Duke Health study says AI could flag ADHD risk years earlier

An AI model built at Duke Health was able to estimate which children might later receive an ADHD diagnosis years before clinicians usually label the condition, but the finding also underscores a harder question: whether earlier flagging will bring earlier help or just earlier stigma.

Researchers at Duke University School of Medicine and the Duke Department of Psychiatry & Behavioral Sciences trained the system on routine electronic health records from more than 140,000 children with and without ADHD. Using a broader electronic health record foundation model pretrained on data from more than 720,000 patients, the tool identified combinations of developmental, behavioral and clinical patterns that often appear long before a formal diagnosis. By age 5, it reached a time-dependent area under the receiver operating characteristic curve of 0.92 at a four-year horizon, a strong measure of prediction performance.

Elliot Hill, the study’s lead author, said the goal was to see whether hidden patterns in medical data could predict which children might later be diagnosed with ADHD, “well before” that diagnosis usually happens. Matthew Engelhard, M.D., Ph.D., the senior author, said the work is meant to help clinicians focus limited time and resources, not replace them. The model was reported to perform consistently across sex, race, ethnicity and insurance status, an important detail in a condition where diagnosis has long tracked unevenly across demographic groups.

That inequity is already visible in national public health data. The Centers for Disease Control and Prevention estimates that 7 million U.S. children ages 3 to 17, or 11.4%, had ever been diagnosed with ADHD in 2022, about 1 million more than in 2016. CDC data also show boys are diagnosed more often than girls, 15% versus 8%, and rates vary by race, ethnicity and state. Those gaps matter because children who go years without a diagnosis can miss academic supports, behavioral care and treatment decisions that may shape long-term school, social and health outcomes.

The American Academy of Pediatrics recommends that primary care clinicians evaluate children ages 4 to 18 who have academic or behavioral problems and symptoms of inattention, hyperactivity or impulsivity, while also screening for conditions that can overlap with ADHD, including learning disorders, anxiety, depression, autism spectrum disorder, tics, sleep disorders and apnea. Duke’s researchers framed their work as a population-health tool for primary care, where records already exist but warning signs are easy to miss. The next test is not whether AI can raise an alert, but whether health systems can turn that alert into timely evaluation, referral and support before families are left waiting years for answers.