A prediction tool works best when it gives busy clinicians a clear reason to look more closely.
Long before a child receives an ADHD diagnosis, the early signals are often already written into routine medical records — scattered notes about sleep, speech, and restlessness that no single appointment connects. Researchers at Duke University have built an artificial intelligence system that reads those patterns across time, predicting ADHD up to four years before a clinical label arrives. The tool does not replace the physician's judgment, but it offers something rarer: the gift of earlier attention, when intervention still has the most room to shape a child's unfolding story.
- Roughly one in ten American children carries an ADHD diagnosis, yet many struggle for years in silence before anyone formally names what they are experiencing.
- By the time most families reach a specialist, school confidence has already eroded, friendships have frayed, and frustration has become a familiar companion — the cost of waiting.
- Trained on records from over 720,000 patients, the AI model achieved a 0.92 accuracy score by age five, identifying not single warning signs but recurring combinations of developmental and behavioral patterns across a child's medical history.
- Researchers are careful to frame the tool as a prompt for closer clinical attention, not a diagnosis — a shorter watchlist for busy pediatricians, not a label to be applied without human judgment.
- Fairness across race, income, and geography remains an unresolved challenge, as a tool that works well in one health system may quietly fail the communities that need it most.
- Real-world validation across diverse populations and healthcare settings stands between this promising signal and the moment it could meaningfully change how children are seen and supported.
A routine pediatric visit — a note about restlessness, a referral to speech therapy — seems unremarkable in the moment. But researchers at Duke University School of Medicine have built an AI system that can gather those scattered early signals and predict which children will eventually be diagnosed with ADHD, sometimes four years before a family ever hears the clinical term.
The stakes are rooted in timing. Many children show signs of struggle long before anyone formally assesses them, losing ground in school and friendships while the pattern goes unnamed. By the time a diagnosis arrives, years of frustration have already shaped how a child sees themselves. Earlier identification opens a different path — one where parent training, classroom accommodations, and coping strategies can take hold while foundational habits are still being formed.
Elliot D. Hill and his team trained their model on records from over 720,000 patients, refining it with data from more than 140,000 children tracked from birth through age nine. The system learned to recognize not isolated warning signs but recurring combinations — speech delays alongside behavioral concerns, psychiatric symptoms woven together with developmental questions. By age five, it achieved a 0.92 accuracy score, a signal strong enough to guide clinical attention. Senior author Matthew Engelhard was clear about its limits: this is not an AI doctor. Risk scores must still be followed by family interviews, teacher reports, and clinical judgment — otherwise a prompt for better care becomes a premature label.
Fairness remains an open and urgent question. The Duke team found balanced performance across sex, race, ethnicity, and insurance status within their own system, but one health network cannot stand in for every clinic, every community, or every resource-scarce setting. A screening tool that works reliably for some children while quietly overlooking others risks deepening the very gaps it aims to close.
The tool, published in Nature Mental Health, points toward smarter, earlier screening. But its real promise depends on what comes after the alert — families with time to learn, schools with room to adjust, and children who receive support before repeated struggle hardens into shame. Privacy protections, equity testing, and real-world validation must come before any pediatrician's office begins using it to shape a child's care.
A child sits in a pediatrician's office for a routine checkup. The visit gets logged in the electronic health record—a note about sleep trouble, maybe a mention of restlessness, a referral to speech therapy. Years later, that same child receives an ADHD diagnosis. Researchers at Duke University School of Medicine have now built an artificial intelligence system that can connect those scattered early signals and predict which children will eventually be diagnosed with attention-deficit/hyperactivity disorder, sometimes as much as four years before a family ever hears the clinical label.
The work matters because timing shapes outcomes. About one in ten American children ages four to seventeen carries an ADHD diagnosis, but many show signs of struggle long before anyone formally assesses them. A child might fidget through kindergarten, lose focus in second grade, or clash with peers without anyone naming the pattern. By the time a family reaches a specialist—if they reach one at all—school confidence has already fractured, friendships have strained, and daily routines have calcified around frustration. Earlier identification could change that arc. Parent training, classroom accommodations, and coping strategies work best when a child is still building foundational habits, not when years of struggle have already reshaped how they see themselves.
Elliot D. Hill and his team trained their model on medical records from over 720,000 patients, then refined it using data from more than 140,000 children whose care had been tracked from birth through age nine. The system learned to recognize not single warning signs but recurring combinations—patterns woven across development, behavior, sleep, medication use, and referrals. A child might have a speech delay and behavioral concerns and repeated visits about attention. Another might show psychiatric symptoms alongside developmental questions. The model learned which sequences of events, taken together, correlated with later ADHD diagnosis. By age five, the tool achieved a 0.92 accuracy score, meaning it sorted children into risk categories with strong predictive power. On a scale where 1.0 would mean perfect separation, 0.92 represents a signal clear enough to guide clinical action.
But the researchers were careful about what that score actually means. A high prediction number does not diagnose ADHD. It does not prove a child has the condition. Instead, it flags children who warrant closer attention—a shorter list for busy pediatricians to watch more carefully before problems accumulate. Matthew Engelhard, a senior author and biostatistics researcher at Duke, emphasized the distinction: "This is not an AI doctor." Clinicians still need to conduct family interviews, gather teacher reports, review developmental history, and exercise clinical judgment. Without that human step, a risk score could become a label instead of a prompt for better care.
Fairness across populations remains an open question. The Duke team tested their model across sex, race, ethnicity, and insurance status within their own health system and found balanced performance—the tool did not systematically miss one group more than another. That is important because screening tools can cause harm when they identify some children reliably while overlooking others, potentially widening existing diagnostic gaps. But one health system cannot prove fairness for every community, every insurance network, or every clinic workflow. Future testing in diverse settings, especially in places where specialist appointments take months and resources are scarce, will determine whether these alerts actually help all children or whether they risk becoming another tool that works better for some families than others.
The real promise lies not in prediction itself but in what comes after. Earlier concern gives families time to learn strategies, helps schools adjust expectations, and lets adults respond consistently at home and classroom before repeated frustration hardens into shame or learned helplessness. The American Academy of Pediatrics supports this approach—family input, school input, medical input working together to shape care. A child who gets support while still building habits has a different trajectory than one who waits years for answers. The tool, published in Nature Mental Health, points toward smarter screening. But privacy protections, fairness testing, and real-world validation remain essential before any pediatrician's office starts using it to flag children for follow-up.
Notable Quotes
We have this incredibly rich source of information sitting in electronic health records.— Elliot D. Hill, Duke University School of Medicine
This is not an AI doctor. Clinicians still need family interviews, teacher reports, developmental history, and judgment before making an ADHD diagnosis.— Matthew Engelhard, M.D., Ph.D., Duke University School of Medicine
The Hearth Conversation Another angle on the story
Why does it matter that we catch this years early? Can't families just wait for the diagnosis when it becomes obvious?
Because by then the damage is done. A child who struggles for years without understanding why starts to believe they're broken. They avoid school, withdraw from friends, develop anxiety. Early support changes that story—it gives them strategies while they're still building confidence.
The model achieved 0.92 accuracy. That sounds nearly perfect. Why the caution?
Because 0.92 is not 1.0. It's a strong signal, but it will miss some children and flag others who don't have ADHD. More importantly, a prediction is not a diagnosis. The AI is a flag, not a verdict. A doctor still has to listen, observe, and decide.
You mentioned fairness concerns. What could go wrong?
If the tool works better for white children than Black children, or better for insured kids than uninsured ones, it becomes another way existing inequalities get baked into medicine. We don't know yet if that's happening outside Duke's system.
So this tool could actually harm some children?
If it's used as a shortcut instead of a starting point, yes. If a teacher or parent sees a risk score and stops listening to the child, or if it becomes a label instead of an invitation to help, it could do real damage.
What would success look like?
A pediatrician sees a flag, takes time to really understand the child, and then offers the family support—training, school accommodations, strategies—before the child has spent years feeling lost. That's when early prediction actually changes lives.