Duke study shows AI predicts ADHD risk in young children

Researchers at Duke University School of Medicine published results showing an AI-based model can estimate a child's risk of later receiving an ADHD diagnosis from routine electronic health records (EHRs). Per Duke Health, the study analyzed EHRs from more than 140,000 children covering birth through early childhood and evaluated model performance for children aged 5 and older. Duke quotes a lead author, Elliot Hill: "We have this incredibly rich source of information sitting in electronic health records." Duke and Axios report the tool is intended to flag children for follow-up rather than to provide a clinical diagnosis. Axios and AzerNEWS report that the underlying system had been trained using EHRs from over 720,000 patients; AzerNEWS additionally reports a standardized evaluation score of approximately 0.92 for predicting a diagnosis within four years when applied at age five.

Editorial analysis - technical context: EHR-based risk models typically rely on longitudinal feature extraction, sparse event sequences, and handling of class imbalance and missingness. Researchers in comparable studies often use representation-learning or sequence models to capture developmental patterns, and performance is commonly reported with discrimination metrics such as area under the receiver operating characteristic curve (AUC) or similar standardized scores. For practitioners, this class of approach requires careful attention to data curation, temporal validity, and subgroup performance before clinical deployment.

Early identification of developmental and mental-health risk from routinely collected health data is an active research area because earlier support can materially affect educational and behavioral outcomes. The Duke team frames the tool as a prioritization aid for primary care, not a replacement for clinical assessment; Axios quotes co-author Matthew Engelhard: "This is not an AI doctor." Public reporting highlights that the model showed consistent performance across sex, race, ethnicity, and insurance status in the study cohort, per Duke Health and Axios.

For practitioners: key open questions include external validation in independent health systems, prospective evaluation for clinical utility, transparency of model inputs and thresholds, and evaluation of fairness across underrepresented subpopulations. Observers should also track integration paths into pediatric workflows, regulatory review if the tool is packaged as a clinical decision support product, and data-governance safeguards around EHR-derived risk labeling.

The Duke work adds to evidence that longitudinal EHR signals can flag developmental risk earlier than typical diagnostic timelines. The study reports strong retrospective performance on a large cohort, but further multi-site validation and careful operational testing are necessary before broad clinical use.