Children with ADHD who take stimulant medication often have brain-scan measures that look like those of healthy children. Unmedicated children look different. This is usually read as the drug repairing the brain. That reading has two weaknesses. First, the data are a single snapshot of separate groups, so the medicated children may simply be the ones who were less impaired to begin with, or who were prescribed differently. Second, the brain measure itself is built from correlations between regional fMRI signals, and drugs can change those correlations without changing how regions communicate. A drug that raises or lowers the strength of one shared input reaching many regions at once will shift dozens of correlations together. A standard analysis then counts that one change as many separate "connections" getting fixed.
This proposal puts the two weaknesses into one test. We model resting fMRI covariance as the sum of three parts: region-specific noise, a small number of shared inputs that each spread over a source region's structural targets as seen in diffusion MRI, and leftover pairwise coupling. We then ask two questions. Which of these three parts carries the medicated-versus-unmedicated difference? And does that difference appear within the same child once they start medication, or is it already present before treatment? The idea of shared inputs comes from single-neuron anatomy showing that many deep-layer cortical neurons "broadcast," meaning one neuron sends branches to several distant targets at once. The causal leverage comes from prospective cohort data in which some children start stimulants between scans.
Core Claim
The apparent stimulant normalization of dynamic connectivity reported by Altered neural flexibility in children with attention-deficit/hyperactivity disorder has three possible sources. That study found unmedicated children had lower whole-brain "neural flexibility," the rate at which regions switch network communities over time ($p = 0.0005$), while medicated children were indistinguishable from controls ($p = 0.74$). The authors themselves note they cannot tell drug effect from prescribing selection. Our claim is that the normalization can be divided, with uncertainty, into three shares:
- a change in shared, one-to-many input variance, which is the expected signature of a drug acting on widely projecting systems;
- non-specific amplitude changes;
- change in specific pairwise coupling.
We further claim that pre-treatment scans of children who later start medication can tell drug effect from selection. Neither division exists in the literature.
The structural ingredient is a one-to-many "fan" built from diffusion MRI. It is tested, not assumed. If fans drawn from the real wiring explain covariance no better than random fans with the same number of targets, the broadcasting component is dropped and the study reduces to a variance-plus-causality analysis.
Background and Positioning
Three findings motivate the shared-input model.
- Neurons that broadcast. Single-axon level morphological analysis of corticofugal projection neurons in mouse barrel field reconstructed single axons and found that 43% of one class of layer-5 output neurons project to two or more subcortical targets, against 10% of layer-6 neurons. Upper brainstem cholinergic neurons send single axons to many nuclei (Zhao et al.).
- Shared signals corrupt correlations. Disambiguating brain functional connectivity showed that much of the change in correlation between rest and task can be explained by additive signals that change variance rather than neural coupling.
- Drugs move imaging measures. Medication shifts imaging and physiological markers in several settings. It biases resting EEG connectivity (Choi et al.), makes prefrontal connectivity findings in schizophrenia variable across studies (Aberrant cortical–subcortical-cerebellar connectivity…), and can mimic pathology in PET imaging (Pascoal et al.).
Combining these findings gives a concrete mechanism by which a drug could "normalize" a correlation-based measure without altering pairwise communication.
Two existing methods supply parts of our approach.
- Joint structure–function modeling. Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome couples each structural node to its own functional counterpart with learned weights. It shows that this coupling varies across regions. We borrow the idea that diffusion structure should constrain functional covariance. We differ in using one-to-many fans rather than one-to-one links, and a decomposable covariance model rather than a predictive network.
- Quantifying confounding. Quantifying Confounding Bias in Neuroimaging Datasets with Causal Inference diagnoses confounding with a one-dimensional hidden factor. It also shows that pooled datasets carry strong site fingerprints: 73.3% accuracy at identifying the source dataset, against 6.7% chance. This constrains us to treat site and scanner as first-class confounders.
Causal approaches based on genetic instruments constrain the design in a different way. Mendelian randomization, which uses inherited genetic variants as natural experiments, resolved direction for depression and white matter (Shen et al.) and for the structural connectome and psychiatric risk (Xiao et al.). It cannot instrument a prescription decision, so we rely on timing within individuals instead.
The ADHD biotype work (Functional imaging derived ADHD biotypes…) shows that medication response varies by connectivity subtype. Its graph has patients as nodes, not brain regions, and its medication subsample has only 44 patients. It therefore motivates heterogeneity analyses but does not address our question.
Research Program
The model writes each subject's regional time-series covariance as
$$\Sigma = D + \sum_{s=0}^{S} \lambda_s\, w_s w_s^{\top} + R,$$
where the symbols are:
- $D$ is a diagonal matrix of region-specific variance;
- $w_s$ is a nonnegative loading vector for shared source $s$, allowed to be nonzero only on regions structurally connected to $s$ in the diffusion connectome;
- $w_0$ is a uniform global loading that absorbs whole-brain hemodynamic and arousal signals;
- $\lambda_s \ge 0$ is the strength of source $s$;
- $R$ is a sparse matrix of residual pairwise coupling.
Fitting proceeds in three steps:
- Fan supports come from each subject's own diffusion connectome where one is available, and otherwise from an age-matched normative connectome.
- Loadings are shared across subjects within a cohort. The parameters $\lambda_s$, $D$ and $R$ vary per subject.
- Estimation is by penalized Gaussian likelihood, with an $\ell_1$ penalty on $R$ and the number of sources $S$ chosen by held-out likelihood.
Before any clinical data are touched, we check that the fitting procedure recovers known broadcast sources and known drug-like changes in $\lambda$. We do this on data simulated from a network model in which those quantities are set by design.
The quantity of interest is a counterfactual flexibility score. We recompute neural flexibility, using the multilayer community method of Yin et al., on time series in which one component at a time is held at the control-group distribution. The share of the medicated–unmedicated gap that closes when a component is held fixed is that component's attribution. Confidence intervals come from bootstrapping over subjects within site.
The causal step uses prospective data. In the ABCD cohort, children with ADHD who begin stimulants between two waves are compared with ADHD children who do not, matched on prior symptoms and site.
- Pre-treatment baseline scans test selection directly. If future initiators already show higher flexibility before any drug, selection explains at least part of the cross-sectional normalization.
- A difference-in-differences contrast on each component estimates the within-person change attributable to medication.
Diffusion structure serves as a negative-control outcome. A largely acute pharmacological effect should move the $\lambda_s$ terms and $D$ but not the structural fans. Group differences in fan-relevant structure at baseline point to selection.
Two dependencies are not documented in the research file and must be verified before committing: that ABCD's medication inventory records stimulant use and the dose taken on the scan day, and that its diffusion data are usable for these subjects.
Baselines
Naive cross-sectional comparison. The do-nothing reference is the Yin et al. analysis itself: medicated against unmedicated against control flexibility in ADHD-200, using their pipeline and their 360-child sample definition. Their reported figures are $p = 0.0005$ for the unmedicated deficit, $p = 0.74$ for medicated against controls, and 77% cross-validated and 74.46% independent-test accuracy for regional flexibility classifying ADHD against controls. Reproducing these is a prerequisite. Our method "beats" this baseline not by higher accuracy but by moving the inferred normalization by an amount outside the bootstrap interval once components are separated or pre-treatment selection is accounted for.
Variance and global-signal correction. This is the strongest simple competitor for the pharmacological question. It applies global-signal regression plus per-region variance normalization, in the spirit of the additive-signal analysis in Disambiguating brain functional connectivity, before computing flexibility. It is exactly our model with $S = 0$, so it is run on identical preprocessed data with the same tuning budget. The comparison metric is held-out covariance log-likelihood on withheld scan segments. That paper reports no exact percentage of connectivity change explained, and no reported result was found on Lacuna for held-out covariance likelihood of structurally constrained factor models.
Random-fan ablation. This isolates the broadcasting ingredient. Each real fan is replaced by a random fan with the same number of targets, preserving each region's degree. The procedure is repeated 1,000 times. Structural fans must exceed the 95th percentile of random-fan held-out likelihood, and must give a different attribution of the medication gap, to justify themselves.
Joint structure–function network. Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome represents the best structure–function fusion in the file. It reports 84.9% accuracy and AUC 0.907 for sex classification on NCANDA, against 83.1% for the best multi-view fusion baseline. We retrain it on NCANDA with the same parcellation, both to confirm our implementation and to check that our decomposition loses little predictive information. The bar is our components predicting sex within two percentage points of it while also being interpretable.
Propensity-score adjustment. For the causal question, we adjust for symptoms and site with a propensity score: a model of each child's probability of being medicated given those measured variables. This is the standard observational competitor. Our pre-treatment and within-person estimates beat it if they diverge from it in a direction that baseline selection explains.
Why not just…?
Why not just regress out global signal and normalize variance? This is the real threat, and the design treats it as the null model. If held-out likelihood and the attribution of the medication gap are statistically unchanged when structural fans are added beyond $w_0$ and $D$, we conclude that broadcasting structure adds nothing at this resolution and report that. Global correction cannot say whether the residual change is spread across specific source territories, such as fans from catecholaminergic or deep cortical sources. That is the distinction between a drug acting on widely projecting systems and a non-specific hemodynamic shift.
Why not just add medication as a covariate, or match on symptoms? A covariate or propensity score adjusts only for what is measured. The confounding by indication that Yin et al. worry about runs through clinician judgment, which is unmeasured. Quantifying Confounding Bias in Neuroimaging Datasets with Causal Inference shows that brain measures retain strong hidden group fingerprints that demographic covariates do not capture: 42.2% dataset identification from demographics alone against 73.3% from morphometry. Comparing pre-treatment scans of future initiators bypasses that problem, and the propensity baseline is run alongside so the disagreement is visible.
Why not just use Mendelian randomization, as for depression and white matter? No genetic variant instruments the decision to prescribe. A genetic score for ADHD liability would estimate the effect of the disorder, not of the drug. That is precisely the quantity to be separated from the drug, so within-person timing is the appropriate source of leverage here.
Why not just feed medication into a large pretrained model? Brain Harmony, a foundation model aligning fMRI with structural geometry, reports 64.65% accuracy on ADNI and 66.67% on ABIDE-II. A better predictor does not answer which part of the signal a drug moves or whether medicated children differed beforehand. We use such models only as a check that our decomposition retains predictive information.
Decisive Experiments
Simulation. Using simulated networks with planted broadcast sources and planted drug-induced changes in $\lambda$, the method must recover source supports with high overlap with the planted sources. It must also attribute a planted drug effect to $\lambda$ rather than $R$ at the scan lengths and sample sizes of ADHD-200 and ABCD. Failure here halts the clinical analyses.
NCANDA validation of fans. NCANDA provides paired diffusion and resting fMRI from 662 youths across 1,976 longitudinal pairs. It tests whether subject-specific fans beat degree-matched random fans on held-out covariance likelihood in a non-clinical sample. It also tests whether fitted $\lambda$ values are more stable within a person across waves than between people.
ADHD-200 attribution. We reproduce the Yin et al. normalization and attribute the medicated–unmedicated gap to the noise, global, fan and pairwise terms. A finding that most of the gap closes when the fan and global terms are held fixed while $R$ does not differ supports a non-specific pharmacological reading. A finding that the gap lives in $R$ supports specific circuit change.
ABCD timing test. This is the decisive causal check, in two parts. First, the pre-treatment flexibility of future stimulant initiators is compared with that of never-initiators with ADHD. Second, a difference-in-differences contrast on each component is estimated, with diffusion structure as the negative control.
The outcomes are read as follows:
- Selection. Normalization present before treatment and no within-person change would show the cross-sectional result reflects selection.
- Drug effect via shared input. A within-person rise confined to fan and global terms would show a real drug effect on shared input rather than on pairwise coupling.
- Specific circuit change. A within-person change in $R$ would be the strongest evidence for specific circuit change.
Each outcome is publishable and changes how the cross-sectional normalization result should be interpreted.
References
- Altered neural flexibility in children with attention-deficit/hyperactivity disorder
- Single-axon level morphological analysis of corticofugal projection neurons in mouse barrel field
- Cardiorespiratory Coupling Analysis Based on Entropy and Cross-Entropy in Distinguishing Different Depression Stages
- Disambiguating brain functional connectivity
- Electroencephalographic Resting-State Functional Connectivity of Benign Epilepsy with Centrotemporal Spikes
- Aberrant cortical–subcortical-cerebellar connectivity in resting-state fMRI as an imaging marker of schizophrenia and psychosis: a systematic review of data-driven whole-brain functional connectivity analyses
- Monoamine oxidase B inhibitor, selegiline, reduces 18F-THK5351 uptake in the human brain
- Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome
- Quantifying Confounding Bias in Neuroimaging Datasets with Causal Inference
- A phenome-wide association and Mendelian Randomisation study of polygenic risk for depression in UK Biobank
- Mendelian randomization analyses uncover causal relationships between brain structural connectome and risk of psychiatric disorders
- Functional imaging derived ADHD biotypes based on deep clustering: a study on personalized medication therapy guidance
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens