AI Can Now Spot ADHD Risk in Children Years Before Diagnosis

For many families, an ADHD diagnosis feels like it arrives either too late or not at all. A child spends years labelled difficult, lazy, or inattentive. Teachers grow frustrated. Parents blame themselves. By the time a formal diagnosis lands, the child may already carry the scars of years of misunderstanding.

A new study from Duke University suggests that artificial intelligence could fundamentally change that timeline — not by performing a new kind of test or scan, but by reading the data that already exists in your child's ordinary medical records.

The findings, published in Nature Mental Health, describe an AI model that can flag children at elevated risk of ADHD diagnosis with striking accuracy — and can do so years before most children currently receive one.

But here's what the headlines tend to gloss over: researchers have been trying to crack this problem for two decades. And why this particular approach matters isn't just that it works — it's why all the previous attempts didn't.

The Problem No One Talks About: The Diagnosis Gap

ADHD affects roughly 5–7% of children worldwide. Yet the average age at diagnosis in the UK sits somewhere between 8 and 10 — and that's for children who receive a diagnosis at all. For girls, for children from lower-income families, and for children from ethnic minority backgrounds, the wait is typically even longer. Some don't get a formal diagnosis until their thirties or forties.

That gap has consequences. Research consistently shows that early support — whether therapeutic, educational, or environmental — produces substantially better long-term outcomes for children with ADHD. Every year of unidentified struggle is a year of potential support that doesn't happen. It's an academic trajectory altered. A sense of self-worth quietly eroded.

The question researchers have been wrestling with is deceptively simple: can we find ADHD earlier, before the child has already started to fall behind?

---

Two Decades of Attempts — and Why They All Hit the Same Wall

The quest to predict or detect ADHD before formal diagnosis didn't start with AI. It started with neuroscience.

Brain imaging (fMRI) was one of the earliest and most ambitious approaches. Researchers discovered that children with ADHD show measurable differences in connectivity between brain networks — particularly reduced connectivity between the dorsal attention network and somatomotor regions. Studies using functional MRI achieved impressive accuracy rates in classifying ADHD versus non-ADHD brains in research settings. The problem? An fMRI scan costs thousands of pounds, requires specialist facilities, and is deeply impractical for a routine paediatric appointment. It also requires children to lie completely still in a noisy machine — something many ADHD children understandably struggle with.

EEG offered a cheaper, more accessible alternative. Electroencephalography measures electrical activity in the brain and has become one of the most studied biomarkers in ADHD research. Multiple studies showed that certain EEG patterns — elevated theta waves, reduced beta activity — correlate with ADHD presentations. Wearable EEG devices brought the technology closer to practical use. But EEG remains primarily a research tool. Signal quality varies enormously between devices and settings, and the diagnostic standards required for clinical use haven't yet been established.

Wearable devices and digital phenotyping represented a more recent wave. A 2023 study published in JAMA Network Open used Fitbit data from the ABCD study cohort to predict ADHD in children based on circadian rhythm features and sleep patterns captured by wrist-worn trackers. The results were promising. But the study required children to wear specific devices consistently — a compliance challenge even in motivated research participants, let alone in routine healthcare.

Game-based and questionnaire AI tools emerged as another strand. Companies developed digital assessments — reaction-time tests, attention tasks, gamified cognitive challenges — that could be administered on a tablet and scored by an algorithm. Some achieved reasonable sensitivity. But they still required a specific assessment moment, and their real-world performance often failed to match their research-setting results.

Each of these approaches shared a common limitation. They all required something new: a new piece of equipment, a new test, a new referral pathway, a new clinical appointment. And anything that requires "new" in healthcare tends to face the same barriers — cost, access, equity, and implementation time.

---

What the Duke Study Does Differently

The Duke Health team, led by data scientist Elliot Hill and senior author Dr Matthew Engelhard, took a fundamentally different approach. Rather than developing a new test, they asked a deceptively simple question: what if the clues are already there?

"We have this incredibly rich source of information sitting in electronic health records," Hill explained. "The idea was to see whether patterns hidden in that data could help us predict which children might later be diagnosed with ADHD, well before that diagnosis usually happens."

To answer that question, the team built what's called a foundation model — a large AI system pre-trained on a vast amount of healthcare data — and then fine-tuned it specifically for ADHD prediction. The pre-training phase used records from more than 720,000 patients. The fine-tuning used a paediatric cohort of more than 140,000 children, tracking them from birth to age 9.

The model wasn't looking for a single red flag. It was looking for patterns — combinations of developmental milestones, behavioural notes, clinical observations, and routine visit data that, in aggregate, tend to precede an ADHD diagnosis by several years. No new test. No specialist equipment. Just the ordinary accumulated record of a child's healthcare history.

The results were striking. By age 5, the model achieved an area under the ROC curve of 0.92 at a four-year predictive horizon. In plain language: it was highly accurate at identifying which children would go on to receive an ADHD diagnosis by age 9, based only on their medical records up to age 5.

Perhaps more importantly, the model performed consistently across different demographic groups — across sex, race, ethnicity, and insurance status. This matters enormously. Previous research tools have often performed best in the populations they were developed in (typically white, male, higher-income), which risks widening existing diagnostic inequalities. The Duke model's equity performance is one of its most significant features.

---

What This Actually Means for Families

It would be easy to read this research and imagine a future where an algorithm diagnoses your child before they even start school. That is not what this study describes, and the researchers are careful to say so.

"This is not an AI doctor," said Dr Engelhard. "It's a tool to help clinicians focus their time and resources, so kids who need help don't fall through the cracks or wait years for answers."

The practical implication is more modest — but still meaningful. A model like this could function as an early warning system within a primary care setting. When a child's electronic health record reaches a certain threshold, it could prompt their GP or paediatrician to ask different questions at the next appointment, or to make a referral for a formal assessment earlier than they otherwise would have.

That's not a small thing. One of the most consistent findings in ADHD research is that children often reach diagnosis only after their difficulties have become impossible to ignore — by which point, years of struggle have already accumulated. A nudge in the right direction at age 5 or 6, rather than age 9 or 10, could change a child's trajectory significantly.

Dr Naomi Davis, one of the study's co-authors and an associate professor in Duke's Department of Psychiatry and Behavioural Sciences, put it plainly: "Children with ADHD can really struggle when their needs aren't understood and adequate supports are not in place. Connecting families with timely, evidence-based interventions is essential for helping them achieve their goals and laying a foundation for future success."

---

The Bigger Picture: A Shift in How We Think About Data

What strikes me most about this research is what it says about the nature of clinical data itself.

For decades, electronic health records have been primarily a documentation tool — a way for clinicians to record what happened, for administrative systems to process billing, for regulatory bodies to audit care. The data existed, but it sat largely dormant in terms of predictive value.

The Duke study is part of a broader movement in clinical AI that asks a different question: not "how do we document care?" but "what patterns are hidden in the care we've already delivered?" The same records that exist simply because a child saw a GP for an ear infection, or had a developmental check at 18 months, or received a referral for speech therapy — those records may collectively contain signals that could redirect a child's life.

That's a genuinely new idea. Not a new scanner. Not a new biomarker. Just a new way of reading the story that's already been written.

---

Important Caveats

The researchers are clear that this work is not yet ready for clinical deployment. The study was conducted within Duke Health's patient population, and the model will need to be validated in different healthcare settings, countries, and populations before it could be considered for routine use. The UK's NHS, with its own electronic health record architecture, would require its own validation work.

There are also important ethical questions about predictive tools in healthcare — particularly tools that concern children. Who owns the prediction? How is it communicated to parents? What are the risks of a false positive — a child labelled "high risk" who doesn't go on to develop ADHD? These questions don't invalidate the research, but they do need careful answers before any tool like this moves into clinical practice.

---

The Bottom Line

The history of AI-based ADHD detection is littered with promising tools that were too expensive, too specialised, or too demanding to ever reach the families who needed them most.

The Duke study is different not because it's more accurate than its predecessors — though it is — but because it works with what already exists. The data is already there. The records are already being created. The question is whether the healthcare system can be organised to read them better.

That's a solvable problem. And for families currently waiting years for answers, it's one worth solving urgently.

---

Dr Malcolm Pye, PhD, is a research psychologist specialising in neurodevelopmental conditions and AI applications in clinical assessment. This article is for informational purposes. If you are concerned your child may have ADHD, speak with your GP or paediatrician about a formal assessment.

---

Take our free online ADHD screening tool: adhdtest.ai/adult-adhd-test/