Multimodal mapping of the brain’s functional connectivity and the adult outcome of attention deficit hyperactivity disorder

Significance

Many children do not simply “outgrow” attention deficit hyperactivity disorder (ADHD). The disorder often persists and affects around one in 40 adults, presenting a major public health challenge. Defining the mechanisms that underpin this variable clinical outcome could stimulate novel approaches to boost recovery in ADHD. We map the brain’s functional architecture in 205 young adults followed clinically since childhood. We find clinically significant inattention persisting from childhood has a disruptive effect on the functional connections within and between the brain’s major networks. These disruptions are similar whether defined through direct observation of neuronal activity or measures of hemodynamic change. By contrast, adults who remit from childhood ADHD showed typical brain connectivity, suggesting convergence toward typical brain function may underpin recovery.

Abstract

We have a limited understanding of why many children with attention deficit hyperactivity disorder do not outgrow the disorder by adulthood. Around 20–30% retain the full syndrome as young adults, and about 50% show partial, rather than complete, remission. Here, to delineate the neurobiology of this variable outcome, we ask if the persistence of childhood symptoms into adulthood impacts on the brain’s functional connectivity. We studied 205 participants followed clinically since childhood. In early adulthood, participants underwent magnetoencephalography (MEG) to measure neuronal activity directly and functional MRI (fMRI) to measure hemodynamic activity during a task-free period (the “resting state”). We found that symptoms of inattention persisting into adulthood were associated with disrupted patterns of typical functional connectivity in both MEG and fMRI. Specifically, those with persistent inattention lost the typical balance of connections within the default mode network (DMN; prominent during introspective thought) and connections between this network and those supporting attention and cognitive control. By contrast, adults whose childhood inattentive symptoms had resolved did not differ significantly from their never-affected peers, both hemodynamically and electrophysiologically. The anomalies in functional connectivity tied to clinically significant inattention centered on midline regions of the DMN in both MEG and fMRI, boosting confidence in a possible pathophysiological role. The findings suggest that the clinical course of this common childhood onset disorder impacts the functional connectivity of the adult brain.

Childhood attention deficit hyperactivity disorder (ADHD) has a highly variable outcome in adulthood, with around 20–30% of those affected retaining the full syndrome and the remainder experiencing either full or partial remission (1, 2). The public health importance of ADHD persisting into adulthood is clear. Costs related to each adult with ADHD amount to between $10,000 and to $30,000 per annum when considering the health care and justice systems along with income-related factors (3). This does not consider the distress that ADHD persisting into adulthood can cause for individuals at work, at home, and with friends. The scientific opportunities of considering the variable clinical course of ADHD are also clear. Defining the neurobiological mechanisms that underpin remission could stimulate novel approaches to boost recovery in ADHD and related disorders.

Here, we ask how the degree of persistence of childhood ADHD symptoms into adulthood impacts the brain’s functional connectivity. We nest this study of adult brain function within a longitudinal study which has followed children with and without ADHD into their early adulthood (2). This design allows us to determine how the clinical course of the disorder, defined prospectively, impacts adult brain function. By contrasting adults whose childhood ADHD persists with those whose symptoms remit, we can elucidate the mechanisms that underpin the clinical outcome of the disorder. In earlier studies, we found that adults whose childhood ADHD had remitted showed white matter microstructural and some neuroanatomic features that did not differ significantly from peers who were never affected by ADHD (2, 4). By extension, we now hypothesize that remission will be associated with typical functional connectivity, whereas persistent symptoms will be tied to anomalous functional connectivity.

We examine the brain’s functional networks and the connections between them by delineating the patterns of coordinated brain activity that occur while a subject is at rest, free of any explicit task demands. Usually, this is accomplished by mapping correlated patterns of changes in the blood oxygenation level-dependent (BOLD) signal detected by functional MRI (fMRI). It also possible to measure neuronal activity directly using magnetoencephalography (MEG). This technique noninvasively detects the magnetic fields evoked by synchronized current flow in neuronal populations. It allows a direct observation of the neuronal oscillations that synchronize brain activity across spatially distinct brain regions, scaffolding cognitive functions pertinent to ADHD, such as attentional processes (5, 6). The exquisite temporal resolution of MEG (on the order of milliseconds) complements the excellent spatial resolution of fMRI. Additionally, similarities have been reported in intrinsic functional networks as mapped by MEG and fMRI in healthy adults (7, 8). However, it is unknown if such cross-modal similarities would also hold in a clinical population of individuals with ADHD, whose core symptoms arguably reflect dysregulation of the brain-intrinsic networks. If symptoms of ADHD were associated with similar network anomalies across modalities, this would bolster the case for a possible pathophysiological role of these networks.

Based on the prior literature, we expect that anomalies in functional connectivity pertinent to the adult outcome of childhood ADHD would center on the default mode network (DMN) (9–11). This network is prominent during introspective processing and is held to be pivotal in the pathophysiology of ADHD (10). Specifically, it has been hypothesized that a loss of the counterbalance between the DMN and networks supporting cognitive processes (e.g., attention and cognitive control) leads to intrusion of the DMN into task-oriented processing. Such dysregulated interactions, in turn, have been tied to core symptoms of ADHD, particularly increased distractibility, behavioral impulsivity, and deficient sustained attention (12–14).

In summary, we determine functional connectivity both electrophysiologically and hemodynamically. We hypothesize that symptoms of ADHD persisting from childhood would be associated with disruptions to functional connectivity, centered on the DMN. By contrast, adults whose symptoms have resolved are predicted to show functional connectivity that resembles that seen in those never affected by the disorder.

Results

In total, 205 individuals participated in this study (mean age = 23.4 y, SD = 3.9 y, 115 males). Of these, 101 had DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition; ref. 15)–defined childhood ADHD; 50 showed persistence of the full syndrome into adulthood (Table S1). Of the 205 participants, 67 had both good-quality fMRI and MEG resting-state data, 80 had fMRI alone, and 58 had MEG alone. Connectivity matrices were defined for the MEG data (considering the consistency of phase coupling between all possible combination of 2,146 regions of a cortical grid) and fMRI data (measuring BOLD signal correlations across the same cortical grid). Next, using a bootstrapped-based independent component analysis (ICA), we reduced these matrices to independent connectivity patterns or “stable components” (Fig. S1). We identified six such stable components in fMRI resting-state data and 56 such stable components in MEG resting-state data (12 in delta bands, 13 in theta bands, 10 in alpha bands, 10 in beta bands, and 11 in gamma bands). Cross-modal similarities were found (Fig. S2). For example, one fMRI component showed strong positive connections within and between the default mode, cognitive control, and attention networks. One of the MEG components similarly showed strong connectivity within and between the default mode, cognitive control, and ventral attention networks. Hence, these components revealed both intranetwork patterns and a complex scaffolding of internetwork connections, similar to some prior reports (16, 17).

Table S1.

Sample characteristics

	MEG				fMRI
Variable	Never affected (n = 58)	Remitted (n = 35)	Persistent (n = 32)	Test statistic (P value)	Never affected (n = 71)	Remitted (n = 35)	Persistent (n = 41)	Test (P value)
Age: Mean (SD), y	23.7 (3.7)	23.5 (3.3)	23.5 (4.0)	F = 0.06, P = 0.95	24.8 (3.7)	24.2 (4.1)	25.0 (4.8)	F = 0.04, P = 0.65
Male N, %; Female N, %	37 (63.8)	24 (68.6)	16 (50.0)	χ(2)2 = 2.7, P = 0.26	38 (53.5)	25 (61.0)	17 (48.6)	χ(2)2 = 1.2, P = 0.54
Inattention (N symptoms and SD)	NA	1.83 (1.42)	6.13 (1.58)	t(65) = 11.7, P < 2 × 10−16	NA	1.76 (1.32)	5.89 (1.80)	t(74) = 11.5, P < 2 × 10−16
Hyperactivity-impulsivity	NA	1.42 (1.44)	3.88 (2.79)	t(65) = 4.57, P < 2.3 × 10−5	NA	1.51 (1.34)	3.94 (2.57)	t(74) = 5.28, P < 1.3 × 10−6

NA, not applicable.

Fig. S1.

Methods overview. (A) The definition of each participant's connectivity matrix. (B) The extraction of group level connectivity patterns—the stable components. (C) Calculation of the degree of each individual's expression of each stable component. DAN, dorsal attention network; VAN, ventral attention network.

Fig. S2.

Similar internetwork patterns of connectivity in fMRI and MEG. (Left) Stable components extracted from fMRI resting state data. (Right) MEG stable components that showed a similar pattern of internetwork connectivity. Thicker lines represent more typical values in the distribution of all network connections in the component, and thinner lines are outliers. Six networks (nodes) are illustrated; MEG coverage of the affective network was spare and so is not shown. CC, cognitive control; DAN, dorsal attention network; VAN, ventral attention network.

Altered Functional Connectivity Patterns and the Adult Outcome of Childhood ADHD.

We then asked if symptoms of adult ADHD persisting from childhood impacted an individual’s expression of each component. Individual expression of each stable component was represented as a beta weight calculated using linear regression. We then correlated these components (the beta weights) with symptoms. In MEG, four of the 56 stable components were significantly associated with adult symptoms of inattention persisting from childhood (Table 1). These correlations ranged from rho = 0.45, P = 1.1 × 10⁻⁴ (for a MEG component in the theta band); to rho = 0.47, P = 5.1 × 10⁻⁵ (beta-frequency component); to rho = 0.48, P = 3.8 × 10⁻⁵ (beta band); to rho = 0 0.54, P = 1.9 × 10⁻⁶ (gamma band). No component showed a significant association with hyperactivity-impulsivity.

Table 1.

Components that had a significant association with symptoms of inattention

Component (connectivity pattern)	Spearman correlation rho (P value)	Kruskal–Wallis χ(2)2 (P value)	Post hoc Mann–Whitney U tests
			Never affected vs. persistent (P value)	Never affected vs. remitted (P value)	Persistent vs. remitted (P value)
MEG, theta	0.45 (1.1 × 10−4)	19 (7.5 × 10−5)	3.9 × 10−5	3.3 × 10−2	1.1 × 10−2
MEG, beta	0.47 (5.1 × 10−5)	11.1 (3.9 × 10−3)	1.8 × 10−3	2.0 × 10−1	2.0 × 10−2
MEG, beta	0.48 (3.8 × 10−5)	29.7 (1.0 × 10−7)	1.6 × 10−7	1.6 × 10−2	1.2 × 10−3
MEG high gamma	0.54 (1.9 × 10−6)	16.5 (2 × 10−6)	7.3 × 10−7	3.2 × 10−1	1.1 × 10−4
fMRI	0.31 (6.2 × 10−3)	7.8 (2.0 × 10−2)	8.0 × 10−3	8.4 × 10−1	2.5 × 10−2

Results of pairwise post hoc comparisons shown in bold were significant at an unadjusted P < 0.05. The remitted and never-affected groups either did not differ significantly (three components) or differed at nominal levels of significance (P < 0.05).

We next conducted categorical analyses of these inattention-related components by classifying the adults into those with persistent ADHD and those with remitted ADHD, thus facilitating the clinical interpretation of the findings. We also drew contrasts against the never-affected control group to determine the degree to which the remitted group showed the predicted typical functional connectivity patterns. Group differences were found in all four of the MEG components significantly associated with inattention. These differences were consistently driven by the persistent group (as can be seen from Fig. 1, which shows heat maps of the associations between each network for each outcome group). Specifically, post hoc pairwise contrasts showed that the persistent group showed significantly higher expression of all four components compared with the never-affected group (all P < 0.001; Table 1) and higher expression compared with the remitted group for two of the stable components (P = 0.01 for the theta component and P = 0.001 for one of the beta components). By contrast, the remitted and never-affected groups showed mostly similar patterns of connectivity within and between the brain’s intrinsic networks (i.e., the two groups did not generally differ significantly in their expression levels of the inattention-related stable components).

Fig. 1.

Group differences in the expression of stable components associated with inattention, detected by MEG (Top) and fMRI (Bottom). The box plots show the expression of each component in the persistent ADHD, remitted, and never-affected groups. Throughout, the persistent ADHD group showed atypically high levels of expression of these components. The adjacent connectivity matrices (“heat maps”) show the strength of intra- and internetwork connectivity for each group. Hotter colors indicate stronger connectivity. The persistent ADHD group differs significantly from the never-affected and remitted ADHD groups. The latter two groups did not differ from one another. DAN, dorsal attention network; VAN, ventral attention network.

We used the same approach to consider the correlation between symptoms and the expression of each component in fMRI. Only one of the six components showed significant association with inattentive (rho = 0.31, P = 0.0062) but not hyperactive-impulsive (rho = 0.15, P = 0.19) symptoms. In categorical contrasts, a group difference [Kruskal–Wallis, χ(2)² = 7.8, P = 0.02] was driven by higher levels of expression of the stable component in the persistent ADHD group compared with both the remitted (P = 0.03) and never-affected (P = 0.008) groups, which did not differ significantly from one another (P = 0.85).

In summary, adult inattention persisting from childhood was tied to disruption in patterns of connectivity within and across brain networks, mapped both hemodynamically (fMRI) and electrophysiologically (MEG). Individuals with persistent inattention showed atypical (higher) expression of connectivity patterns between the brain’s networks in both modalities, whereas those who remitted did not differ from those who were never affected.

The Functional Architecture of ADHD-Related Stable Components.

We further characterized the four inattention-associated components detected by MEG against the 52 MEG components that were unrelated to symptoms. Informed by current models of ADHD, we focused on patterns of connectivity within the DMN and between the DMN and other networks [as defined by Yeo et al. (18)].

We found that the four MEG inattention-related components showed a significant alteration in the ratio of such intra- to internetwork connectivity. Specifically, the components related to inattention showed relatively less intra-DMN connectivity and more internetwork connectivity, a difference which reached significance in a group rank test (P = 0.02; Fig. 2).

Fig. 2.

Patterns of network connectivity in components related to inattention (A) against those that are not (B). The thickness of the arrows corresponds to the strength of the connection. Inattention-related components showed more internetwork connectivity than unrelated components, particularly between the default mode and ventral attention network (VAN) and cognitive control network. DAN, dorsal attention network. (C) Scatterplot shows that the four inattention-related MEG components had a significantly lower ratio of intranetwork/internetwork default mode connections than those not associated with inattention.

Finally, we tested for similarities in the spatial structure of the MEG and fMRI inattention-related components. For every cortical region and each component, we calculated a metric summarizing the connectivity between that cortical region and all others. High values of this metric indicated greater deviation away from mean connectivity. Ranking regions by this metric, we found that the left precuneus and bilateral posterior cingulate showed the most atypical connectivity in MEG and fMRI modalities; overlap was less prominent in lateral cortical regions (Fig. 3). Both the precuneus and posterior cingulate are regarded as key components of the DMN (9).

Fig. 3.

Similarities in the spatial structure of the MEG and fMRI inattention-related components. Cortical regions are ranked according to their connectivity with all other regions (high “hot” color indicates greater deviation away from the mean connectivity). Atypically connected regions that overlapped in fMRI and MEG localized to midline structures (left precuneus and bilateral posterior cingulate gyrus) that are core components of the DMN.

Discussion

We show that clinically significant symptoms of adult ADHD were tied to similar disruptions to the brain’s functional connectivity, mapped using both resting-state fMRI and MEG data. Specifically, adult inattention persisting from childhood was associated with an atypical balance of connectivity within the DMN and between the DMN and task-positive networks. These patterns of connectivity, tied to symptom severity, had a similar spatial structure, whether defined hemodynamically or electrophysiologically. We also find that adult remission from childhood ADHD was associated with essentially typical functional connectivity. As the adults in our study had been followed clinically since childhood, the findings inform our understanding of recovery from the disorder.

The Impact of Adult Inattention on Neural Oscillations.

While fMRI data acquired during the resting state have been widely used to parse the brain’s functional architecture in ADHD, only a handful of studies have used MEG. Prior MEG studies have exclusively focused on adults with persistent ADHD, and, consistent with our findings, they report atypical neural oscillations in beta-, delta-, and theta-frequency bands (19, 20). Neural oscillations scaffold cognitive domains, and some of these cognitive domains are impaired in ADHD, such as working memory and the processing of temporal information. For example, working memory capacity is predicted by peak theta-band activity, working memory load correlates with gamma-band power, and interventions that boost working memory also augment beta-band activity (6, 21–23). Coupling across the frequency bands is also implicated in working memory, specifically modulation of the gamma band amplitude by the theta band phase (24–26). ADHD is also linked with the aberrant processing of temporal information, with errors in time perception, reproduction, and estimation (27). Neural oscillations convey temporal information, and the encoding of time intervals and temporal prediction has been linked to theta-, beta-, and delta-range activity (28–31). Imaging brain activity during working memory and temporal processing tasks would delineate more clearly the links between symptoms, cognitive deficits, and anomalous neural oscillations.

We found that adults with persistent inattention show an atypical balance of intra- to internetwork connectivity pertaining to the DMN. As discussed earlier, such disruptions within the DMN and its interactions with task-positive networks are associated with core cognitive deficits in ADHD, such as increased distractibility, motor disinhibition, and deficient sustained attention (12–14). These connectivity patterns change with age, with most studies finding developmental increases in functional connectivity within the DMN along with maturational change in its interconnections with task-positive networks (17, 32). Indeed, the extent of deviation from the patterns of normative network maturation has been found to predict both deficient sustained attention and the diagnosis of childhood ADHD (16, 17). Here, we extend this finding into adulthood by showing that the degree of deviation from adult norms in the balance of connectivity within and beyond the DMN is tied to the degree of persistence of childhood inattention.

Cross-Modal Similarities in the Patterns of Functional Connectivity Tied to Inattention.

We find similarities in the spatial structure of the fMRI- and MEG-defined connectivity patterns related to inattention. This is consistent with prior studies of typically developing adults that find resting-state neural oscillations as measured by MEG in theta, alpha, and beta bands coalesce into networks that have a similar spatial structure to networks defined using fMRI (7, 8). We extend these findings by showing that cross-modal similarities can also be detected in a neuropsychiatric disorder. We found that in both MEG and fMRI, the patterns of abnormal connectivity related to inattention were centered on the midline regions of the DMN. Differences in the functional networks uncovered by MEG and fMRI were also apparent and to be expected, given factors such as the different spatial and temporal resolutions of each modality.

Models of the Adult Outcome of Childhood ADHD.

Functional connectivity in adults whose childhood ADHD symptoms had remitted did not differ significantly from their never-affected peers, both hemodynamically and electrophysiologically. This finding of “typical” function in remitted ADHD is consonant with our prior demonstration that remission is also associated with more typical neuroanatomy and white matter microstructural properties (2, 4). Our current findings are less consistent with the concept of remission arising through the recruitment of new brain systems that help the individual overcome the core symptoms (33). Such compensation would entail some atypical, albeit functionally beneficial, changes in functional networks on those who remit, which we did not find.

Inattention, rather than hyperactivity-impulsivity, tracked with anomalies in the DMN and its interconnectivity with other networks, consistent with pathophysiological models that link default mode disruption with inattention. Associations between brain function and hyperactivity-impulsivity may be better detected through cognitive probes which directly investigate pertinent cognitive domains. Indeed, using a subsample of this group, we have found links between hyperactive-impulsive symptoms persisting from childhood and anomalous activation of frontostriatal-cerebellar circuitry during motor inhibition (34).

The study has its limitations. First, some adults with persistent ADHD were taking psychostimulant medication. To mitigate acute effects, all participants stopped psychostimulant medication the day before scanning. In addition, we repeated all analyses excluding those on psychostimulant treatment, and the overall pattern of results remained the same (Table S2). We also did not collect childhood functional data as the study started before the widespread use of resting-state fMRI and MEG as tools for parsing intrinsic functional connectivity. Meta-analyses report intrinsic connectivity differences between children with and without ADHD (11). However, such group-level analyses do not exclude the possibility that some of the children within the ADHD group do not differ from typicals. Only longitudinal data can rule out the possibility that children with ADHD who are destined for remission as adults have a typical childhood functional architecture that is carried into adulthood.

Table S2.

Medication analysis

Connectivity pattern (IC)	Spearman correlation rho (P value)	Kruskal–Wallis χ(2), n = 103)2 (P value)	Post hoc Mann–Whitney U tests
			Never affected vs. persistent (P value)	Never affected vs. remitted (P value)	Persistent vs. remitted (P value)
MEG, theta	0.32 (3.1 × 10−2)	6.72 (3.5 × 10−2)	<4.7 × 10−2	<4.7 × 10−2	=2.4 × 10−1
MEG, beta	0.47 (1.1 × 10−3)	6.27 (4.3 × 10−2)	<2.8 × 10−2	=1.4 × 10−1	=1.2 × 10−1
MEG, beta	0.47 (1.2 × 10−3)	14.42 (7.4 × 10−4)	<5.5 × 10−4	<3.5 × 10−2	>2.5 × 10−2
MEG high gamma	0.56 (6.9 × 10−5)	17.93 (1.3 × 10−4)	<4.5 × 10−5	=4.3 × 10−1	>3.0 × 10−4
fMRI	0.41 (1.5 × 10−3)	6.72 (3.5 × 10−2)	<1.1 × 10−2	=8.4 × 10−1	>3.7 × 10−2

Analyses shown are repeated, excluding participants taking medication, only for components that had previously shown a significant association with symptoms of inattention. Diagnostic group differences and post hoc pairwise comparisons are also given. Symbols are given to indicate the relationship between group medians.

While it is often argued that anomalies in connectivity cause symptoms, the reverse relationships may also be in play. For example, children growing up with ADHD are likely, by virtue of their symptoms, to have a partly distinct set of environmental exposures and experiences compared with those who are unaffected that could impact the development of the functional connectome. A relevant population-based, longitudinal imaging study found that ADHD symptoms in early childhood were more predictive of some white matter microstructural properties 2 y later than the inverse relationship (35). However, evidence against the concept that symptoms cause connectivity anomalies comes from the finding that unaffected siblings show similar, if attenuated, anomalies in white matter tract microstructure to their siblings who have ADHD (36). Neither study examined the functional connectome. In short, longitudinal studies using cross-lagged panel and similar designs are needed to delineate fully the temporal links between symptoms and the functional connectome.

By studying the adult outcome of childhood ADHD, we show how persisting inattention is tied to anomalous connectivity centered on the DMN, both electrophysiologically and hemodynamically. Such steps toward understanding the neurobiological mechanisms that underpin remission in some individuals might guide the development of interventions to promote recovery in all.

Methods

Participants were drawn from studies at the intramural programs of the NIH. Of the 101 participants with childhood ADHD, 90 had their initial diagnosis made at the NIH using the Parent Diagnostic Interview for Children and Adolescents (37). The other 11 had a community diagnosis, which we confirmed through both collateral informants and reference to records. The assessment of adult symptoms of ADHD was through the clinician-administered ADHD Rating Scale, version IV, which provides prompts appropriate for late adolescent and young adult groups (38). Two clinicians (P.S. and W.S.) rated each of the nine possible symptoms of inattention and nine symptoms of hyperactivity/impulsivity. In line with the DSM-5, persistent ADHD was defined as the presence of five or more symptoms of inattention and/or five or more symptoms of hyperactivity/impulsivity. The Structured Clinical Interview for DSM Axis I Disorders was used to assess for other psychiatric disorders. Medication histories were obtained from participants. Contrasts were made against 104 subjects who never had ADHD, who were drawn from a study of typical brain development. These subjects, referred to as “never affected,” were free of current Axis I DSM-IV mental disorders. For all participants, intelligence quotient (IQ) was estimated using age-appropriate versions of the Wechsler intelligence scales. Exclusion criteria were a full-scale IQ of <80, evidence of neurological disorders known to affect brain structure, current substance dependence, or psychotic disorders. The Institutional Review Boards of the National Human Genome Research Institute and National Institute of Mental Health approved the research protocol, and written informed consent was obtained from participants.

MEG Acquisition and Preprocessing.

Resting-state data were acquired in the MEG scanner, with participants instructed to relax for 240 s with their eyes closed. MEG data were acquired using a 275-gradiometer, whole-helmet MEG system (CTF Systems) at 600 Hz. The locations of three fiducial points (nasion and left/right auricular) were determined for signal source localization. A combination of CTF SAMtools (https://www.ctf.com/products/meg/ctf/software.htm), MNE-Python (39), and custom scripts were used in analyses.

To control data quality, head position with respect to the MEG helmet was recorded at the beginning and end of the session, and data were retained only if there was less than 0.5 cm of total movement. Next, data were visually inspected for known artifacts (e.g., movement, muscle related noise), and boundaries before and after each occurrence of an artifact were marked. The data outside artifact boundaries were then split into 13.65-s contiguous epochs to achieve a good balance between frequency resolution and connectivity estimation (40). Only subjects with at least five clean epochs (68.25 s) were used in the analysis. As a result of these measures, 56 of the initial 181 datasets were excluded, leaving 125 datasets for analysis. There was no group difference (remitted vs. persistent ADHD vs. never affected) in the amount of data analyzed [Kruskal–Wallis, χ(2)² = 0.90, P = 0.64] or correlation between the amount of data analyzed and symptom counts (n = 67; inattention: rho = 0.07, P = 0.58; hyperactivity/impulsivity: rho = 0.16, P = 0.20). Furthermore, there was no group difference in the pre/post head position displacement [Kruskal–Wallis, χ(2)² = 1.56, P = 0.46] or correlation between this displacement and symptom counts (n = 67; inattention: rho = 0.14, P = 0.25; hyperactivity/impulsivity: rho = −0.07, P = 0.55).

Before source localization, MEG data were low-pass-filtered at 100 Hz and third-order gradient software compensation was applied. Localization used a beamformer method [synthetic aperture magnetometry (41)].

Covariance matrices were estimated in six frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), low gamma (30–55 Hz), and high gamma (65–100 Hz). Virtual electrode weights were calculated for positions in an equidistant 8-mm grid constrained to the cortex (grid defined in the same template used for fMRI registration). The frequency-dependent activation in each cortical location was estimated by applying the weight matrix to the MEG data of each clean epoch. This preprocessing produced for each individual and each frequency band an activation matrix of location by epoch by time. Further details are given in SI Methods.

fMRI Acquisition and Preprocessing.

Resting-state gradient echo planar images were acquired on a 3-T GE Signa scanner for 315 s (General Electric). Participants were instructed to lie in the scanner at rest, looking at a fixation cross. We acquired gradient echo planar images [repetition time = 2,500 ms; echo time = 27 ms; flip angle = 90°; 44 axial contiguous interleaved slices per volume; 2.8-mm slice thickness; field of view = 22 cm; 64 × 64 acquisition matrix; single-voxel volume = 3.4 mm, 3.4 mm, 2.8 mm] and aligned to a high-resolution, T1-weighted anatomical image [magnetization prepared rapid acquisition gradient recalled echo sequence (MP RAGE): 124 axial slices, 1.3-mm slice thickness, field of view = 22 cm, 224 × 224 acquisition matrix].

To control quality, functional volumes registering more than 0.2 mm of motion were excluded (removing the volume just before and after the movement). Volumes that had more than 10% of voxels as outliers were also censored. Each volume was then registered to the volume which had the minimum outlier fraction. Variables measuring head movement (pitch, roll, and yaw) were also regressed out before the connectivity analysis. As a result of these quality measures, we retained 147 of the original 155 datasets for analysis, with an average 287.1 ± 29.8 s of data after preprocessing. There was no difference between the groups (remitted vs. persistent ADHD vs. never affected) in the amount of data analyzed [Kruskal–Wallis, χ(2)² = 3.17, P = 0.20] or correlation between the amount of data analyzed and symptom counts (n = 76; inattention: rho = −0.14, P = 0.24; hyperactivity/impulsivity: rho = 0.005, P = 0.97). Furthermore, there was no group difference in amount of head movement estimated over the data analyzed [Kruskal–Wallis, χ(2)² = 3.32, P = 0.19] or correlation between the amount of movement and symptom counts (n = 76; inattention: rho = 0.13, P = 0.28; hyperactivity/impulsivity: rho = 0.07, P = 0.55).

The functional image was nonlinearly registered to a Montreal Neurological Institute (MNI) template (42), and ANATICOR (43) was used to remove the first three principal components of the lateral ventricles mask and the mean of a white matter mask, both extracted using FreeSurfer (surfer.nmr.mgh.harvard.edu/). The result of this fMRI preprocessing was a functional image with a time series of residuals in each of 67,821 voxels.

Mapping Functional Connectivity.

The goal of the analysis was to detect connectivity patterns in the brain associated with symptoms of ADHD using data-driven methods. In brief, first, connectivity matrices were defined for the MEG and fMRI data. Second, ICA reduced these matrices to independent connectivity patterns (called stable components here). Finally, we determined the degree to which each individual expressed each of these stable components (a methods overview is provided in Fig. S1).

Stage 1: Defining Resting-State fMRI and MEG Connectivity Matrices.

We laid an 8 × 8 × 8-mm grid with 2,146 locations constrained to the cerebral cortex for the MEG data and calculated all pairwise connections [2,301,585 possible connections (nconn)]. For the fMRI data, we placed the same grid across the cerebral cortex (again with 2,146 locations) to facilitate the comparison of MEG and fMRI connectivity matrices. As both MEG and fMRI were registered to the same MNI template, the same 2,146 locations were defined for all subjects. Each location was mapped to one of the seven canonical resting-state networks provided by Yeo et al. (18): visual, somatomotor, dorsal attention, ventral attention, affective, cognitive, and default mode. Thus, we could determine the canonical network from which each connection originated and ended.

For MEG, we quantified connectivity using a measure of the consistency of phase coupling between two signals over time: the imaginary part of the coherency (44). Specifically, for each frequency band, we computed the imaginary coherency between each possible combination of localized sources. Imaginary coherency ranges from 1 to −1, where 0 represents no synchronization and 1/−1 represents perfect phase coupling between two signals depending on which one is lagging/leading the interaction. As this approach projects the coherency onto the imaginary axis, it estimates phase synchrony without zero-lag contributions, thus removing potential contributions due to field spread and source localization leakage (45). These analyses produced matrices for each subject (one for each frequency band) that characterize the connectivity between every possible pairing of the 2,146 locations of the cortical grid. For fMRI, we similarly generated connectivity matrices between the locations of the same grid. Specifically, we averaged the residualized time series in each location of the grid and calculated Pearson correlations between these time series.

Stage 2: Extracting Stable Components (Connectivity Patterns).

We applied ICA to the connectivity matrices to extract independent components of spatial connectivity. Before defining these components, we first regressed out age, sex, movement, and movement^2 from each connection. In line with prior analyses, we then added the mean effect due to inattention and hyperactivity symptoms to augment the variance associated with these effects of interest (17). We reduced the data to 15 components for fMRI and for each MEG band in line with previous analyses (to give a total 105 components) (17). We then identified which of these components were stable using an ICA-based approach [ICASSO (46)], which estimates a stability metric between 0 and 1 (with 1 indicating perfect stability) (SI Methods). We ran 1,000 iterations using different initializations and bootstrappings of the data, and kept all components that had a stability metric of 0.85 or greater. This returned six stable components in fMRI and 56 stable components in MEG.

Stage 3: Individual Expression of Each Stable Component and Association with Symptoms.

Individual expression of each stable component was represented as a beta weight calculated using linear regression (Fig. S1). Our main question was whether an individual’s expression of these stable components (the beta weights) was associated with the severity of symptoms of inattention and hyperactivity/impulsivity. In these analyses, we controlled for multiple testing by applying a Bonferroni correction for the number of components in each modality (six for fMRI and 56 for MEG). We also conducted categorical analyses, contrasting adults with persistent ADHD against adults with remitted ADHD and the never-affected control group.

Stage 4: Interrogation of Components Associated with Inattentive Symptoms.

We next examined the components that were shown in stage 3 to be associated with adult outcome, and we contrasted these components against those unrelated to outcome. As discussed earlier, ADHD has been linked to a loss in the typical balance between connectivity within the DMN and connectivity between the DMN and other networks. We thus asked if the stable components we found to be linked to outcome showed such atypical patterns of intra- and internetwork connectivity. Specifically, we averaged all network connections where both locations mapped to the DMN (intra), and then all connections where one location mapped to the DMN but the other (inter) did not. Each component was assigned this ratio of intra- to internetwork connectivity, and a Mann–Whitney U test was used to test whether this ratio varied according to component association with outcome.

Finally, we asked whether there were spatial similarities between MEG and fMRI disorder-relevant stable components. Using the stereotaxic atlas of Talairach and Tournoux (47) to define 60 cortical subregions, we estimated the average connectivity weight for every stable component for those regions. Thus, for each cortical region, we calculated the average strength of the connection between that region and all others. We ranked the regions based on their absolute connectivity strength and compared the spatial structure of the fMRI- and MEG-derived stable components.

SI Methods

MEG Acquisition and Preprocessing.

All data not marked as artifacts were used to estimate the synthetic aperture magnetometry (SAM) (42) covariance matrices, band-passed within each of the six bands. Each subject’s anatomical MRI scan was registered to the MNI template, and a forward model of multiple local spheres (search radius = 9 mm) was used to calculate weights to predefined locations in an 8 × 8 × 8-mm grid. A few locations in the grid (specifically along the ventral side of the brain, mostly parts of the affective network) were poorly sampled by the MEG sensors and received SAM weights of 0. So, while MEG is well suited for the characterization of most cortical networks, it has poor coverage of the affective network, which lies in medial/inferior temporal and orbitofrontal regions (from the surface sensors). Thus, many locations within the affective network received zero weights in source localization procedures, and overall connections involving the network amount to 1% of the total connections. While we include the affective network in the main analyses, due to this sparse coverage, it is not included in the figures. Finally, the SAM solution was regularized using a Backus–Gilbert multiplier of 4, and weights were generated for normalized (signal-to-noise ratio) source projection.

Mapping Functional Connectivity.

Stage 2: Extracting stable components (connectivity patterns).

We used ICASSO (47) to reduce the dimensionality of the data to 15 components. Specifically, we ran 1,000 iterations employing different resamples with replacement (bootstrapping) of the data and random starts. Each iteration estimates the desired latent space with 15 dimensions [independent components (ICs)], which can be conceptualized as 15 individual points in an nconn-dimensional space. ICASSO reports the centroid of each cloud of points as the final estimate of each of the 15 different ICs.

If an IC is stable across iterations, it follows that the cloud of points (different iterations) around the centroid will be tight and easily separable among the 15 ICs. ICASSO implements a stability metric quality index (Iq) which combines those two characteristics into a single number which goes from 1 (tight clusters, easily separable) to 0 (very spread cloud, hard to separate from other clusters). For the current analysis, we focus on ICs with an Iq of at least 0.85 to assure that the ICs analyzed are stable in the dataset.

The result of the ICA analysis is 15 different connectivity patterns common across participants (specifically, a 15 × nconn matrix). Each of these connectivity patterns contains weights indicating the degree to which connectivity features covary across subjects. We have a total of seven such 15 × nconn matrices (one fMRI and six MEG), for a total of 105 total connectivity patterns (ICs). We only kept 62 of the 105 ICs that had Iqs above 8.5. Those 62 ICs were split as follows: 6 fMRI, 5 delta, 14 theta, 8 alpha, 10 beta, 5 gamma, and 14 high gamma.

Fig. S2 shows some of the fMRI stable components and closely matching MEG components. In each plot, we computed the average absolute connectivity between networks. The data were then rescaled so that the minimum value was 0 and the maximum value was 1. The inverse of those results is shown, such that thinner lines represent more atypical values in the distribution of internetwork connections.

Stage 3: Defining the degree to which each individual expresses each stable component.

Individual expression of each stable component was calculated using linear regression. Specifically:

$$S_{i,nconn}={\beta }_{i,k}{C}_{k,nconn},$$

where S_i,nconn is the connectivity matrix for subject i and C_{k, nconn} is the connectivity pattern for stable component k. This yields a vector of nsubj beta values for each stable IC, where nsubj is the number of subjects in the imaging modality. If β_i,k is close to 1, participant i’s data very closely match the connectivity pattern in the stable component k. Conversely, if β_i,k is close to −1, participant i’s data are less closely related to the connectivity pattern captured in the stable component k.

The beta weights, the degree to which each individual expresses each stable component, were further interrogated. Our main question was whether individual expression of these stable components was correlated with the severity of adult symptoms of inattention and hyperactivity/impulsivity (Spearman correlation). To control for multiple testing, Bonferroni corrections were applied for each modality (six for fMRI and 56 for MEG).

Nonparametric categorical analyses of ICs related to symptom counts were further conducted by classifying the adults as having persistent or remitted ADHD, and drawing contrasts against the never-affected control group.

SI Results

We repeated the statistical analysis after removing participants taking medication. All results shown in the main text held in this reduced sample (Table S2).