AI-Powered Early Detection of ADHD in Children Using Health Records

Attention-deficit/hyperactivity disorder (ADHD) affects millions of children worldwide, yet many receive a diagnosis years after early signs first appear. This delay can mean missed opportunities for interventions that may improve long-term outcomes. A new study from Duke Health researchers demonstrates that artificial intelligence (AI) tools can analyze routine electronic health records (EHRs) to predict a child’s risk of developing ADHD years before a typical clinical diagnosis. By identifying patterns in everyday medical data, the approach could help flag children who might benefit from earlier evaluation and support.

AI Predicts ADHD Risk Using Electronic Health Records

The Duke Health study, published in Nature Mental Health, developed a neural network model capable of estimating both the likelihood of an ADHD diagnosis and the time remaining until that diagnosis. The model was trained on de-identified EHR data, which includes structured information such as lab results and unstructured data like physician notes. Unlike traditional diagnostic methods, which rely on behavioral assessments and parent or teacher reports, this AI-driven approach leverages existing medical records to detect subtle patterns that may precede a formal diagnosis.

“Early intervention is critical for children with ADHD, as it can significantly improve academic performance, social functioning, and overall quality of life,” said Dr. Scott Kollins, a senior author of the study and professor of psychiatry and behavioral sciences at Duke University School of Medicine. “Our model doesn’t replace clinical judgment, but it could serve as an early warning system, helping clinicians prioritize which children may need closer monitoring or additional screening.”

How the Model Works

The AI tool analyzes a combination of structured and unstructured data from EHRs. Structured data includes measurable information such as height, weight, blood pressure, and medication prescriptions. Unstructured data—comprising roughly 80% of EHR content—consists of physician notes, which often contain observations about a child’s behavior, developmental milestones, and family history. The model identifies correlations between these data points and later ADHD diagnoses, even when those connections might not be immediately apparent to human reviewers.

For example, the model may detect that children who frequently visit primary care providers for complaints such as sleep disturbances, anxiety, or difficulty concentrating are more likely to receive an ADHD diagnosis later in childhood. It may also flag patterns in medication use, such as prescriptions for asthma or allergies, which have been linked to higher rates of ADHD in some studies. By synthesizing these disparate data points, the AI generates a risk score that clinicians can use to guide follow-up care.

Potential Benefits of Early Prediction

ADHD is one of the most common neurodevelopmental disorders in children, affecting an estimated 6 million children in the United States alone. However, the average age of diagnosis is around 7 years old, despite the fact that symptoms often emerge as early as age 3 or 4. Delays in diagnosis can lead to academic struggles, behavioral challenges, and strained family relationships. Early identification, allows for interventions such as behavioral therapy, classroom accommodations, and, when appropriate, medication, all of which can help children manage symptoms more effectively.

The Duke Health study highlights several potential advantages of AI-driven early prediction:

• Reducing diagnostic delays by flagging high-risk children for earlier evaluation.
• Helping clinicians differentiate between ADHD and other conditions with overlapping symptoms, such as anxiety or learning disabilities.
• Enabling more personalized care plans by identifying which children may respond best to specific interventions.
• Supporting public health efforts by providing data-driven insights into ADHD prevalence and risk factors.

Limitations and Ethical Considerations

While the study’s findings are promising, the researchers caution that the model is not a diagnostic tool on its own. “AI can identify patterns, but it cannot replace the nuanced judgment of a trained clinician,” said Dr. Kollins. “The goal is to augment clinical decision-making, not to automate it.”

The study also raises important ethical questions. For instance, false positives—children flagged as high-risk who never develop ADHD—could lead to unnecessary stress for families or even stigmatization. The model’s reliance on EHR data means its accuracy depends on the quality and completeness of those records. Children from underserved communities, who may have less consistent access to healthcare, could be underrepresented in the data, potentially limiting the model’s generalizability.

Another concern is the potential for bias in the training data. If historical EHRs reflect disparities in how ADHD is diagnosed across racial, ethnic, or socioeconomic groups, the AI model could inadvertently perpetuate those biases. The Duke Health team emphasized the need for ongoing validation and refinement of the model to ensure it performs equitably across diverse populations.

Broader Implications for Pediatric Care

The Duke Health study is part of a growing trend of using AI to improve pediatric care. Similar tools are being developed to predict risks for conditions such as autism, asthma, and type 1 diabetes. For example, researchers at Stanford Medicine recently trained a large language model to review medical charts for signs that children with ADHD received appropriate follow-up care after starting new medications. That study, published in Pediatrics, found that AI could efficiently identify gaps in care, such as missed follow-up appointments or inadequate monitoring of medication side effects.

“The power of these AI tools lies in their ability to process vast amounts of data quickly and consistently,” said Dr. Yair Bannett, lead author of the Stanford study and assistant professor of pediatrics. “What would take a human researcher months to review can be done by AI in a matter of hours, freeing up clinicians to focus on direct patient care.”

However, experts caution that AI should be viewed as a complement to, rather than a replacement for, clinical expertise. “AI can highlight trends and flag potential concerns, but it cannot understand the full context of a child’s life,” said Dr. Heidi Feldman, senior author of the Stanford study and a professor of developmental and behavioral pediatrics. “The human element—listening to parents, observing the child, and considering the family’s unique circumstances—remains essential.”

What Comes Next?

The Duke Health team plans to conduct further validation studies to assess the model’s performance in real-world clinical settings. They are also exploring ways to integrate the tool into electronic health record systems, allowing it to generate risk scores automatically during routine visits. If successful, such integration could make early ADHD prediction a standard part of pediatric care.

In the meantime, the researchers emphasize the importance of continued collaboration between AI developers, clinicians, and families. “This is not about replacing doctors or parents,” said Dr. Kollins. “It’s about giving them an additional tool to make more informed decisions, sooner.”

For families, the study offers a glimmer of hope. Early intervention has long been recognized as a key factor in improving outcomes for children with ADHD. If AI can help bridge the gap between the first signs of concern and a formal diagnosis, it could transform the lives of millions of children and their families.

As the field of AI-driven healthcare continues to evolve, studies like this one underscore the potential for technology to enhance, rather than replace, human expertise. The challenge lies in ensuring that these tools are developed and deployed responsibly, with a focus on equity, transparency, and patient-centered care.