AI Model Predicts ADHD Risk Years Before Diagnosis, Duke Study Shows

Millions of children carry the weight of undiagnosed ADHD into their school years, their struggles often mistaken for laziness or defiance, their brains working differently while no one yet understands why. By the time a diagnosis arrives—sometimes years after the first signs appeared—crucial windows for intervention have already closed. A child's academic confidence has eroded. Social relationships have frayed. The foundation for later success has already begun to crack.

Researchers at Duke Health have now demonstrated that artificial intelligence can change this timeline. In a study published in Nature Mental Health, they showed that machine learning models trained on routine electronic health records can identify which children are likely to develop ADHD years before a clinician would typically make that diagnosis. The finding opens a path toward catching these children earlier, when evidence-based support can still reshape their trajectories.

The work began with a simple observation: electronic health records contain a wealth of information that no single doctor has time to fully synthesize. Elliot Hill, a data scientist in Duke's Department of Biostatistics & Bioinformatics, and his colleagues wondered whether patterns hidden in that data—combinations of developmental milestones, behavioral notes, clinical events—might reveal which children were at risk. To test the idea, they analyzed records from more than 140,000 children, some later diagnosed with ADHD and some who were not. They trained an AI model to learn from medical histories spanning birth through early childhood, teaching it to recognize the subtle signatures that often precede an ADHD diagnosis by years.

The model proved remarkably accurate. When tested on children age five and older, it consistently identified future ADHD risk with strong predictive power. Crucially, its performance held steady across different demographic groups—boys and girls, children of different races and ethnicities, families with different insurance coverage. The tool did not discriminate; it worked.

But the researchers were careful about what they claimed. This is not a diagnostic tool. It does not replace a clinician's judgment or a formal evaluation by a specialist. Instead, it functions as a spotlight, helping primary care providers direct their attention and resources toward children who might benefit from closer monitoring or earlier referral for assessment. As Matthew Engelhard, senior author and a physician-scientist at Duke, put it: the goal is to ensure that children who need help do not slip through the cracks or spend years waiting for answers.

The potential impact hinges on a simple fact: early identification matters. Children diagnosed with ADHD sooner tend to have better academic outcomes, stronger social relationships, and improved health trajectories when they receive timely, evidence-based interventions. Naomi Davis, another author on the study, emphasized that children with undiagnosed ADHD often struggle profoundly when their needs go unrecognized. Connecting families with appropriate support early can be transformative.

The researchers acknowledge that more work lies ahead. Before such tools enter clinical practice, they will need validation in real-world settings, testing in different healthcare systems, and careful examination of how they might affect clinical workflows and decision-making. The promise is real, but the path from research to routine care is neither short nor guaranteed. Still, for the millions of children currently waiting years for answers, the possibility that a computer algorithm might help their doctor see them sooner represents something worth pursuing.