OpenAI Study Finds o3 Deep Research Helped Identify 18 Diagnoses in 376 Unsolved Rare Paediatric Cases

Imagine waiting years — sometimes an entire childhood — for an answer to what is making your child ill. For families living with rare diseases, that wait is often the reality. A study published on 18 June 2026 in the journal NEJM AI suggests that AI may be able to help chip away at that backlog, at least in some cases.

The research, a collaboration between OpenAI, Boston Children’s Hospital, and Harvard, looked at 376 previously unsolved paediatric cases. These weren’t straightforward presentations. Every one of them had already been through genetic testing and expert review — and still had no diagnosis. Using OpenAI’s o3 Deep Research model, clinicians went back through the case files and came away with 18 new confirmed diagnoses. That’s a 4.8% additional diagnostic yield, which may sound modest, but for the families involved, it represents an answer they had been waiting years to receive.

What o3 Deep Research Actually Did

It’s worth being clear about what the AI was and wasn’t doing here. The study describes o3 Deep Research as an assistive tool — not an autonomous diagnostician. The model worked through clinical features, inheritance patterns, genetic variant evidence, and published medical literature to propose hypotheses. Doctors then reviewed those hypotheses, and in many cases sought laboratory confirmation before any diagnosis was recorded.

So the AI was doing something closer to a very thorough literature review and pattern-matching exercise, surfacing possibilities that a busy clinician might not have time to chase down manually. Every diagnosis still required a human expert to sign off.

The cases covered a range of conditions. Ten of the 18 new diagnoses were neurodevelopmental disorders. Four involved neuromuscular disease. Two cases concerned sudden unexpected death in children, and two involved early childhood psychosis. These are some of the most complex and distressing presentations in paediatric medicine, and they illustrate exactly why re-analysis matters as medical knowledge advances.

The Research Partners Behind the Study

Boston Children’s Hospital’s Manton Center for Orphan Disease Research was central to the collaboration. The Manton Centre focuses specifically on orphan diseases — conditions so rare that they often fall outside mainstream research funding and clinical attention. Partnering with Harvard researchers and OpenAI, the team used the re-analysis project to test whether AI could meaningfully support the kind of deep-dive case review that rare disease families desperately need.

One report suggested the collaboration between OpenAI and Boston Children’s had been underway since early 2025, though this timing isn’t independently confirmed across all sources covering the study. A funding figure of around £39 million (reported as US$50 million in some coverage) was also cited in connection with OpenAI’s commitment to the hospital’s work, but this could not be independently verified from the published study itself, so treat that figure with caution.

What the Researchers Say

The research team is careful not to oversell the findings. The study is retrospective — it looked back at old cases rather than testing the model on incoming patients in real time. That distinction matters. It doesn’t show that o3 Deep Research can diagnose children autonomously in routine clinical practice, and nobody involved is claiming it can.

What the study does argue is that AI can help clinicians revisit complex rare disease backlogs and surface new possibilities as the published medical literature grows. Rare disease medicine is especially suited to this kind of re-analysis, because new gene-disease associations are being described constantly, and cases that were genuinely unsolvable five years ago may now have answers.

Alan Beggs, director of the Manton Center for Orphan Disease Research at Boston Children’s Hospital, said: “This study demonstrates how AI can serve as a powerful tool to help clinicians uncover diagnoses for patients who have been waiting years for answers.”

The framing from the clinical side is consistent: AI as a research assistant, not a replacement for specialist judgement.

The Broader Picture for Rare Disease Diagnosis

Rare diseases affect roughly one in 17 people at some point in their lives, and many of those conditions are paediatric. The so-called “diagnostic odyssey” — the years some families spend moving from specialist to specialist without a clear answer — is a well-documented problem in healthcare systems around the world, including the NHS.

The promise of AI-assisted re-analysis is that it could help reduce that odyssey. But it won’t happen overnight, and it won’t happen without specialist clinical infrastructure to review whatever the AI surfaces. The 18 diagnoses in this study came through a rigorous process at one of the world’s leading children’s hospitals, with expert human review at every stage.

And that’s the honest caveat here. The technology is promising. The early results are real. But scaling this to routine clinical use — with the safeguards that patients deserve — is a different challenge entirely.

What This Means for Kent Residents

There is no confirmed involvement from NHS Kent and Medway or any Kent hospital in this research. But for families here in Kent who are living through a rare disease diagnostic odyssey with a child, this study is a meaningful development to follow. If AI-assisted case re-analysis methods like those used at Boston Children’s Hospital are eventually adopted within NHS specialist pathways, it could in principle benefit patients referred through UK rare disease networks — including those from Kent. For now, it’s a research finding rather than an available service, but it’s the kind of work that shapes where NHS diagnostic medicine heads next.