A Linear Structural Equation Model for Covert Verb Generation Based on Independent Component Analysis of fMRI Data from Children and Adolescents

Article (PDF Available)inFrontiers in Systems Neuroscience 5:29 · June 2011with50 Reads
DOI: 10.3389/fnsys.2011.00029 · Source: PubMed
Abstract
Human language is a complex and protean cognitive ability. Young children, following well defined developmental patterns learn language rapidly and effortlessly producing full sentences by the age of 3 years. However, the language circuitry continues to undergo significant neuroplastic changes extending well into teenage years. Evidence suggests that the developing brain adheres to two rudimentary principles of functional organization: functional integration and functional specialization. At a neurobiological level, this distinction can be identified with progressive specialization or focalization reflecting consolidation and synaptic reinforcement of a network (Lenneberg, 1967; Muller et al., 1998; Berl et al., 2006). In this paper, we used group independent component analysis and linear structural equation modeling (McIntosh and Gonzalez-Lima, 1994; Karunanayaka et al., 2007) to tease out the developmental trajectories of the language circuitry based on fMRI data from 336 children ages 5-18 years performing a blocked, covert verb generation task. The results are analyzed and presented in the framework of theoretical models for neurocognitive brain development. This study highlights the advantages of combining both modular and connectionist approaches to cognitive functions; from a methodological perspective, it demonstrates the feasibility of combining data-driven and hypothesis driven techniques to investigate the developmental shifts in the semantic network.
SYSTEMS NEUROSCIENCE
Moreover, a priori model testing based on fewer free parameters
presents significant challenges for an investigation that deals with
the developmental trajectories of skills sub-serving verb genera-
tion. In particular, such limitations significantly impact our ability
to test specific models (e.g., regionally weighted or focal network
models) of language development using neuroimage data as input
to connectionist approaches for neurocognitive modeling. In the
connectionist approaches a system behavior is captured by adjust-
ing the weights on connections between elements in the network
to investigate how the statistical structure of inputs influences the
behavior of the network (Plaut et al., 1996). Therefore, with more
parameters (degrees of freedom) one is better positioned to capture
any development shifts in neurocognitive modeling.
The utility of independent component analysis (ICA) for exam-
ining changes in brain networks associated with age and brain
development has recently been demonstrated in the context of
resting state (Stevens et al., 2009a) as well as active neurocogni-
tive processes (Stevens et al., 2009a,b) such as language function
(Schmithorst et al., 2006; Karunanayaka et al., 2007, 2010, 2011;
Kim et al., 2011). Unlike model-based approaches, ICA is a data-
driven technique capable of detecting additional task-related neu-
ral networks that exhibit activity with different temporal behavior
(Calhoun et al., 2001a). This approach has significant advantages
IntroductIon
Functional brain imaging methods have recently emerged as means
of investigating connectivity and the dynamic flow of information
across neural networks sub-serving cognitive functions (McIntosh
and Gonzalez-Lima, 1994; McIntosh et al., 1994; Friston et al., 2003;
Penny et al., 2004a,b). These methods measure, e.g., the gener-
ated electrical/magnetic fields (EEG/MEG) or the hemodynamic
response associated with neural activity (fMRI). The functional
data analysis methods frequently focus on identifying areas of acti-
vation under different behavioral conditions with less attention
paid to the behavior of the underlying network (Friston et al., 1995).
Until recently, fMRI studies have employed model-based
approaches predicated upon a priori knowledge of an applied
stimulus and the brains response [hemodynamic response function
(HRF)] to the stimulus (Bandettini et al., 1993; Worsley and Friston,
1995). Such models are typically based on canonical forms for the
HRF and do not reflect individual variations or account for differ-
ences between individuals of different age, sex, or pathologies. We
have previously discussed that this statistical approach may not cap-
ture the complexity of brain networks supporting a language task
such as covert verb generation (Karunanayaka et al., 2010). Several
methods have been proposed to circumvent this drawback by avoid-
ing assumptions about the shape of the HRF (Ollinger et al., 2001).
A linear structural equation model for covert verb generation
based on independent component analysis of fMRI data from
children and adolescents
Prasanna Karunanayaka
1
, Vincent J. Schmithorst
2
, Jennifer Vannest
2
, Jerzy P. Szaflarski
2,3
, Elena Plante
4
and Scott K. Holland
2
*
1
Center for NMR Research, Department of Radiology, The Milton S. Hershey Medical Center, The Pennsylvania State University College of Medicine, Hershey, PA, USA
2
Pediatric NeuroImaging Research Consortium, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA
3
Center for Imaging Research, Department of Neurology, University of Cincinnati, Cincinnati, OH, USA
4
Department of Speech and Hearing Sciences, University of Arizona, Tucson, AZ, USA
Human language is a complex and protean cognitive ability. Young children, following well defined
developmental patterns learn language rapidly and effortlessly producing full sentences by the
age of 3 years. However, the language circuitry continues to undergo significant neuroplastic
changes extending well into teenage years. Evidence suggests that the developing brain adheres
to two rudimentary principles of functional organization: functional integration and functional
specialization. At a neurobiological level, this distinction can be identified with progressive
specialization or focalization reflecting consolidation and synaptic reinforcement of a network
(Lenneberg, 1967; Muller et al., 1998; Berl et al., 2006). In this paper, we used group independent
component analysis and linear structural equation modeling (McIntosh and Gonzalez-Lima,
1994; Karunanayaka et al., 2007) to tease out the developmental trajectories of the language
circuitry based on fMRI data from 336 children ages 5–18 years performing a blocked, covert
verb generation task. The results are analyzed and presented in the framework of theoretical
models for neurocognitive brain development. This study highlights the advantages of combining
both modular and connectionist approaches to cognitive functions; from a methodological
perspective, it demonstrates the feasibility of combining data-driven and hypothesis driven
techniques to investigate the developmental shifts in the semantic network.
Keywords: brain development, functional neuroimaging, language, pediatric, child, fMRI
Edited by:
Silvina G. Horovitz, National Institutes
of Health, USA
Reviewed by:
Vince D. Calhoun, University of New
Mexico, USA
*Correspondence:
Scott K. Holland, Pediatric
NeuroImaging Research Consortium,
Childrens Hospital Medical Center,
3333 Burnet Avenue, ML 5033,
Cincinnati, OH 45229, USA.
e-mail: scott.holland@cchmc.org
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 1
Original research article
published: 01 June 2011
doi: 10.3389/fnsys.2011.00029
prenatal period (Wada et al., 1975; Chi et al., 1977), which suggests a
structural basis for early left hemisphere lateralization of related func-
tions (Foundas et al., 1994). While functional asymmetries are not
present at birth (Kotilahti et al., 2010) the anatomical asymmetries,
in association with genetic factors, may underlie later development of
functional asymmetries (Szaflarski et al., 2002; Francks et al., 2007).
In fact, previous reports from the parent project that generated this
data set (Szaflarski et al., 2006a,b; Holland et al., 2007) and others
(Wood et al., 2004; Chou et al., 2006) have indicated that the initial
left lateralization in this area strengthens with age with maximum
left lateralization achieved around the age of 20–25 years followed
by gradual decrease in the observed asymmetries with increasing age
(Szaflarski et al., 2002, 2006a). Therefore, in this study we expected
to confirm the age-related changes in the networks (inter and intra)
that sub-serve verbal abilities.
Because the noun must be held in working memory as the verbs
are generated, we expect that the temporal cortex must be connected
to a fronto-parietal network that is routinely activated in studies
involving working memory (Chein et al., 2003). This includes acti-
vation of the inferior frontal gyrus, dorsolateral prefrontal cortex,
and parietal cortex. More specifically, the superior temporal cortex
is functionally connected to the inferior frontal gyrus (through the
arcuate fasciculus) as the anatomical connections between these two
regions are well established (Catani et al., 2005). The dorsolateral pre-
frontal (executive control) and parietal (sustained attention to words)
cortices modulate activity in this region through either the superior
branch of the arcuate fasciculus (Catani et al., 2005) or the superior
longitudinal fasciculus. Because working memory shows age-related
improvement, we would expect that the associated neural regions will
also show age-related changes. Furthermore, the protracted period
of development of the frontal lobes (Giedd et al., 1999; Schmithorst
et al., 2002; Giedd, 2004; including connections with Brodmanns
areas 17, 18, 31, and 32) may make the associated cognitive func-
tions, the underlying regions, and the connections with these regions
particularly dynamic through the course of childhood.
To generate a verb that is plausibly related to the noun, the
child must select semantic concepts that are associated with the
meaning of the noun. On the output side, semantic retrieval is
likely to engage the middle and inferior temporal regions (seman-
tic knowledge) and the hippocampi (information retrieval). The
semantic concepts must be coded into phonological form, typi-
cally thought of as the second stage of the word retrieval process
(Binder et al., 2008). Because semantic associations refine over
the course of childhood (McDonald, 1997; McGregor et al., 2002;
Beitchman et al., 2008), it is likely that activation in both of these
areas will change with age. The phonological form is further coded
into subvocal speech (Thompson-Schill et al., 1997). This suggests
a second activation by inferior frontal gyrus for subvocal phono-
logical encoding as well as contributions by the insula for speech
coordination in subvocal naming. In the covert verb generation task
the speech motor network is still engaged but must be inhibited so
that words are not spoken overtly (Skipper et al., 2005). We would
also expect age-related changes in the neural networks supporting
these cognitive components and the connections between them.
There is evidence to suggest that the developing brain adheres to
two rudimentary principles of organization: functional integration
and functional specialization (Berl et al., 2006). At a neurobiological
when compared to the model-based methods that may not identify
brain areas with temporal behavior that is not correlated with the
experimental design matrix. However, ICA generates a considerable
number of components that may not necessarily be part of the stud-
ied network (Calhoun et al., 2001b). To address this issue, we have
incorporated several additional steps in our ICA method that make
the results of our study more targeted and objective (Karunanayaka
et al., 2010). In particular, we have adopted a theory-driven focus
based on the Wernicke–Geschwind model of the language network
with the aim of investigating developmental shifts in the verb gen-
eration circuitry in children from 5 to 18 years of age (Geschwind,
1965a; Anderson et al., 1999). Inclusion of such a focus yields a
biologically plausible network model for covert verb generation
predicted by the methods proposed here, which is more inclusive
and specific in comparison to models extracted using general linear
modeling (Yuan et al., 2006; Holland et al., 2007).
In the present paper, we explore the age dependency of the
connections between the nodes of the language network that
sub-serve the verb generation task. The verb generation model
discussed here is based on nodes generated from a group ICA of
fMRI data obtained from 336 children ranging in age from 5 to
18 years: which has been discussed in detail in one of our previ-
ous publications (Karunanayaka et al., 2010). The current analysis
takes the previously described ICA analysis further by investigating
interactions within the identified verb generation network using
linear structural equation modeling (LSEM). The initial investiga-
tion of verb data using group ICA specifically dealt with: (1) ICA
decomposition of verb data; (2) age effects in the task-relatedness
of each individual IC map (at individual network level) using an
a priori criterion (e.g., correlation with the task reference) and
Bayesian formalism; (3) brief description of the steps leading to the
present model (Karunanayaka et al., 2010). In the current analysis,
the previously described model is further expanded with input
functions (processing of presented nouns) and output functions
(retrieval and covert verb production) together with hypothesized
connections within and between them. In addition, a theory-driven
focus has been proposed taking the biological plausibility of the
verb model into account; evaluated against the literature with the
focus on the language circuitry. Thus, the current investigation
provides an elegant methodology capable of providing unique
insight into the framework of neurocognitive brain development
in children: by combining and extending previously published
ICA results (Karunanayaka et al., 2010) with some of the figures
reproduced in this paper for convenience and completeness. Thus,
in the present work the emphasis is placed on the neurocogni-
tive brain development and on estimating the age dependency of
inter- and intra-network connectivity predicted using LSEM and
correlation analysis.
The verb generation task begins with an auditory presentation of
a noun: requires the listener to process the nouns phonological form,
and attach meaning to that form (Hickok and Poeppel, 2004). Based
on our previous description, this process begins with an input via
the superior temporal gyrus (Karunanayaka et al., 2010). Processing
in this area extends posteriorly from primary auditory cortex, as the
act of accessing word’s meaning is presumed to activate a broader
network (Levelt et al., 1999; Pulvermuller, 2001). The posterior supe-
rior temporal cortex is known to be structurally asymmetric from the
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 2
generation networks extending the finding of our previous study
(Karunanayaka et al., 2010). The previously described group ICA
provides a complete recipe of the prerequisite steps involved in ICA
decomposition to identify the key elements underlying a biologi-
cally plausible neural network that sub-serve a specific neurocogni-
tive task (Schmithorst and Brown, 2004; Schmithorst et al., 2006;
Karunanayaka et al., 2010, 2011; Kim et al., 2011). A complete age
and sex breakdown of the included subjects (native, monolingual,
English speakers) is detailed in Table 1. Based on the Edinburgh
Handedness Inventory (Oldfield, 1971), 311 subjects were right-
handed, 24 left-handed, and 1 ambidextrous. All subjects were
prescreened for any conditions which would prevent an MRI scan
from being acquired (Karunanayaka et al., 2010). Out of 336 sub-
jects, 331 received the Wechsler Preschool and Primary Scale of
Intelligence (WPPSI-R, ages below 6) or the Wechsler Intelligence
Scale for Children [Third Edition (WISC–III, ages 6–16 years);
Wechsler, 1991] or the Wechsler Adult Intelligence Scale, Third
Edition (WAIS–III, ages 17 and 18 years; Wechsler, 1997).
Similarly, 330 subjects received the Oral and Written Language
Scales (Carrow-Woolfolk, 1996). The age range for all subjects
was 4.92–18.92 years; Mean Wechsler Full-scale IQ = 111.6 ± 13.84
(range = 70–147); Mean OWLS = 107.7 ± 14.3 (range = 66–151).
FunctIonal ImagIng
All images were acquired using a Bruker 3T Medspec (Bruker
Medizintechnik, Karlsruhe, Germany) imaging system. An MRI-
compatible audiovisual system was used for presentation of the
stimuli. Details of the techniques used to obtain fMRI data from
younger children are discussed elsewhere (Byars et al., 2002).
EPI–fMRI scan parameters were: TR/TE = 3000/38 ms; 125 kHz;
FOV = 25.6 cm × 25.6 cm; matrix = 64 × 64; slice thickness = 5 mm.
Twenty-four slices were acquired, covering the entire cerebrum.
One hundred ten whole-brain volumes were acquired (with the
first 10 being dummy scans) in 5 min 30 s. Techniques detailed
elsewhere (Byars et al., 2002) were used to acclimatize the subjects
to the MRI procedure and make them comfortable inside the scan-
ner. A whole-brain T1 weighted MP-RAGE scan was also acquired
for anatomical co-registration.
Verb generatIon task
The fMRI paradigm of silent verb generation (Holland et al., 2001,
2007) is a 30-s on–off block design. All stimuli were presented
using MacStim (White Ant Software, Melbourne, VIC, Australia)
at a rate of one noun every 5 s, for six stimuli during each 30 s
epoch. During the active epochs, the subjects silently generated
level, this distinction can be identified with progressive specialization
or focalization reflecting the consolidation of synaptic reinforce-
ment of a network (Lenneberg, 1967; Muller et al., 1998; Berl et al.,
2006). In this paper, we present a unified framework and examine
the developmental trajectories in the language circuitry based on
fMRI data using complementary modeling approaches. As previ-
ously (Karunanayaka et al., 2010), we employ ICA, a data-driven
method, to identify spatially coherent activation patterns. In the
current investigation, we extend these analyses by applying correla-
tion analysis and LSEM to model connectivity between these spatial
distributions. Several, other approaches have previously been pro-
posed to investigate network interactions following ICA analyses.
Stevens et al.s (2007) used dynamic causal modeling (DCM) to
search for the presence of a meaningful causal structure among
selected IC time courses in an event related fMRI study of visual
Go/No–Go task. Another study examined the functional network
connectivity (FNC) between schizophrenia patients and healthy
controls based on the temporal dependency among ICA compo-
nents (Jafri et al., 2008). Demirci et al. (2009) extended this analysis
one step further by incorporating Granger causality test (GCT) to
investigate causal relationships between brain activation networks;
we also have recently implemented Granger causality analysis to
investigate the connections within the epileptic network (Szaflarski
et al., 2010). Several, other investigations have further highlighted
the usefulness of combining ICA with Granger causality on sim-
ulated, single subject and group data (Londei et al., 2006, 2007,
2010). Some of the above mentioned methods are relatively sophis-
ticated and more suitable for investigating specific group differences
between healthy and patient populations. However, the emphasis
of the current analysis is on investigating the overall developmental
trends associated with the language circuitry and presenting the
findings in the framework of a theoretical model for neurocog-
nitive brain development. Thus, given the large sample size, the
simplicity of the proposed partially data-driven approach can be
considered more suitable for understanding the global network
structure (including connectivity) associated with complex verbal
language tasks in general, and verbal fluency tasks in particular.
materIals and methods
subjects
One hundred sixty-five boys and 171 girls took part in the study
following Cincinnati Childrens Hospital Institutional Review
Board approval. Informed consent was obtained from parent
or guardian, an assent was also obtained from subjects 8 years
and older. Exclusion criteria were: previous neurological illness;
learning disability; head trauma with loss of consciousness; cur-
rent or past use of psychostimulant medication; pregnancy; birth
at 37 weeks gestational age or earlier; or abnormal findings at
a routine neurological examination performed by an experi-
enced pediatric neurologist. All subjects were considered healthy
based on neurological, psychological, and structural measures
(Holland et al., 2007). Subjects included in this report were also
included in our previous studies focusing on verb generation in
children (Holland et al., 2001, 2007; Karunanayaka et al., 2010)
and adults (Szaflarski et al., 2006a). While this report includes
fMRI data from the same subjects, it describes an entirely new
analysis of connectivity ( causality) associated with covert verb
Table 1 | Age and gender breakdown of the study population (165 boys
and 171 girls).
Age in years 5 6 7 8 9 10 11 12 13 14 15 16 17 18
SEX
M 9 8 9 17 14 12 17 18 17 9 10 9 13 3
F 7* 12 17 10 11 12 11 15 21 11 11 10 12 11
*Includes one girl 4 years 11 months. The ethnic background of the subjects was:
302 Caucasian, 21 African–American, 2 Asian, 3 Hispanic, 1 Native American, 2
Asian/European, and 5 Multi-Ethnic.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 3
basis for completion of the task through the age of the oldest sub-
jects (18 years). By concatenating the data from the entire cohort
and searching for the components (or networks) that are com-
mon across the age group, we are able to identify the persistent
structural elements (network nodes) underlying the fMRI verb
generation task consisting of seven IC networks and shown in
Figure 1. Table 2 contains a summary of the respective activation
foci for each of the components. Coordinates listed for each IC
correspond to the center of mass of each individual spatial element
contained in the IC map.
To summarize, the selection of IC maps is based on three criteria:
(1) power spectral analysis at the task frequency; (2) phase and
(3) relevance of the spatial maps to the theoretical model of verb
generation. Thus, by following the above mentioned criteria, the
results of this ICA analysis can be replicated by other researchers
in the field.
From this point on, we focus on estimating the changes in con-
nectivity between elements of the model identified by ICA using
LSEM which is a unique contribution of this work.
appropriate verbs such as drink or fill, to aurally presented nouns
such as cup. Subjects were asked to tap their fingers in response to a
modulated tone presented at 5 s intervals during the control epochs.
The control task was specifically designed to control for sublexical
auditory processing and also to divert subjects to stop generating
verbs into the control epochs. The fMRI task was selected such that
children as young as 5 years old would be readily able to perform
the task without any difficulty.
group Ica
A complete description of the group ICA methodology for verb
generation fMRI data has been discussed in detail elsewhere
(McKeown et al., 1998; Calhoun et al., 2001a; Schmithorst et al.,
2006; Karunanayaka et al., 2010). Basic steps involved in ICA
decomposition are briefly mentioned here for the purpose of com-
pleteness. ICA is a data-driven analysis technique that does not
rely on any prior knowledge of the task performed and is capable
of identifying spatially independent components that have similar
time courses. The power of group ICA in making statistical infer-
ences from fMRI data has been presented in several investigations
(Calhoun et al., 2001a; Schmithorst and Brown, 2004; Schmithorst
and Holland, 2006; Karunanayaka et al., 2010).
The ICA decomposition entails several preprocessing steps [nor-
malizing (mean centering) and 40 retained principal components
(PCA)] at the single subject level. The data from all subjects are then
concatenated into a single dataset before a second PCA reduction
resulting in 50 retained components. Finally, 25 runs of the Fast
ICA algorithm (Hyvarinen, 1999b) are combined with hierarchical
agglomerative clustering (Himberg et al., 2004) to estimate and
validate the independent component maps sub-serving covert verb
generation. Performing multiple runs (when combined with hier-
archical agglomerative clustering) ensures that our analysis resulted
in the most reliable components even after taking into account the
stochastic nature of the Fast ICA algorithm. Although, ICA can be
used to remove motion-related artifacts, individual motion has
been fully characterized before performing the ICA decomposition.
A detailed analysis of motion (including task-related movement)
related to this task is discussed elsewhere (Yuan et al., 2009).
The task-relatedness of each IC map is then investigated using
the associated IC time course by examining the spectral power at
the task frequency and the phase of the IC time course relative to
the task reference function as detailed previously (Karunanayaka
et al., 2010). It should be noted that, by definition, spatial ICA
requires independence only in the spatial domain and not in
the time domain. Thus, an analysis performed in one domain
(e.g., time) can be followed by analysis in another domain (e.g.,
spatial) without adding any undue bias to subsequent statisti-
cal manipulations. Finally, a voxel-wise random effects analysis
(one-sample t test) is performed on selected individual IC maps
in the spatial domain to determine the cortical regions active in
the entire cohort. To further clarify this step, if one were to reverse
the domains of the preceding analysis (e.g., spatial followed by
time), the end result would be the same because of the above
mentioned symmetry. An assumption inherent in this approach
is that the structural components of the network for verb genera-
tion are in place by the age of the youngest subjects in our cohort
(5 years) and continue to get fine tuned to form the structural
FIGURE 1 | Seven task-related spatial independent components maps are
shown in panels a-g. These ICs are computed using group ICA analysis of 336
children ages 5–18 performing the task of covert verb generation (Karunanayaka
et al., 2010). Slice range: Z = −25 to +50 mm (Talairach coordinates). Three
corresponding single subject IC maps are shown at bottom (g, b, d). These individual
spatial maps and the associated time courses (Figure 2B) are estimated using a
back propagation algorithm following the ICA decomposition at the group level and
used in the subsequent LSEM analysis. All images are in radiologic orientation.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 4
of each group IC map (Karunanayaka et al., 2007). Specifically,
representative average time courses were extracted from the func-
tional data set (i.e., real signal intensities) based on these function-
ally defined ROIs. It is important to remember that ROIs derived
from spatial IC maps often include multiple anatomical brain areas,
as outlined in Table 2. For all of the IC maps shown in Figure 1,
except for IC d, ROIs were defined separately for the left and right
hemisphere components of the IC. Based on these ROIs, as men-
tioned earlier, extracted real signal time courses from the functional
data set were then used for the between hemispheres intra-network
connectivity computations.
A variety of models can be tested in SEM to capture relationships
among variables and can provide a quantitative test for a hypoth-
esized theoretical model. SEM takes the entire variance–covariance
structure into consideration when evaluating models. Furthermore,
SEM is a generalization of regression, path and confirmatory fac-
tor models that have been extensively used in psychology, eco-
nomics and other social sciences. The model estimation in SEM
involves minimizing the difference between the observed variance–
covariance structure and the one predicted by the implied model.
However, when using SEM to model brain activity no distinction
is made between the neuronal and the hemodynamic levels (Penny
et al., 2004b) which can be considered a drawback of the method.
In the presented model, which is based on Figure 1, we only
evaluated the feed-forward connections. As noted above, repre-
sentative time courses for each of the components (elements) in the
LSEM are comprised of individual IC time courses from the previ-
ously performed ICA decomposition. The individual LSEM(s) were
then solved for optimal path coefficients using the Amos software
(Arbuckle, 1989) which utilizes an iterative maximum likelihood
method. These optimal path coefficients (connection strengths)
correspond to the solution of the structural equations where the
difference between the observed and the predicted covariance
matrix is a minimum. Finally, we evaluated the goodness of fit
between the predicted and the implied covariance matrices using
the χ
2
distribution with m (m + 1) n degrees of freedom (m
corresponds to the number of elements and n corresponds to the
number of coefficients in the LSEM respectively). The details of
LSEM implementation for fMRI data have been discussed else-
where (McIntosh and Gonzalez-Lima, 1994; Solodkin et al., 2004;
Karunanayaka et al., 2007; Dick et al., 2010). The LSEM itself was
used (constrained by the proposed verb generation model discussed
in the introduction) in a semi-exploratory manner when selecting
the final LSEM. Advantages of alternative methods for brain activity
modeling (such as DCM) have also been discussed by other authors
(Friston et al., 2003; Penny et al., 2004b). Recently, an extended ver-
sion of SEM called unified structural equation modeling (uSEM;
Smith et al., 2010) has been proposed capable of estimating con-
temporaneous as well as lagged effects simultaneously (Stoeckel
et al., 2009). An automatic search procedure has also been proposed
to uSEM making it entirely data-driven by increasing its flexibil-
ity substantially (Kim and Horwitz, 2009). However, DCM is still
appears to be the most statistically sophisticated approach that
incorporates neuronal hemodynamic relationship into a dynamic
model of BOLD activities using Bayesian estimation (Friston et al.,
2003; Friston and Stephan, 2007; Sarty, 2007). Thus, given the fact
that the relationship between BOLD signal and neuronal activity
lInear structural equatIon modelIng
Linear Structural Equation Modeling is a statistical method mainly
used for hypothesis testing regarding causal influences among meas-
ured or latent variables. In addition, SEM is capable of statistically
testing a variety of theoretical models that hypothesize how sets of
variables define constructs and how these constructs are related to each
other. In terms of neuroimaging, SEM relates to effective connectivity
that captures causal relationships (directionality) in terms of path coef-
ficients in the model. This approach differs from a typical functional
connectivity analysis that can only determine the degree to which two
brain regions co-vary (Friston et al., 1997). As mentioned elsewhere,
our group ICA decomposition is based on the methods developed by
Calhoun et al. (2001b) and is designed to evaluate individual IC maps
and corresponding time courses based on group results. In other words,
in this method individual IC time courses are estimated using a back
propagation method which is followed by the ICA decomposition at
the group level. In this paper, we use these individual IC time courses as
input to estimate LSEM(s) at the subject level in order to examine the
effective connectivity within the network model for verb generation.
A second level, intra-network functional connectivity analysis
was also performed using representative real signal intensity average
time courses from ROIs defined based on the spatial distribution
Table 2 | Activation foci (Talairach coordinates) for the ICA components
displayed in Figure 1.
Anatomical region BA Talairach
X, Y, Z
A
R. parahippocampal gyrus 30/35 22, 41, 5
L. parahippocampal gyrus 30/35 26, 41, 5
R. inferior temporal gyrus 19/37 42, 57, 5
L. inferior temporal gyrus 19/37 46, 57, 5
R. medial temporal gyrus 19/39 34, 69, 20
L. medial temporal gyrus 19/39 26, 73, 25
B
Cuneus 17 2, 77, 10
C
R. inferior frontal gyrus 44 34, 11, 10
L. inferior frontal gyrus 44 34, 11, 10
D
L. medial temporal gyrus 21 54, 41, 5
L. inferior frontal gyrus 45/46 46, 27, 15
L. inferior/medial frontal gyrus 44/9 42, 7, 35
L. middle frontal gyrus 6/8 6, 23, 45
L. angular gyrus/inferior parietal lobule 39/40 30, 65, 40
E
R. superior temporal gyrus 22 50, 29, 5
L. superior temporal gyrus 22 54, 45, 10
F
R. inferior frontal gyrus 45/47 30, 31, 0
L. Inferior frontal gyrus 45/47 38, 23, 0
G
R. insula Insula 38, 11, 0
L. insula 38, 11, 0
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 5
IC based on known functional neuroanatomy (knowledge-base)
ascribed to each Brodmanns area encompassed by the component.
As mentioned previously, a theory-driven focus (Geschwind, 1965a;
Anderson et al., 1999) complements data-driven methods such
as ICA by way of corroborating prior hypotheses about cognitive
functions sub-serving the verb generation fMRI task.
Depending on the modularity (or function), IC modules are
then connected to one another to form the LSEM. In this study,
LSEM is directly derived from the covert verb generation model as
discussed in the Section “Introduction. For studies of developmen-
tal changes within a network, LSEM of an fMRI task can investigate
what changes in functional connectivity explain the neural basis of
development in language networks. This physiological approach
should be guided by the weak constraint that anatomical proximity
and connectivity of brain regions are incorporated in the model
(Karunanayaka et al., 2007). Alternatively, a cognitive approach
can also be implemented to investigate how functional/effective
connectivity changes are related to cognitive development. The
emphasis of the current analysis is inline with the latter approach
where the effective connectivity changes between IC modules sub-
serving covert verb generation are investigated.
Finally, a second level Pearson correlation analysis was per-
formed on path coefficients in the LSEM to investigate any age
effects associated with the proposed cognitive model for covert
verb generation.
results
Six out of the seven IC maps shown in Figure 1 were detected in all
25 IC runs while the component shown in Figure 1a was detected
in 17 IC runs assuring high reliability (Karunanayaka et al., 2010)
and defines the covert verb generation network for each subject
included in the study. The maps in the lower row (individual sub-
ject level) of Figure 1 shows three corresponding individual sub-
ject level IC maps with corresponding IC time courses: estimated
following the ICA decomposition at the group level and used in
the subsequent subject level LSEM analysis. Figure 2A shows two
of the corresponding average time courses for IC maps shown in
Figures 1a,d. Figure 2B shows the individual IC time courses for
these networks in two subjects: used as the input to the LSEM
computations. The phase progression of the average time courses
from leading to lagging the task reference time course (indicated by
dark and light gray background) is clearly visualized in Figure 2A.
The developmental trajectories, network behavior (lateraliza-
tion, task-relatedness, etc.) and the language functions attributed
to each IC have been discussed in detail elsewhere (Karunanayaka
et al., 2010). The highly left-lateralized IC map shown in Figure 1d
(with lateralization index equal to 1) was identified previously as
capturing most of the left-dominance observed in a standard GLM
analysis for the covert verb generation task (Holland et al., 2007).
To perform the intra-network connectivity analysis for this left-
lateralized network, four separate ROIs were defined in the left
hemisphere as shown and labeled in Figure 3. As explained above,
only the real signal time courses from activated regions (refer to
Table 2 for further details) inside the colored circles are included in
the ROI analysis. The connection between (1) left middle temporal
gyrus (LMTG) (3) left middle inferior frontal gyrus (LMIFG)
showed significant age dependent connectivity changes (r = 0.15,
is poorly understood (de Marco et al., 2009), LSEM may be a very
effective method for making inferences about changes in the causal
structure from fMRI time series data.
In addition, several methods have been employed to obtain rep-
resentative time courses for the components included in a SEM
analysis: one popular method being the maximum active voxel rep-
resentation (Jennings et al., 1998; Goncalves et al., 2001) which we
employed previously to investigate developmental trends associated
with the narrative story comprehension in children (Karunanayaka
et al., 2007). However, in the current analysis, IC time courses were
used to evaluate individual LSEMs to investigate the verb genera-
tion task in children. A brief description of the differences between
the two methods are included in the section below and also in the
Section “Discussion.
bIologIcal constraInts
Several principles have guided the process of constructing a biologi-
cally plausible linear structural equation model for verb generation.
As the first step (described above), ICA was used as a data-driven
descriptor of neural elements involved in performing the fMRI
paradigm. The second step involved a Fourier method in the time
domain to determine which ICs were most task-related by testing
the correlation between the fundamental frequency of each IC time
course and the task frequency. The third step involved constructing
a biologically plausible LSEM using the knowledge of the sequence
of neurocognitive functions involved in the task with IC mod-
ules as building blocks (Karunanayaka et al., 2007). The IC maps
require only independence in the spatial domain allowing highly
correlated temporal structures to form the theoretical basis for the
current SEM analysis. Finally, the phase of the Fourier transform
of the associated IC time courses and the known neuroanatomi-
cal constraints were also taken into consideration when imposing
connections between the model elements.
Some individual ICs out of the seven selected, contain more
than one Brodmanns area even though the representative time
course for the IC represents all of the voxels included in the spatial
map. This is because ICA reveals a set of chronoarchitectonically
identified areas (Bartels and Zeki, 2004) or functionally connected
regions that may span several Brodmanns areas. If a given cognitive
task recruits only one of the observed regions in a given map, then
there will be another component separated out by ICA containing
only that region. However, if two distinct cognitive functions have
very similar time course, they may well be grouped into a single
ICA component. This is a limitation of correlational analysis. Still,
under certain minimal assumptions, the spatial independence of
IC maps can be equated with their modularity, establishing a cor-
respondence between the IC component and a specific cognitive
task (Duann et al., 2002; Calhoun et al., 2004). A limitation of this
assumption is our inability to determine spatial independence of
components with absolute certainty due to the finite number of
voxels in fMRI experiments. However, this limitation may have only
a minimal effect on the current investigation because of the excel-
lent signal to noise ratio provided by the large number of subjects
in the study. Therefore, we argue that it is reasonable to assume that
each IC map constitutes a module (a cognitive functional unit) in
the proposed LSEM. Depending on the spatial distribution of the
IC (activation), a specific language function can be assigned to each
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 6
As described in the Section “Materials and Methods, an LSEM
was constructed using the functional IC maps with reference to
the literature for prior knowledge (i.e., knowledge-base) about the
known neuroanatomy of the brain regions involved in the language
circuitry. This LSEM was further refined based on the hypothesized
cognitive functions associated with the brain regions encompassed
within each spatial IC map, forming the basis for the proposed
theoretical cognitive model for covert verb generation as shown
in Figure 5. Of note is that the connections between brain regions
may not be explicitly included in the proposed model if they are
implied by inclusion within a single IC. For example, IC d includes
frontal, temporal, and parietal regions. The cartoon in Figure 5
demonstrates this aspect of the model by using an extended ROI
spanning these lobes to illustrate the spatial extent of IC d.
Table 3 shows the average value of each standardized path coef-
ficient and the age-related changes in path coefficients computed
for the LSEM shown in Figure 5. A similar figure (model) was
included in a previous study by Karunanayaka et al. (2010); though
that diagram did not include the path coefficients computed here
as a parameter expressing brain connectivity. As mentioned earlier,
the focus of the present analysis is on developmental changes in
connectivity within the neural circuitry of language; therefore, we
p = 0.007). The Functional connectivity between (1) LMTG (4)
left angular gyrus (LANG) showed no significant age effects.
Similarly, the functional connectivity between (4) LANG (2)
left inferior frontal gyrus (LIFG) showed significant age effects
(r = 0.143, p = 0.0089) while the functional connectivity between (4)
LANG (3) LMIFG did not. Finally, the functional connectivity
between (3) LMIFG (2) LIFG showed a highly significant age
effect (r = 0.18, p = 0.002).
Similarly, we also examined the inter-hemispheric functional
connectivity based on individual spatial IC maps. The IC map
shown in Figure 1c showed a highly significant age effect (r = 0.3,
p = 2.457e 008) in the connectivity between the hemispheres
(Figure 4). Similarly, the IC map shown in Figure 1f also showed
significant age effect (r = 0.132, p = 0.015) in inter-hemispheric
connectivity. However, the IC shown in Figure 1e (bilateral supe-
rior temporal gyri; BA 22) did not exhibit significant age effects
in functional connectivity between the left and right hemispheres
(Figure 4). Similarly, IC modules a, b, and g also did not exhibit
any age-related inter-hemispheric functional connectivity changes.
Thus, for these components, we have not included the results of
the above mentioned inter-hemispheric functional connectivity
analysis.
FIGURE 2 | (A) Associated averaged time courses from two group IC
networks shown in Figure 1. Horizontal axis is time and the vertical axis is
intensity (pseudo). Gray and white background indicates the timing of the
task reference function; (B) associated IC time courses from two subjects
(red and blue) corresponding to Figures 1a,d networks. These IC time
courses correspond to similar individual subject networks as shown in the
lower row (individual subject level) of Figure 1. These IC time courses are
used in subject level LSEM evaluations.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 7
attributed to the youngest subjects having higher than average IQ
(Karunanayaka et al., 2010). To be more specific, when the children
between the ages 5 and 8 years were excluded, the weak correlation
between age an IQ did not reach significance and consequently we
have not included IQ as a covariate in the analysis.
dIscussIon
Methods for network connectivity analysis based on functional
neuroimaging data are developing rapidly as a means of expand-
ing our understanding of neurocognitive function beyond what
the neo-phrenology or functional blobology of fMRI have been
able to reveal (Friston et al., 2003; Schmithorst and Holland, 2007;
Schmithorst et al., 2007; Rajapakse et al., 2008; Dick et al., 2010).
ICA is an ideal preliminary step for network connectivity analysis
because it is able to detect areas that exhibit task-related behavior
which might not correlate highly with an a priori model or refer-
ence function. In the present analysis, we began with ICA of verb
generation data which detected activations in multiple networks
with different temporal signatures. Multiple activation time courses
detected in the same brain regions (specifically frontal and temporal
cortex) provide direct evidence of their participation in multiple
cognitive aspects of the verb generation task. ICA provided the basis
for construction of a LSEM for the network that sub-serves verb
generation task and allowed us to use this standard statistical meth-
odology to explore the age dependency of the relationships among
cognitive modules revealed by the ICA analysis (Karunanayaka
et al., 2007).
The theoretical framework guiding this research focuses on
investigating the developing brain from a network perspective
and lays the foundation for deciphering any developmental trends
as interactions between underlying networks. Starting with the
Wernicke–Geschwind model for the language network, we used
a data-driven approach to analyze results from an fMRI experi-
ment in a large sample of children over a wide age range in order
to extract key network elements supporting verb generation. This
classical model guided our thinking about how to connect mod-
ules identified by group ICA results as having a strong correlation
with the task behavior. We then examined the network structure to
identify developmental trajectories that correlate with age and abil-
ity of children to think and reason at increasing levels of maturity
(Schmithorst et al., 2006, 2007). We have shown elsewhere, how ICA
can be used to explore developmental changes in brain activation
patterns associated with individual neural networks supporting
covert verb generation (Karunanayaka et al., 2010). The current
analysis takes this approach one step further by incorporating
LSEM to the investigations of the theories of brain development
using the regionally weighted or focal network models (Berl et al.,
2006). Although, these hypothesized brain developmental models
draw support from current neuroimaging literature, our analysis
seems to favor the regionally weighted model of normal language
development.
Independent component analysis by itself is not capable of reveal-
ing the precise cognitive correlates of the identified components
(Schmithorst et al., 2006). Instead, this data-driven method must
be utilized to identify spatial distributions (IC maps) from fMRI data.
As with GLM-based analyses, the function of the detected regions
must be inferred and should be constrained by prior knowledge of
examined changes in the path coefficients estimated by the LSEM
as a function of age. The following path coefficients exhibited age-
related changes: The path coefficient between IC e IC f showed
an increase in connectivity with age (r = 0.13, p < 0.017). The path
coefficient between IC e IC d showed a modest (identified with
a trend) age-related connectivity decrease (r = 0.111, p < 0.044).
However, the path coefficient between IC f IC d exhibited a
highly significant age-related increase in connectivity (r = 0.18,
p < 0.00088). Figure 6 graphically displays the corresponding
standardized path coefficients that showed statistically significant
age-related changes. These values are italicized in Table 3.
For the group of children included in this study, subject age was
significantly correlated with the full-scale IQ (Spearmans r = 0.18,
p < 0.0008). This small but significant negative correlation is mainly
FIGURE 3 | Regions included (only the active areas) in the intra-
component functional connectivity analysis for IC d are
1
medial temporal
gyrus (LMTG),
2
inferior frontal gyrus (LIFG),
3
middle inferior frontal gyrus
(LMIFG),
4
angular gyrus (LANG). Each brain region will be represented by
the average activation within that ROI across time. Slice range: Z = 25 to
+50 mm (Talairach coordinates). All images are in radiologic orientation (left in
the picture is right in the brain).
FIGURE 4 | Graphical representation of the age dependence of functional
connectivity between left and right hemispheres corresponding to IC
maps shown in Figures 1C,E. IC c exhibits a highly significant functional
connectivity between the left and the right inferior frontal gyrus.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 8
(Karunanayaka et al., 2010). Given the limitations [(Wright’s rules;
Write, 1934) and the number of nodes in the model] in evaluating
LSEM(s), a careful consideration must be given before selecting either
approach. In general, any theoretical model for language related
cognitive functions will be a compromise between the complexity
of the neural system sub-serving language comprehension and the
interpretability of the resulting models. Complex models can account
for intricate dependencies (both anatomical and functional) but the
interpretability of the resulting models would be severely compro-
mised (McIntosh and Gonzalez-Lima, 1994; Dick et al., 2010).
As suggested by Dick et al. (2010), one approach would be to use
the hypotheses being tested as guiding the constraining aspects of the
model development. An alternative, more appealing approach would
be to model brain functions in terms of interactions between underly-
ing sub-networks, inline with the method we have proposed in this
paper. To circumvent inherent drawbacks of the second approach, in
addition to the theory-driven focus, we incorporated a secondary cor-
relation analysis specifically to investigate the within network behavior
sub-serving covert verb generation in children (Friston et al., 1997).
The functional connectivity results of IC d revealed unique fea-
tures related to semantic processing circuitry in children. Several
studies have implicated activation in the middle temporal gyrus
the functional neuroanatomy. However, once the spatial distributions
are known, depending on the complexity either a physiological or
a cognitive approach can be employed for the connectivity analysis
FIGURE 5 | The proposed covert verb generation model based on group ICA
maps shown in Figure 1. This model is based on our previous publication
(Karunanayaka et al., 2010; Figure 4). The brain cartoon shows the approximate
locations of each IC map from Figure 1. Transparent ellipses indicate regions
located medially within the brain and not visible from the lateral surface whereas
opaque ellipses correspond to regions that are mainly located on the lateral surface
of the brain. IC d is represented in both frontal and temporal–parietal regions as
reflected in the distributed nature of this left-lateralized network. The network is
divided into word processing (shown in blue) and word generation modules (shown
in green). The SEM block diagram at bottom shows how these brain networks are
graphically connected forming the basis for the cognitive model for the covert verb
generation task. Only the Feed Forward Connections are evaluated.
Table 3 | The age-related changes in the standardized path coefficients (r
and p value) for the SEM shown in Figure 5 are shown in column 2 as
Pearson correlations between the path coefficient and age. Column 3
shows the average value of each standardized path coefficient for the entire
age range of 5–18 years included in the analysis. Path coeffiences with a
significant age correlations are highlighted in bold font.
Connection r value, p value Avg. value of Std.
path coefficient
IC e IC f 0.1295, 0.0175 0.31
IC e IC a 0.0429, 0.4326 0.16
IC e IC d 0.1804, 0.0444 0.27
IC f IC d 0.1804, 0.000888 0.33
IC d IC a 0.07917, 0.1475 0.17
IC a IC b 0.1036, 0.0577 0.36
IC a IC g 0.0232, 0.6660 0.23
IC g IC c 0.0385, 0.4807 0.57
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 9
lateralization and localization over the course of language devel-
opment (Ahmad et al., 2003; Gaillard et al., 2003). Our previous
findings of increasing left lateralization of IFG activation with
age for the verb generation task in children are consistent with
the functional connectivity findings showing decreasing left–right
connectivity with age suggesting that the left hemisphere is able
to act more autonomously in support of word generation as the
brain matures (Holland et al., 2007). This interpretation is also
consistent with the regionally weighted model of normal language
development. Further, this finding alone can explain the differ-
ences between young and old subjects in language recovery after
left-hemispheric injury with the ability of the language functions
to shift to the right hemisphere in the early (prenatal and early
postnatal injury) but dependence on the left-hemispheric regions
for aphasia recovery in late life stroke (Tillema et al., 2008; Saur
et al., 2010).
The inter-hemispheric functional connectivity between the
posterior aspects of superior temporal gyrus (IC e) showed no
age effects. The time courses for IC e and IC c described above
have shown the highest increase in task-relatedness (developmental
trend) as detailed in a previous study involving the same subject
population (Karunanayaka et al., 2010). However, the age depend-
ence of these networks differs in terms of the inter-hemispheric
connectivity as seen in Figure 4, with no significant age trend found
in the posterior network encompassed by IC e (BA22).
Although the relationship between structural maturation and
functional activation is rather complex, the present functional con-
nectivity data provides additional evidence in support of language
lateralization being dominated by the inferior frontal brain regions.
While one recent study did not observe any asymmetries in lan-
guage lateralization in newborns (Kotilahti et al., 2010), this study
also found a more uniform involvement of the left hemisphere in
speech processing indicating that left-hemispheric specialization
for language processing may already be present at birth. Dehaene-
Lambertz et al. (2002) also found that left lateralization of lan-
guage function was present in posterior brain regions in infants as
young as 3 months of age. These findings are inline with previously
reported left lateralization of language functions noted in 6- to
in the acquisition of semantic representations (Blumenfeld et al.,
2006; Booth et al., 2007). Similarly, research in adults suggests that
more activation in the inferior frontal cortex is associated with more
effortful retrieval or greater selection demands (Seger et al., 2000;
Gurd et al., 2002; Whatmough et al., 2002; Booth et al., 2007). Age
effects seen in the functional connectivity between these two regions
suggest that the selection demands imposed on the inferior frontal
gyrus increase with age. This may be due to the fact that the present
verb generation task does not impose restriction on the number
of verbs a subject can generate for a given noun. Evidence suggests
that this design is successful in minimizing the amount of variance
attributable to performance (Gaillard et al., 2003).
The functional connectivity between LMTG LANG showed
no significant age effects. The inferior parietal cortex has been
implicated in feature integration and semantic categorization to
form a coherent concept so that semantic relationships between
words can be determined (Grossman et al., 2003; Karunanayaka
et al., 2010). The demand for such processes may be at a minimum
for this task (ceiling effect) since we developed this fMRI task in
such a manner that even the youngest children in our study can
perform this task easily. Nevertheless, Booth et al. (2007) have sug-
gested that the inferior parietal lobule may have distinct areas for
processing semantic versus phonological information. This may
explain observed age effects in functional connectivity: between
LANG LMIFG with no age effects and between LMIFG LIFG
with highly significant age effects.
The significant decrease in functional connectivity with age
between right and left hemisphere elements of IC c implies a sub-
stantial change in the degree to which the left and right brain
regions (inferior frontal gyrus) co-vary. Note that structural and
functional asymmetries have also been found in the prenatal and
early postnatal brain (Wada et al., 1975; Chi et al., 1977; Dehaene-
Lambertz et al., 2002) suggesting a bias for left hemisphere language
lateralization very early in life. The anatomical data suggest that
early brain development may lead to an underlying architecture
that preferentially supports language within the left hemisphere: a
normal variant of the focal network model (Berl et al., 2006). This
neuroanatomical bias is hypothesized to be related to functional
FIGURE 6 | Standardized path coefficients corresponding to the SEM shown in Figure 5 that showed significant changes with age are plotted as a
function of age in months.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 10
verb generation in the developing brain. As mentioned previously,
compared to DCM, the current investigation only models contem-
poraneous connections without taking into account the neuronal
hemodynamic relationships explicitly (Penny et al., 2004b). The
emphasis is, therefore, on the overall network behavior confirm-
ing or facilitating the generation of new hypothesis. The current
investigation focused on the overall connectivity pattern shedding
more insight into several networks that need further investigations
using more sophisticated methods like uSEM, DCM, or Granger
causality (Stevens et al., 2007; Jafri et al., 2008; Demirci et al., 2009;
Londei et al., 2010; Smith et al., 2010). SEM is useful in this regard
in that it provides a quantitative measure of overall model fit
which allows the optimum set of path coefficients to be identified
objectively. These coefficients can then be examined as a function
of age to determine how connection strengths change with brain
development. Finally, LSEM was also used as an exploratory tool
in the proposed theoretical model in a highly restrictive manner.
By introducing a theory-driven focus we partially avoided evaluat-
ing models of different structures. However, model selection (or
identifying the true network structure) is a challenging statistical
problem that has received increased attention in the neuroimaging
community in recent times (Zheng and Rajapakse, 2006; Rajapakse
and Zhou, 2007). We have already developed a Spectral Bayesian
Network method (based on Model Averaging) to identify the most
plausible models based on fMRI data, which is inline with our
long-term objective of developing statistical methods capable of
confirming (or rejecting) existing theoretical models for cognitive
development.
lImItatIons
Study limitations inherent in covert verb generation task have
been discussed in detail elsewhere (Szaflarski et al., 2006a,b;
Karunanayaka et al., 2010). Therefore, we will only review addi-
tional limitations pertaining to the analyses employed in this paper.
In this study, we have only focused our attention on task-related
networks even though considerable amount of intrinsic fluctua-
tions are typically inherent in fMRI time courses. ICA, in general,
tends to over specify the problem imposing severe limitations on
our computational ability for connectivity analysis. Implementing
objective methods to select non-task-related components to be
included in the connectivity analysis is non-trivial. On the other
hand, including such components is very subjective making inter-
pretations difficult. The SEM should also be limited to a reasonable
number of nodes (maximum of 1015) as any data set can be fit-
ted to models with increasing complexity. Thus, we have adopted
a theory-driven focus coupled with proper selection processes to
guide the analysis and interpretations circumventing above men-
tioned drawbacks. Therefore, we had no option but to limit the
analysis to task-related components. However, if one can overcome
the computational (and methodological) limitations, DCM might
be more suitable to investigate intrinsic connectivity that is affected
by the context of the task in ways which do not show up as a strictly
task-related modulation of the time course.
As previously mentioned, ICA is a data-driven technique and,
therefore, its use obviates conventional statistical approaches
to hypothesis testing. Consequently, one extension of this data
analysis method would be to incorporate constraints at the ICA
12-month-old children (Minagawa-Kawai et al., 2007) and later
studies of language lateralization in older children, adolescents and
young adults (Holland et al., 2001; Szaflarski et al., 2006a).
The proposed LSEM for verb generation is hypothesized to sup-
port both word processing and word generation. However, only
the networks included in the word processing module exhibited
age dependent effective connectivity changes. Each of these net-
works represents a unique spatial distribution with corresponding
time course that sub-serves specific functions of the network (e.g.,
working memory, visual imagery, or acoustic word recognition). As
mentioned earlier, although the spatial distributions of IC maps are
independent, the corresponding time courses are allowed to have
highly correlated temporal structures.
According to the focal network theory, the underlying neural
network structure for language processing is generally well estab-
lished by the age of 5 (Ahmad et al., 2003) with first evidence of
network structure seen already in newborns (Kotilahti et al., 2010).
Therefore, it is reasonable to assume that interactions between
functions such as coordination of speech articulation, subvocal
word production, and visual imagery at network level are well
established for this group of children. However, based on our
results, there is ample evidence to suggest that the within network
(intra-network) behavior is undergoing a continuous process of
dynamic change. As discussed in detail elsewhere (Karunanayaka
et al., 2010), the areas of a distributed network can change the
degree of engagement making it a more efficient component of the
normally developing network. This forms the basis for the region-
ally weighted model and the differences in weights may account
for the observed normal variations in cognitive skill level, use of
different cognitive strategies and changes in the biological substrate
for a function (Berl et al., 2006). This picture is consistent with the
intra- component functional connectivity results observed for IC
c, d, and e. (Karunanayaka et al., 2010).
As mentioned above, module IC d is the most left-lateralized
part of the network for this task and is presumed to be associated
with semantic representations of the nouns that are being heard
(Karunanayaka et al., 2010). All connections to this module are
age dependent. This module may also sub-serve working memory
required by the verb generation task. Several studies have reported
age-dependent BOLD signal and connectivity changes mainly in
the frontal areas of the brain (Gaillard et al., 2000; Schlaggar et al.,
2002; Schmithorst et al., 2002; Schapiro et al., 2004). We suggest
that these later aspects of development are captured by the observed
connectivity changes within the word processing module in our
proposed model. Finally, in terms of the regionally weighted model,
these changes can be interpreted as increasing the participation of
this left-lateralized network supporting phonological and semantic
expressive functions as part of covert verb generation.
The biological relevance of the model derives from two sources.
First, the highly task-related elements of the model are selected
based on the data-driven ICA results. Secondly, the biological plau-
sibility originated with the close correspondence to the Wernicke–
Geschwind model and has been evaluated against the literature for
the neural circuitry of language; especially for the semantic process-
ing network (Kim et al., 2011). These biological underpinnings for
our model give us confidence that the proposed model is indeed rel-
evant to the cognitive and biological processes taking place during
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 11
comparison of five fMRI protocols
for mapping speech comprehension
systems. Epilepsia 49, 1980–1997.
Blumenfeld, H. K., Booth, J. R., and
Burman, D. D. (2006). Differential
prefrontal-temporal neural correlates
of semantic processing in children.
Brain Lang. 99, 226–235.
Booth, J. R., Bebko, G., Burman, D. D.,
and Bitan, T. (2007). Children with
reading disorder show modality inde-
pendent brain abnormalities during
semantic tasks. Neuropsychologia 45,
775–783.
Beitchman, J. H., Jiang, H., Koyama, E.,
Johnson, C. J., Escobar, M., Atkinson,
L., Brownlie, E. B., and Vida, R. (2008).
Models and determinants of vocabu-
lary growth from kindergarten to
adulthood. J. Child. Psychol. Psychiatry
49, 626–634.
Berl, M. M., Vaidya, C. J., and Gaillard,
W. D. (2006). Functional imaging of
developmental and adaptive changes
in neurocognition. Neuroimage 30,
679–691.
Binder, J. R., Swanson, S. J., Hammeke,
T. A., and Sabsevitz, D. S. (2008). A
of the Wernicke-Geschwind model.
Brain Lang. 70, 1–12.
Arbuckle, J. (1989). Analysis of moment
structures. Am. Stat. 43, 66–67.
Bandettini, P. A., Jesmanowicz, A., Wong,
E. C., and Hyde, J. S. (1993). Processing
strategies for time-course data sets in
functional MRI of the human brain.
Magn. Reson. Med. 30, 161–173.
Bartels, A., and Zeki, S. (2004). The
chronoarchitecture of the human
brain-natural viewing conditions
reveal a time-based anatomy of the
brain. Neuroimage 22, 419–433.
reFerences
Ahmad, Z., Balsamo, L. M., Sachs, B. C.,
Xu, B., and Gaillard, W. D. (2003).
Auditory comprehension of language
in young children: neural networks
identified with fMRI. Neurology 60,
1598–1605.
Anderson, J. M., Gilmore, R., Roper, S.,
Crosson, B., Bauer, R. M., Nadeau, S.,
Beversdorf, D. Q., Cibula, J., Rogish,
M. III, Kortencamp, S., Hughes, J. D.,
Gonzalez Rothi, L. J., and Heilman, K.
M. (1999). Conduction aphasia and
the arcuate fasciculus: a reexamination
dependent changes in network connectivity. Connections between
networks such as word recognition, phonological working memory
and semantic processing exhibited age dependent connectivity
changes to varying degrees. We also highlighted the importance
of combining network analysis at both the local and global levels.
The intra-network connectivity analysis of the highly left-lat-
eralized network (IC d) provided convincing evidence for ongo-
ing within network connectivity changes. These changes highlight
the importance of the degree of engagement of each brain region
within a network and can be considered a part of normal devel-
opment. The results show the capacity of ICA to detect additional
task-related spatial modules and their developmental trajectories.
Age dependence found in both the path coefficients and in the
task-relatedness of this network support the idea that this network
will be used increasingly for covert verb generation as children
mature. Combined with the results of our previous investigations,
the current analysis further endorses the regionally weighted model
for neurocognitive development that highlights the importance of
interactions between brain regions within a network.
The main emphasis of the current analysis was on investigat-
ing the age dependent connectivity trends in language networks in
the developing brain using functional neuroimaging data and an
inferential statistical method capable of testing hypothesis based
upon existing theoretical models for cognitive functions. Modeling
cognitive functions will always be a compromise between the com-
plexity of the underlying neural systems and the interpretability
of proposed models. However, by carefully selecting appropriate
analysis methods in an incremental manner of complexity, theories
of cognitive development can be investigated in detail. This new
approach to network modeling reveals a more complex develop-
mental trajectory than proposed by models currently in use, e.g.,
the Wernicke–Lichtheim–Geschwind theory (Geschwind, 1965a,b)
and has the potential to lead to a deeper understanding of devel-
opmental changes in cognitive function.
acknowledgments
The study was supported by a grant from the U.S. National Institute
of Child Health and Development, R01-HD38578. The authors
acknowledge the assistance of Dr. Anna Byars, Ph.D., in the admin-
istration of the Wechsler IQ tests, and Drs. Richard Strawsburg,
M.D. and Mark Schapiro, M.D., for performing the neurological
examinations. Dr. Szaflarski is currently supported by NINDS K23
NS052468. Thanks to Meghan O’Connell for applying her medical
illustration skills to Figure 5.
decomposition step to guide the analysis and increase the predic-
tive power (Lu and Rajapakse, 2005). Finally, the verb genera-
tion task utilized here was not specifically designed to acquire
in-scanner performance data. Consequently, performance effects
on the connectivity coefficients cannot be completely discounted
even though the fMRI task design enabled the youngest children
in the study to complete the task without any difficulty. Since,
performance can be related to IQ, including IQ as a covariate can
produce overcorrected, anomalous, and counterintuitive findings
about neurocognitive functions (Dennis et al., 2009). It has also
been shown that IQ should only be used as a covariate in those
rare circumstances where selection bias has produced problems of
non-representativeness in the sample (Dennis et al., 2009). Clearly,
such a condition was not present here although we observed a
small negative correlation between age and IQ. This was mainly
due to our youngest subjects having higher than average IQ scores.
Furthermore, in one of our previous connectivity studies (narrative
story comprehension) with the same population, the effects of the
age × IQ interaction term were investigated using a multivariate
regression model and were found not to confound the age-related
tendencies associated with SEM path coefficients (Karunanayaka
et al., 2007). This performance-related limitation can be addressed
in the future by collecting intra-scanner performance data using
either sparse fMRI data collection (Schmithorst and Holland, 2004)
or block-design task with forced responses (Szaflarski et al., 2002).
Such a design will also allow real-time performance on the task to
be monitored and potentially included as a covariate in the analysis
of age dependence in connectivity. Recently we have shown that
brain activation during covert verb generation correlates with the
number of verbs generated during an overt phase of verb generation
during the same task (Vannest et al., 2010). While both overt and
covert verb generation produced similar patterns of activation, the
correlation with performance suggests that performance could also
be related to connectivity in the language networks sub-serving the
tasks. This question could be specifically addressed with a modi-
fied overt verb generation task in which the number of responses
is explicitly controlled as a design parameter.
conclusIon
A theoretical model for covert verb generation was investigated
using fMRI data from a large cohort of children and adolescents
between the ages 5–18 years undergoing fMRI study with such a
task. Previously identified, spatially independent and task-related
networks (IC maps) were combined with SEM to investigate age
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 12
Hyvarinen, A. (1999b). Sparse code
shrinkage: denoising of nongaussian
data by maximum likelihood estima-
tion. Neural Comput. 11, 1739–1768.
Jafri, M. J., Pearlson, G. D., Stevens, M.,
and Calhoun, V. D. (2008). A method
for functional network connectivity
among spatially independent resting-
state components in schizophrenia.
Neuroimage 39, 1666–1681.
Jennings, J. M., McIntosh, A. R., and
Kapur, S. (1998). Mapping neural
interactivity onto regional activity:
an analysis of semantic processing
and response mode interactions.
Neuroimage 7, 244–254.
Karunanayaka, P., Kim, K. K., Holland,
S. K., and Szaflarski, J. P. (2011). The
effects of left or right hemispheric
epilepsy on language networks inves-
tigated with semantic decision fMRI
task and independent component
analysis. Epilepsy Behav. 20, 623–632.
Karunanayaka, P., Schmithorst, V. J.,
Vannest, J., Szaflarski, J. P., Plante, E.,
and Holland, S. K. (2010). A group
independent component analysis of
covert verb generation in children: a
functional magnetic resonance imag-
ing study. Neuroimage 51, 472–487.
Karunanayaka, P. R., Holland, S. K.,
Schmithorst, V. J., Solodkin, A., Chen,
E. E., Szaflarski, J. P., and Plante, E.
(2007). Age-related connectivity
changes in fMRI data from children
listening to stories. Neuroimage 34,
349–360.
Kim, J., and Horwitz, B. (2009). How well
does structural equation modeling
reveal abnormal brain anatomical
connections? An fMRI simulation
study. Neuroimage 45, 1190–1198.
Kim, K. K., Karunanayaka, P., Privitera, M.
D., Holland, S. K., and Szaflarski, J. P.
(2011). Semantic association investi-
gated with functional MRI and inde-
pendent component analysis. Epilepsy
Behav. 20, 613–622.
Kotilahti, K., Nissila, I., Näsi, T., Lipiäinen,
L., Noponen, T., Meriläinen, P.,
Huotilainen, M., and Fellman, V.
(2010). Hemodynamic responses to
speech and music in newborn infants.
Hum. Brain Mapp. 31, 595–603.
Lenneberg, E. H. (1967). Biological
Foundations of Language. New York:
Wiley.
Levelt, W. J., Roelofs, A., and Meyer, A. S.
(1999). A theory of lexical access in
speech production. Behav. Brain Sci.
22, 1–38; discussion 38–75.
Londei, A., D’Ausilio, A., Basso, D., and
Belardinelli, M. O. (2006). A new
method for detecting causality in
fMRI data of cognitive processing.
Cogn. Process. 7, 42–52.
Londei, A., D’Ausilio, A., Basso, D.,
Sestieri, C., Del Gratta, C., Romani,
G. L., and Olivetti Belardinelli, M.
C. B. (2003). Developmental aspects
of language processing: fMRI of verbal
fluency in children and adults. Hum.
Brain Mapp. 18, 176–185.
Geschwind, N. (1965a). Disconnexion
syndromes in animals and man. I.
Brain 88, 237–294.
Geschwind, N. (1965b). Disconnexion
syndromes in animals and man. II.
Brain 88, 585–644.
Giedd, J. N. (2004). Structural magnetic
resonance imaging of the adolescent
brain. Ann. N. Y. Acad. Sci. 1021, 77–85.
Giedd, J. N., Blumenthal, J., Jeffries, N. O.,
Castellanos, F. X., Liu, H., Zijdenbos,
A., Paus, T., Evans, A. C., and Rapoport,
J. L. (1999). Brain development during
childhood and adolescence: a longi-
tudinal MRI study. Nat. Neurosci. 2,
861–863.
Goncalves, M. S., Hall, D. A., Johnsrude,
I. S., and Haggard, M. P. (2001). Can
meaningful effective connectivi-
ties be obtained between auditory
cortical regions? Neuroimage 14,
1353–1360.
Grossman, M., Koenig, P., Glosser, G.,
DeVita, C., Moore, P., Rhee, J., Detre,
J., Alsop, D., Gee, J., and fMRI Study,
Functional Magnetic Resonance
Imaging. (2003). Neural basis for
semantic memory difficulty in
Alzheimer’s disease: an fMRI study.
Brain 126(Pt 2), 292–311.
Gurd, J. M., Amunts, K., Weiss, P. H.,
Zafiris, O., Zilles, K., Marshall, J. C.,
and Fink, G. R. (2002). Posterior
parietal cortex is implicated in con-
tinuous switching between verbal
fluency tasks: an fMRI study with
clinical implications. Brain 125(Pt 5),
1024–1038.
Hickok, G., and Poeppel, D. (2004). Dorsal
and ventral streams: a framework for
understanding aspects of the func-
tional anatomy of language. Cognition
92, 67–99.
Himberg, J., Hyvarinen, A., and Esposito,
F. (2004). Validating the independent
components of neuroimaging time
series via clustering and visualization.
Neuroimage 22, 1214–1222.
Holland, S. K., Plante, E., Weber Byars, A.,
Strawsburg, R. H., Schmithorst, V. J.,
and Ball, W. S. Jr. (2001). Normal fMRI
brain activation patterns in children
performing a verb generation task.
Neuroimage 14, 837–843.
Holland, S. K., Vannest, J., Mecoli,
M., Jacola, L. M., Tillema, J. M.,
Karunanayaka, P. R., Schmithorst, V.
J., Yuan, W., Plante, E., and Byars, A. W.
(2007). Functional MRI of language
lateralization during development in
children. Int. J. Audiol. 46, 533–551.
Hyvarinen, A. (1999a). Fast and robust
fixed-point algorithms for independ-
ent component analysis. IEEE Trans.
Neural Netw. 10, 626–634.
Dennis, M., Francis, D. J., Cirino, P. T.,
Schachar, R., Barnes, M. A., and
Fletcher, J. M. (2009). Why IQ is not
a covariate in cognitive studies of
neurodevelopmental disorders. J. Int.
Neuropsychol. Soc. 15, 331–343.
Dick, A. S., Solodkin, A., and Small, S. L.
(2010). Neural development of net-
works for audiovisual speech com-
prehension. Brain Lang. 114, 101–114.
Duann, J. R., Jung, T. P., Kuo, W.-J., Yeh,
T.-C., Makeig, S., Hsieh, J.-C., and
Sejnowski, T. J. (2002). Single-trial
variability in event-related BOLD
signals. Neuroimage 15, 823–835.
Foundas, A. L., Leonard, C. M., lmore,
R., Fennell, E., and Heilman, K. M.
(1994). Planum temporale asym-
metry and language dominance.
Neuropsychologia 32, 1225–1231.
Francks, C., Maegawa, S., Laurén, J.,
Abrahams, B. S., Velayos-Baeza, A.,
Medland, S. E., Colella, S., Groszer, M.,
McAuley, E. Z., Caffrey, T. M., Timmusk,
T., Pruunsild, P., Koppel, I., Lind, P. A.,
Matsumoto-Itaba, N., Nicod, J., Xiong,
L., Joober, R., Enard, W., Krinsky, B.,
Nanba, E., Richardson, A. J., Riley, B.
P., Martin, N. G., Strittmatter, S. M.,
Möller, H. J., Rujescu, D., St Clair, D.,
Muglia, P., Roos, J. L., Fisher, S. E.,
Wade-Martins, R., Rouleau, G. A., Stein,
J. F., Karayiorgou, M., Geschwind, D. H.,
Ragoussis, J., Kendler, K. S., Airaksinen,
M. S., Oshimura, M., DeLisi, L. E., and
Monaco, A. P. (2007). LRRTM1 on
chromosome 2p12 is a maternally
suppressed gene that is associated
paternally with handedness and schizo-
phrenia. Mol. Psychiatry 12, 1129–1139.
Friston, K. J., Buechel, C., Fink, G. R.,
Morris, J., Rolls, E., and Dolan, R. J.
(1997). Psychophysiological and mod-
ulatory interactions in neuroimaging.
Neuroimage 6, 218–229.
Friston, K. J., Harrison, L., and Penny, W.
(2003). Dynamic causal modelling.
Neuroimage 19, 1273–1302.
Friston, K. J., Holmes, A. P., Poline,
J. B., Grasby, P. J., Williams, S. C.,
Frackowiak, R. S., and Turner, R.
(1995). Analysis of fMRI time-series
revisited. Neuroimage 2, 45–53.
Friston, K. J., and Stephan, K. E. (2007).
Statistical Parametric Mapping: The
Analysis of Functional Brain Images,
eds K. J. Friston, J. T. Ashburner, S. J.
Kiebel, T. E. Nichols, and W. D. Penny
(Amsterdam: Academic Press).
Gaillard, W. D., Hertz-Pannier, L., Mott,
S. H., Barnett, A. S., LeBihan, D., and
Theodore, W. H. (2000). Functional
anatomy of cognitive development:
fMRI of verbal fluency in children and
adults. Neurology 54, 180–185.
Gaillard, W. D., Sachs, B. C., Whitnah, J.
R., Ahmad, Z., Balsamo, L. M., Petrella,
J. R., Braniecki, S. H., McKinney, C.
M., Hunter, K., Xu, B., and Grandin,
Byars, A. W., Holland, S. K., Strawsburg,
R. H., Bommer, W., Dunn, R. S.,
Schmithorst, V. J., and Plante, E.
(2002). Practical aspects of conducting
large-scale functional magnetic reso-
nance imaging studies in children. J.
Child Neurol. 17, 885–890.
Calhoun, V. D., Adali, T., McGinty, V.
B., Pekar, J. J., Watson, T. D., and
Pearlson, G. D. (2001a). fMRI acti-
vation in a visual-perception task:
network of areas detected using the
general linear model and independ-
ent components analysis. Neuroimage
14, 1080–1088.
Calhoun, V. D., Adali, T., Pearlson, G. D.,
and Pekar, J. J. (2001b). A method for
making group inferences from func-
tional MRI data using independent
component analysis. Hum. Brain
Mapp. 14, 140–151.
Calhoun, V. D., Pekar, J. J., and Pearlson,
G. D. (2004). Alcohol intoxica-
tion effects on simulated driving:
exploring alcohol-dose effects on
brain activation using functional
MRI. Neuropsychopharmacology 29,
2097–2017.
Carrow-Woolfolk, E. (1996). Oral and
Written Language Scales. Circle Pines,
MN: American Guidance Service.
Catani, M., Jones, D. K., and Ffytche, D. H.
(2005). Perisylvian language networks
of the human brain. Ann. Neurol. 57,
8–16.
Chein, J. M., Ravizza, S. M., and Fiez, J.
A. (2003). Using neuroimaging to
evaluate models of working memory
and their implications for language
processing. J. Neurolinguistics 16,
315–339.
Chi, J. G., Dooling, E. C., and Gilles, F. H.
(1977). Left-right asymmetries of the
temporal speech areas of the human
fetus. Arch. Neurol. 34, 346–348.
Chou, T. L., Booth, J. R., Burman, D.
D., Bitan, T., Bigio, J. D., Lu, D., and
Cone, N. E. (2006). Developmental
changes in the neural correlates of
semantic processing. Neuroimage 29,
1141–1149.
Dehaene-Lambertz, G., Dehaene, S., and
Hertz-Pannier, L. (2002). Functional
neuroimaging of speech perception in
infants. Science 298, 2013–2015.
de Marco, G., Devauchelle, B., and
Berquin, P. (2009). Brain functional
modeling, what do we measure with
fMRI data? Neurosci. Res. 64, 12–19.
Demirci, O., Stevens, M. C., Andreasen,
N. C., Michael, A., Liu, J., White,
T., Pearlson, G. D., Clark, V. P., and
Calhoun, V. D. (2009). Investigation
of relationships between fMRI brain
networks in the spectral domain using
ICA and Granger causality reveals
distinct differences between schizo-
phrenia patients and healthy controls.
Neuroimage 46, 419–431.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 13
Hammeke, T. A. (2002). Language
lateralization in left-handed and ambi-
dextrous people: fMRI data. Neurology
59, 238–244.
Szaflarski, J. P., DiFrancesco, M.,
Hirschauer, T., Banks, C., Privitera,
M. D., Gotman, J., and Holland, S. K.
(2010). Cortical and subcortical con-
tributions to absence seizure onset
examined with EEG/fMRI. Epilepsy
Behav. 18, 404–413.
Szaflarski, J. P., Holland, S. K., Schmithorst, V.
J., and Byars, A. W. (2006a). fMRI study
of language lateralization in children and
adults. Hum. Brain Mapp. 27, 202–212.
Szaflarski, J. P., Schmithorst, V. J., Altaye,
M., Byars, A. W., Ret, J., Plante, E., and
Holland, S. K. (2006b). A longitudinal
functional magnetic resonance imag-
ing study of language development in
children 5 to 11 years old. Ann. Neurol.
59, 796–807.
Thompson-Schill, S. L., D’Esposito, M.,
Aguirre, G. K., and Farah, M. J. (1997).
Role of left inferior prefrontal cortex
in retrieval of semantic knowledge:
a reevaluation. Proc. Natl. Acad. Sci.
U.S.A. 94, 14792–14797.
Tillema, J. M., Byars, A. W., Jacola, L. M.,
Schapiro, M. B., Schmithorst, V. J.,
Szaflarski, J. P., and Holland, S. K. (2008).
Cortical reorganization of language
functioning following perinatal left
MCA stroke. Brain Lang. 105, 99–111.
Vannest, J., Rasmussen, J. M., Eaton,
K. P., Patel, K., Schmithorst, V.,
Karunanayaka, P., Plante, E., Byars,
A., and Holland, S. (2010). fMRI
Activation in language areas reflects
childrens verb generation perfor-
mance. Neuropediatrics 2010 (neuped-
04-2010-0577-SC). 41, 235–239.
Wada, J. A., Clarke, R., and Hamm, A.
(1975). Cerebral hemispheric asym-
metry in humans. Cortical speech
zones in 100 adults and 100 infant
brains. Arch. Neurol. 32, 239–246.
Wechsler, D. (1991). Wechsler Intelligence
Scale for Children, 3rd Edn. San
Antonio, TX: The Psychological
Corporation.
Wechsler, D. (1997). Wechsler Adult
Intlelligence Scale, 3rd Edn. San
Antonio, TX: The Psychological
Corporation.
Whatmough, C., Chertkow, H., Murtha,
S., and Hanratty, K. (2002). Dissociable
brain regions process object meaning
and object structure during picture
naming. Neuropsychologia 40, 174–186.
Wood, A. G., Harvey, A. S., Wellard, R.
M., Abbott, D. F., Anderson, V., Kean,
M., Saling, M. M., and Jackson, G. D.
(2004). Language cortex activation
in normal children. Neurology 63,
1035–1044.
Worsley, K. J., and Friston, K. J. (1995).
Analysis of fMRI time-series revisited-
again. Neuroimage 2, 173–181.
utilized for narrative comprehension
in children: a functional magnetic
resonance imaging study. Neuroimage
29, 254–266.
Schmithorst, V. J., Holland, S. K., and
Plante, E. (2007a). Development
of effective connectivity for nar-
rative comprehension in children.
Neuroreport 18, 1411–1415.
Schmithorst, V. J., Holland, S. K., and
Plante, E. (2007b). Object identifi-
cation and lexical/semantic access
in children: a functional magnetic
resonance imaging study of word-
picture matching. Hum. Brain Mapp.
28, 1060–1074.
Schmithorst, V. J., Wilke, M., Dardzinski,
B. J., and Holland, S. K. (2002).
Correlation of white matter diffusivity
and anisotropy changes with age dur-
ing childhood: a cross-sectional diffu-
sion tensor imaging study. Radiology
222, 212–218.
Seger, C. A., Desmond, J. E., Glover, G. H.,
and Gabrieli, J. D. (2000). Functional
magnetic resonance imaging evidence
for right-hemisphere involvement in
processing unusual semantic relation-
ships. Neuropsychology 14, 361–369.
Skipper, J. I., Nusbaum, H. C., and Small,
S. L. (2005). Listening to talking
faces: motor cortical activation dur-
ing speech perception. Neuroimage
25, 76–89.
Smith, J. F., Alexander, G. E., Chen, K.,
Husain, F. T., Kim, J., Pajor, N., and
Horwitz, B. (2010). Imaging sys-
tems level consolidation of novel
associate memories: a longitudinal
neuroimaging study. Neuroimage
50, 826–836.
Solodkin, A., Hlustik, P. E., Chen, E., and
Small, S. L. (2004). Fine modulation
in network activation during motor
execution and motor imagery. Cereb.
Cortex 14, 1246–1255.
Stevens, M. C., Kiehl, K. A., Pearlson, G. D.,
and Calhoun, V. D. (2007). Functional
neural networks underlying response
inhibition in adolescents and adults.
Behav. Brain Res. 181, 12–22.
Stevens, M. C., Pearlson, G. D., and
Calhoun, V. D. (2009a). Changes in the
interaction of resting-state neural net-
works from adolescence to adulthood.
Hum. Brain Mapp. 30, 2356–2366.
Stevens, M. C., Skudlarski, P., Pearlson, G.
D., and Calhoun, V. D. (2009b). Age-
related cognitive gains are mediated
by the effects of white matter develop-
ment on brain network integration.
Neuroimage 48, 738–746.
Stoeckel, L. E., Kim, J., Weller, R. E., Cox,
J. E., and Cook, E. W. III, Horwitz,
B. (2009). Effective connectivity of
a reward network in obese women.
Brain Res. Bull. 79, 388–395.
Szaflarski, J. P., Binder, J. R., Possing, E.
T., McKiernan, K. A., Ward, B. D., and
causal models. Neuroimage 23(Suppl.
1), S264–S274.
Plaut, D. C., McClelland, J. L., Seidenberg,
M. S., and Patterson, K. (1996).
Understanding normal and impaired
word reading: computational princi-
ples in quasi-regular domains. Psychol.
Rev. 103, 56–115.
Pulvermuller, F. (2001). Brain reflections
of words and their meaning. Trends
Cogn. Sci. (Regul. Ed.) 5, 517–524.
Rajapakse, J. C., Wang, Y., Zheng, X., and
Zhou, J. (2008). Probabilistic frame-
work for brain connectivity from
functional MR images. IEEE Trans.
Med. Imaging 27, 825–833.
Rajapakse, J. C., and Zhou, J. (2007).
Learning effective brain connectiv-
ity with dynamic Bayesian networks.
Neuroimage 37, 749–760.
Sarty, G. E. (2007). Computational
brain activity maps from fMRI time-
series images. New York: Cambridge
University Press.
Saur, D., Ronneberger, O., Kümmerer, D.,
Mader, I., Weiller, C., and Klöppel, S.
(2010). Early functional magnetic
resonance imaging activations predict
language outcome after stroke. Brain
133(Pt 4), 1252–1264.
Schapiro, M. B., Schmithorst, V. J., Wilke,
M., Byars, A. W., Strawsburg, R. H.,
and Holland, S. K. (2004). BOLD fMRI
signal increases with age in selected
brain regions in children. Neuroreport
15, 2575–2578.
Schlaggar, B. L., Brown, T. T., Lugar, H.
M., Visscher, K. M., Miezin, F. M.,
and Petersen, S. E. (2002). Functional
neuroanatomical differences between
adults and school-age children in the
processing of single words. Science 296,
1476–1479.
Schmithorst, V. J., and Brown, R. D.
(2004). Empirical validation of the
triple-code model of numerical pro-
cessing for complex math operations
using functional MRI and group
independent component analysis of
the mental addition and subtraction of
fractions. Neuroimage 22, 1414–1420.
Schmithorst, V. J., and Holland, S. K.
(2004). Event-related fMRI technique
for auditory processing with hemody-
namics unrelated to acoustic gradient
noise. Magn. Reson. Med. 51, 399–402.
Schmithorst, V. J., and Holland, S. K.
(2006). Functional MRI evidence for
disparate developmental processes
underlying intelligence in boys and
girls. Neuroimage 31, 1366–1379.
Schmithorst, V. J., and Holland, S. K.
(2007). Sex differences in the develop-
ment of neuroanatomical functional
connectivity underlying intelligence
found using Bayesian connectivity
analysis. Neuroimage 35, 406–419.
Schmithorst, V. J., Holland, S. K., and
Plante, E. (2006). Cognitive modules
(2007). Brain network for passive
word listening as evaluated with ICA
and Granger causality. Brain Res. Bull.
72, 284–292.
Londei, A., D’Ausilio, A., and Basso, D.,
Sestieri, C., Gratta, C. D., Romani,
G. L., and Belardinelli, M. O. (2010).
Sensory-motor brain network con-
nectivity for speech comprehension.
Hum. Brain Mapp. 31, 567–580.
Lu, W., and Rajapakse, J. C. (2005).
Approach and applications of con-
strained ICA. IEEE Trans. Neural Netw.
16, 203–212.
McDonald, J. L. (1997). Language acqui-
sition: the acquisition of linguistic
structure in normal and special
populations. Annu. Rev. Psychol. 48,
215–241.
McGregor, K. K., Friedman, R. M., Reilly,
R. M., and Newman, R. M. (2002).
Semantic representation and naming
in young children. J. Speech Lang. Hear.
Res. 45, 332–346.
McIntosh, A. R., and Gonzalez-Lima, F.
(1994). Structural equation modeling
and its application to network analy-
sis in functional brain imaging. Hum.
Brain Mapp. 2, 22.
McIntosh, A. R., Grady, C. L., Ungerleider,
L. G., Haxby, J. V., Rapoport, S. I., and
Horwitz, B. (1994). Network analysis
of cortical visual pathways mapped
with PET. J. Neurosci. 14, 655–666.
McKeown, M. J., Makeig, S., Brown, G. G.,
Jung, T. P., Kindermann, S. S., Bell, A.
J., and Sejnowski, T. J. (1998). Analysis
of fMRI data by blind separation into
independent spatial components.
Hum. Brain Mapp. 6, 160–188.
Minagawa-Kawai, Y., Mori, K., Naoi, N.,
and Kojima, S. (2007). Neural attune-
ment processes in infants during the
acquisition of a language-specific
phonemic contrast. J. Neurosci. 27,
315–321.
Muller, R. A., Rothermel, R. D., Behen,
M. E., Muzik, O., Mangner, T. J., and
Chugani, H. T. (1998). Developmental
changes of cortical and cerebellar
motor control: a clinical positron
emission tomography study with
children and adults. J. Child Neurol.
13, 550–556.
Oldfield, R. C. (1971). The assessment and
analysis of handedness: the Edinburgh
inventory. Neuropsychologia 9, 97–113.
Ollinger, J. M., Corbetta, M., and Corbetta,
M. (2001). Separating processes within
a trial in event-related functional MRI.
Neuroimage 13, 218–229.
Penny, W. D., Stephan, K. E., Mechelli, A.,
and Friston, K. J. (2004a). Comparing
dynamic causal models. Neuroimage
22, 1157–1172.
Penny, W. D., Stephan, K. E., Mechelli, A.,
and Friston, K. J. (2004b). Modelling
functional integration: a comparison
of structural equation and dynamic
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 14
Front. Syst. Neurosci. 5:29. doi: 10.3389/
fnsys.2011.00029
Copyright © 2011 Karunanayaka,
Schmithorst, Vannest, Szaflarski, Plante
and Holland. This is an open-access
article subject to a non-exclusive license
between the authors and Frontiers Media
SA, which permits use, distribution and
reproduction in other forums, provided
the original authors and source are cred-
ited and other Frontiers conditions are
complied with.
conducted in the absence of any commer-
cial or financial relationships that could be
construed as a potential conflict of interest.
Received: 24 May 2010; accepted: 29 April
2011; published online: 01 June 2011.
Citation: Karunanayaka P, Schmithorst
VJ, Vannest J, Szaflarski JP, Plante E and
Holland SK (2011) A linear structural
equation model for covert verb generation
based on independent component analysis
of fMRI data from children and adolescents.
(2006). The impact of early childhood
lead exposure on brain organization:
a functional magnetic resonance
imaging study of language function.
Pediatrics 118, 971–977.
Zheng, X., and Rajapakse, J. C. (2006).
Learning functional structure
from fMR images. Neuroimage 31,
1601–1613.
Conflict of Interest Statement: The
authors declare that the research was
Write, S. (1934). The method of path coef-
ficients. Ann. Math. Stat. 5, 161–215.
Yuan, W., Altaye, M., Ret, J., Schmithorst,
V., Byars, A. W., Plante, E., and
Holland, S. K. (2009). Quantification
of head motion in children during
various fMRI language tasks. Hum.
Brain Mapp. 30, 1481–1489.
Yuan, W., Holland, S. K., Cecil, K. M.,
Dietrich, K. N., Wessel, S. D., Altaye,
M., Hornung, R. W., Ris, M. D.,
Egelhoff, J. C., and Lanphear, B. P.
Karunanayaka et al. A developmental model for language
Frontiers in Systems Neuroscience www.frontiersin.org June 2011 | Volume 5 | Article 29 | 15
    • "with the earliest accounts of expressive language disturbance following left inferior frontal injury (e.g., Broca, 1861; see also, Dronkers, 1996) and subsequent clinically informed models of gross language representation championed by Geschwind (1970 Geschwind ( , 1972). Localization is also consistent with previous neuroimaging studies of expressive language in both children (e.g., Karunanayaka et al., 2010 Karunanayaka et al., , 2011) and adults (e.g., McCarthy et al., 1993; Szaflarski et al., 2006). In general, the left inferior and middle frontal cortex has been implicated in semantic processing and word retrieval ; the insular cortex is involved in speech or articulatory planning (Price, 2012). "
    [Show abstract] [Hide abstract] ABSTRACT: Resting-state functional magnetic resonance imaging (fMRI) is a promising tool for neuroscience and clinical studies. However, there exist significant variations in strength and spatial extent of resting-state functional connectivity over repeated sessions in a single or multiple subjects with identical experimental conditions. Reproducibility studies have been conducted for resting-state fMRI where the reproducibility was usually evaluated in predefined regions-of-interest (ROI). It was possible that reproducibility measures strongly depended on the ROI definition. In this work, this issue was investigated by comparing data-driven and predefined ROI-based quantification of reproducibility. In the data-driven analysis, the reproducibility was quantified using functionally connected voxels detected by a support vector machine (SVM)-based technique. In the predefined ROI-based analysis, all voxels in the predefined ROIs were included when estimating the reproducibility. Experimental results show that 1) a moderate to substantial within-subject reproducibility and a reasonable between-subject reproducibility can be obtained using functionally connected voxels identified by the SVM-based technique; 2) in the predefined ROI-based analysis, an increase in ROI size does not always result in higher reproducibility measures; 3) ROI pairs with high connectivity strength have a higher chance to exhibit high reproducibility; 4) ROI pairs with high reproducibility do not necessarily have high connectivity strength; 5) the reproducibility measured from the identified functionally connected voxels is generally higher than that measured from all voxels in predefined ROIs with typical sizes. The findings 2) and 5) suggest that conventional ROI-based analyses would under-estimate the resting-state fMRI reproducibility.
    Full-text · Article · Oct 2015
    • "with the earliest accounts of expressive language disturbance following left inferior frontal injury (e.g., Broca, 1861; see also, Dronkers, 1996) and subsequent clinically informed models of gross language representation championed by Geschwind (1970 Geschwind ( , 1972). Localization is also consistent with previous neuroimaging studies of expressive language in both children (e.g., Karunanayaka et al., 2010 Karunanayaka et al., , 2011) and adults (e.g., McCarthy et al., 1993; Szaflarski et al., 2006). In general, the left inferior and middle frontal cortex has been implicated in semantic processing and word retrieval ; the insular cortex is involved in speech or articulatory planning (Price, 2012). "
    [Show abstract] [Hide abstract] ABSTRACT: Using noninvasive neuroimaging, researchers have shown that young children have bilateral and diffuse language networks, which become increasingly left-lateralized and focal with development. Connectivity within the distributed pediatric language network has been minimally studied, and conventional neuroimaging approaches do not distinguish task-related signal changes from those that are task-essential. Here, we propose a novel multimodal method to map core language sites from patterns of information flux. We retrospectively analyze neuroimaging data collected in two groups of children, ages 5 to 18 years, performing verb generation in fMRI (n = 343) and MEG (n = 21). The fMRI data were conventionally analyzed and the group activation map parcellated in order to define node locations. Neuronal activity at each node was estimated from MEG data using a linearly constrained minimum variance beamformer, and effective connectivity within canonical frequency bands computed using the phase slope index (PSI) metric. We observed significant (p ≤ 0.05) effective connections in all subjects. The number of suprathreshold connections was significantly and linearly correlated with participant age (r = 0.50, n = 21, p ≤ 0.05), suggesting that core language sites emerge as part of the normal developmental trajectory. Across frequencies, we observed significant effective connectivity among proximal left frontal nodes. Within the low frequency bands, information flux was rostrally-directed within a focal, left frontal region, approximating Broca's area. At higher frequencies, we observed increased connectivity involving bilateral perisylvian nodes. Frequency-specific differences in patterns of information flux were resolved through fast (i.e., MEG) neuroimaging.
    Full-text · Article · Oct 2015
    • "Such is the case between groups (Kim and Horwitz, 2009) and between experimental conditions, which entails a difficult situation to manage in statistical terms. Despite these comments, effective connectivity models based on SEMs have proven useful for their verification, and many are the papers which can be considered as good praxis from a statistical point of view (Rowe, 2010; Carballedo et al., 2011; Karunanayaka et al., 2011; Inman et al., 2012), in addition to the modifications and extensions generated (Chen et al., 2011), or the use of Extended Unified SEMs (Gates et al., 2011). The choice of DCM models seems an interesting alternative to the SEM models, given that their statistical properties make them somewhat more malleable. "
    [Show abstract] [Hide abstract] ABSTRACT: The study of orthographic errors in a transparent language like Spanish is an important topic in relation to writing acquisition. The development of neuroimaging techniques, particularly functional magnetic resonance imaging (fMRI), has enabled the study of such relationships between brain areas. The main objective of the present study was to explore the patterns of effective connectivity by processing pseudohomophone orthographic errors among subjects with high and low spelling skills. Two groups of 12 Mexican subjects each, matched by age, were formed based on their results in a series of ad hoc spelling-related out-scanner tests: a high spelling skills (HSSs) group and a low spelling skills (LSSs) group. During the f MRI session, two experimental tasks were applied (spelling recognition task and visuoperceptual recognition task). Regions of Interest and their signal values were obtained for both tasks. Based on these values, structural equation models (SEMs) were obtained for each group of spelling competence (HSS and LSS) and task through maximum likelihood estimation, and the model with the best fit was chosen in each case. Likewise, dynamic causal models (DCMs) were estimated for all the conditions across tasks and groups. The HSS group’s SEM results suggest that, in the spelling recognition task, the right middle temporal gyrus, and, to a lesser extent, the left parahippocampal gyrus receive most of the significant effects, whereas the DCM results in the visuoperceptual recognition task show less complex effects, but still congruent with the previous results, with an important role in several areas. In general, these results are consistent with the major findings in partial studies about linguistic activities but they are the first analyses of statistical effective brain connectivity in transparent languages.
    Full-text · Article · May 2015
Show more