An AI Spotted ADHD in Kids' Medical Records. The Headline Is Hiding the Real Story.

TL;DR

• Researchers at Duke University trained an AI on the medical records of more than 700,000 patients, then fine-tuned it on 140,000 children. It flagged ADHD up to four years before a formal diagnosis, getting it right roughly 92 times out of 100.
• The paper was published in Nature Mental Health on 27 April 2026. The code is public. The work was funded by the US National Institute of Mental Health.
• The model performed evenly across boys and girls, across racial and ethnic groups, and across kids on private vs. public insurance. That part is genuinely unusual and matters.
• This is not a tool your child's paediatrician can use yet. It has only been tested on old records, not on real children in real time.
• The real news is that the kind of AI that powers ChatGPT now works on children's medical histories. ADHD is the demonstration. The implication is much wider.

What actually happened

A team led by Elliot Hill and Matthew Engelhard at Duke trained a large AI model on the electronic health records of more than 720,000 patients — the medications, visits, diagnoses, lab results, and notes that pile up over a childhood. Then they fine-tuned it on the records of about 140,000 children specifically, and asked a simple question: looking only at what was in a child's chart at age four, age five, age six, can the model predict who will eventually be diagnosed with ADHD?

It could. At the four-year horizon, the model was right about 92 times out of 100 — a level of accuracy that, in medical AI, is meaningfully strong rather than borderline. It held up across boys and girls, across white, Black, Hispanic, and Asian children, and across families on different kinds of insurance.

The paper went up in Nature Mental Health on 27 April. The code is on GitHub. The funding came from a $15 million NIMH grant.

What it actually means

The instinctive read — AI can now diagnose ADHD — is the wrong one. Two things are actually going on, and the second is bigger.

The first is narrow. A model can look at the breadcrumbs in a child's medical history — the ear infections, the sleep problems, the early speech referrals, the asthma visits, the patterns nobody is consciously tracking — and detect a signal that points toward ADHD years before the formal label arrives. That is useful, but it is not new in kind. Researchers have been building risk-prediction models for years.

The second is wider. The model the Duke team used is a foundation model — the same family of AI that powers ChatGPT and the assistant in your phone. Until very recently, that approach was almost entirely confined to language and images. What this paper quietly demonstrates is that the same architecture works on the messy, sparse, longitudinal data of a child's medical record. ADHD is the showcase. Depression, anxiety, autism, eating disorders, and early psychosis all sit inside the same data. So the question stops being can AI predict ADHD? and becomes which childhood condition is next?

That is the part the headlines are missing.

What this isn't

It isn't a diagnostic tool. The model has only been tested retrospectively — on old records, with the answer already known. It has never been pointed at a real child in a real clinic to see whether its prediction holds up going forward. Those are different problems, and medical AI has a long history of looking great in the first test and stumbling in the second.

It isn't ready for your paediatrician. There is no app, no plug-in, no clinical workflow. The code is public, but using it would require a hospital's data science team, ethics approvals, and prospective validation studies that haven't been done.

It isn't a verdict. A 92-out-of-100 accuracy figure means the model is also wrong some of the time — both flagging children who never go on to be diagnosed and missing some who are. In a screening tool used at population scale, those errors translate into real children, real labels, and real consequences. The maths matters; the maths after the maths matters more.

And it isn't only about ADHD diagnosis. ADHD is one of the most over- and under-diagnosed conditions in paediatrics, depending on which country, postcode, gender, and race you happen to live in. Telling a model to predict the current diagnosis is partly telling it to reproduce the current biases in who gets diagnosed. The Duke team's even performance across groups is encouraging on this front, but it doesn't make the underlying problem disappear.

The quieter story

The most interesting finding in the paper isn't the headline accuracy number. It's that the model performed about the same across sex, race, ethnicity, and insurance status.

Medical AI has a known failure pattern: trained mostly on data from wealthier, whiter, better-insured patients, then deployed everywhere. The result is tools that work well for the people they were built around and worse for everyone else. That has been the single biggest drag on trust in clinical AI for the better part of a decade.

A model that holds its accuracy across demographic groups in a 140,000-child cohort is not the end of that problem. But it is one of the cleaner counter-examples to date. If this generalises, it will matter more in five years than the ADHD result itself.

Who this affects

Parents. Not yet. Nothing changes about how your child is screened, evaluated, or treated this year. If your paediatrician already takes early concerns seriously and refers thoughtfully, this work doesn't add or subtract from that. If you're worried about ADHD, the path is unchanged: a real evaluation, a real clinician.

Children with already-complex medical histories. Eventually, kids with a lot of early healthcare contact — frequent visits, multiple referrals, comorbid conditions — are the most likely to benefit from a tool like this, because there is more signal in their charts. They are also the most likely to be over-flagged. Both at once.

Paediatricians and family doctors. A version of this, in five to ten years, may show up as a quiet flag in your child's chart at a routine visit — this child's pattern resembles those who are later diagnosed with ADHD; consider screening. That is the realistic deployment, not autonomous diagnosis.

The wider mental-health field. This is the audience that should be paying closest attention. If the same technique works for adolescent depression or early psychosis, the prediction window — years before a formal diagnosis — is exactly where prevention has the most leverage and the most ethical hazard.

What's still unresolved

• Prospective testing. The model has never been used on a real child in real time. Every clinical AI looks better in the rear-view mirror than through the windscreen.
• What you do with the prediction. A four-year-early flag is only useful if there's a meaningful action to take. For ADHD, the evidence base for early behavioural intervention exists but is uneven, and medication in very young children is contested. The system downstream of the prediction is the bottleneck, not the prediction itself.
• Calibration. Accuracy is one number. Calibration — whether a 70 percent risk really means 70 percent — is another, and the paper's reporting on it is partial.
• Generalisation beyond Duke. The cohort is one health system. Different populations, different documentation habits, different prescribing cultures may all move the numbers.
• The label problem. The model is being trained to predict who gets diagnosed, which is not quite the same as who has ADHD. In a country where diagnosis rates vary by a factor of two between neighbouring states, that gap is doing real work.

Bottom line

A Duke team has shown that the kind of AI behind your phone's assistant can read a child's medical history and flag ADHD years before a clinician does, with even performance across the demographic lines that usually break these models. That is a genuine result, and the foundation-model approach almost certainly transfers to other childhood mental-health conditions. None of it changes anything in your paediatrician's office this year, and the harder problem — what to actually do with an early prediction without over-labelling a generation of children — has barely been started.

Sources

• Nature Mental Health, Hill, Engelhard et al., 27 April 2026 — primary publication. Tier 1.
• Duke Health institutional press release, 27 April 2026. Tier 1.
• Axios Raleigh, coverage of the Duke study, April 2026. Tier 2.
• Neuroscience News, summary of the Nature Mental Health paper. Tier 2.
• ClinicalTrials.gov / NIMH grant register, $15M foundation-model paediatric mental health award. Tier 1.