How OpenAI Is Helping Doctors Diagnose Rare Childhood Diseases That Stumped Experts for Years
For families living with a child whose illness has no name, the wait for answers can stretch for years, sometimes decades. A new study published in NEJM AI on June 18, 2026, shows that AI diagnosis of rare childhood diseases may be getting closer to reality, after researchers used OpenAI’s o3 Deep Research model to crack 18 previously unsolved cases.
The study, carried out by researchers from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University, and OpenAI, involved reanalyzing 376 cases that had already gone through extensive specialist review with no conclusion. The model was given de-identified packets of each patient’s clinical features, genomic data, and family history, then asked to propose the most plausible molecular explanation and show its reasoning. After expert review, additional testing, and clinical confirmation in a certified laboratory, physicians established diagnoses in 18 of those cases, a 4.8% additional diagnostic yield on top of what years of prior expert analysis had produced.
To be clear about what happened here: the AI did not diagnose anyone. Every result still required qualified clinicians to evaluate the evidence, order appropriate tests, and confirm findings. What the model did was act as a powerful reasoning layer, connecting scattered genomic signals, clinical notes, inheritance patterns, and scientific literature into reviewable hypotheses that human experts could then interrogate.
The AI diagnosis of rare childhood diseases was tested across four patient groups: children with neurodevelopmental conditions (10% yield), people with rare neuromuscular disease (6.6% yield), cases of sudden unexpected death in pediatrics (1% yield), and children and adolescents with early psychosis (13.3% yield). One of the 18 solved cases involved Kyra, a woman who had spent nearly two decades without a diagnosis after her mother first noticed muscle weakness when she was nine years old. The AI workflow helped link her condition to a variant in HSPB8, pointing to a form of myofibrillar myopathy. A genetic counselor called her just before her 28th birthday with the answer.
The model also demonstrated unexpected flexibility. In one early psychosis case, it inferred a structural deletion in chromosome 22 that was not listed in the input data at all, connecting low-quality sequencing signals to the patient’s cardiac, immune, and psychiatric features. Follow-up testing confirmed the deletion, a finding associated with DiGeorge syndrome. In another case, it flagged a possible novel mechanistic link between a variant in the gene S1PR1 and vitiligo, a connection that still requires experimental validation but illustrates how AI can pull together clues from structural biology and immunology into concrete, testable hypotheses.
The Manton Center will now lead the next phase of the work, supported by a grant from the OpenAI Foundation, aimed at building a platform-agnostic AI copilot that can help clinical teams analyze rare disease cases more quickly and consistently. As Alan Beggs, director of the Manton Center, put it: “Researchers like Catherine and me can’t possibly keep 8,000 different diseases in our heads. That’s the power of AI.”
