Duke University AI Model Analyzes Routine Health Records to Predict ADHD Risk in Young Children

New AI tool from Duke University analyzes routine electronic health records to flag ADHD risk in children years before a standard clinical diagnosis.

By: AXL Media

Published: Apr 27, 2026, 6:40 AM EDT

Source: Information for this report was sourced from EurekAlert

Unlocking Insights from Routine Medical Data

A significant advancement in pediatric mental health has emerged from Duke University, where data scientists have harnessed the power of artificial intelligence to spot early indicators of ADHD. Published in Nature Mental Health, the study demonstrates that routine electronic health records contain a wealth of untapped information that can predict a child's developmental trajectory. Lead author Elliot Hill explains that the research focused on identifying hidden patterns within everyday medical data, allowing the team to estimate a child's risk well before a formal diagnosis usually occurs. This approach utilizes information already collected during standard healthcare visits, requiring no additional testing for the patient.

Specialized Training on Massive Patient Datasets

To develop the predictive tool, the Duke Health team analyzed the medical histories of more than 140,000 children. The AI model was trained to recognize specific combinations of behavioral, clinical, and developmental events recorded from birth through early childhood. By comparing the records of children who were eventually diagnosed with ADHD against those who were not, the algorithm learned to flag subtle clusters of symptoms and medical visits that often precede a diagnosis. This large-scale data analysis provides a level of pattern recognition that would be nearly impossible for individual clinicians to track over several years of disparate health encounters.

Consistency Across Diverse Demographics

The reliability of the AI model was a primary focus of the research, with the tool demonstrating high accuracy in children aged five and older. Crucially, the model maintained consistent performance levels across various patient characteristics, including sex and insurance status. The study specifically noted that the predictive accuracy remained steady across different racial and ethnic groups, ensuring that the tool could be applied equitably in diverse clinical settings. This consistency is vital for a screening tool intended to function as a universal safety net, preventing children from falling through the regulatory cracks regardless of their background or socioeconomic standing.