Toward Systems Neuroscience of ADHD: A Meta-Analysis of 55 fMRI Studies

We searched the following databases: PubMed, Ovid (including PsycINFO and Ovid MEDLINE), EMBASE, Web of Science (Science Citation Index Expanded, Social Sciences Citation Index, and Arts and Humanities Citation Index), ERIC, CINAHL, and NeuroSynth (www.neurosynth.org). Details of the search strategy are reported in section A1 of the data supplement that accompanies the online edition of this article.

Studies were included if they 1) used a diagnosis of ADHD according to DSM-IV, DSM-IV-TR, or ICD-10, 2) used a typically developing comparison group, 3) reported data as three-dimensional coordinates in stereotactic space, and 4) used between-group contrasts.

Studies were excluded if they 1) used a neuroimaging method other than fMRI; 2) included participants with ADHD symptoms but without a formal diagnosis of ADHD; 3) assessed the effect of medication without reporting fMRI data at baseline or after washout; 4) reported only within-group contrasts; 5) conducted a priori region-of-interest analyses (as these violate the assumption, under the null hypothesis, that the likelihood of locating activated foci is equal at every voxel); 6) reported only deactivations (this occurred in only one study [12], which was thus not comparable to the others); or 7) included adults with ADHD in partial remission, as it has not been established whether the neuronal correlates of individuals with ADHD in partial remission are similar to those with the full syndrome.

Two authors (S.C. and C.C.) independently searched the literature, examined the retrieved articles, and extracted and cross-checked data. Initial disagreements on seven of 2,287 screened articles were resolved by consensus. We extracted demographic information, ADHD diagnostic criteria and subtype, psychiatric comorbidities, medication status, three-dimensional coordinates, tasks, and contrasts.

We conducted an activation likelihood estimation meta-analysis using GingerALE, version 2.1.1 (www.brainmap.org/ale/). Activation likelihood estimation allows the detection of quantitative interstudy consistencies in activation by generating maps of activation likelihood estimates. In fMRI studies, the precise localization of specific activation coordinates is limited by substantial intersubject anatomical variability. Within studies, this is imperfectly addressed by Gaussian smoothing. Accordingly, activation foci are best considered as localization probability distributions centered at the reported coordinates. Based on this logic, in activation likelihood estimation meta-analysis, foci are first transformed into probability distributions using three-dimensional Gaussian functions with width expressed in millimeters at half the maximum value (referred to as full width at half maximum). Second, a whole-brain map is created by assigning each voxel a value equal to the probability that at least one of the activation points will be found within the voxel. This value is referred to as the activation likelihood estimation for each voxel. Third, to differentiate the voxels in the map that represent signal (i.e., nonrandom clustering of foci) from those that represent noise (i.e., random clustering), activation likelihood estimation values are compared with a null hypothesis distribution generated by permutation analysis (see www.brainmap.org/ale/).

For our meta-analysis, coordinates reported in Talairach space were transformed to Montreal Neurological Institute (MNI) coordinates using the icbm2tal (Lancaster) transformation (13). Moreover, since nearby coordinates cannot be assigned unequivocally to different regions, when coordinates associated with multiple contrasts from the same task were less than 12 mm apart, we excluded all but one (e.g., for a set of four coordinates within 12 mm of each other, all but the fourth peak were excluded). Statistical significance was determined using a permutation test (5,000 permutations) of randomly generated foci, corrected for multiple comparisons. Per Eickhoff et al. (14), full width at half maximum was calculated based on the number of participants in each study. The threshold for final activation likelihood estimation maps was set at p<0.05 using the false discovery rate with an extent threshold greater than 200 mm³ (the GingerALE default) and overlaid onto the MNI 152 template. As recommended (15), anatomical labels were assigned after direct examination of anatomy (16).

We performed focused meta-analyses of studies contrasting 1) children (age<18 years) with ADHD and comparison children across all tasks; 2) adults (age≥18 years) with ADHD and comparison adults across all tasks; 3) all stimulant-naive participants with ADHD (regardless of age) and comparison subjects (studies were included in this subanalysis only if all participants were stimulant naive); 4) all comorbidity-free individuals with ADHD (regardless of age) and comparison subjects; all individuals with ADHD (children and adults) and comparison subjects in 5) inhibition tasks, 6) working memory tasks, and 7) vigilance and attention tasks. As shown in Table 1, the number of retrieved studies with relevant foci was insufficient to perform separate meta-analyses of studies assessing 1) paradigms other than inhibition, working memory, or vigilance/attention; 2) individual tasks in children and adults separately; 3) individual paradigms in participants who were stimulant naive or without psychiatric comorbidities; and 4) ADHD > comparison subjects for working memory or vigilance and attention tasks. Additionally, for comparability with the previous meta-analysis (4), we performed a meta-analysis across all pertinent studies reporting results of between-group contrasts, regardless of participant characteristics and the specific paradigm tested (an “omnibus” meta-analysis).

To test whether results of the focused meta-analyses differed statistically, we performed subtraction analyses using the contrast studies procedure in GingerALE. We compared ADHD adults with ADHD children; stimulant-naive individuals with stimulant-treated individuals; and participants with comorbid mental disorders with comorbidity-free participants in the contrasts comparison subjects > ADHD and ADHD > comparison subjects. For “all participants, inhibition tasks” compared with “all participants, working memory tasks,” “all participants, inhibition tasks” compared with “all participants, vigilance/attention tasks,” and “all participants, working memory tasks” compared with “all participants, vigilance/attention tasks,” we could only examine the contrast comparison subjects > ADHD.

We related the ADHD hypo- and hyperactivated regions from our meta-analysis to seven reference networks defined by Yeo et al. (9) on the basis of a data-driven analysis of resting state functional imaging data collected from 1,000 participants. Those seven robustly replicable networks, which are limited to cortical regions, include the frontoparietal, the dorsal and ventral attentional, the somatomotor, the visual, the limbic, and the default networks. We first determined the network in which each voxel of the ADHD-related hypo- or hyperactivated regions was located by computing the number of significant voxels from the comparison subjects > ADHD and ADHD > comparison subjects contrasts that overlapped each of the network masks (downloaded from http://surfer.nmr.mgh.harvard.edu/fswiki/CorticalParcellation_Yeo2011). We then performed a chi-square analysis contrasting the proportions of hypo- and hyperactivated voxels in the seven networks.

Figure 1 summarizes the search results according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart (17). Details of included and excluded studies are provided in section A2 of the online data supplement. The search yielded 55 eligible studies, 16 of them assessing adults and 39 assessing children. Additional characteristics of the studies are summarized in Table 1. While we endeavored to count individuals from the same sample only once, the total number of participants reported in Table 1 (741 with ADHD and 801 comparison subjects) is approximate because some research groups reported results with partially overlapping samples. References of the included studies are provided in section A3 of the online data supplement.

The meta-analysis focused on children (Table 2 and Figure 2) revealed significant ADHD-related hypoactivation in frontal regions and putamen bilaterally and in right parietal and right temporal regions. ADHD-related hyperactivation was observed in the right angular gyrus, middle occipital gyrus, posterior cingulate cortex, and midcingulate cortex. In the meta-analysis restricted to adults (Table 2 and Figure 2), a different pattern emerged. Adults with ADHD showed significant hypoactivation relative to comparison subjects in the right central sulcus, precentral gyrus, and middle frontal gyrus. ADHD-related hyperactivation was observed in a region with a peak in the right angular and middle occipital gyri.

In stimulant-naive participants (see Table S1 and Figure S1 in the online data supplement), ADHD-related hypoactivation was observed in several frontal regions bilaterally, the right superior temporal gyrus, right posterior cingulate cortex, right postcentral gyrus, the putamen bilaterally, and the right thalamus. Only one significant cluster of ADHD-related hyperactivation was observed, with a peak located in the right superior longitudinal fasciculus underlying the insula.

When considering comorbidity-free participants (see Table S1 and Figure S1 in the online data supplement), ADHD-related hypoactivated regions were located in the frontal regions and the putamen bilaterally, the right superior temporal gyrus, and the right occipital pole. ADHD-related hyperactivation was observed in the left inferior frontal gyrus, left Heschl’s gyrus, and several right posterior regions.

In analyses limited to specific tasks (see Table S1 in the online data supplement), ADHD-related hypoactivation in studies examining inhibition paradigms included several frontal regions bilaterally as well as the right superior temporal gyrus, the left inferior occipital gyrus, the right thalamus, and the midbrain. The contrast ADHD > comparison subjects yielded two regions with peaks in the deep right parieto-occipital cortex and right intermediate frontal sulcus. The analysis restricted to working memory tasks revealed significant ADHD-related hypoactivation in the left inferior frontal gyrus and anterior insula and in the right middle frontal gyrus. In vigilance or attention tasks, we observed significant ADHD-related hypoactivation only in the right paracingulate gyrus.

The omnibus meta-analysis (Figure 2; see also Table S2 in the online data supplement) largely recapitulated the results of the meta-analysis focused on children, with additional ADHD-related hypoactivations in the inferior frontal gyri, right central sulcus, and right posterior parietal lobe and hyperactivations in Heschl’s gyrus as well as additional loci in inferior frontal gyri.

When we performed planned subtraction analyses (adults compared with children; stimulant-naive compared with stimulant-treated individuals; comorbid compared with comorbidity-free participants; and comparisons among specific tasks), no significant differences that survived whole-brain false discovery rate correction were found. However, with an exploratory threshold set at p<0.05, uncorrected, we observed differences in the following analyses: adults with ADHD compared with children with ADHD for the comparison subjects > ADHD, all tasks contrast; stimulant-naive individuals compared with stimulant-treated individuals for the comparison subjects > ADHD, all tasks contrast, and the ADHD > comparison subjects, all tasks contrast; comorbid participants compared with participants without comorbidity for the comparison subjects > ADHD, all tasks contrast and the ADHD > comparison subjects, all tasks contrast (see Table S3 in the online data supplement).

Figure 3 shows the number of voxels located in each of the seven networks described by Yeo et al. (9), expressed as a percentage of the total number of significant voxels. In children, ADHD-related hypoactivation was predominantly located in the ventral attention (44%) and frontoparietal (39%) networks, whereas hyperactivation was predominant in the default (37%), ventral attention (23%), and somatomotor (22%) networks. The overall distribution of hypo- and hyperactivations by network differed significantly (χ²>100, df=5, p<0.0001). In adults, voxels hypoactivated in ADHD were almost exclusively located in the frontoparietal network (97%), whereas hyperactivated voxels were found in the visual (41%), dorsal attention (33%), and default (26%) networks. The overall distribution of hypo- and hyperactivations by network differed significantly (χ²>100, df=5, p<0.0001). Results for the omnibus meta-analyses largely overlapped those for the meta-analysis focused on children (Figure 3 and Table S2 in the online data supplement). Considering comorbidity-free participants (Figure S2 in the online data supplement), voxels hypoactivated in ADHD were predominantly in the ventral attention (30%), frontoparietal (29%), and default (21%) networks, while most of the ADHD-related hyperactivation was in the default (44%) and somatomotor (26%) networks. In medication-naive participants, voxels hypoactivated in ADHD were mostly in the frontoparietal (27%), default (24%), and ventral attention (18%) networks. Considering the opposite contrast (ADHD > comparison subjects), few voxels were hyperactivated, all in the somatomotor system, which accounts for the unanimity reported in Figure S2 in the online data supplement.

Given the lack of data for the contrast ADHD > comparison subjects for most of the specific tasks, we do not report assignment to the canonical networks for the meta-analyses restricted to specific tasks.

ADHD is increasingly being conceived as a disorder underpinned by dysfunctions in multiple large-scale brain networks (7, 8). This perspective facilitates identification of several broad themes that emerged across our multiple meta-analytic instances. These include hypoactivation in the frontoparietal executive control network, putamen, and ventral attention network, which is consistent with the classical model of ADHD as a disorder of deficient fronto-striatal activation (32). However, we also detected substantial hyperactivation, particularly in the default network and visual circuits, which supports a model of ADHD based on the interrelationships among neural networks. Together with more recent reports (12, 23), our results of default network hyperactivation are consistent with the hypothesis that the inconsistency that characterizes many individuals with ADHD results from faulty interregulation between the default network and task-positive circuits such as the frontoparietal, ventral, or dorsal attention networks (7). The next generation of functional imaging studies should examine the temporal dependencies between behavioral indices of attentional lapses (e.g., episodically prolonged response times) in relation to their brain imaging correlates with the goal of capturing the deviations in the interplay between the default and task-positive networks.

Our results also provide meta-analytic support to views positing ADHD as a disorder characterized not only by functional deficiencies but also by possible compensatory mechanisms, such as hyperactivation in visual regions. Such putative compensatory mechanisms can be difficult to observe through clinical measures alone, but become evident through neuroimaging (33). Awareness and documentation of brain compensatory mechanisms may eventually yield a clinical benefit from neuroimaging. This would be analogous to the use of neurocognitive assessments to identify particular strengths to best formulate a comprehensive clinical treatment plan.

We failed to find support for the involvement of regions related to motivation and emotion, such as the ventral striatum (34), orbitofrontal, or amygdala/hippocampus (35), despite increasing recognition of motivational and emotional dysfunction in models of ADHD (36). We cannot rule out type II error, as three studies assessing reward (37–39) and two on emotional processing (40, 41) did not meet our inclusion criteria. In addition, the orbitofrontal cortex and the medial temporal lobes are challenging brain regions to examine with fMRI because of susceptibility artifact and signal dropout. We also note the lack of apparent involvement of the cerebellum, which has been implicated in ADHD by multiple volumetric and functional studies and by theoretical models emphasizing the role of cerebellar dysfunction in contributing to deficits in monitoring the frequency and timing of events (6). In our meta-analysis, several peaks in the cerebellum did not reach our cluster size threshold for significance. We suspect that high intersubject variability in cerebellar geometry relative to stereotaxic space mitigated the emergence of cerebellar findings across studies.

Several limitations of our meta-analysis should be taken into account. The first relates to selection criteria. To minimize confounding factors such as differences in diagnostic procedures or analytical approaches, we excluded approximately half the screened studies. Still, the studies included were heterogeneous, for example, with respect to the method used to correct for multiple comparisons. This is notable because activation likelihood estimation does not take into account interstudy differences in statistical thresholds. Second, separate meta-analyses could not be performed by ADHD subtype or sex, because separate results for male and female participants and ADHD subtypes are not usually reported in the literature. This is unfortunate since patterns of fMRI activation in ADHD can differ by sex (42) and, possibly, by subtype (43). Third, the meta-analytic approach we adopted allows a quantitative summary of positive results but cannot take into account negative findings. Effect sizes in fMRI may be confounded by many factors, such as movement covariates, and there is no agreement on how they should be handled. Thus, activation likelihood estimation should be considered a summary of the spatial distribution of positive results, rather than a true meta-analysis. Fourth, it was generally not possible to determine the extent to which overlapping samples were reported across studies. Thus, the total number of participants represents an upper bound. The open sharing of fMRI data (e.g., via www.OpenfMRI.org) should obviate this problem in the future. Finally, fMRI data are fundamentally limited. Besides only indirectly reflecting neuronal activity (44), fMRI data cannot define an absolute or quantitative baseline state of activation (45), as they always depend on the differences in signal between two conditions. Positron emission tomography provides absolute quantification (e.g., see 46, 47), but current ethical constraints limit its application to adults.