An OpenAI model helps reopen undiagnosed rare disease cases

An OpenAI o3 Deep Research model helped Boston Children's Hospital diagnose 18 rare genetic disease cases, achieving a 4.8% yield on previously unsolved cases.

A study published in NEJM AI, conducted by the Manton Center at Boston Children's Hospital, Harvard, and OpenAI, tasked an OpenAI reasoning model, o3 Deep Research, with the reanalysis of 376 previously examined but undiagnosed cases of rare genetic diseases in children. Using de-identified clinical and genomic data, the model surfaced candidate explanations supported by evidence. After review by specialists, additional tests, and laboratory confirmation, physicians made a diagnosis in 18 cases, representing a 4.8% gain on cases that experts had been unable to resolve.

The model's role remains circumscribed: it neither diagnosed anyone nor made clinical decisions. It generated interrogable hypotheses, linking phenotype, mode of inheritance, variants, and literature, which reviewers evaluated using the ACMG/AMP framework for laboratories, before confirmation and reporting back to families. The logic is based on a simple observation: the genome doesn't change, but knowledge surrounding it evolves, so an old case can become decipherable again.

OpenAI deems the yield modest but significant, for cases already extensively reviewed, with significant variations from one cohort to another. Seven of the eighteen diagnoses were in fact known elsewhere but absent from the consulted record, and the model sometimes opened new avenues, including a hypothesis on vitiligo yet to be validated. The study remains retrospective and did not measure time, cost, or false positives; OpenAI specifies that it does not validate any use of its models for making a diagnosis. The prospective follow-up will be conducted by the Manton Center thanks to a grant from the OpenAI Foundation.