced in clinical populations such as attention-
deficit/hyperactivity disorder, variation exists even among
healthy individuals. Variability exists, too, in the type of
attention with which people succeed or struggle. One
person, for example, might be unable to resist distraction
for long periods of time but easily switch between tasks,
whereas another might have no trouble maintaining un-
interrupted focus but suffer productivity costs when shift-
ing from one task to the next. Dissociations of this type
have been observed in action video game players, who,
compared with nongamers, show better visual-selective
attention (Green & Bavelier, 2003) but not more efficient
allocation of attention in response to external cues (Green
& Bavelier, 2012).
Although phenomenology suggests that focusing on a
lecture, braking quickly to avoid an obstacle in the road,
and scanning a crowd for a friend tax different forms of
attention, do these behaviors in fact rely on distinct
mechanisms? As suggested previously, psychology re-
search has long demonstrated that attention is not a sin-
gle monolithic ability but rather can be divided into
several subtypes (Chun, Golomb, & Turk-Browne,
2011). One recent study, for example, tested more than
200 participants on at least nine common attention tasks
and used cross-correlations to identify a “general attention
factor”related to intelligence, inhibition, and task switch-
ing, as well as several highly specific factors related to spa-
tial orienting, attentional capture, and inhibition of return
(Huang, Mo, & Li, 2012).
An earlier, influential model divided attention into three
major components: alerting, or preparing and maintaining
alertness and vigilance; orienting, or directing overt or
covert attention to a stimulus; and executive control, or
detecting and resolving cognitive conflict (Petersen &
Posner, 2012; Posner & Petersen, 1990). Behavioral studies
show that alerting, orienting, and executive control are
largely uncorrelated within participant (Fan, McCandliss,
Fossella, Flombaum, & Posner, 2005; Fan, McCandliss,
Sommer, Raz, & Posner, 2002; but see MacLeod et al.,
2010). Supporting these distinctions, neuroimaging stud-
ies have found that these components are related to ac-
tivity in different brain regions (Petersen & Posner, 2012;
1
Yale University,
2
Yale School of Medicine
© 2017 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 30:2, pp. 160–173
doi:10.1162/jocn_a_01197
Fan et al., 2005; Posner & Petersen, 1990) and integrity in
different white matter tracts ( Yin et al., 2012; Niogi,
Mukherjee, Ghajar, & McCandliss, 2010), as well as to cor-
tical thickness (Westlye, Grydeland, Walhovd, & Fjell,
2011), white matter asymmetry (Yin et al., 2013), and func-
tional connectivity in the dorsal attention and default mode
networks (Madhyastha, Askren, Boord, & Grabowski, 2015;
Visintin et al., 2015). Despite these models, however, the
comprehensive set of attentional components and their
underlying neural mechanisms are still debated.
A new approach in cognitive neuroscience, connectome-
based predictive modeling (CPM; Shen et al., 2017), may
help elucidate the independent factors of attention
(Rosenberg, Finn, Scheinost, Constable, & Chun, 2017).
This approach uses models based on functional brain
connectivity, or coordinated activity in spatially distinct
regions of the brain, to predict behavior in previously un-
seen individuals. One benefit of making predictions from
functional connectivity is that connectivity can be mea-
sured in resting-state fMRI data—that is, data collected
as participants simply lie still in the MRI scanner. Further-
more, from a broad perspective, models that make indi-
vidualized predictions from brain data can offer both
theoretical insights and translational applications ( Woo,
Chang, Lindquist, & Wager, 2017; Gabrieli, Ghosh, &
Whitfield-Gabrieli, 2015).
Using CPM, our group developed a model that predicts
how well novel individuals sustain attention based on
their unique functional connectivity patterns (Rosenberg,
Finn, et al., 2016). This model, the sustained attention
connectome-based predictive model, comprises two
functional networks whose strength is related to perfor-
mance on a challenging sustained attention task, the
gradCPT (Esterman, Noonan, Rosenberg, & Degutis,
2013; Rosenberg, Noonan, DeGutis, & Esterman, 2013).
Even during rest, high-attention network strength predicts
better gradCPT performance, and low-attention network
strength predicts worse gradCPT performance. The sus-
tained attention CPM has generalized to predict attention-
deficit/hyperactivity disorder symptom severity in children
(Rosenberg, Finn, et al., 2016) and stop signal task perfor-
mance in adults (Rosenberg, Zhang, et al., 2016), and is
sensitive to changes in attention function resulting from
pharmacological intervention (Rosenberg, Zhang, et al.,
2016). This model’s success demonstrates that sustained
attention can be measured while a person is not actively
paying attention, providing novel evidence that sustained
attentional abilities are reflected in the intrinsic functional
organization of the brain.
Although the sustained attention CPM shows promise
for quantifying the ability to maintain focus over time,
again, more than a century of behavioral research demon-
strates that sustained attention is just one aspect of human
attention (Chun et al., 2011). To investigate whether other
proposed facets of attention are (1) related to the sus-
tained attention factor described previously and (2) re-
flected in the brain’s intrinsic functional organization,
here we apply CPM to predict individual differences in
alerting, orienting, and executive control from task-based
and resting-state functional connectivity. Practically, this
effort represents a first step toward developing a suite of
models that predict a person’s overall attentional abilities.
Theoretically, it may shed light on the relationships (or
lack thereof) between proposed components of attention
and the functional networks that support them.
METHODS
Participants
Fifty-six participants from Yale University and the sur-
rounding community were scanned during Attention
Network Task (ANT) performance and rest. Eight individ-
uals were excluded for excessive head motion, defined
a priori as >2 mm translation, >3° rotation, or >0.15 mm
mean frame-to-frame displacement (Rosenberg, Zhang,
et al., 2016) in three or more of the six task runs, and four
were excluded for missing data from three or more task runs
because of insufficient scan time or technical issues such as
a failure to record button presses. The remaining 44 partic-
ipants, all of whom were right-handed and had normal or
corrected-to-normal vision, were included in the sub-
sequent analyses (29 women, ages 18–37 years, mean
age = 23.9 years). All participants gave written informed
consent in accordance with the Yale University human
subjects committee and were paid for their participation.
Task Paradigm
ANT trials began with a 200-msec cue period. On central-
cue trials, an asterisk appeared in the center of the
screen. On spatial-cue trials, an asterisk appeared above
or below a central fixation cross. (This cross was present
on the screen for the entire experiment.) On no-cue
trials, no cue appeared at all. After a variable SOA of
0.3–11.8 sec (mean = 2.8 sec), five arrows appeared
above or below the central fixation cross. Participants
were instructed to press a button with their right index
finger if the center arrow (the third in the row of five)
pointed to the right of the screen and to press a button
with their left index finger if the center arrow pointed to
the left. On congruent trials, the central or target arrow
pointed in the same direction as the surrounding distrac-
tor arrows. On incongruent trials, the target pointed in the
opposite direction as the four arrows that flanked it. After
a button press or 2 sec, the arrows disappeared, and a
variable intertrial interval (ITI; 5–17 sec, mean = 8 sec)
began.
Participants performed the ANT during six runs of fMRI
data collection. Each run consisted of two buffer trials
followed by 36 task trials. Task trials were divided equally
between the cue and target conditions, and each run in-
cluded six examples of each of the six trial types (3 cue
types: central-cue vs. spatial-cue vs. no-cue × 2 target
Rosenberg et al. 161
types: congruent vs. incongruent). Trial order was counter-
balanced within and across runs. Timing parameters, trial
order, and counterbalancing procedures exactly replicate
Fan et al. (2005), except that, in that article, the ITIs ranged
from 3 to 15 sec (mean = 6 sec). Here, we used the same
distribution of ITIs but increased the length of each by 2 sec
to facilitate functional connectivity analyses, which benefit
from longer time windows over which to calculate
correlations.
Behavioral Analysis
The ANT purports to measure three types of attention—
alerting, orienting, and executive control—by comparing
RTs on different trial types (Fan et al., 2002, 2005). For
each participant, alerting is measured as the difference
in mean correct-trial RT to no-cue and central-cue trials.
Because central cues provided information about when a
target will appear, we expect positive alerting scores—that
is, faster RTs to central-cue than no-cue trials. Orienting is
measured as the difference in mean correct-trial RT to
central-cue and spatial-cue trials. Unlike central cues, which
provide information about the timing but not spatial loca-
tion of targets, spatial cues always indicated whether the
upcoming target would appear above or below the central
fixation cross. Thus, we expect positive orienting scores or
faster RTs to spatial-cue than central-cue trials. Finally, ex-
ecutive control is measured as the difference in mean
correct-trial RT to incongruent and congruent trials. We
expect faster RTs to trials in which the distractor arrows
point in the same direction as the target arrow than to
those in which they point in the opposite direction,
and so positive executive control scores.
In addition to measuring alerting, orienting, and exec-
utive control, we assessed ANT accuracy as the percent-
age of correct responses. Because accuracy was near
ceiling, however, we used RT coefficient of variation (the
standard deviation divided by the mean of correct-trial
RTs) as a more sensitive measure of overall task perfor-
mance ( Wojtowicz, Omisade, & Fisk, 2013; Lundervold
et al., 2011; Adólfsdóttir, Sørensen, & Lundervold, 2008;
Kelly, Uddin, Biswal, Castellanos, & Milham, 2008). On
cognitive and attentional tasks, higher intraindividual
response variability indicates less successful sustained
attention or attentional control (Rosenberg et al., 2013;
Sonuga-Barke & Castellanos, 2007; MacDonald, Nyberg,
& Bäckman, 2006; Castellanos et al., 2005; Stuss, Murphy,
Binns, & Alexander, 2003). On the ANT in particular, greater
RT standard deviation has been associated with attention
impairments in attention-deficit/hyperactivity disorder
(Adólfsdóttir et al., 2008) and multiple sclerosis ( Wojtowicz
et al., 2013).
Imaging Parameters and Preprocessing
Experimental sessions began with a high-resolution ana-
tomical scan, followed by two 6-min resting-state scans
and six 7:05-min ANT runs. Participants were instructed
to fixate on a cross presented in the center of the screen
during rest runs and to respond to arrow stimuli as
quickly and accurately as possible during task runs.
fMRI data were collected at the Yale Magnetic Reso-
nance Research Center on a 3T Trio TIM system (Siemens,
Erlangen, Germany). Resting-state runs included 360
whole-brain volumes, and task runs included 425 volumes.
Both were acquired using a multiband EPI sequence with
the following parameters: repetition time = 1000 msec,
echo time = 30 msec, flip angle = 62°, acquisition matrix =
84 × 84, in-plane resolution = 2.5 mm
2
, 51 axial-oblique
slices parallel to the AC-PC line, slice thickness = 2.5 mm,
multiband 3, acceleration factor = 2. MPRAGE parameters
were repetition time = 2530 msec, echo time = 2.77, flip
angle = 7°, acquisition matrix = 256 × 256, in-plane res-
olution = 1.0 mm
2
, slice thickness = 1.0 mm, 176 sagittal
slices. A 2-D T1-weighted image with the same slice pre-
scription as the EPI images was also collected for regis-
tration purposes.
Data were analyzed using BioImage Suite ( Joshi et al.,
2011) and custom scripts in MATLAB (The MathWorks,
Natick,MA).Motioncorrectionwasperformedusing
SPM8. Nuisance regressors including linear and quadratic
drift, mean signal from cerebrospinal fluid, global signal,
and a 24-parameter motion model (six motion parame-
ters, six temporal derivatives, and their squares) were re-
moved from the data. Data were temporally smoothed
with a zero mean unit variance Gaussian filter (cutoff
frequency = 0.12 Hz).
Global signal regression was included as a preprocess-
ing step because of its well-established role in reducing
the confounding effects of motion in functional connec-
tivity data (Ciric et al., 2017; Power, Plitt, Laumann, &
Martin, 2017; Power, Schlaggar, & Petersen, 2015; Power
et al., 2014; see Motion Controls section for detail). One
concern about global signal regression, that it may induce
false negative functional connections (edges), is not rele-
vant here because we do not interpret edge sign. Instead,
in taking an individual differences approach, we only con-
sider the relative strength of edges across participants.
Data Exclusion
In the final set of 44 participants, runs with excessive
head motion, defined a priori as >2 mm translation,
>3° rotation, or 0.15 mm mean frame-to-frame displace-
ment (Rosenberg, Finn, et al., 2016; Rosenberg, Zhang,
et al., 2016), were excluded from analysis. One task run
was excluded from three participants, and two task runs
were excluded from one participant for excessive motion.
Because of technical issues or insufficient scan time, two
additional participants were missing 205 and 81 volumes,
respectively, from one task run; five participants were
missing one task run; one participant was missing two
task runs; and one participant was missing three task
runs. All participants had two rest runs with acceptable
162 Journal of Cognitive Neuroscience Volume 30, Number 2
levels of motion. All exclusion was performed before
functional connectivity data were analyzed.
Network Construction
Whole-brain functional connectivity networks were defined
as described previously (Rosenberg, Finn, et al., 2016;
Rosenberg, Zhang, et al., 2016; Finn et al., 2015). Briefly,
network nodes were defined with the Shen 268-node
functional brain atlas (Shen, Tokoglu, Papademetris, &
Constable, 2013; https://www.nitrc.org/frs/?group_id=51).
The atlas was warped from MNI space into individual-
subject space with a concatenation of linear and non-
linear registrations between the functional images, 2-D
and 3-D anatomical scans, and the MNI brain. Transfor-
mations, which were estimated using intensity-based reg-
istration algorithms in BioImage Suite (New Haven, CT),
were calculated independently, combined into a single
transform, and inverted.
For each participant, a task matrix was calculated using
data concatenated across task runs, and a rest matrix was
calculated using data concatenated across resting-state runs.
The first eight frames (8 sec) of each run were excluded
from analysis. A mean time course for each node was then
calculated by averaging the time courses of all voxels in the
node. The time courses of every pair of nodes were corre-
lated, and the resulting Pearson correlation coefficients
were Fisher z-transformed to yield symmetric 268 × 268
functional connectivity matrices. Cells in these matrices rep-
resent functional connections or edges.
We generated one task matrix rather than six condition-
specific matrices (i.e., no-cue, center-cue, spatial-cue,
congruent-target, and incongruent-target matrices) for
each individual for several reasons. This approach maxi-
mizes the amount of data used to calculate correlations
and thus the reliability of functional connectivity esti-
mates (Noble et al., 2017; Shah, Cramer, Ferguson, Birn,
& Anderson, 2016; Birn et al., 2013). It also facilitates test-
ing models on resting-state data, from which condition-
specific connectivity patterns cannot be calculated.
Furthermore, tasks like the ANT may serve as attentional
“stress tests”that put participants in a task-engaged state,
perturb attention-related brain circuitry, and enhance in-
dividual differences in behaviorally relevant patterns of
functional connectivity (Finn et al., 2017). Here, we inves-
tigated whether functional connectivity observed during
task-engaged and task-free states predicts individual differ-
ences in attention. Future work could explore whether, for
example, functional connectivity patterns observed during
single trials or trial types of a cognitive or attentional task
predict behavior.
Connectome-based Predictive Modeling
We used a recently developed technique, CPM (Shen
et al., 2017; Rosenberg, Finn, et al., 2016; Finn et al.,
2015), to predict individual differences in alerting, orient-
ing, executive control, percent accuracy, and RT variability
on the ANT. CPM includes three steps: feature selection,
model building, and model validation.
First, we selected one of the five behavioral variables of
interest: accuracy, RT variability, alerting, orienting, and
executive control. We then set aside data from one par-
ticipant. Using task matrices from the remaining 43 par-
ticipants, we identified functional connections related to
the selected performance measure by correlating every
edge in the matrices with behavior across participants.
We used Spearman’s rank correlation at this step because
two of the five behavioral scores (accuracy and executive
control) were not normally distributed (Jarque-Bera test
pvalues < .041). Correlation coefficients were thresh-
olded at p< .05 and separated into a positive tail
(edges greater in individuals with higher behavioral
scores) and a negative tail (edges greater in individuals
with lower behavioral scores). We applied an edge
threshold of p< .05 rather than p< .01, which was
applied previously (Rosenberg, Finn, et al., 2016), be-
cause approximately 27.3% of all possible connections
in the brain were excluded from analysis due to corre-
lations with motion parameters (see Motion Controls
section). The choice of this feature selection threshold
does not affect the validity of predictions, because cross-
validation is a built-in guard against false-positive results.
Positive and negative network strengths, summary
measures that represent overall functional connectivity
strength, were calculated for each training subject by sum-
ming edges in the tails (i.e., taking a dot product between
the positive and negative tail masks and each training sub-
ject’staskmatrix).
Next, we defined a linear model relating positive and
negative network strength to behavior in the training
set. Because Spearman’s correlations and rank were
used, this model included coefficients for positive and
negative network strength but not an intercept (Shen
et al., 2017). The left-out subject’s positive and negative
network strengths, measured during task or rest, were
input into the model to generate a predicted behavioral
score.
We repeated this procedure so that each participant
was left out of the training set once and measured
models’predictive power by correlating the predicted
values and observed behavioral scores, controlling for
motion (see Motion Controls section below). Spearman’s
rho (r
s
) was used to evaluate model performance because
it is less sensitive to the effect of outliers than Pearson’sr
and because CPM predictions are best considered relative
rather than absolute (Rosenberg, Finn, et al., 2016). Per-
mutation (i.e., randomization) testing was used to assess
significance because standard r-to-pconversions assume
that the number of degrees of freedom is equal to the
number of participants −2, and this assumption is vio-
lated in leave-one-out cross-validation because folds are
not independent (e.g., Rosenberg, Finn, et al., 2016). To
Rosenberg et al. 163
perform permutation testing, we randomly shuffled partic-
ipants’behavioral scores 1000 times and ran these shuffled
values through our prediction pipeline to generate 10 null
distributions (a task-based distribution and a resting-state
distribution for each of our five behavioral variables).
pValues associated with each model were based on the
corresponding null distribution with the formula p=
(1 + the number of permutation r
s
values greater than
or equal to the observed r
s
value)/1001.
Finally, the entire pipeline was repeated to define one
CPM for each of the five behavioral measures.
Motion Controls
Because head motion can confound functional connectiv-
ity analyses, we investigated whether measures of head
motion were correlated with behavior across participants
(Table 1).
Given the nonzero correlations between some head
motion parameters and behavior, we excluded any edge
in the task and rest matrices that was correlated with
motion across participants. That is, we correlated the
strength of every edge in the task matrices with maxi-
mum displacement, maximum rotation, and mean
frame-to-frame displacement during task across partici-
pants using Spearman’s correlation. We also correlated
the strength of every edge in the rest matrices with mo-
tion measured during rest. Edges related to motion at
p< .05 were removed, or masked, in every participant’s
task and rest matrices. This step resulted in the removal
of 9772 edges from the matrices. (When global signal re-
gression is not included as a preprocessing step but the
rest of the pipeline is identical, 17,688 edges [or 49.44%]
of all possible functional connections are excluded based
on the same criteria. Thus, global signal regression effec-
tively reduces the relationship between head motion and
functional connectivity in this data set.) 26,006 edges or
72.7% of the original 35,778 edges remained. Given the
large number of multiple comparisons, this step almost
certainly resulted in the removal of edges related to
motion parameters by chance. However, given the sub-
stantial confounding effects of head motion on functional
connectivity estimates (van Dijk,Sabuncu,&Buckner,
2012), this conservative edge exclusion step is justified.
Finally, after eliminating fMRI runs with excessive head
motion, correcting for head motion during preprocess-
ing, and excluding edges related to motion parameters
across participants, we controlled for head motion at
the model evaluation step. Specifically, we evaluated
model predictions with Spearman’s partial correlations
between observed and predicted behavioral scores, con-
trolling for motion parameters. Partial correlations be-
tweenobservedscoresandtask-basedpredictions
included maximum displacement, rotation, and mean
frame-to-frame displacement during task runs. Correla-
tions between observed scores and rest-based predic-
tions included these three measures from both task
and resting-state runs, because models were trained
using task data and tested using rest.
Predictive Network Anatomy
To investigate network anatomy, we first defined com-
mon networks for each behavioral measure that was suc-
cessfully predicted from functional connectivity. That is,
because network models were defined using leave-one-
subject-out cross-validation, there were as many different
models as there were rounds of cross-validation. Because
our data set included 44 participants, there were 44
unique models per behavioral measure. For each mea-
sure of behavior, we retained edges that appeared in
every round of cross-validation—those edges that ap-
peared in all 44 positive tails and all 44 negative tails—
for further analysis (Rosenberg, Finn, et al., 2016). This
Table 1. Correlation between Motion Parameters and Behavior
During Task Runs During Rest Runs
Maximum
Displacement
Maximum
Rotation
Mean Frame-to-Frame
Displacement
Maximum
Displacement
Maximum
Rotation
Mean Frame-to-Frame
Displacement
Accuracy −.04 −.18 −.09 −.29 −.22 −.21
RT variability .18 .26 .08 .31* .23 .09
Alerting .07 .22 .27 .00 .10 .25
Orienting −.18 −.09 −.05 −.26 .01 −.10
Executive
control .32* .29 .22 .17 .21 .22
Cells contain Spearman’s correlation coefficients. Maximum displacement is the average maximum displacement in millimeters across runs; maxi-
mum rotation is the average maximum rotation in degrees across runs. Mean frame-to-frame displacement is the average displacement in millimeters
from one frame to the next.
*p< .05 uncorrected.
164 Journal of Cognitive Neuroscience Volume 30, Number 2
resulted in a common positive network (edges that consis-
tently predicted higher behavioral scores) and a common
negative network (edges that consistently predicted lower
behavioral scores) for each measure. Although we could
have visualized edges that appeared in any round of
leave-one-out cross-validation, we elected to apply this
conservative edge retention step to identify edges most
consistently related to behavior and facilitate concise
visualization.
We visualized common positive and negative network
anatomy by grouping nodes into macroscale brain re-
gions, which included prefrontal, motor, parietal, tempo-
ral, occipital, and limbic cortex, as well as the insula,
cerebellum, subcortex, and brainstem (Finn et al., 2015).
Sustained Attention CPM
Previous work using CPM demonstrated that two large-
scale functional connectivity networks, the high-attention
network and the low-attention network, predict sustained
attention performance in several data sets (Rosenberg,
Finn, et al., 2016; Rosenberg, Zhang, et al., 2016). To test
whether this model, the sustained attention CPM, pre-
dicts ANT performance in the current data set, we applied
it to each of 41 participant’s task and rest matrices. (Data
from three participants who had previously participated
in the study used to define the sustained attention CPM
were excluded from this analysis.) In other words, we
calculated the dot product between the high- and low-
attention network masks and each task or rest matrix to
measure high- and low-attention network strength for
each participant. Edges removed from ANT matrices for
relationships with motion were excluded from analysis,
leaving 478 of the original 757 high-attention network
edges and 412 of the original 630 low-attention network
edges. We input these network strength values into the
sustained attention CPM, a linear model with coefficients
for high- and low-attention network strength, to generate
a predicted behavioral score. This predicted score cor-
responds to how well that participant would hypotheti-
cally perform on a challenging continuous performance
task—in essence, a measure of their overall ability to
sustained attention (Rosenberg, Finn, et al., 2016).
We assessed predictive power with Spearman’s partial
correlations between sustained attention CPM predic-
tions, observed ANT performance, and head motion
parameters. Partial correlations between behavior and
sustained attention CPM predictions from task data
included maximum displacement, rotation, and mean
frame-to-frame displacement during task runs; partial cor-
relations between behavior and sustained attention CPM
predictions from resting-state data included these mea-
sures during rest runs. pvalues were calculated with two-
tailed permutation tests. To generate null distributions,
ANT scores and motion measures were shuffled together
across participants and correlated with observed high-
and low-attention network strength values 100,000 times.
Relationships between Predictive Networks
To investigate the relationship between networks pre-
dicting overall ANT performance (accuracy and RT vari-
ability) and the sustained attention networks defined
previously (Rosenberg, Finn, et al., 2016), we counted
the number of edges common to each ANT and sus-
tained attention network pair. We assessed the statistical
significance of edge overlap with the hypergeometric
cumulative density function (Rosenberg, Zhang, et al.,
2016). This was implemented in MATLAB with p=1−
hygecdf(x,M,K,N), where x= number of overlapping
edges, M= 26,006, the total number of edges, K= num-
ber of edges in the first network of interest, N= number
of edges in the second network of interest, and p= the
probability of observing up to xof Kpossible edges in
Ndrawings without replacement from a full set of M
edges (The MathWorks Inc., 2016). For this analysis,
high-attention network sizewas478(thenumberof
edges that could have possibly overlapped with ANT net-
works due to motion-related edge removal), and low-
attention network size was 412.
Significance of the overlap between three networks
was assessed with 100,000 permutation tests. For each
permutation, we randomly generated six networks from
26,006 possible edges whose size matched the high- and
low-attention, high- and low-accuracy, and low- and high-
RT variability networks. We counted the number of edges
that appeared in the high-attention, high-accuracy, and
low-RT variability networks as well as the number of
edges that appeared in the low-attention, low-accuracy,
and high-RT variability networks. Significance of ob-
served overlap was computed as p= (1 + the number
of permutations on which three-network overlap was
greater than or equal to the observed overlap)/100,001.
To account for the fact that accuracy and RT variability
networks were not independent because they were de-
fined using inversely correlated behavioral variables, we
confirmed that results did not change if we performed
this analysis with the constraint that the smaller of the
high-accuracy/low-RT variability and low-accuracy/high-
RT variability network pairs be a subset of the larger
(i.e., the most extreme possible scenario).
We expected significant overlap between networks that
predicted more successful attention (the high-attention
network defined previously and networks predicting
higher accuracy and lower RT variability) and the net-
works that predicted worse attention (the low-attention
network defined previously and networks predicting
lower accuracy and higher RT variability).
RESULTS
Task Performance
Mean accuracy was 95.69% (SD = 5.88%), mean RT was
754 msec (SD = 108 msec), and RT variability (coefficient
of variation) was 21.66% (SD = 5.03%). As predicted,
Rosenberg et al. 165
alerting scores were positive, meaning that participants
were faster to respond to center-cue than to no-cue trials
(M= 39.99 msec, SD = 34.38 msec). Average orienting
score was also positive, indicating that participants were fas-
ter to respond on spatial-cue than center-cue trials (M=
57.38 msec, SD = 34.85 msec). Positive executive control
scores (M= 120.10 msec, SD = 57.04 msec) indicated
that, as expected, participants were faster to respond to
congruent than incongruent trials (Figure 1).
Previous work arguing that alerting, orienting, and
executive control measure independent components of
attention has demonstrated that the scores are largely
uncorrelated within participants (Fan et al., 2002, 2005).
We replicated this result and observed a relationship be-
tween higher alerting and executive control scores and
worse overall ANT performance. In addition, we saw a
strong negative correlation between accuracy and RT
variability, providing further evidence that the coefficient
of variation is a sensitive measure of overall attentional
performance (Table 2).
Behavioral Predictions
We used CPM to predict individual differences in accuracy,
RT variability, alerting, orienting, and executive control.
Models trained and tested on task-based functional connec-
tivity matrices significantly predicted novel individuals’
accuracy, RT variability, and executive control scores, but
not alerting or orienting scores (Table 3, Column 1).
To test whether models generalized to predict behav-
ior from resting-state functional connectivity patterns,
weappliedmodelstrainedontask-baseddatatoleft-
out participants’rest matrices. The alerting model suc-
cessfully predicted scores from resting-state data alone
(Table 3, Column 2). Interestingly, the alerting model was
not significant in the task-based data (although results
Figure 1. Histogram of ANT performance.
Table 2. Correlations between Behavioral Measures
Accuracy RT Variability Alerting Orienting Executive Control
Accuracy 1
RT variability −.59** 1
Alerting −.005 .32* 1
Orienting .10 .08 .18 1
Executive control −.32* .55** .06 −.14 1
Cells show Spearman’s correlation coefficients.
*p< .05 uncorrected.
**p< .05 Bonferroni-corrected for 10 comparisons.
166 Journal of Cognitive Neuroscience Volume 30, Number 2
trend toward significance). Furthermore, although accu-
racy and RT variability models were the most successful
on task-based data, they did not generalize to resting-state
data. Thus, models that perform well on task data may not
generalize to rest, whereas models that perform less well
on task data may successfully generalize to predict behav-
ior from functional connectivity observed at rest.
Predictive Network Anatomy
We examined the functional anatomy of networks that
significantly predicted ANT performance in novel partici-
pants: the accuracy, RT variability, alerting, and executive
control networks (Figure 2). For each behavioral mea-
sure, there are two predictive networks: one of edges
positively correlated with behavior in every round of
leave-one-subject-out cross-validation and one of edges
negatively correlated with behavior in every round of
leave-one-out cross-validation. By definition, positive net-
works are stronger in individuals with higher behavioral
scores, whereas negative networks are stronger in indi-
viduals with lower behavioral scores.
The anatomy of the accuracy, variability, and executive
control networks was broadly similar, an expected finding
given that the behavioral measures themselves are cor-
related in this data set. For all three measures, contralateral
and ipsilateral intracerebellar and intratemporal connec-
tions predicted worse attention (lower accuracy; higher
variability, and higher executive control scores, which cor-
respond to lower executive control abilities), and contra-
lateral and ipsilateral cerebellar–temporal connections
predicted better attention. Cerebellar–temporal edges
were the most common edge in the high-accuracy, low-
variability, and low-executive control score networks,
making up over 9.8% of the edges in each. Subcortical–
cerebellar edges were most common in the low-accuracy
network (11.6% of all edges), intracerebellar edges were
most common in the high-variability network (8.6%), and
subcortical–prefrontal edges were most common in the
high-executive control score network (6.9%). The impor-
tance of the cerebellum in networks that predict ANT
performance underscores a growing appreciation for its
role in attention and high-level cognition (Buckner, 2013;
Castellanos & Proal, 2012; Stoodley, 2012), and the out-
sized role of intracerebellar connections, in particular,
replicates a trend previously observed in the sustained
attention CPM’s low-attention network (Rosenberg, Finn,
et al., 2016).
The alerting network shared several anatomical trends
with the accuracy, variability, and executive control net-
works. In particular, intratemporal connections predicted
higher scores, whereas temporal–cerebellar connections
predicted lower scores. Temporal–motor and occipital–
parietal connections also predicted higher alerting
scores. Previous work has found that fMRI activity in pa-
rietal regions of the dorsal visual stream is related to
alerting (Petersen & Posner, 2012; Fan et al., 2005;
Posner & Petersen, 1990); here, connections between
the occipital and parietal cortices may also reflect coordi-
nated dorsal stream activity. Interestingly, these connections
predict individual differences in alerting during rest, sug-
gesting that resting-state functional connectivity between
dorsal stream regions may support attentional abilities re-
lated to arousal and vigilance. Finally, the most common
connections in the low-alerting score network were be-
tween the cerebellum and temporal lobes, and the most
common connections in the high-alerting score network
were between temporal and motor cortex, potentially reflect-
ing the coordination of responses during task performance.
External Validity: Sustained Attention
CPM Predictions
As a demonstration of external validity, the sustained
attention CPM, predefined in Rosenberg, Finn, et al.
(2016), generalized to predict several measures of ANT per-
formance from task-based and resting-state functional con-
nectivity. Predictions from task matrices were positively
correlated with accuracy and negatively correlated with
RT variability and executive control scores, and predic-
tions from rest matrices were negatively correlated with
variability (Table 4). Inverse correlations between sus-
tained attention CPM predictions and variability and exec-
utive control scores were expected given that higher
predictions correspond to better attention (that is, better
predicted performance on a sustained attention task),
whereas higher variability and executive control scores cor-
respond to worse attention. The relationship between sus-
tained attention CPM predictions and ANT performance,
Table 3. Correlations between Predicted and Observed
Behavioral Scores
Predictions from
Task-based
Functional
Connectivity
Predictions from
Resting-state
Functional
Connectivity
Accuracy r
s
= .62, p= .001* r
s
= .12, p= .31
RT variability r
s
= .63, p= .001* r
s
= .25, p= .10
Alerting r
s
= .28, p= .10 r
s
= .31, p= .036*
Orienting r
s
= .18, p= .23 r
s
= .19, p= .16
Executive
control r
s
= .35, p= .042* r
s
= .05, p= .42
To further control for potentially confounding effects of motion, we performed
Spearman’s partial correlations between observed and predicted behavioral
scores and motion. Correlations between observed scores and task-based predic-
tions (left column) included maximum displacement, rotation, and mean frame-
to-frame displacement during task runs. Correlations between observed scores
and resting-state predictions (right column) included these three measures from
both task and resting-state runs, because models were trained using task data and
tested using rest. Predictions were made with general linear models with two pre-
dictors, positive and negative network strength. pvalues were determined with
permutation testing.
*Significant at p< .05.
Rosenberg et al. 167
especially RT variability, suggests that significant variance
in performance on this task is explained by an overall sus-
tained attention factor.
Relationship between Predictive Networks
As predicted, we observed significant overlap between
the high-attention, high-accuracy, and low-RT variability
networks. We also observed significant overlap between
the low-attention, low-accuracy, and high-RT variability
networks (Figure 3). Given that accuracy and variability
were inversely correlated in this data set, the overlap be-
tween networks predicting these measures was expected.
The high- and low-attention networks, however, were de-
fined to predict a different attention task in a completely
independent data set. Along with results showing that
these networks generalize to predict ANT performance,
their similarity to networks explicitly defined to predict
ANT behavior suggests that a common sustained atten-
tion factor supports the ability to successfully complete
the ANT and the gradCPT, the task used to define the
sustained attention CPM.
Figure 2. Networks predicting
ANT performance. Each point
on the circle represents a node
in the Shen et al. (2013) brain
atlas. Lines between the
nodes represent functional
connections; opaque lines
represent the 10% of edges with
the strongest correlation to
behavior, averaged across leave-
one-out folds. Nodes are
grouped by hemisphere and
then by macroscale brain
region. Networks whose
strength predicts better
performance on the ANT
(higher accuracy, less erratic
responses, and lower alerting
and executive control scores)
are shown in red, and networks
whose strength predicts worse
ANT performance are shown
in blue. Matrix plots show
differences in the number of
edges between each pair of
macroscale regions, calculated
by subtracting the number
of edges in “unsuccessful
attention”network from the
number in the corresponding
“successful attention”network.
Orienting networks are not
visualized because orienting
scores were not significantly
predicted from task-based or
resting-state functional
connectivity.
168 Journal of Cognitive Neuroscience Volume 30, Number 2
There was no significant overlap in the unexpected
directions. That is, the following network pairs only had
one edge in common: high-attention and low-accuracy,
high-attention and high-RT variability, low-attention and
high-accuracy, and low-attention and low-RT variability.
This degree of overlap was not significant in any case
(p> .99). The edges appearing in the low-attention
and high-accuracy networks and the low-attention and
low-RT variability networks were the only two over-
lapping edges not pictured in Figure 3 due to visualiza-
tion constraints.
Edges that appear in all three “successful attention”
networks (high-attention, high-accuracy, and low-RT var-
iability) and all three “unsuccessful attention”networks
(low-attention, low-accuracy, and high-RT variability)
may represent core functional systems supporting the
ability to sustain attention to task. With the exception
of temporal–motor connections, connections between
occipital, temporal, motor, and cerebellar regions were
morecommoninthesuccessfulthantheunsuccessful
core attention network. Given that predictive networks
were defined as participants performed visuomotor
attention tasks, these edges may represent response co-
ordination. Intratemporal and intracerebellar connec-
tions, on the other hand, were more common in the
unsuccessful core attention network (Figure 4).
DISCUSSION
In their influential model of attention, Posner and Petersen
propose that attending and responding to the environ-
ment is accomplished with three distinct processes: re-
maining alert to upcoming stimuli, orienting toward
salient or behaviorally relevant stimuli, and identifying
and resolving cognitive conflicts (Petersen & Posner,
2012; Posner & Petersen, 1990). Although previous work
has identified fMRI activity correlates of alerting, orienting,
and executive control and has related individual differ-
ences in these abilities to functional and structural features
of the brain, these relationships have not been leveraged
to predict abilities in novel individuals. To investigate
whether models based on functional brain connectivity
predict individual differences in alerting, orienting, and
executive control, we applied CPM to performance on
the ANT.
Models trained on 43 participants’task-based con-
nectivity and applied to the left-out participant’stask
data successfully predicted overall ANT performance,
measured by accuracy and RT variability, and executive
control scores. In addition, models generalized to pre-
dict left-out participants’alerting scores from resting-
state connectivity alone. Thus, signatures of the ability
to prepare for upcoming behaviorally relevant stimuli
(measured by alerting scores) and detect and resolve
cognitive conflict (measured by executive control
scores) are reflected in patterns of coordinated activity
across the cortex, subcortex, and cerebellum during
task performance or rest. Furthermore, predictions
from resting-state data suggest that the intrinsic func-
tional architecture supporting alerting is distinct from
that supporting the other components of attention.
Orienting scores were not significantly predicted for
novel individuals. It is perhaps unsurprising that orient-
ing results did not follow the same pattern as other
Table 4. Correlations between Sustained Attention CPM Predictions
and ANT Behavior
Predictions from Task-based
Functional Connectivity
Predictions from Resting-state
Functional Connectivity
Accuracy r
s
= .49, p= .0024* r
s
= .09, p= .5844
RT variability r
s
=−.68, p= 9.99 × 10
−6
*r
s
=−.36, p= .0288*
Alerting r
s
=−.28, p= .0899 r
s
=−.08, p= .6166
Orienting r
s
=−.12, p= .4741 r
s
=−.05, p= .7782
Executive control r
s
=−.34, p= .0385* r
s
=−.16, p= .3244
Cells show Spearman’s partial correlations between predicted and ob-
served behavioral scores, controlling for head motion. Forty-one novel
participants (out of 44) who had not previously participated in Rosenberg,
Finn, et al. (2016) were included in this analysis. Negative correlations
between RT variability and ANT component scores were expected, given
that higher sustained attention CPM predictions correspond to better
attention, whereas higher variability and alerting, orienting, and exec-
utive control scores correspond to worse attention.
*Significant at p< .05.
Figure 3. Overlap between predictive networks. Colored circles
represent six networks predicting overall attentional performance.
Circles are sized according to the total number of edges in each
network, and labels indicate the number of edges in each cell. The
significance of overlap between each ANT and sustained attention
network pair was determined with the hypergeometric cumulative
density function. For this calculation, we considered network overlap
the total number of edges appearing in both networks. For example,
168 edges (91 + 77) appeared in both the high-attention and low-RT
variability networks. Based on permutation testing, both instances of
three-network overlap are also significant. *p< 9.99 × 10
−6
.
Rosenberg et al. 169
measures of attention, as evidence suggests that orient-
ing abilities are dissociated from a general attention fac-
tor that underlies performance on many attention tasks
(Huang et al., 2012). One possible reason that orienting
models failed to generalize is that orienting scores them-
selves are unreliable. Recently, Wang and colleagues
found that orienting score reliability was the lowest of
the three ANT components (Wang et al., 2015), and a
meta-analysis of 15 ANT studies estimated the reliability or
orienting scores at r= .32, considered low (alerting score
reliability, however, was also considered low; MacLeod
et al., 2010). Another possibility is that the neural mecha-
nisms of orienting are simply not reflected in functional
connectivity patterns consistently across individuals. Func-
tional connectivity analyses, rather, may be best suited for
ongoing processes such as sustaining attention and main-
taining alertness rather than attention reorienting, a more
transient process.
It is worth pointing out that the predictive power of all
models introduced here, not just the orienting model, is
limited by a variety of factors. Theoretical ceilings on pre-
dictive power are set by the reliability of behavioral and
imagingmeasuresaswellasthehypothetical“ground
truth”relationship between individual differences in
attention and pairwise BOLD signal time-series correla-
tions. Practical limits are also imposed by study features
such as sample size, participant population, data quality,
scan duration, and modeling method. Future work can ex-
plore ways to optimize models for individualized attention
predictions, perhaps by leveraging large, open-access data
sets with appropriate behavioral measures, applying alter-
native prediction algorithms, and/or considering multiple
imaging and behavioral measures simultaneously.
To investigate the relationship between ANT perfor-
mance and a general sustained attention factor, the sus-
tained attention CPM, a model defined in a completely
Figure 4. Core attention networks. A, B, and C show the 91 edges present in all three “successful attention”networks (high-attention, high-
accuracy, and low-RT variability) and the 85 edges present in all three “unsuccessful attention”networks (low-attention, low-accuracy, and high-RT
variability). The successful attention network is visualized in red and the unsuccessful network in blue. (A) Lines represent edges. Spheres, or
nodes, are sized according to the number of edges in which they participate and colored according to the network in which they have more
connections. (B) Differences in the number of edges between regions. Cells are colored according to the difference in the number of edges in
the successful and unsuccessful attention networks. (C) Each semicircle represents a hemisphere of the brain. Nodes are grouped by brain
region. Lines, colored by network membership, represent edges.
170 Journal of Cognitive Neuroscience Volume 30, Number 2
independent data set, was applied to the current ANT
data. Predictions from task-based data were positively
correlated with ANT accuracy and inversely correlated
with RT variability and executive control scores. Predic-
tions from resting-state data were inversely correlated
with RT variability and were, surprisingly, even more ac-
curate than predictions of the leave-one-subject-out
model explicitly defined to predict erratic responding.
(One potential explanation for this result is that the sus-
tained attention CPM was defined using data collected
as participants in a previous study performed a taxing
go/no-go task, which could have amplified individual dif-
ferences in behavior and underlying network connectiv-
ity more effectively than did the ANT, resulting in a model
with stronger predictive power.) In other words, when the
sustained attention CPM predicted that a participant had
strong sustained attentional abilities, that person would
go on to perform well on the ANT. He or she would also
show low executive control scores, or relatively small dif-
ferences in RT between trials with distracting flanker stim-
uli that were congruent or incongruent with a central
target. Anatomically, there was significant overlap between
the sustained attention CPM’s high-attention network and
the networks predicting better overall ANT performance
and significant overlap between the low-attention network
and the networks predicting worse ANT performance. To-
gether, these results underscore the fact that overall ANT
performance, and potentially performance on a diverse
range of psychological tasks, can be largely explained by a
person’s ability to sustain attention. Furthermore, the rela-
tionship between sustained attention network strength and
executive control, but not alerting, further reinforces the in-
dependence of the alerting component from other factors
of attention. Given that alerting scores are thought to mea-
sure the preparation and maintenance of focus whereas ex-
ecutive control scores are thought to measure conflict
detection and resolution, this perhaps counterintuitive re-
sult should be investigated in future studies.
The overlap between the sustained attention and ANT
networks—the core “successful”and “unsuccessful”atten-
tion networks visualized in Figure 4—may constitute a
neuromarker of sustained attention function with im-
proved generalizability. In other words, by constraining
model features to edges that predict the ability to focus
on multiple attention tasks, we may be homing in on net-
works with the most robust and reliable relationships to
behavior. Thus, new “core”or “overlap”models may gen-
eralize to novel data sets more successfully than models
defined using a single data set. To our knowledge, the cur-
rent study represents the first attempt to consider the fea-
tures in common between distinct but related predictive
network models and may be an important approach in
building and improving new models of attention and cog-
nition. Moreover, together with existing models of the
ability to sustain attention (Kessler, Angstadt, & Sripada,
2016; Rosenberg, Finn, et al., 2016) and suppress distrac-
tors (Poole et al., 2016), the alerting and executive control
models represent preliminary steps toward developing a
set of models that predict different components of atten-
tion function from a single connectivity matrix (Rosenberg
et al., 2017). Future studies should test the predictive
power of current network models of attention using novel
data sets to validate and refine them.
In summary, we demonstrated for the first time that
components of a highly influential model of attention
can be predicted from brain data for novel individuals.
Specifically, patterns of functional connectivity observed
during attention task performance predict individual par-
ticipants’overall performance and executive control abili-
ties, and patterns of connectivity observed during the
resting state predict alerting. In concert with the observed
relationship between sustained attention CPM predictions
and individual differences in executive control, our results
provide evidence that, whereas executive control may be
closely related to sustained attention, the functional infra-
structure of alerting is distinct and can be measured at
rest. In the future, the CPM approach may be applied to
investigate other factors of attention and cognition and
shed light on the relationships between them.
Acknowledgments
This work was supported by the Yale FAS MRI Program funded
by the Office of the Provost and the Department of Psychology,
a Nation Science Foundation Graduate Research Fellowship
to M. D. R., and National Institutes of Health MH108591 to
M. M. C., EB009666 to R. T. C., and T32 DA022975 to D. S.
Reprint requests should be sent to Monica D. Rosenberg, Depart-
ment of Psychology, Yale University, 2 Hillhouse Avenue, New
Haven, CT 06511, or via e-mail: monica.rosenberg@yale.edu.
REFERENCES
Adólfsdóttir, S., Sørensen, L., & Lundervold, A. J. (2008). The
Attention Network Test: A characteristic pattern of deficits
in children with ADHD. Behavioral and Brain Functions,
4, 9.
Birn, R. M., Molloy, E. K., Patriat, R., Parker, T., Meier, T. B.,
Kirk, G. R., et al. (2013). The effect of scan length on the
reliability of resting-state fMRI connectivity estimates.
Neuroimage, 83, 550–558.
Buckner, R. L. (2013). The cerebellum and cognitive function:
25 years of insight from anatomy and neuroimaging. Neuron,
80, 807–815.
Castellanos, F. X., & Proal, E. (2012). Large-scale brain systems
in ADHD: Beyond the prefrontal-striatal model. Trends in
Cognitive Sciences, 16, 17–26.
Castellanos, F. X., Sonuga-Barke, E. J. S., Scheres, A., Di Martino,
A., Hyde, C., & Walters, J. R. (2005). Varieties of attention-
deficit/hyperactivity disorder-related intra-individual variability.
Biological Psychiatry, 57, 1416–1423.
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.
Ciric, R., Wolf, D. H., Power, J. D., Roalf, D. R., Baum, G. L.,
Ruparel, K., et al. (2017). Benchmarking of participant-level
confound regression strategies for the control of motion
Rosenberg et al. 171
artifact in studies of functional connectivity. Neuroimage,
154, 174–187.
Esterman, M., Noonan, S. K., Rosenberg, M., & Degutis, J.
(2013). In the zone or zoning out? Tracking behavioral and
neural fluctuations during sustained attention. Cerebral
Cortex, 23, 2712–2723.
Fan, J., McCandliss, B. D., Fossella, J., Flombaum, J. I., & Posner,
M. I. (2005). The activation of attentional networks.
Neuroimage, 26, 471–479.
Fan, J., McCandliss, B. D., Sommer, T., Raz, A., & Posner,
M. I. (2002). Testing the efficiency and independence of
attentional networks. Journal of Cognitive Neuroscience,
14, 340–347.
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.
Gabrieli,J.D.E.,Ghosh,S.S.,&Whitfield-Gabrieli,S.
(2015). Prediction as a humanitarian and pragmatic
contribution from human cognitive neuroscience. Neuron,
85, 11–26.
Green, C. S., & Bavelier, D. (2003). Action video game modifies
visual selective attention. Nature, 423, 534–537.
Green, C. S., & Bavelier, D. (2012). Learning, attentional
control, and action video games. Current Biology, 22,
R197–R206.
Huang, L., Mo, L., & Li, Y. (2012). Measuring the
interrelations among multiple paradigms of visual
attention: An individual differences approach. Journal of
Experimental Psychology: Human Perception and
Performance, 38, 414–428.
Joshi, A., Scheinost, D., Okuda, H., Belhachemi, D., Murphy, I.,
Staib, L., et al. (2011). Unified framework for development,
deployment and robust testing of neuroimaging algorithms.
Neuroinformatics, 9, 69–84.
Kelly, C. A. M., Uddin, L. Q., Biswal, B. B., Castellanos, F. X., &
Milham, M. P. (2008). Competition between functional brain
networks mediates behavioral variability. Neuroimage, 39,
527–537.
Kessler,D.,Angstadt,M.,&Sripada,C.(2016).Growth
charting of brain connectivity networks and the
identification of attention impairment in youth. JAMA
Psychiatry, 73, 481–489.
Lundervold, A. J., Adolfsdottir, S., Halleland, H., Halmøy, A.,
Plessen, K., & Haavik, J. (2011). Attention Network Test in
adults with ADHD—the impact of affective fluctuations.
Behavioral and Brain Functions, 7, 27.
MacDonald, S. W. S., Nyberg, L., & Bäckman, L. (2006). Intra-
individual variability in behavior: Links to brain structure,
neurotransmission and neuronal activity. Trends in
Neurosciences, 29, 474–480.
MacLeod, J. W., Lawrence, M. A., McConnell, M. M., Eskes,
G.A.,Klein,R.M.,&Shore,D.I.(2010).Appraising
the ANT: Psychometric and theoretical considerations of
the Attention Network Test. Neuropsychology, 24,
637–651.
Madhyastha, T. M., Askren, M. K., Boord, P., & Grabowski, T. J.
(2015). Dynamic connectivity at rest predicts attention task
performance. Brain Connectivity, 5, 45–59.
Niogi, S. N., Mukherjee, P., Ghajar, J., & McCandliss, B. D.
(2010). Individual differences in distinct components of
attention are linked to anatomical variations in distinct white
matter tracts. Frontiers in Neuroanatomy, 4, 2.
Noble, S., Scheinost, D., Finn, E. S., Shen, X., Papademetris, X.,
McEwen, S. C., et al. (2017). Multisite reliability of MR-based
functional connectivity. Neuroimage, 146, 959–970.
Petersen, S. E., & Posner, M. I. (2012). The attention system of
the human brain: 20 years after. Annual Review of
Neuroscience, 35, 73–89.
Poole, V. N., Robinson, M. E., Singleton, O., DeGutis, J.,
Milberg, W. P., McGlinchey, R. E., et al. (2016). Intrinsic
functional connectivity predicts individual differences in
distractibility. Neuropsychologia, 86, 176–182.
Posner, M. I., & Petersen, S. E. (1990). The attention system of the
human brain. Annual Review of Neuroscience, 13, 25–42.
Power, J. D., Mitra, A., Laumann, T. O., Snyder, A. Z., Schlaggar,
B. L., & Petersen, S. E. (2014). Methods to detect,
characterize, and remove motion artifact in resting state fMRI.
Neuroimage, 84, 320–341.
Power, J. D., Plitt, M., Laumann, T. O., & Martin, A. (2017).
Sources and implications of whole-brain fMRI signals in
humans. Neuroimage, 146, 609–625.
Power, J. D., Schlaggar, B. L., & Petersen, S. E. (2015). Recent
progress and outstanding issues in motion correction in
resting state fMRI. Neuroimage, 105, 536–551.
Rosenberg, M. D., Finn, E. S., Scheinost, D., Constable, R. T.,
& Chun, M. M. (2017). Characterizing attention with
predictive network models. Trends in Cognitive Sciences,
21, 290–302.
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., Noonan, S., DeGutis, J., & Esterman, M. (2013).
Sustaining visual attention in the face of distraction: A novel
gradual-onset continuous performance task. Attention,
Perception & Psychophysics, 75, 426–439.
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.
Shah,L.M.,Cramer,J.A.,Ferguson,M.A.,Birn,R.M.,&
Anderson, J. S. (2016). Reliability and reproducibility of
individual differences in functional connectivity acquired
during task and resting state. Brain and Behavior, 6,
e00456.
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.
Sonuga-Barke, E. J. S., & Castellanos, F. X. (2007). Spontaneous
attentional fluctuations in impaired states and pathological
conditions: A neurobiological hypothesis. Neuroscience and
Biobehavioral Reviews, 31, 977–986.
Stoodley, C. J. (2012). The cerebellum and cognition: Evidence
from functional imaging studies. Cerebellum, 11, 352–365.
Stuss,D.T.,Murphy,K.J.,Binns,M.A.,&Alexander,
M. P. (2003). Staying on the job: The frontal lobes
control individual performance variability. Brain, 126,
2363–2380.
The MathWorks, Inc. (2016). MATLAB (R2016a). Retrieved from
scholar.google.com/scholar?hl=en&btnG=Search&q=intitle:
MATLAB+The+Language+of+Technical+Computing#0.
van Dijk, K. R. A., Sabuncu, M. R., & Buckner, R. L. (2012). The
influence of head motion on intrinsic functional connectivity
MRI. Neuroimage, 59, 431–438.
172 Journal of Cognitive Neuroscience Volume 30, Number 2
Visintin, E., De Panfilis, C., Antonucci, C., Capecci, C., Marchesi,
C., & Sambataro, F. (2015). Parsing the intrinsic networks
underlying attention: A resting state study. Behavioral Brain
Research, 278, 315–322.
Wang, Y.-F., Jing, X.-J., Liu, F., Li, M.-L., Long, Z.-L., Yan, J. H.,
et al. (2015). Reliable attention network scores and
mutually inhibited inter-network relationships revealed by
mixed design and non-orthogonal method. Scientific
Reports, 5, 10251.
Westlye, L. T., Grydeland, H., Walhovd, K. B., & Fjell, A. M.
(2011). Associations between regional cortical thickness and
attentional networks as measured by the Attention Network
Test. Cerebral Cortex, 21, 345–356.
Wojtowicz, M., Omisade, A., & Fisk, J. D. (2013). Indices of
cognitive dysfunction in relapsing-remitting multiple
sclerosis: Intra-individual variability, processing speed, and
attention network efficiency. Journal of the International
Neuropsychological Society, 19, 551–558.
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.
Yin, X., Han, Y., Ge, H., Xu, W., Huang, R., Zhang, D., et al.
(2013). Inferior frontal white matter asymmetry correlates
with executive control of attention. Human Brain Mapping,
34, 796–813.
Yin, X., Zhao, L., Xu, J., Evans, A. C., Fan, L., Ge, H., et al. (2012).
Anatomical substrates of the alerting, orienting and executive
control components of attention: Focus on the posterior
parietal lobe. PLoS One, 7, e50590.
Rosenberg et al. 173
