AI Finds Three Brain Biotypes Hidden Beneath Childhood ADHD and Anxiety

Attention-deficit/hyperactivity disorder and anxiety disorders are among the most common psychiatric conditions of childhood, and they frequently arrive together in the same child. Yet clinicians have long known that two children sharing the same diagnosis can look strikingly different in the clinic and respond differently to treatment. A new study published in Nature Mental Health tackles that heterogeneity head-on, using a deep learning framework to show that the brain activity of thousands of nine- to eleven-year-olds organizes itself into reproducible brain–behavior dimensions that cut across diagnostic categories. From those dimensions, the researchers derived three biologically grounded biotypes — externalizing–disinhibited, internalizing–distressed and mixed dysregulation — each with a distinct symptom profile, network organization and pattern of brain development. The work, led by Yong Jiao and Jing Jiang of the University of Iowa together with Yu Zhang, Xiaoyu Tong and Kilian M. Pohl of Stanford University, with Gregory A. Fonzo of the University of Texas at Austin, Ian H. Gotlib of Stanford and Theodore D. Satterthwaite of the University of Pennsylvania, suggests that the messy overlap of childhood ADHD and anxiety may conceal an orderly structure that deep learning can finally reveal.

The evidence base is unusually large for pediatric neuroimaging. The team analyzed resting-state functional MRI data from 3,508 children aged 9 to 11 years enrolled in the Adolescent Brain Cognitive Development (ABCD) Study, the largest long-term study of brain development in the United States, drawing on Release 5.1 of the ABCD data and the ABCD-BIDS Community Collection. Functional connectivity — the statistical correlation of spontaneous activity between brain regions while a child simply rests in the scanner — served as the primary brain signal. The researchers parcellated each child’s brain into network nodes using established whole-brain parcellation schemes, including the Shen 268-node atlas and cortical parcellations in the Yeo tradition, then computed connectivity between every pair of regions. Crucially, they did not treat the brain data in isolation. The framework integrated connectivity with clinically salient cognitive and behavioral measures, including NIH Toolbox cognition assessments, Wechsler intelligence testing and Achenbach child-behavior checklists, so that the dimensions the model discovered were anchored simultaneously in neural statistics and in real clinical meaning.

Technically, the pipeline is called DeCoDE, short for deep contrastive decoding of brain–behavior dimensions, and it is publicly available on GitHub. At its core sits a contrastive learning strategy in the spirit of contrastive principal component analysis and the contrastive variational autoencoder: the model is trained to isolate the features that are enriched in the clinical population relative to a background distribution, rather than merely capturing the dominant variance shared by all children. This matters because the most salient differences between children with ADHD, anxiety or both may be subtle relative to the enormous common signal of typical development. The model then links the salient functional connectivity features to the behavioral measures through a deep generalized canonical correlation analysis, a neural extension of Hotelling’s classical canonical correlation that finds linear combinations of two sets of variables maximally correlated with each other. Hyperparameter optimization was handled with Optuna, training ran on PyTorch with Lightning, and the statistical follow-up used SciPy, scikit-learn and statsmodels, with longitudinal mixed-effects modeling in R. The result is a framework that is deep enough to learn nonlinear structure but constrained enough to remain interpretable — the researchers can trace each discovered dimension back to specific connectivity patterns and specific behavioral scales.

What emerged from the analysis were two reproducible canonical dimensions linking brain and behavior. The first reflects cognitive control: it is carried by connectivity within and between the frontoparietal control network, the cingulo-opercular system and midcingulate regions that neuroscientists have long associated with task progress monitoring, integrative control dynamics and the regulation of impulses. The second reflects emotion–attention regulation: it involves the salience network anchored in the insula and anterior cingulate, the dorsal and ventral attention systems, and circuitry of the mood-disorder neurocircuitry tradition involving amygdala-linked prefrontal and limbic connections. Children differ along each dimension continuously, and the two dimensions together explain how symptoms of inattention, hyperactivity, worry and distress co-organize in individual brains. Rather than asking which diagnosis a child has, the framework asks where the child sits on two brain-anchored axes — an approach that echoes earlier canonical connectivity-dimension work in depression and in general population samples but pushes it into a deep learning regime with far larger data.

The two dimensions then yielded three biotypes. The externalizing–disinhibited biotype is dominated by low cognitive-control dimension scores and carries profiles of impulsivity, inattention and hyperactivity. The internalizing–distressed biotype is dominated by the emotion–attention regulation dimension and carries anxious arousal, worry and internalizing distress. The mixed dysregulation biotype combines abnormalities on both dimensions and carries the heaviest symptom burden. These are not arbitrary clusters: the biotypes showed distinct functional network organization, and — critically — distinct structural neurobiology. Longitudinal mixed-effects models of gray matter trajectories revealed that the biotypes differ in how their brains develop, with structural volumetric differences consistent with altered pace of brain maturation, a phenomenon that environmental and clinical influences on developmental pace have been shown to produce. The biotypes also mapped onto canonical cortical network organization, connecting the discovered dimensions to the intrinsic functional connectivity architecture of the human cerebral cortex estimated in the Yeo parcellation tradition.

Reproducibility was tested in two ways. First, the biotypes generalized to 224 age-matched participants from the Healthy Brain Network (HBN), an independent open transdiagnostic resource for pediatric mental health and learning disorders run by the Child Mind Institute in New York. The same three biotypes, derived from the same canonical dimensions, reappeared in a cohort collected at a different site with different scanners — a demanding test given the notorious site effects of multi-site MRI, which the team addressed with harmonization procedures in the ComBat tradition developed for multi-site diffusion tensor imaging. Second, and perhaps most striking, the biotypes were detectable before symptom onset. Children who later developed full clinical profiles already showed the characteristic brain–behavior dimension signatures at baseline, indicating preclinical expression of the biotypes. This is the pediatric analogue of the biomarker logic that has transformed adult depression research, where resting-state connectivity subtypes were first defined in Nature Medicine in 2017 and where frontostriatal salience network expansion was recently linked to depression in Nature in 2024.

The preclinical finding carries the study’s most consequential implication: early risk stratification. If a child’s position on the two brain–behavior dimensions can be measured at age nine or ten, before symptoms crystallize, clinicians may eventually be able to identify which children are on a disinhibited, distressed or mixed trajectory and intervene accordingly. The intervention literature already suggests that matching matters. Cognitive behavioral therapy is the best-supported treatment for pediatric anxiety, while organizational-skills remediation and behavioral parent training target the executive deficits of ADHD. A mixed-dysregulation child may need a different combination than a purely disinhibited or purely distressed child, and mechanism-informed intervention — choosing treatment by circuit profile rather than by checklist diagnosis — is exactly what the biotype framework is designed to motivate. The authors are careful to note that the study is observational and that the biotypes are research tools, not yet clinical instruments; but the pathway from dimension to stratification to matched intervention is now visible.

The study also speaks to a broader movement in computational psychiatry. Earlier work by members of this team identified psychiatric disorder subtypes from functional connectivity in resting-state EEG, decoded anxiety–impulsivity subtypes in preadolescent internalizing disorders from ABCD data, and revealed the structure of neuroanatomical variation within autism using contrastive machine learning in Science in 2022. The present study extends that program from EEG to fMRI, from single disorders to transdiagnostic populations, and from subtype discovery to longitudinal developmental validation. It joins a lineage of linked psychopathology-connectivity dimension work published in Nature Communications in 2018 and contrastive functional connectivity dimension work in depression published in Cell Reports Medicine in 2025. What distinguishes the new contribution is the combination of scale — 3,508 ABCD children plus 224 HBN participants — with the contrastive deep learning design that keeps the discovered dimensions interpretable and clinically anchored rather than black-box.

Limitations remain. The biotypes were derived from cross-sectional baseline data with longitudinal follow-up of brain structure, not from longitudinal prediction of symptom onset in a fully prospective design. The ABCD and HBN cohorts are controlled-access resources, and participant-level data cannot be redistributed, which means independent replication requires data-use agreements with the NIH Brain Development Cohorts Data Hub and the Healthy Brain Network Data Portal. The framework’s generalization beyond nine- to eleven-year-olds, beyond North American cohorts, and beyond ADHD and anxiety remains untested. And the clinical utility of biotype-based stratification — whether it actually improves outcomes compared with diagnosis-based treatment selection — is a question for future randomized studies. Still, the study’s message is clear and its data are open: the DeCoDE code is on GitHub, the demonstration pipeline is documented, and the statistical machinery is standard. The barrier to replication is now institutional rather than technical.

For a field that has spent decades arguing about where ADHD ends and anxiety begins, the study offers a reframing that may prove viral in its simplicity: the boundaries that matter may not be diagnostic at all. Two brain-anchored dimensions — cognitive control and emotion–attention regulation — suffice to carve the heterogeneous population of children with ADHD and anxiety into three biologically coherent groups that replicate across cohorts, differ in brain development and are visible before symptoms fully emerge. If the framework holds up, the future of pediatric mental health diagnosis may look less like a checklist of disorders and more like a brain-based map of trajectories — one on which every child’s position can be plotted, tracked over development, and used to choose the right intervention at the right time. The study, published in Nature Mental Health on 16 October 2026, was supported by the National Institutes of Health, the Eagles Autism Foundation and the Stanford Knight Initiative for Brain Resilience, and it was peer reviewed by Haiping Lu, Pew-Thian Yap and anonymous reviewers.

Subject of Research: Deep learning of brain–behavior dimensions identifies transdiagnostic biotypes in youth with ADHD and anxiety disorders

Article Title: Deep learning of brain–behavior dimensions identifies transdiagnostic biotypes in youth with ADHD and anxiety disorders

Article References: Jiao, Y., Tong, X., Fonzo, G. A., Gotlib, I. H., Pohl, K. M., Satterthwaite, T. D., Jiang, J., & Zhang, Y. (2026). Deep learning of brain–behavior dimensions identifies transdiagnostic biotypes in youth with ADHD and anxiety disorders. Nature Mental Health. https://doi.org/10.1038/s44220-026-00740-3

Image Credits: AI Generated

DOI: 10.1038/s44220-026-00740-3

Keywords: deep learning, ADHD, anxiety disorders, functional connectivity, biotypes, ABCD study, youth mental health, transdiagnostic psychiatry, Deep, learning, brain, behavior