Summary: A new study shows that artificial intelligence can reliably estimate a child’s future risk of developing attention-deficit/hyperactivity disorder (ADHD) years before a clinical diagnosis is typically made. By mining subtle patterns in routine electronic health records (EHRs) collected from birth through early childhood, the AI detects combinations of developmental, behavioral, and clinical markers that may be missed during brief primary care visits.
Designed as a clinical safety net rather than a diagnostic tool, this AI-based approach aims to ensure that children at elevated risk receive earlier evaluation and timely support during critical periods of development.
Key Facts
- The dataset: Researchers studied the medical histories of over 140,000 children to build a robust comparative baseline of those with and without ADHD.
- Early detection: The AI model analyzes data from birth and becomes highly accurate at estimating future ADHD risk by age 5, well before the average age of diagnosis.
- Equitable performance: The model maintained consistent accuracy across sex, race, ethnicity, and insurance status, indicating potential to reduce disparities in ADHD identification and care.
- Support, not diagnosis: The tool is explicitly intended to flag children who should receive prioritized screening or referral; it is not intended to replace clinicians or establish a diagnosis.
- Improved outcomes: Earlier identification can enable evidence-based interventions sooner, which is linked to better academic, social, and long-term health outcomes for children with ADHD.
Source: Duke University
Attention-deficit/hyperactivity disorder affects millions of children, and many remain undiagnosed for years despite early signs that, if recognized, could prompt timely support.
In a recent study, researchers at Duke Health demonstrate that AI models trained on standard EHR data can estimate a child’s likelihood of receiving an ADHD diagnosis years in advance. By automatically scanning routine clinical records, the method can highlight children who may benefit from closer monitoring, earlier screening, or referral to specialists.
Published in Nature Mental Health on April 27, the study highlights how much predictive power already exists in information routinely recorded during health care visits. The researchers emphasize that these insights are intended to support clinical decision-making in primary care by identifying children who might otherwise be overlooked.
“We have an incredibly rich source of information sitting in electronic health records,” said Elliot Hill, lead author and data scientist in the Department of Biostatistics & Bioinformatics at Duke University School of Medicine. The team set out to determine whether patterns hidden in those records could forecast which children will later receive an ADHD diagnosis, well before such a diagnosis typically occurs.
The investigators trained a specialized AI model on longitudinal EHR data spanning birth through early childhood for more than 140,000 pediatric patients with and without ADHD. The model learned to recognize sequences and combinations of developmental delays, behavioral concerns, sleep disturbances, frequent behavioral visits, and other clinical events that, when occurring together, formed a “risk signature” for later ADHD diagnosis.
By age 5 the model showed especially strong predictive accuracy. Importantly, its performance remained robust across diverse demographic groups, including different sexes, races, ethnicities, and insurance statuses—an observation that supports the tool’s potential to promote more equitable screening and follow-up.
The researchers are careful to stress that the AI does not make clinical diagnoses. Instead, it functions as a triage and prioritization tool: flagging children who may benefit from targeted screening, additional discussion during pediatric visits, or timely referral to developmental or behavioral specialists.
“This is not an AI doctor,” said Matthew Engelhard, M.D., Ph.D., senior author and faculty member in Duke’s Department of Biostatistics & Bioinformatics. “It’s designed to help clinicians focus their time and resources so that children who need evaluation don’t fall through the cracks or wait years for answers.”
Earlier identification and screening could shorten the time to diagnosis and intervention, which is associated with better educational, social, and health outcomes. The authors also acknowledge the need for further validation studies and careful implementation research before integrating such tools into routine clinical workflows.
“When children’s needs are understood early and appropriate supports are in place, families and educators have a better chance to help those children succeed,” said Naomi Davis, Ph.D., associate professor in the Department of Psychiatry and Behavioral Sciences and coauthor of the study.
In addition to Hill, Engelhard, and Davis, the study lists De Rong Loh, Benjamin A. Goldstein, and Geraldine Dawson among its authors. Funding was provided by grants from the National Institute of Mental Health and the National Center for Advancing Translational Sciences.
Key Questions Answered
Q: What kinds of “hidden patterns” does the AI detect?
A: The model evaluates the timing and combination of events—such as specific developmental delays, sleep problems, repeated behavioral complaints, or psychiatric comorbidities—that may each seem minor in isolation but together form a predictive signal for later ADHD.
Q: Will AI start diagnosing children with ADHD?
A: No. The tool is intended to help clinicians prioritize which children need closer attention and screening. It supports clinical workflows but does not replace clinician judgment or formal diagnostic evaluation.
Q: Why is AI helpful when we already have doctors?
A: Primary care visits are often brief and fragmented across providers. AI can rapidly review years of EHR data to surface relevant trends and historical events that a clinician may not have time to manually assemble during a visit.
Editorial Notes
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by editorial staff.
About this AI and ADHD research news
Author: Stephanie Lopez
Source: Duke University
Contact: Stephanie Lopez – Duke University
Image credit: Neuroscience News
Original Research: Closed access. “Fetal and postnatal metal metabolism-related changes in brain function are associated with childhood behavioral deficits” by Elliot D. Hill, De Rong Loh, Naomi O. Davis, Benjamin A. Goldstein, Geraldine Dawson & Matthew Engelhard. DOI: 10.1038/s44220-026-00628-2
Abstract
Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental condition that can adversely affect long-term outcomes. Early detection is critical, yet demographic and clinical disparities often delay diagnosis.
In this work, researchers pretrained an EHR foundation model using records from a large cohort and then fine-tuned it to predict the probability and timing of ADHD diagnosis from birth up to age 9 in a pediatric cohort of over 140,000 patients.
By age 5, the model achieved a time-dependent area under the receiver operating characteristic curve of 0.92 at a 4-year horizon, and it maintained stable performance across demographic subgroups. Feature importance analysis linked predicted ADHD risk to developmental, behavioral, and psychiatric conditions documented in EHRs. These findings suggest that EHR-based predictive models could help clinicians identify children at risk for ADHD in a more timely and equitable manner.