OpenAI study helps solve 18 rare pediatric disease cases

An OpenAI-assisted reanalysis of 376 previously unsolved pediatric rare-disease cases produced 18 confirmed diagnoses, a result that points to real clinical utility without turning the model into a bedside diagnostician. The company said its o3 Deep Research reasoning model helped surface evidence-linked leads from de-identified clinical and genomic records, then left every medical decision to specialists at Boston Children’s Hospital and Harvard University.

The study, reported in NEJM AI, matters because rare-disease care often stalls even after extensive genetic testing and expert review. OpenAI said roughly half of people with rare diseases still go undiagnosed, in part because clues are scattered across dense records, millions of variants and a fast-changing scientific literature. In this case, the model did not issue diagnoses or recommend treatment. It generated candidate explanations for clinicians to review, and physicians pursued additional testing before confirming the answers through established medical processes.

OpenAI described the outcome as an extra diagnostic yield of 4.8% beyond earlier specialist analysis. That is a modest percentage, but in rare-disease medicine, a single diagnosis can end years of uncertainty for one family and point clinicians toward a specific management plan, surveillance strategy or inherited risk.

The results also temper the hype. One secondary report said seven of the 18 diagnoses were rediscoveries of known variants, a reminder that the value of AI in this setting may lie less in discovering wholly new biology than in helping doctors reconnect old data with newer evidence. The model’s job was not to replace expert judgment but to accelerate the reanalysis of hard cases that had already passed through standard genetic testing and specialist review without an answer.

Boston Children’s Hospital’s Manton Center for Orphan Disease Research, which took part in the collaboration, focuses on rare genetic conditions and families affected by them. The hospital has separately said it has helped diagnose more than 40 rare disease cases through AI-enabled workflows in its broader clinical environment, suggesting this study is part of a larger push to make rare-disease reanalysis more scalable as gene-disease knowledge keeps expanding.

For now, the clearest lesson is institutional rather than futuristic: AI can help narrow the search, but physicians still decide which lead is real. In rare pediatric genomics, that distinction is the difference between hype and a usable clinical tool.