Brain Topogr (2016) 29:429–439

DOI 10.1007/s10548-015-0463-1

ORIGINAL PAPER

Connectivity Analysis and Feature Classification in Attention

Deficit Hyperactivity Disorder Sub-Types: A Task Functional
Magnetic Resonance Imaging Study
Bo-yong Park1 • Mansu Kim2 • Jongbum Seo3 • Jong-min Lee4 • Hyunjin Park5,6

Received: 26 May 2015 / Accepted: 16 November 2015 / Published online: 24 November 2015
Ó Springer Science+Business Media New York 2015

Abstract Attention deficit hyperactivity disorder Diagnostic and Statistical Manual of Mental Disorders
(ADHD) is a pervasive neuropsychiatric disorder. Patients (DSM) scores. Significant connectivity differences between
with different ADHD subtypes show different behaviors ADHD subtypes were identified mainly in the frontal,
under different stimuli and thus might require differential cingulate, and parietal cortices and partially in the tempo-
approaches to treatment. This study explores connectivity ral, occipital cortices and cerebellum. Classifier accuracy
differences between ADHD subtypes and attempts to for distinguishing between ADHD subtypes was 91.18 %
classify these subtypes based on neuroimaging features. A for both gambling punishment and emotion task paradigms.
total of 34 patients (13 ADHD-IA and 21 ADHD-C sub- Linear prediction under the two task paradigms showed
types) underwent functional magnetic resonance imaging significant correlation with DSM hyperactive/impulsive
(fMRI) with six task paradigms. Connectivity differences score. Our study identified important brain regions from
between ADHD subtypes were assessed for the whole brain connectivity analysis based on an fMRI paradigm using
in each task paradigm. Connectivity measures of the gambling punishment and emotion task paradigms. The
identified regions were used as features for the support regions and associated connectivity measures could serve
vector machine classifier to distinguish between ADHD as features to distinguish between ADHD subtypes.
subtypes. The effectiveness of connectivity measures of the
regions were tested by predicting ADHD-related Keywords Connectivity  ADHD  ADHD subtypes 
Task fMRI  SVM classifier

Electronic supplementary material The online version of this

article (doi:10.1007/s10548-015-0463-1) contains supplementary
material, which is available to authorized users. Introduction
& Hyunjin Park Attention deficit hyperactivity disorder (ADHD) is a per-
[email protected]
vasive neuropsychiatric disorder frequently associated with
1
Department of Electronic, Electrical and Computer failure of executive and reward related functions (Baroni
Engineering, Sungkyunkwan University, Suwon, Korea and Castellanos 2015). Patients with ADHD show inat-
2
Graduate School of Human ICT Convergence, tention, hyperactivity, and impulsive behaviors. ADHD
Sungkyunkwan University, Suwon, Korea patients are grouped into three subtypes: inattentive (IA),
3
Department of Biomedical Engineering, Yonsei University, hyperactive/impulsive (HI), and combined (C) types
Wonju, Korea according to the Diagnostic and Statistical Manual of
4
Department of Biomedical Engineering, Hanyang University, Mental Disorders IV (DSM-IV), which this study adopted
Seoul, Korea (American Psychiatric Association 1994). A recent criteria,
5
School of Electronic and Electrical Engineering, DSM-V, are similar to DSM-IV in regards to ADHD
Sungkyunkwan University, Suwon, Korea except that the term ‘‘subtypes’’ was replaced with ‘‘pre-
6
Center for Neuroscience Imaging Research (CNIR), Institute sentations’’ (American Psychiatric Association 2013).
for Basic Science, Suwon, Korea Previous studies have demonstrated that patients of

123
430 Brain Topogr (2016) 29:429–439

different ADHD subtypes react differently to learning They observed that motor network was the primary region
problems (Barnard et al. 2010). ADHD-IA patients are whose connectivity could distinguish among ADHD sub-
easily distracted and have a difficult time focusing on a types as connectivity alteration occurred in the motor
given task. ADHD-HI patients tend to interrupt or intrude network for the ADHD hyperactive/impulsive type.
on others. ADHD-C patients suffer from issues of both Another study used task fMRI to compare differences
ADHD-IA and ADHD-HI patients. The importance of among ADHD subtypes (Solanto et al. 2009). They
managing these behaviors and improving academic observed that ADHD-IA patients showed greater activation
achievement for such patients has increased as ADHD has in the middle frontal, temporal, and parietal cortices and
become more prevalent (Faraone and Sergeant 2003). less activation in medial occipital cortex than ADHD-C
Behavioral and medication therapies including stimulant patients. Raw data from imaging modalities cannot be used
and non-stimulant medications are used for ADHD treat- to directly extract relevant information for ADHD research.
ment (Pelham and Gnagy 1999; Castle et al. 2007; Willcutt Connectivity analysis focuses on how activities in one
et al. 2012). They found ADHD subtypes showed limited region correlate with activities in another region and thus
differences in therapeutic efficacy for different ADHD allows the observation of the whole brain as a complex,
treatment options, while all subtypes showed significant interconnected network (Watts and Strogatz 1998; Bassett
enhancement of the inattention and hyperactive/impulsive and Bullmore 2006). Well-established theories based on
symptoms (Willcutt et al. 2012). Researchers have reported network graphs can be applied to quantify the connectivity
that applying indiscriminate treatment without considering within a given brain (Stam and Reijneveld 2007; Bullmore
ADHD subtype could result in a negative influence on and Sporns 2009; Rubinov and Sporns 2010). Several
patient behavior or academic achievement (Barnard et al. network parameters including betweenness centrality and
2010). They have also argued that properly understanding degree have been adopted to quantify brain network pat-
and classifying ADHD subtypes should be considered prior terns (Rubinov and Sporns 2010). Betweenness centrality
to treatment (Barnard et al. 2010). Many researchers have is the number of shortest paths between any combinations
attempted to identify differentiating behavior patterns of nodes that run through that node and degree is the
among ADHD subtypes (Carlson et al. 1986; Baeyens et al. number of links connected to a node (Rubinov and Sporns
2006). Several studies have reported that patients of dif- 2010). These two measures largely reflect importance of a
ferent ADHD subtypes behaved differently when asked to given node.
perform certain tasks (Conzelmann et al. 2009; Gong et al. We aim to: (1) explore connectivity differences between
2014). ADHD-IA and ADHD-C patients showed different adult ADHD-IA and ADHD-C patients using task fMRI;
responses to gambling punishment and emotion tasks (2) apply a machine learning framework, support vector
(Conzelmann et al. 2009; Gong et al. 2014). ADHD-IA machine (SVM), to distinguish between ADHD-IA and
patients showed decreased response in gambling punish- ADHD-C patients using connectivity measures of identi-
ment task and increased response in emotion task than fied brain regions with significant group-wise differences
ADHD-C patients (Conzelmann et al. 2009; Gong et al. (Klöppel et al. 2012; dos Santos Siqueira et al. 2014); (3)
2014). These studies suggest that patient performance on test the effectiveness of identified regions and connectivity
specific tasks is a helpful feature to distinguish among measures to predict clinically meaningful scores, such as
ADHD subtypes. DSM inattention, hyperactive/impulsive, and combined
Recent widespread adoption of neuroimaging has scores. We hypothesized that (1) brain regions with sig-
allowed routine inspection of ADHD patient brains using nificant group-wise differences could be used to classify
various imaging modalities. Many neuroimaging studies between ADHD subtypes, (2) certain task paradigms were
have explored differences among ADHD subtypes using more suitable for distinguishing ADHD subtypes, and (3)
magnetic resonance imaging (MRI) and single photon DSM scores could be predicted using brain regions and
emission computed tomography (SPECT) (Lorberboym associated connectivity measures.
et al. 2004; dos Santos Siqueira et al. 2014). MRI as a
neuroimaging tool has been widely adopted, as it can
provide both structural and functional information about Methods
the brain. Resting-state functional MRI (fMRI) measures
blood oxygenation level-dependent (BOLD) signal when a Subjects and Imaging Data
subject is not performing any task, while task fMRI mea-
sures the BOLD signal when a subject is performing a We obtained task fMRI data from the Human Connectome
specific task. A previous study adopted resting-state fMRI Project (HCP) database (Van Essen et al. 2013). Images
to distinguish among ADHD subtypes, but the classifica- were scanned by a Siemens Skyra 3T scanner housed at
tion accuracy was sub-par (dos Santos Siqueira et al. 2014). Washington University in St. Louis. Imaging parameters

123
Brain Topogr (2016) 29:429–439 431

were as follows: image matrix = 104 9 90; pixel resolu- preceded by a 3-s task cue (‘‘shape’’ or ‘‘face’’) and was
tion = 2.0 mm isotropic; slice thickness = 2.0 mm; repe- 21 s in duration including the cue. Each of the two runs
tition time (TR) = 720 ms; echo time (TE) = 33.1 ms; included three face blocks and three shape blocks, with 8 s
and field of view (FOV) = 180 mm. Each task fMRI uti- of fixation at the end of each run.
lized a different number of slices as follows: emo- In the gambling task, participants played a card guessing
tion = 405; gambling = 253; motor = 284; relational game (Delgado et al. 2000; May et al. 2004). Participants
processing = 232; social cognition = 274; and working knew the card number was in the range of 1–9 and were
memory = 405. A total of 34 subjects were grouped into asked to guess whether a card number was more or \5.
ADHD-IA (n = 13) and ADHD-C (n = 21) subtypes. Feedback was shown after guessing. If the participant won
ADHD subtypes were classified based on DSM-IV scores the game, a green up-arrow with ‘‘$1’’ appeared. If they
that HCP database provided (American Psychiatric Asso- lost the game, a red down arrow with ‘‘-$0.50’’ appeared.
ciation 1994). The DSM scores in HCP database are If the card number was 5, a gray double-headed arrow with
restricted. Researchers need to apply for permission to number ‘‘5’’ appeared. Fixations between task blocks were
access the DSM scores (Van Essen et al. 2013). The ratio considered as the control state. One block consisted of
of ADHD-IA and ADHD-C patients in our study was eight trials including 1.5 s of card guessing, 1 s of feed-
similar to the natural incidence rate (Baeyens et al. 2006). back, and 1 s of ITI. There were 15 s of fixation between
The mean age was 27.77 with standard deviation (SD) 4.11 blocks. Each block was either mostly reward or mostly
for ADHD-IA patients. The mean age was 27.95 with SD loss. The reward-leaning block consisted of six rewards
2.94 for ADHD-C patients. The sex ratio, age, and DSM along with one neutral and one loss, two neutrals, or two
inattention scores between ADHD-IA and ADHD-C losses. The loss-leaning block consisted of six losses along
patients did not show significant differences (p = 0.24, with one neutral and one reward, two neutrals, or two
p = 0.88, and p = 0.70, respectively), while DSM hyper- rewards. Four blocks were included in each run.
active/impulsive and combined scores showed significant In the motor task, participants were asked to tap their
differences (p \ 0.001 for both scores) (Table 1). fingers, squeeze their toes, or move their tongue (Buckner
et al. 2011). The task cue was shown for 3 s, and move-
Task Paradigm ments were continued for 12 s. There was a 15-s control
fixation after two tasks. Ten blocks were included in each
We employed a total of six task paradigms, brief expla- run.
nations of which are given below. In the emotion task, In the relational processing task, participants were pre-
participants were presented with blocks of trials that either sented with two pairs of objects on the computer screen
asked them to decide which of two faces presented on the (Smith et al. 2007). One pair was located at the top of the
bottom of the screen matched the face at the top of the screen, and the other pair was located at the bottom of the
screen (i.e. task state) or which of two shapes presented at screen. Participants first decided what dimension differed
the bottom of the screen matched the shape at the top of the in the top pair of objects (shape or texture) and then
screen (i.e. control state) (Tessitore et al. 2005; Stein et al. decided whether the bottom pair of objects also differed in
2007). The faces had either an angry or fearful expression. that same dimension. For the control state, participants
Trials were presented in blocks of six trials of the same task were presented with two objects at the top of the screen and
(face or shape), with the stimulus presented for 2 s fol- one object at the bottom of the screen. There was a word in
lowed by a 1-s inter-task interval (ITI). Each block was the middle of the screen representing either shape or

Table 1 Demographic data of ADHD-IA and ADHD-C patients (Mean and standard deviation are reported)
Characteristic ADHD-IA (n = 13) ADHD-C (n = 21) t value/ v2 DOF p value

Sex (Male:Female)a 6:7 14:7 1.39 1 0.24

Age (years) 27.77 (4.11) 27.95 (2.94) -0.15 32 0.88
DSM inattention score 8.46 (0.52) 8.67 (1.85) -0.39 32 0.70
DSM hyperactivity/impulsive score 2.77 (0.93) 6.52 (1.81) -6.92 32 \0.001
DSM combined score 11.23 (0.73) 15.19 (2.50) -5.53 32 \0.001
ADHD-IA Attention deficit hyperactivity disorder inattentive, ADHD-C attention deficit hyperactivity disorder combined, DSM diagnostic and
statistical manual of mental disorders, DOF degree of freedom
a
Chi square statistic

123
432 Brain Topogr (2016) 29:429–439

texture. Participants decided whether the bottom object brain mask into the individual subject space. The temporal
matched either of the top two objects on that dimension. pre-processing was performed using AFNI software (Cox
There were four trials in one block including 3.5 s of 1996). Six rigid-body motion parameters, white matter, and
stimulus and 0.5 s of ITI for the task state. There were five cerebrospinal fluid contributions were regressed out using
trials in one block including 2.8 s of stimulus and 0.4 s of the 3dDeconvolve package. Band-pass filtering was
ITI for control state. There was a 16-s fixation after one applied in the range of 0.009–0.08 Hz in order to remove
pair of task and control block. Six blocks were included in slow drift using the 3dFourier package. Task fMRI data
each run. were divided according to explanatory variables (EVs).
In the social cognition task, participants viewed a short One set of data consisted of fMRI volumes from the first
video of objects including squares, circles, and triangles 6 s of task onset to the first 2 s of task offset. This approach
(Castelli et al. 2000). The objects either interacted in some is known to reduce the delay in hemodynamic response
way or moved randomly on the screen. Interactive move- (Koshino et al. 2005). The HCP database provides two
ment was considered as a task state, and random movement distinct phase encodings: ‘‘left-to-right’’ and ‘‘right-to-
was considered as a control state. Participants judged left.’’ We adopted concatenated time series from two
whether or not the objects interacted. There were 23 s of encoding runs and the possible time series fluctuations
stimulus and 15 s of fixation in one block. Five blocks were between encoding runs were corrected by regressing out
included in each run. the signal contributions from the white matter and cere-
In the working memory task, participants viewed blocks brospinal fluid.
of trials that consisted of pictures of places, tools, faces,
and body parts (Peelen and Downing 2005; Drobyshevsky Network Construction
et al. 2006). Participants performed 2-back and 0-back
tasks, as described below. A series of pictures was dis- Connectivity analysis requires regions of interest (ROIs)
played on the screen, and participants determined if the to be specified so that correlations among them can be
current picture was the same as the one presented two explored. These ROIs might be specified by transferring
pictures previous (2-back). The 0-back task required information from a pre-defined parcellation of the brain.
searching for a target picture that was pre-specified at the We parcellated the whole brain into 384 distinct regions
start of the block. One block consisted of 10 trials using functional information (Craddock et al. 2012).
including 2 s of stimulus and 0.5 s of ITI. There was a 15-s Given a set of ROIs, we defined connectivity using nodes
fixation after a block. Eight blocks were included in each and edges (Stam and Reijneveld 2007; He et al. 2007;
run. Bullmore and Sporns 2009). Nodes were pre-defined
parcellations, and edges were defined as correlation val-
Image Pre-Processing ues between nodes. The edge values were entered as
individual elements of a matrix, which was then referred
The HCP database provided spatially pre-processed fMRI to as the correlation matrix. We adopted undirected and
data (Jenkinson et al. 2002; Glasser and Van Essen 2011; unweighted edges for simple network construction. The
Van Essen et al. 2012; Glasser et al. 2013). Image pre- correlation matrix was converted into binary form by
processing steps were performed using FSL and FreeSurfer applying a fixed-density threshold. The correlation
software (Fischl 2012; Jenkinson et al. 2012). Distortions matrices were computed with a wide range of densities
due to gradient nonlinearity were corrected using the gra- (1–50 %). The threshold of 41 %, the minimum value at
dient_nonlin_unwarp package in FreeSurfer. Head motion which all nodes were connected, was adopted. Subsequent
correction was performed by realigning time series vol- analyses were based on correlation matrix at the chosen
umes to the single-band reference image obtained at the threshold. Fisher’s r-to-z transformation was performed
beginning of the data acquisition with six-degree of free- on all correlation matrices.
dom (DOF) transformation using the FLIRT package in
FSL. Motion-corrected images were registered onto the T1- Connectivity Analysis
weighted structural image, first with six-DOF FLIRT with
a boundary-based registration (BBR) cost function and then We divided the whole time series into time series with task
with bbregister in FreeSurfer for fine tuning. The co-reg- and time series without task using timing information of
istered images were registered onto the standard Montreal task onset and duration for both left-to-right and right-to-
Neurological Institute (MNI) space with nonlinear trans- left runs. Each time series contributed separately to task
formation. Intensity normalization of the 4D dataset was state correlation and control state correlation matrices
applied with a whole-brain global mean of 10,000. The (Fig. 1a, b). We quantified characteristics of brain net-
skull was removed from the image by bringing an MNI works using degree as a connectivity measure. Degree is

123
Brain Topogr (2016) 29:429–439 433

Fig. 1 Overall flow of the study. a Time series with task and time from each node. d Group-wise differences were tested in term of
series without task were extracted separately. b Correlation matrix changes of degree value (Ddegree) between task and control state.
was generated using correlation of time series between nodes. e SVM classifier framework with linear kernel was applied to
c Degree, the number of links connected to a node, was calculated distinguish between ADHD-IA and ADHD-C groups

defined as the number of edges connected to a node and is a 1999). A linear kernel was used for the SVM. Readers
local parameter that quantifies the importance of a given concerned with the technical details of SVM may consult
node in terms of network organization (Fig. 1c). Degree the following reference (Vapnik 1999). With limited sub-
was computed for each node in the network at the mini- jects available, we adopted the leave-one-out cross vali-
mum sparsity at which all nodes were connected. Brain dation method for separating training and test data. For
regions showing significant group-wise differences in example, given 13 ADHD-IA and 21 ADHD-C cases, we
change of degree value (Ddegree) between task and control assigned one case as the test case and used the remaining
state were identified (Fig. 1d). We performed permutation 33 cases as the training data for the SVM classifier. The
tests to correct for the multiple comparison issue. We process was repeated 34 times choosing a different test
randomly assigned 13 subjects to ADHD-IA group and case each time. A hypothetically biased classifier would
assigned the remaining 21 subjects to ADHD-C group 5000 have baseline accuracy of 0.62 (=21/34) not 0.5 due to
times (Smith et al. 2013). Differences in degree values different number of training samples. The SVM classifier
were deemed significant if they did not belong to the 99 % seeks a decision boundary that can effectively separate
of the null distribution derived from the permutation tests sample near the decision boundary. One group might have
(p \ 0.01, corrected). The whole procedure was repeated more training sample but that does not necessarily mean
for the six task paradigms. the group would have more samples near the decision
boundary (Vapnik 1999). Thus the SVM classifier accounts
Classification of ADHD Subtypes for un-equal number of training samples. Several studies
applied the SVM framework to classify two different
The differences in degree values of the identified regions groups with different number of subjects (Klöppel et al.
were fed into an SVM classifier framework to distinguish 2008; dos Santos Siqueira et al. 2014). Classifier accuracy
between ADHD-IA and ADHD-C groups (Fig. 1e) (Vapnik was computed by comparing the classifier outcome with

123
434 Brain Topogr (2016) 29:429–439

original ADHD subtype assignment. The whole procedure Results

was repeated for six task paradigms.
Differences in Connectivity
Predicting ADHD-Related DSM Score
We performed the same set of analyses described in the
The differences in degree values of the identified regions Methods section to compare ADHD and normal subjects.
were further tested for their ability to predict clinical This was to establish baseline results so that ADHD sub-
meaningful DSM inattention, hyperactive/impulsive, and types could be better compared. Our primary purpose was
combined scores. We performed multiple linear regression to distinguish between ADHD subtypes (i.e. IA and com-
to predict ADHD-related DSM scores (i.e. DSM inatten- bined types) and thus the results of the analyses between
tion, hyperactive/impulsive, and combined scores). Age, ADHD and normal subjects were given in the supplement.
sex, and Ddegree of the identified regions were used as Differences in degree values between task and control
predictors for the regression. The general form of the state were calculated within each group (ADHD-IA and
multiple linear regression model was as follows: DSM ¼ ADHD-C groups). Degree values were calculated at sparsity
P
a þ b1  Age þ b2  Sex þ Ni¼1 bi  Ddegreei , where a was of 41 %, the minimum value at which all nodes were con-
a constant, bi were estimated coefficients, and N was the nected. Differences in degree values were then compared
number of identified regions. Multiple linear regression between groups. Brain regions with significant (p \ 0.05)
prediction was performed with a leave-one-out cross vali- group-wise differences for each task paradigm were identi-
dation approach. A linear equation was generated from the fied (Table 2). We identified the following notable regions:
training data set. DSM scores were predicted by assigning medial- and inferior-frontal gyrus, anterior- and posterior-
predictor variables from the left-out test case. Quality of cingulate cortex, inferior parietal lobule, postcentral gyrus,
prediction was assessed with p- and r-values between precuneus, superior- and middle-temporal gyrus, middle
actual and predicted DSM scores. The whole procedure occipital gyrus, cuneus, lingual gyrus, and cerebellum. The
was repeated for the six task paradigms. frontal lobe, cingulate cortex, and parietal cortex control

Table 2 Regions that showed

Task Regions Degree difference p value, corrected
significant group-wise
differences in degree values ADHD-IA ADHD-C
between task and control states
(Mean and standard deviation Emotion Left postcentral 10.85 (36.45) -19.57 (26.56) 0.0076
are reported) Right inferior temporal 17.31 (19.08) -5.57 (18.86) 0.0018
Right cuneus -26.92 (36.00) 8.43 (30.12) 0.0040
Right lingual -16.38 (26.88) 11.19 (26.20) 0.0070
Gambling reward Left medial frontal 15.23 (26.21) -9.48 (25.03) 0.0096
Left postcentral 11.00 (20.12) -18.95 (32.38) 0.0038
Left superior temporal 4.23 (23.93) -29.90 (32.13) 0.0032
Left middle occipital 15.77 (20.04) -7.24 (25.51) 0.0076
Gambling punishment Left medial frontal 20.54 (30.52) -13.10 (23.24) 0.0006
Left inferior frontal 14.38 (38.75) -17.71 (28.12) 0.0076
Left precuneus -22.85 (38.91) 13.05 (30.98) 0.0042
Left superior temporal 0.31 (24.59) -32.52 (32.59) 0.0020
Right cerebellum -23.54 (30.83) 1.57 (18.59) 0.0018
Motor Left middle frontal 21.92 (19.85) -2.57 (26.44) 0.0050
Right inferior frontal -14.31 (22.72) 12.71 (24.23) 0.0034
Left middle temporal -32.15 (32.01) 3.52 (28.43) 0.0010
Relational processing Left anterior cingulate 22.69 (23.76) -4.33 (30.04) 0.0094
Social cognition Right precentral -13.15 (33.58) 19.05 (32.69) 0.0094
Right middle frontal -18.46 (32.55) 18.10 (37.79) 0.0064
Working memory Right middle temporal -24.85 (30.95) 10.95 (33.51) 0.0034
ADHD-IA Attention deficit hyperactivity disorder inattentive, ADHD-C attention deficit hyperactivity
disorder combined

123
Brain Topogr (2016) 29:429–439 435

Table 3 Classifier accuracy for each task paradigm. Task paradigms Classifier accuracy was computed by comparing the clas-
with the best classifier performance are listed in italics sifier outcome with original ADHD subtype assignment
Task Accuracy (%) according to DSM. Classifier accuracy values were
91.18 % for both gambling punishment and emotion task
Emotion 91.18
paradigms. Classifier accuracy was lower than 90 % for the
Gambling reward 85.29 other task paradigms (Table 3). Identified regions under the
Gambling punishment 91.18 gambling punishment and emotion task paradigms are
Motor 85.29 shown in Fig. 2. Our findings confirmed previous non-
Relational processing 70.59 neuroimaging studies that found different behavior patterns
Social cognition 76.47 between ADHD-IA and ADHD-C patients during gam-
Working memory 67.65 bling punishment and emotion tasks (Conzelmann et al.
2009; Gong et al. 2014). The classification accuracy in this
study based on task fMRI (over 90 %) was much higher
attention and executive functions, which are closely linked to than a previous study using resting state fMRI (61 %) to
ADHD (Bush et al. 2005; Schneider et al. 2006). The distinguish between ADHD subtypes (dos Santos Siqueira
occipital cortex is known to have a role in attention pro- et al. 2014).
cessing (Schneider et al. 2006; Oldehinkel et al. 2013). The
identified brain regions of this study correspond to those Predicting ADHD-Related DSM Scores
previously noted to be involved in ADHD (Bush et al. 2005;
Schneider et al. 2006; Oldehinkel et al. 2013). We found that some regions and the associated differences
in degree values were effective at distinguishing between
Classification of ADHD Subtypes ADHD subtypes under gambling punishment and emotion
task paradigms. The effectiveness of the differences in
The differences of degree values of the identified regions degree values of the identified regions were further tested
were used as features for the SVM classifier in order to to determine if they could predict ADHD-related DSM
distinguish between ADHD-IA and ADHD-C groups. The scores under the two task paradigms. Age, sex, and Dde-
whole procedure was repeated for the six task paradigms. gree of the identified regions were used as predictors for

Fig. 2 Brain regions that

showed significant group-wise
differences in degree of task and
control state for a gambling
punishment and b emotion tasks

123
436 Brain Topogr (2016) 29:429–439

Fig. 3 Comparison of actual and predicted DSM hyperactive/impulsive scores for a gambling punishment task and b emotion task

the regression for DSM inattention, hyperactive/impulsive, occipital gyrus, cuneus, lingual gyrus, and cerebellum
and combined scores. Quality of prediction was assessed involved in ADHD. These identified regions corresponded to
with p- and r-values between actual and predicted DSM the regions previously reported as related to psychopathol-
scores. The regions and the associated features showed ogy of ADHD.
significant correlation with DSM hyperactive/impulsive We additionally performed the comparison between
score (r = 0.4063, p = 0.0171, for gambling punishment ADHD and normal subjects as reported in the supplement.
paradigm; r = 0.4197, p = 0.0135, for emotion task We identified one set of regions reflecting differences
paradigm (Fig. 3). between ADHD subtypes and another set of regions
reflecting differences between ADHD and normal subjects.
The two comparison sets had common identified regions as
Discussion the following; CFP network, occipital cortices, and cere-
bellum. Subcortical regions including insula, thalamus, and
One of the objectives of this study was to explore connec- caudate were only identified in the comparison between
tivity differences between ADHD subtypes using task fMRI. ADHD and normal subjects, while temporal cortices were
Brain regions with significant group-wise differences in only identified in the comparison between ADHD sub-
terms of degree difference were identified. The cingulo- types. Temporal cortices were reported to be related to
front-parietal (CFP) network including dorsolateral-, and behaviors including hyperactivity and thus it could play an
ventrolateral-prefrontal, cingulate, and parietal cortices is important role in distinguishing between ADHD subtypes
considered as the primary brain network related to ADHD, as (Schneider et al. 2010; dos Santos Siqueira et al. 2014).
they play a critical role in ADHD psychopathology (Sch- Subcortical regions are partly responsible for the attention
neider et al. 2006; Bush 2010; Hoekzema et al. 2014). These system and have been reported as regions that could dis-
regions perform functions in attention, working memory, tinguish between ADHD and normal subjects (Baeyens
executive control, and response inhibitory control (Bush et al. 2006; Schneider et al. 2006; Bush 2010). We were
et al. 2005; Schneider et al. 2006). Previous studies have able to replicate the regions in this study.
reported dysfunction in frontal, anterior cingulate, and Existing non-imaging studies have also reported that
parietal cortices in ADHD patients (Schneider et al. 2006; patients of different ADHD subtypes responded differently
Oldehinkel et al. 2013). Our analysis identified ROI as to gambling punishment, emotion stimulus, and stimulant
medial- and inferior-frontal gyrus, anterior- and posterior- treatment (Conzelmann et al. 2009; Barnard et al. 2010;
cingulate cortex, inferior parietal lobule, postcentral gyrus, Gong et al. 2014). ADHD-IA patients did not show obvious
and precuneus. The temporal and occipital cortices and brain responses, while ADHD-C patients showed large
cerebellum are known to be partly responsible in attention brain responses in gambling punishment task (Gong et al.
processing and previous studies have demonstrated dys- 2014). For emotion task, ADHD-IA patients showed sim-
functions in these regions in patients with ADHD (Schneider ilar startled responses compared with normal subjects,
et al. 2006; Oldehinkel et al. 2013). Our analysis also iden- while ADHD-C patients did not show obvious startled
tified the superior- and middle-temporal gyrus, middle responses to the pleasant task and hyper-responses to the

123
Brain Topogr (2016) 29:429–439 437

unpleasant task (Conzelmann et al. 2009). These studies between ADHD subtypes were above 80 % for the emotion
showed that there were notable differences in responses (=91.18 %) and gambling punishment (=91.18 %) task
between ADHD-IA and ADHD-C patients under gambling paradigms. We identified many regions (e.g. CFP and
punishment task and emotion task paradigms. These temporal regions) related to emotion and gambling pun-
notable differences in response between ADHD subtypes ishment tasks and thus classifiers under these tasks were
could be one of the reasons why the classification perfor- expected to perform well. Similar attempts to distinguish
mances between ADHD subtypes were good for these two between ADHD subtypes based on resting-state fMRI fared
task paradigms (Conzelmann et al. 2009; Gong et al. 2014). worse (around 60 % accuracy) than our attempts using task
We were able to corroborate existing findings with neu- fMRI (91.18 % for both gambling punishment and emotion
roimaging. We found brain regions in CFP network under task) (dos Santos Siqueira et al. 2014). There are major
gambling punishment and emotion task paradigms that differences in active brain circuitry between task and
might be used for distinguishing ADHD subtypes. ADHD control states; therefore, resting-state fMRI might not
subtypes were well classified using the identified regions provide enough information to distinguish between ADHD
and associated features. Gambling punishment task para- subtypes. Our approach using task fMRI is better suited to
digm is closely associated with reward related system and characterizing ADHD, as information from both task and
emotion task paradigm is partly associated with attention control states is included. The differences in degree values
system as the brain circuits overlap between emotion and of the identified regions were successfully used to predict
attentional regulation system (Delgado et al. 2000; May ADHD-related DSM scores under the two task paradigms.
et al. 2004; Baroni and Castellanos 2015). The identified These results confirmed that the identified regions and
brain regions of our study regulate reward related and imaging features were indeed meaningful to distinguish
attention systems and they might be used to explain the between ADHD subtypes.
behavioral differences between ADHD subtypes (Schnei- Our study has following limitations. We obtained neu-
der et al. 2006; Bush 2010; Hoekzema et al. 2014). Con- roimaging data from a research database. The database did
nectivity features of these regions were used for ADHD not include large number of cases that met our selection
subtype classification and the performance was 91.18 % criteria. Further research with more samples and inclusion
accuracy for both gambling punishment and emotion task of hyperactive subtype is necessary to fully explore dif-
paradigms. One might hypothesize that these regions could ferences among ADHD subtypes. The database did not
explain behavioral differences between ADHD subtypes contain intelligence related scores and thus intelligence
but these regions are related to various behaviors including scores could not be factored into the regression model. In
hyperactivity. Further research controlling for confounding addition, our study used only one imaging modality. Brain
factors is needed to verify this hypothesis. Our results networks can be assessed using not only fMRI, but also
based on neuroimaging were consistent with previous study other imaging modalities including diffusion tensor imag-
that distinguished ADHD subtypes using behavioral pat- ing. Multi-modal analysis of the brain network allows
terns (Conzelmann et al. 2009; Gong et al. 2014). incorporation of complementary information derived from
The identified brain regions and the associated connec- different modalities to better quantify differences between
tivity measures were used as features in the classifier. ADHD subtypes. We adopted the degree measure to
Classifier performance is heavily influenced by the features determine connectivity of the brain network. There are
which it is trained upon. Classifier performs well if the test many other promising connectivity parameters including
data are similar to those of training data. We also per- betweenness centrality, closeness centrality, and eigen-
formed the classification between ADHD and normal vector centrality that might have better sensitivity for dis-
subjects as reported in the supplement. The performance of tinguishing between ADHD subtypes (dos Santos Siqueira
classifier, either for distinguishing between ADHD sub- et al. 2014).
types or distinguishing between ADHD and normal sub-
jects, was largely task dependent. Classification
performances were good if the identified regions were Conclusion
correlated with the task paradigms under which classifi-
cation occurs. The classification accuracy to distinguish We explored connectivity differences in degree values
between ADHD and normal subjects were above 80 % for between ADHD-IA and ADHD-C patients using task
the motor (=80.88 %) and working memory (=91.18 %) fMRI. The identified regions and associate features were
task paradigms. We identified many regions (e.g. CFP and fed into the SVM classifier framework to distinguish
cingulate regions) related to motor and working memory between ADHD subtypes. Classification accuracy was
tasks and thus classifiers under these tasks were expected to excellent (i.e. 91.18 %) for gambling punishment and
perform well. The classification accuracy to distinguish emotion task paradigms. The effectiveness of identified

123
438 Brain Topogr (2016) 29:429–439

regions and associate features were further tested by per- Castelli F, Happé F, Frith U, Frith C (2000) Movement and mind: a
forming correlation with ADHD-related DSM scores. The functional imaging study of perception and interpretation of
complex intentional movement patterns. Neuroimage
regions and the associated features showed significant 12:314–325. doi:10.1006/nimg.2000.0612
correlation with DSM hyperactive/impulsive score Castle L, Aubert RE, Verbrugge RR et al (2007) Trends in medication
(r = 0.4063, p = 0.0171, for gambling punishment para- treatment for ADHD. J Atten Disord 10:335–342. doi:10.1177/
digm; r = 0.4197, p = 0.0135, for emotion task para- 1087054707299597
Conzelmann A, Mucha RF, Jacob CP et al (2009) Abnormal affective
digm). The identified regions were related to reward related responsiveness in attention-deficit/hyperactivity disorder: sub-
and attention systems and these might be good candidates type differences. Biol Psychiatry 65:578–585. doi:10.1016/j.
for imaging biomarkers to explain behavioral differences biopsych.2008.10.038
between ADHD subtypes. Further research need to accu- Cox RW (1996) AFNI: software for analysis and visualization of
functional magnetic resonance neuroimages. Comput Biomed
rately quantify the effectiveness of these potential Res 29:162–173
biomarkers. Craddock RC, James GA, Holtzheimer PE et al (2012) A whole brain
fMRI atlas generated via spatially constrained spectral cluster-
Acknowledgments This study was supported by the Institute for ing. Hum Brain Mapp 33:1914–1928. doi:10.1002/hbm.21333
Basic Science (grant number IBS-R015-D1). This study was also Delgado MR, Nystrom LE, Fissell C et al (2000) Tracking the
supported by the National Research Foundation of Korea (Grant hemodynamic responses to reward and punishment in the
number NRF-2013R1A2A2A04016262). Data were provided by the striatum. J Neurophysiol 84:3072–3077
Human Connectome Project, WU-Minn Consortium (Principal Dos Santos SA, Biazoli Junior CE, Comfort WE et al (2014)
Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) Abnormal functional resting-state networks in ADHD: graph
funded by the 16 NIH Institutes and Centers that support the NIH theory and pattern recognition analysis of fMRI data. Biomed
Blueprint for Neuroscience Research and by the McDonnell Center Res Int 2014:1–10. doi:10.1155/2014/380531
for Systems Neuroscience at Washington University. Drobyshevsky A, Baumann SB, Schneider W (2006) A rapid fMRI
task battery for mapping of visual, motor, cognitive, and
emotional function. Neuroimage 31:732–744. doi:10.1016/j.
neuroimage.2005.12.016
References Faraone S, Sergeant J (2003) The worldwide prevalence of ADHD: is
it an American condition? World Psychiatry 2:104–113
American Psychiatric Association (1994) Diagnostic and statistical Fischl B (2012) Free surfer. Neuroimage 62:774–781. doi:10.1016/j.
manual of mental disorders, 4th edn. American Psychiatric Press, neuroimage.2012.01.021
Washington, DC Glasser MF, Van Essen DC (2011) Mapping human cortical areas
American Psychiatric Association (2013) Diagnostic and Statistical in vivo based on myelin content as revealed by T1- and T2-
Manual of Mental Disorders, 5th edn. American Psychiatric weighted MRI. J Neurosci 31:11597–11616. doi:10.1523/
Press, Washington, DC JNEUROSCI.2180-11.2011
Baeyens D, Roeyers H, Vande WJ (2006) Subtypes of attention- Glasser MF, Sotiropoulos SN, Wilson JA et al (2013) The minimal
deficit/hyperactivity disorder (ADHD): distinct or related disor- preprocessing pipelines for the Human Connectome Project.
ders across measurement levels? Child Psychiatry Hum Dev Neuroimage 80:105–124. doi:10.1016/j.neuroimage.2013.04.127
36:403–417. doi:10.1007/s10578-006-0011-z Gong J, Yuan J, Wang S et al (2014) Feedback-related negativity in
Barnard L, Stevens T, To YM et al (2010) The importance of ADHD children with two subtypes of attention deficit hyperactivity
subtype classification for educational applications of DSM-V. disorder. PLoS One 9:e99570. doi:10.1371/journal.pone.0099570
J Atten Disord 13:573–583 He Y, Chen ZJ, Evans AC (2007) Small-world anatomical networks
Baroni A, Castellanos FX (2015) Neuroanatomic and cognitive in the human brain revealed by cortical thickness from MRI.
abnormalities in attention-deficit/hyperactivity disorder in the Cereb Cortex 17:2407–2419. doi:10.1093/cercor/bhl149
era of ‘‘high definition’’ neuroimaging. Curr Opin Neurobiol Hoekzema E, Carmona S, Ramos-Quiroga JA et al (2014) An
30:1–8. doi:10.1016/j.conb.2014.08.005 independent components and functional connectivity analysis
Bassett DS, Bullmore E (2006) Small-world brain networks. Neuro- of resting state fMRI data points to neural network dysreg-
science 12:512–523. doi:10.1177/1073858406293182 ulation in adult ADHD. Hum Brain Mapp 35:1261–1272.
Buckner RL, Krienen FM, Castellanos A et al (2011) The organization of doi:10.1002/hbm.22250
the human cerebellum estimated by intrinsic functional connectiv- Jenkinson M, Bannister P, Brady M, Smith S (2002) Improved
ity. J Neurophysiol 106:2322–2345. doi:10.1152/jn.00339.2011 optimization for the robust and accurate linear registration and
Bullmore E, Sporns O (2009) Complex brain networks: graph motion correction of brain images. Neuroimage 17:825–841.
theoretical analysis of structural and functional systems. Neuro- doi:10.1006/nimg.2002.1132
science 10:186–198. doi:10.1038/nrn2575 Jenkinson M, Beckmann CF, Behrens TEJ et al (2012) FSL.
Bush G (2010) Attention-deficit/hyperactivity disorder and attention Neuroimage 62:782–790. doi:10.1016/j.neuroimage.2011.09.015
networks. Neuropsychopharmacology 35:278–300. doi:10.1038/ Klöppel S, Stonnington CM, Chu C et al (2008) Automatic
npp.2009.120 classification of MR scans in Alzheimer’s disease. Brain
Bush G, Valera EM, Seidman LJ (2005) Functional neuroimaging of 131:681–689. doi:10.1093/brain/awm319
attention-deficit/hyperactivity disorder: a review and suggested Klöppel S, Abdulkadir A, Jack CR et al (2012) Diagnostic
future directions. Biol Psychiatry 57:1273–1284. doi:10.1016/j. neuroimaging across diseases. Neuroimage 61:457–463. doi:10.
biopsych.2005.01.034 1016/j.neuroimage.2011.11.002
Carlson CL, Lahey BB, Neeper R (1986) Direct assessment of the Koshino H, Carpenter PA, Minshew NJ et al (2005) Functional
cognitive correlates of attention deficit disorders with and connectivity in an fMRI working memory task in high-
without hyperactivity. J Psychopathol Behav Assess 8:69–86. functioning autism. Neuroimage 24:810–821. doi:10.1016/j.
doi:10.1007/BF00960874 neuroimage.2004.09.028

123
Brain Topogr (2016) 29:429–439 439

Lorberboym M, Watemberg N, Nissenkorn A et al (2004) Technetium Smith SM, Vidaurre D, Beckmann CF et al (2013) Functional
99m ethylcysteinate dimer single-photon emission computed connectomics from resting-state fMRI. Trends Cogn Sci
tomography (SPECT) during intellectual stress test in children 17:666–682. doi:10.1016/j.tics.2013.09.016
and adolescents with pure. J Child Neurol 19:91–96 Solanto MV, Schulz KP, Fan J et al (2009) Event-Related fMRI of
May JC, Delgado MR, Dahl RE et al (2004) Event-related functional inhibitory control in the predominantly inattentive and combined
magnetic resonance imaging of reward-related brain circuitry in subtypes of adhd: clinical investigative study. J Neuroimaging
children and adolescents. Biol Psychiatry 55:359–366. doi:10. 19:205–212. doi:10.1111/j.1552-6569.2008.00289.x
1016/j.biopsych.2003.11.008 Stam CJ, Reijneveld JC (2007) Graph theoretical analysis of complex
Oldehinkel M, Francx W, Beckmann CF et al (2013) Resting state networks in the brain. Nonlinear Biomed Phys 1:1–19. doi:10.
FMRI research in child psychiatric disorders. Eur Child Adolesc 1186/1753-4631-1-3
Psychiatry 22:757–770. doi:10.1007/s00787-013-0480-0 Stein M, Simmons A, Feinstein J, Paulus M (2007) Increased
Peelen MV, Downing PE (2005) Within-subject reproducibility of amygdala and insula activation during emotion processing in
category-specific visual activation with functional MRI. Hum anxiety-prone subjects. Am J Psychiatry 164:318–327
Brain Mapp 25:402–408. doi:10.1002/hbm.20116 Tessitore A, Hariri AR, Fera F et al (2005) Functional changes in the
Pelham WE, Gnagy EM (1999) Psychosocial and combined treat- activity of brain regions underlying emotion processing in the
ments for ADHD. Ment Retard Dev Disabil Res Rev 5:225–236. elderly. Psychiatry Res Neuroimaging 139:9–18. doi:10.1016/j.
doi:10.1002/(SICI)1098-2779 pscychresns.2005.02.009
Rubinov M, Sporns O (2010) Complex network measures of brain Van Essen DC, Glasser MF, Dierker DL et al (2012) Parcellations and
connectivity: uses and interpretations. Neuroimage 52:1059–1069. hemispheric asymmetries of human cerebral cortex analyzed on
doi:10.1016/j.neuroimage.2009.10.003 surface-based atlases. Cereb Cortex 22:2241–2262. doi:10.1093/
Schneider M, Retz W, Coogan A, et al (2006) Anatomical and cercor/bhr291
functional brain imaging in adult attention-deficit/hyperactivity Van Essen DC, Smith SM, Barch DM et al (2013) The WU-Minn
disorder (ADHD): a neurological view. Eur Arch Psychiatry Clin human connectome project: an overview. Neuroimage 80:62–79.
Neurosci 256:i32–41. doi:10.1007/s00406-006-1005-3 doi:10.1016/j.neuroimage.2013.05.041
Schneider MF, Krick CM, Retz W et al (2010) Impairment of fronto- Vapnik VN (1999) An overview of statistical learning theory. IEEE
striatal and parietal cerebral networks correlates with attention Trans Neural Netw 10:988–999. doi:10.1109/72.788640
deficit hyperactivity disorder (ADHD) psychopathology in Watts D, Strogatz S (1998) Collective dynamics of ‘‘small-world’’
adults: a functional magnetic resonance imaging (fMRI) study. networks. Nature 393:440–442
Psychiatry Res Neuroimaging 183:75–84. doi:10.1016/j.pscy Willcutt EG, Nigg JT, Pennington BF et al (2012) Validity of DSM-
chresns.2010.04.005 IV attention-deficit/hyperactivity disorder symptom dimensions
Smith R, Keramatian K, Christoff K (2007) Localizing the rostrolat- and subtypes. J Abnorm Psychol 121:991–1010. doi:10.1037/
eral prefrontal cortex at the individual level. Neuroimage a0027347.Validity
36:1387–1396. doi:10.1016/j.neuroimage.2007.04.032

123