Abstract
Adults with childhood-onset attention-deficit hyperactivity disorder (ADHD) show altered whole-brain connectivity. However, the relationship between structural and functional brain abnormalities, the implications for the development of life-long debilitating symptoms, and the underlying mechanisms remain uncharted. We recruited a unique sample of 80 medication-naive adults with a clinical diagnosis of childhood-onset ADHD without psychiatric comorbidities, and 123 age-, sex-, and intelligence-matched healthy controls. Structural and functional connectivity matrices were derived from diffusion spectrum imaging and multi-echo resting-state functional MRI data. Hub, feeder, and local connections were defined using diffusion data. Individual-level measures of structural connectivity and structure-function coupling were used to contrast groups and link behavior to brain abnormalities. Computational modeling was used to test possible neural mechanisms underpinning observed group differences in the structure-function coupling. Structural connectivity did not significantly differ between groups but, relative to controls, ADHD showed a reduction in structure-function coupling in feeder connections linking hubs with peripheral regions. This abnormality involved connections linking fronto-parietal control systems with sensory networks. Crucially, lower structure-function coupling was associated with higher ADHD symptoms. Results from our computational model further suggest that the observed structure-function decoupling in ADHD is driven by heterogeneity in neural noise variability across brain regions. By highlighting a neural cause of a clinically meaningful breakdown in the structure-function relationship, our work provides novel information on the nature of chronic ADHD. The current results encourage future work assessing the genetic and neurobiological underpinnings of neural noise in ADHD, particularly in brain regions encompassed by fronto-parietal systems.
Introduction
Adult attention-deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder characterized by inattentive and hyperactive-impulsive symptoms beginning in early childhood1. Identifying the neural underpinnings of adult ADHD is an ongoing research endeavor, critical to the definition of neural mechanisms supporting clinical outcomes of childhood-onset ADHD and the development of novel targeted interventions2.
Neuroimaging work has provided important insights into altered structural3–5 and functional6,7 brain connectivity underpinning ADHD pathophysiology, and suggest that network interactions, rather than regional abnormalities, contribute to phenotypic expression of the disorder8. Anatomically, results have been mixed. Recent studies showing no changes in the ADHD connectome9, whereas others have pointed to various abnormalities in white matter tracts including the corpus callosum and posterior circuits related to the limbic and occipital systems, the fronto-striato-cerebellar connections, and the pathways linking the default-mode and fronto-parietal hub regions4,5,10.
Complementing findings from diffusion MRI, resting-state functional magnetic resonance image (rs-fMRI) studies have highlighted that both diagnosis and symptoms of ADHD are linked to reduced segregation between the activity of control networks supporting external task engagement and the default-mode network6,7,11. Reduced functional connectivity within, and between, the default-mode, sensory, and control networks has also been reported both in children and adults with ADHD6,7,10,11.
Emerging evidence suggests that patterns of functional connectivity are constrained by their anatomical underpinning: The connectome12,13. Structural and functional brain network alterations in adult ADHD partially overlap, but the direct link between these structure-function aberrations has not been formally explored. Here, we used multi-echo rs-fMRI and diffusion spectrum imaging (DSI) to investigate possible changes in whole-brain structure-function coupling in a large sample of well-characterized, medication-naïve adults with childhood-onset ADHD and matched healthy controls11. Based on previous findings11 and the hypothesis that psychiatric conditions are primarily pathologies of brain hubs14, we expected significant departures from the typical structure-function coupling in ADHD. Specifically, a breakdown in the structure-function association is likely to occur in connections involving brain hubs that belong to the control and default-mode brain networks14,15. To investigate a likely underlying neural mechanism of this deficit16, we adopted whole-brain computational modeling. Our model explicitly tested the hypothesis that increased heteroscedasticity in the levels of intrinsic neural noise drives the expected breakdown in the structure-function coupling. Heteroscedasticity occurs when the variance of explanatory variables – neural noise level – is not constant across brain regions. The hypothesis tested by our model is grounded in previous work suggesting that ADHD symptoms are linked to a pathological increase in baseline neural noise17–20.
Methods
Sample
We recruited 80 medication-nai□ve adults with childhood-onset ADHD aged 18–39 years (mean 26.7 years), who fulfilled DSM-IV-TR criteria for the current diagnosis of ADHD. This carefully phenotyped sample allows the unequivocal assessment of structural and functional brain networks in the absence of common confounds in ADHD research including other developmental delays, medication exposure and intellectual disabilities. Results from the clinical sample were benchmarked against the findings of 123 age- (mean 25.7 years), sex-, and IQ-matched healthy controls. Participants were assessed at the Department of Psychiatry, National Taiwan University Hospital (NTUH), Taipei, Taiwan.
This study has been approved by the Research Ethics Committee of NTUH (201401024RINC; ClinicalTrials.gov number: NCT02642068) and all participants provided written informed consent. Details regarding the recruitment procedure are described in our previous work11 (Supplementary Methods).
MRI acquisition and preprocessing
Brain imaging data were acquired with a Siemens 3T Tim Trio scanner equipped with a 32-channel head coil. Details regarding preprocessing multi-echo resting-state data are described elsewhere11 (Supplementary Methods). In short, the pipeline included: quality control, comprehensive data denoising using multi-echo independent components analysis (ME-ICA v3.0)21, coregistration to individual anatomical images, non-linear normalization to MNI space, and filtering (0.01∼0.1 Hz).
DSI data underwent an initial quality control procedure to ensure acceptable levels of in-scanner head motion, which were estimated by signal loss22. This quality control step resulted in a final sample of 78 ADHD adults and 118 healthy controls (Table 1). DSI data were then reconstructed using the q-space diffeomorphic reconstruction approach23 implemented in the software DSI Studio (Supplementary Methods).
Further quality control analyses showed that micro-head movements (mean framewise displacement)24 for rs-fMRI and signal dropout counts22 for DSI, were not significantly different between ADHD and controls (p = 0.35 and 0.54, respectively).
Structural and functional brain network construction
We generated whole-brain structural (SC) and functional (FC) connectivity matrices for each individual, based on a common and recently validated cortical parcellation25 (Fig. 1A). Fourteen additional subcortical structures from the Harvard-Oxford atlas were added to the parcellation, resulting in 214 total regions (Schaefer-214 henceforth; Supplementary Table 1). Individual whole-brain tractography maps were combined with the pre-defined anatomical boundaries defined by this Schaefer-214 parcellation to generate a weighted SC matrix (Fig. 1B). Each edge of the network corresponds to the total number of normalized streamlines that interconnect any two brain regions, adjusted for the interregional fiber length26. For resting-state data, regional time-series were calculated as the mean across voxels within each region included in the brain parcellation. For each individual, Pearson’s correlations were calculated between the time-series of all regions to calculate FC. Finally, a Fisher z-transformation was applied to the FC matrices.
Connection classes
We identified hub regions according to an aggregate ranking across multiple metrics including degree, strength, subgraph centrality, and betweenness27,28. The top 15% composite scores (N = 32, Supplementary Table 1&2) were used to identify hub regions within each individual; all other nodes were assigned as periphery nodes. Hub connections were defined as edges that connected any two hub nodes. Feeder connections linked hub nodes to periphery nodes, and local connections linked periphery nodes (Fig. 1C)15,29.
Structure-function relationships
Brain network structure-function relationships were conducted in line with previous research15. First, non-zero SC values within each individual connectome were isolated and normalized using a rank-based inverse Gaussian transformation30. The resulting SC values were correlated with corresponding FC values (i.e., the same edges), within each individual. This analysis produced a single Pearson’s r value that summarized the global structure-function association for each individual31. These values were used to populate group distributions and were subsequently contrasted using between-group statistics. This entire procedure was completed at the level of the whole network and within each respective connection class: hubs, feeders, and local edges.
Previous work investigating resting-state networks, including data from the current cohort11, has highlighted the key role of control, default-mode, and sensory networks in adult ADHD6,7. Based on these results, we also tested for specific changes in SC-FC coupling within these networks. To ensure that a sufficient number (minimum of 50) of edges was used to infer structure-function relationship, control networks were defined as the combination of fronto-parietal, alongside dorsal and ventral attention affiliations from the adopted parcellation, while sensory connections included both visual and somatomotor affiliations. Default-mode connections were as in the original parcellation. Once SC-FC coupling was estimated within each network, the mean r values (Control-ADHD) were presented within and between each network.
Relationship between structure-function coupling and behavioral symptoms of ADHD
Given the notion that measures of ADHD symptoms are continuously distributed in the general population32,33, we investigated brain-behavior relationships across both ADHD and control groups (Fig. 1C). Inattention and hyperactivity-impulsivity symptoms based on the parent-rated Swanson, Nolan, and Pelham, IV (SNAP-IV)34 and self-rated Adult ADHD Self-Report Scale (ASRS)35 (Table 1) were used in the analysis. These four symptom items (two from each measure) were transformed using a rank-based inverse Gaussian, then entered into a principal component analysis to reduce the dimensionality of the data. The first component, accounting for 81% of the variance, was then correlated with the structure-function coupling of the whole sample (Supplementary Table 3).
Statistical comparisons between groups
To ensure that the general structural network density did not explain between-group differences, summed binary and weighted degrees were compared between groups. Average connection weights within each connection class were compared between each group. In addition, the network based statistic (NBS)36 was used to explore any possible differences in SC between controls and ADHD (5000 permutations, threshold t = 3). ADHD-associated alterations of FC using NBS have been reported in our initial study on this sample11.
Mann–Whitney U tests were used to identify possible differences in the structure-function association between control and ADHD groups. Bonferroni correction (family-wise error rate, FWE) for multiple comparisons was applied to follow-up statistics, with αFWE < 0.05 indicating statistical significance. Statistical analyses were performed in MATLAB (Mathworks) with code available online (https://github.com/ljhearne/ADHDSCFC).
Computational modeling: Assessing the neural factors driving structure-function breakdown
The adopted whole-brain computational model incorporates SC to represent the strength of connections between brain regions. In addition to the weights specified in the empirical SC matrix, structural connections are scaled by a global coupling parameter. This parameter can then be varied systematically to simulate and compare the global dynamics emerging from the model with the empirical FC derived from the rs-fMRI data.
We chose a simple stochastic linear model of the Ornstein-Uhlenbeck type37–39. The main motivations behind this choice were that the model: (i) allows us to simulate whole-brain patterns of FC from SC matrices; (ii) enables tests of the hypothesis that increased heteroscedasticity of neural noise levels results in a breakdown in structure-function coupling; (iii) can be considered a generic linearization of more complex models with a stable fixed point (a mathematical approach at the core of e.g. dynamic causal modeling for fMRI40); and (iv) permits a direct analytical derivation of FC from empirical SC without the need of computationally demanding numerical simulations. The model equation is: where xi is the activity of the i-th region; c is the global coupling strength which rescales the strength of structural connections of the system; Wij is the connectivity weight to region i from region j (as specified by the empirical SC matrix); σi is the intrinsic noise amplitude/level of the i-th region, and defines the size of random increments σidWi in the dynamics of the region, and N is the total number of regions in the connectome. Previous modeling studies38,39 have considered the noise levels to be constant across the whole network (i.e., the homoscedastic case in which all σi are identical). In light of previous suggestions17–20, we hypothesized that heteroscedasticity across a specific subset of brain regions (hubs or periphery) would have a detrimental impact on SC-FC decoupling. To test our hypothesis and determine in which connection classes heteroscedasticity has the largest impact, we systematically analyzed varying degrees of heteroscedasticity in the noise levels in distinct subsets of regions independently (hub and periphery regions). A comprehensive description of the modeling can be found in the Supplementary Methods.
Results
Similar structural connectivity between groups
Results showed no difference in weighted (p = 0.89, z = 0.13), or unweighted (p = 0.24, z = −1.19) summed degree across groups. Likewise, the whole-brain network-based statistics comparing ADHD and healthy control groups revealed no significant differences in structural connectivity between the groups (ADHD > controls, p = 0.63; controls > ADHD, p = 0.78). Next, we sought to investigate potential differences in classes of structural connections, namely hubs, feeders, and local connections. No significant group differences were observed when comparing mean connection strength within hub (p = 0.86, z = −0.17), feeder (p = 0.77, z = −0.29), or local connections (p = 0.23, z = 1.21).
Structure and function coupling in ADHD is reduced in feeder connections
When considering all edges within the network, results indicated a significant difference in SC-FC coupling (p = 0.01, z = 2.51, Fig. 2A). We then assessed the contribution to this effect of each connection class (hub, feeder or local). Results showed that compared to controls, ADHD had a significantly lower SC-FC association in feeder connections (pFWE = 0.005, z = 3.10) but not in hub (pFWE = 1, z = 0.55) or local (pFWE = 0.33, z = 1.60) connections (Fig. 2A).
Feeder structure-function decoupling in control, default-mode, and sensory brain networks
To further explore the anatomical specificity of the observed deficits in structure-function coupling, we isolated feeder connections that belonged to control, default-mode, or sensory (merging somatomotor and visual) networks. As per the previous analysis, we correlated SC and FC values for connections within and between the selected brain networks. This resulted in a three-by-three matrix for both ADHD and healthy control groups that represented the degree of SC-FC coupling within and between control, default mode, and sensory networks. The largest reduction in SC-FC associations in ADHD compared to healthy controls were located in connections between control and sensory networks (Fig. 2B).
The magnitude of structure-function decoupling correlates with the severity of ADHD symptoms
Individual symptom scores captured by PCA linearly correlated with indices of structure-function coupling in feeder connections, such that lower structure-function coupling was associated with more severe ADHD symptoms (p = 0.0004, r = −0.25, Fig. 2C).
Control analyses
A number of tests were conducted to establish the reliability of our findings. To ensure that our chosen brain parcellation had little bearing on the results41, we repeated the analyses in two other, independent brain parcellations: Shen-21342 and Brainnetome-24443. The reported effects were all successfully replicated (Supplementary Table 2). Using these alternative brain parcellations, we also found that adults with ADHD exhibited weaker structure-function coupling in hub connections. However, the effect size of these between-group differences was consistently smaller than the effect in feeder connections.
Noise in hubs and periphery as a neural mechanism for structure-function breakdown
Finally, we sought a neural mechanism for how altered structure-function relationships could emerge in the absence of significant differences in the connectome. In particular, we aimed to use computational modeling to explain our finding of selective deficits in feeder connection SC-FC coupling (leaving hub and local connections relatively unscathed). We systematically explored two scenarios with noise heteroscedasticity – i.e., increased heterogeneity in the intrinsic neural noise levels σi across brain regions.
In the first scenario, we analyzed the simple case of heterogeneity between hubs and periphery (σH ≠ σP) for hub nodes (H) and peripheral regions (P), maintaining σH and σP constant within each class of regions. Exploring ranges of σH and σP (Fig. 3A-C) we analyzed the changes in SC-FC coupling for the three classes of connections (hub, feeder, and local). We found that feeder connections were the most susceptible to subtle imbalances between intrinsic noise levels in hub and periphery regions, reflected in the quick decrease in SC-FC coupling (Fig. 3B). On the contrary, hub and local connections exhibited only small changes (Fig. 3A&C). Specifically, a small imbalance such that σH < σP, with σP 10% larger than σH, produced a slight (< 2%) reduction in SC-FC coupling in hubs compared to the homogenous σH = σP case, similar to the empirically observed slight decrease for hub connections in Fig. 2A (< 2%). Conversely, a 10% imbalance in the opposite direction (σH > σP) yielded a negligible (~0.3%) increase in hub SC-FC coupling. The increased sensitivity of feeder connections was demonstrated by the same 10% imbalance (σH < σP) resulting in a 4% decrease in SC-FC coupling for feeder connections compared to the homogenous case. Importantly, an imbalance of approximately 50% (σH < σP) was required to obtain the 10% decrease in SC-FC coupling empirically observed in ADHD feeder connections (Fig. 2A). This larger imbalance also resulted in a < 2% reduced SC-FC coupling in hub connections, again in accordance with empirical results. Thus, larger differences between mean noise amplitude levels in hubs and periphery led to greater SC-FC decoupling specific to feeder connections, mirroring the selective deficits observed in ADHD.
In the second scenario, we modeled the more realistic case where the noise levels (σi) within hubs and periphery also varied from region to region. This allowed us to examine whether heteroscedasticity within hubs and/or periphery regions could contribute to the observed disruption of SC-FC coupling in ADHD. We systematically explored ranges of variance (Var[σH] and Var[σP)] for noise levels normally distributed around means (E[σH] and E[σP)],set here such that E[σP] is 10% larger than E[σH] in line with the above results for hub connections (comparing Fig. 3A to Fig. 2A). We found that connections within a region class (i.e., hub-hub or periphery-periphery) are resilient to increased variability of intrinsic noise levels in the opposite type. Indeed, the SC-FC coupling in hub connections (Fig. 3D) and local connections (Fig. 3F) remained almost constant for increased noise variability in peripheral and hub regions, respectively. However, feeder connections (Fig. 3E) are clearly susceptible to changes in noise level heterogeneity within either hub or periphery regions, which implies an increased sensitivity to heteroscedasticity could also contribute to the disruption of SC-FC coupling in ADHD.
Discussion
The present study provides evidence of a clinically significant breakdown in brain structure-function (SC-FC) coupling in medication-naive adults with childhood-onset ADHD. In line with the hypothesis that hub regions are critically vulnerable to brain pathology14,15,44, ADHD was associated with a marked SC-FC decoupling in connections linking brain hubs to peripheral regions (feeders) within and between control and sensory networks. Results from our modeling work further suggest that such decoupling could be linked to: (i) an imbalance in noise amplitudes in hubs and the periphery (e.g., increased ‘unreliability’ in signals originating from the periphery) and, (ii) higher peripheral heteroscedasticity (i.e., the peripheral noise is more diverse and more difficult for the hubs to filter out). Altogether, results from this work propose a novel neural mechanism explaining structure-function decoupling in brain connectivity underpinning the chronic manifestation of ADHD symptoms.
Structural networks are thought to place significant constraints on FC and local brain activity 12,16,31. The decoupling between FC and its structural basis is therefore thought to represent a key index of brain network pathology in psychiatric illnesses including schizophrenia15,45,46. Our results are in line with the general notion that a structure-function breakdown in psychiatric illnesses involves anatomically defined hub brain regions14. The observed association with behavior, indicating that reduced structure-function coupling in feeder connections is related to higher severity of ADHD symptomology, provides support for the clinical relevance of this deficit in ADHD. By using a parsimonious model explaining the emergence of functional connectivity from underlying anatomical connectivity, we found that increased heteroscedasticity in intrinsic noise levels, either in hubs or periphery, has a strong detrimental effect in feeder connections, and to a lesser extent in hub-hub connections. Physiologically, reduced SC-FC coupling due to increased neural noise heteroscedasticity in peripheral regions can be understood as brain hubs being unable to average out peripheral functional disruptions. This adds weight to the notion that ADHD symptoms may arise from increased neural noise in the resting-state activity of associative brain regions19,20. Our findings are also compatible with ADHD involving a deficit in catecholaminergic systems regulating neural signals (neural gain, 17), and methylphenidate-induced reductions in neural noise19,47.
Our empirical findings showed that feeder connections are the most affected by the decoupling between function and anatomy. Feeder connections comprise long-range anatomical routes allowing efficient communication between remote brain regions belonging to different brain networks 29. We here found that connections within control networks, as well as between regions comprising control and sensory networks, contributed to the overall reduction in structure-function association in ADHD. These findings are in agreement with previous neuroimaging studies in ADHD6,7,48,49 and healthy controls50,51, highlighting the key role of these connectivity patterns to support normal and pathological attention and inhibitory processes. We also note that altered patterns of FC, and SC-FC decoupling, can occur in the absence of deficits in SC45. In fact, whereas white matter connections are predictors of FC31, the opposite is not always true52.
The absence of significant group differences in the structural connectome is at odds with some previous reports3,4. Due to the sample size and the quality of the data, it is unlikely that the negative finding reported here is due to a lack of statistical power in detecting meaningful differences in the ADHD connectome. Moreover, our result is consistent with recent work showing the existence of FC abnormalities with preserved white matter proprieties in ADHD53. The discrepancy between our findings and earlier literature3 may be explained by non-neural factors. For example, the absence of significant differences between the ADHD and control connectomes reported here may reflect our emphasis on comparable levels of head motion between the two groups; a critical factor that has recently been shown to produce spurious group differences in ADHD3,54. Our cohort of medication-naive adults with established childhood-onset ADHD in the absence of co-occurring psychiatric conditions may also contribute to this negative finding, as psychostimulant exposure55 and comorbidity56 have been reported to affect SC in ADHD. Although our results cannot completely exclude the presence of altered white matter integrity in ADHD, they suggest that any such differences are small overall, and the manifestation of ADHD symptoms is underpinned by functional deregulations and related decoupling in SC-FC.
By combining functional and diffusion-weighted imaging with computational modeling, our study has advanced the understanding of neural mechanisms that underpin chronic ADHD symptoms. More specifically, our work showed that a clinically meaningful function-structure decoupling in ADHD is likely to be related to increased neural noise heterogeneity between hubs and periphery regions. This knowledge is consistent with the positive effect of current pharmacological interventions for ADHD and provides neurobiological support for future clinical research focusing on reducing periphery-to-hub noise amplitude ratio and peripheral noise heteroscedasticity using targeted interventions including brain stimulation.
Conflict of Interest
The authors declare no conflict of interest.
Acknowledgments
This work was supported by the Ministry of Technology and Science, Taiwan (MOST103-2314-B-002-021-MY3), the National Health Research Institutes, Taiwan (NHRI-EX103-10008PI), and National Taiwan University Hospital (NTUH103-S2458, NTUH104-S2761). L.C. and J.A.R. are supported by the Australian National Health Medical Research Council (L.C., 1099082 and 1138711; J.A.R., 1145168 and 1144936).
Footnotes
↵§ Shared first authorship