o3 Deep Research surfaces 18 diagnoses across 376 rare cases

The practical pressure point in rare-disease medicine is not only sequencing; it is returning to old results after new gene-disease links appear. OpenAI said work with Boston Children's Hospital and Harvard, published in NEJM AI, used o3 Deep Research to help clinicians reanalyze 376 previously unresolved pediatric cases and surface leads that became 18 new diagnoses after expert review and clinical confirmation.

In the source tweet, OpenAI described the study as o3 Deep Research helping clinicians revisit "previously unsolved rare pediatric disease cases." A follow-up emphasized that every result went through human adjudication and clinical confirmation, framing the model as a reasoning aid for fragmented phenotypes, variants, and literature rather than an autonomous diagnostician.

That distinction matters. Rare-disease families often spend years in a diagnostic odyssey, and a negative result from an earlier genome analysis can become newly informative when the literature changes. Periodic reanalysis is valuable, but it is labor-intensive. A model that can propose ranked hypotheses, cite literature, and connect clinical features to variants could make that loop more scalable if hospitals can audit the evidence trail.

The linked NEJM AI paper says the model was used through the OpenAI API or ChatGPT, while clinicians applied ACMG/AMP interpretation rules and confirmatory testing. The next thing to watch is not a headline accuracy number alone. The more important benchmark is operational: how many cases can be safely reopened, how often the system misses plausible leads, and whether hospitals can preserve accountability when AI compresses weeks of literature review into a much shorter clinical workflow.