Duke University researchers say a new artificial intelligence tool can flag which young children are most likely to develop attention-deficit/hyperactivity disorder years before they typically get a diagnosis, using only standard electronic health records.
The system is built on a foundation model that was pretrained on hundreds of thousands of electronic health records, then fine-tuned on long-term data from more than 140,000 children. The team stresses that the tool is meant to help clinicians decide who should get a closer look and earlier support, not to replace anyone’s clinical judgment.
The study, published April 27 in Nature Mental Health, reports that the model reached a time-dependent area under the receiver operating characteristic curve of 0.92 at a four-year horizon when predicting ADHD by age 5. In plain language, that score indicates the algorithm was good at separating higher risk from lower risk children inside the study groups.
In a press release from Duke Health, lead author Elliot Hill noted that “detecting ADHD by age 5 is well before the diagnosis usually happens.” Co-author Naomi Davis said that catching concerns sooner could give families a shot at supports that may change long-term outcomes.
How the Model Works
The research team pretrained an electronic health record foundation model using data from more than 720,000 patients. They then fine-tuned it with records from about 140,000 children, tracking health information from birth through age 9 to see which patterns tended to show up before an ADHD diagnosis.
According to Nature Mental Health, the model kept its performance across sex, race, ethnicity and insurance status in the study’s tests. The paper details what information went into the system and how the team checked its predictions.
What It Means for Families and Clinicians
Researchers are clear that this is a screening aid, not a shortcut to diagnosis. “This is not an AI doctor,” co-author Dr. Matthew Engelhard said in the Duke Health release, emphasizing the need for clinician oversight.
Earlier risk flags could change who gets evaluated first in a crowded system. An estimated 11 percent of U.S. children have received an ADHD diagnosis, and just over half of those diagnosed take medication, according to 2022 figures from the CDC. The Duke team argues that if pediatricians know which kids are at higher statistical risk, they can prioritize assessments and connect families with services sooner.
Next Steps and Caveats
Even with strong results in retrospective data, the authors and outside experts warn that predictors built from health records can mirror existing inequities in the system. They say the model needs to be tested prospectively and in many different clinical settings before anyone uses it to steer real-world care.
Independent coverage has highlighted those cautions, pointing to the need for more validation, ongoing monitoring and direct clinician involvement if the tool ever moves toward routine use. MedicalXpress summarizes the concerns and the authors’ own calls for additional study.
The project code is publicly available so that other researchers can inspect it and try to reproduce the findings, which the authors say should help surface any blind spots or safety problems. The repository is hosted on GitHub. For now, both the research team and outside commentators describe the work as a promising step forward, but not yet a screening tool that pediatric clinics should plug into everyday practice.
