Artificial intelligence has helped diagnose 18 children at a Boston hospital whose rare illnesses had defeated teams of human specialists, in what researchers describe as a breakthrough for families who had spent years searching for answers.
The study, published on Wednesday, 18 June 2026 in the peer-reviewed journal NEJM AI, used OpenAI’s o3 Deep Research model to reanalyse 376 de-identified paediatric cases that had previously undergone genetic testing and expert review without yielding a diagnosis. The model identified new diagnoses in 18 of those cases — a diagnostic yield of 4.8 per cent. Given that these were cases already extensively analysed by human experts, the researchers consider the result significant. “Considering how many times these had already been analysed, that’s a huge number, and each one means an answer for a family,” said Catherine Brownstein, a lead researcher from the Manton Center for Orphan Disease Research at Boston Children’s Hospital, who called the AI “a total game changer”.
The AI model, released in April 2025, helped identify 10 cases of rare neurodevelopmental diseases, four neuromuscular disorders, two cases of sudden unexpected death in paediatrics, and two cases of early-onset psychosis. John Brownstein, chief innovation officer at Boston Children’s Hospital, said the challenge had been one of cognitive limits rather than effort. “We combine genetic information, phenotypic information, literature search, and the reasoning of AI to deliver diagnoses to families that were once left without any answers,” he said. The hospital has integrated AI across its operations, with more than a third of employees using AI tools daily, reportedly saving approximately 60,000 hours and over $7 million in redeployed labour. Its broader AI efforts have resulted in more than 40 rare disease diagnoses previously considered unsolvable.
How the AI model works
At the heart of the study is the way the o3 Deep Research model processes and synthesises information. Unlike conventional diagnostic tools that rely on matching symptoms to a fixed database, the model can learn from few examples and generalise rules — a capability that researchers describe as a fundamental shift in AI. For each of the 376 cases, the AI was fed doctors’ notes, patients’ symptoms, genomic data and details of genes that might be responsible. It then searched and integrated relevant medical literature, linking scattered clues that had not been connected during earlier manual reviews. The model operates as a “secondary reader” or “collaborating geneticist”, generating evidence-linked hypotheses for specialists to investigate further. Humans then reviewed the model’s answers, conducted additional testing and provided the final clinical confirmation. OpenAI explicitly states that its technology should not be used for self-diagnosis.
Suyash Shringarpure, a researcher at OpenAI focusing on the health sector and another author of the study, explained the time constraints that hamper human researchers. “A researcher can only spend so much time on a single case. Maybe a case remained unsolved when it came to them first, but a year later, a paper was published that clarifies the link between the gene and the disease,” he said. The AI model can continuously incorporate newly published research, capturing links that may have emerged after a human analyst moved on. Of the 18 new diagnoses, seven were “rediscoveries” — diagnoses already established elsewhere but not reflected in the local research workflow, highlighting the difficulty of integrating information across different clinical records and healthcare systems.
Boston Children’s Hospital screens the genomes of patients affected by rare diseases, which collectively affect 30 million people in the United States. The Manton Center for Orphan Disease Research, where Brownstein serves as scientific director of the Gene Discovery Core, works to understand the causes of these conditions. The study drew on the centre’s repository of undiagnosed cases. OpenAI has committed $50 million to the hospital’s AI initiatives.
A patient’s story
Kyra Benton was one of the patients to finally receive a diagnosis through the AI-assisted approach. She began exhibiting concerning symptoms at the age of nine — walking on her tiptoes and struggling to run with a normal gait. Her health declined over the years as doctors struggled to identify the underlying cause. Just before she turned 20 last year, researchers diagnosed her with myofibrillar myopathy, a progressive genetic neuromuscular disorder.
“Quite frankly, I’m the type of person that’s not all that much in favour of AI,” she said. “On the other hand, I do acknowledge that it does have its advantages.”