Overview
- A paper published June 18 in NEJM AI reports that researchers at Boston Children’s, Harvard and OpenAI ran 376 previously unsolved pediatric genomic cases through OpenAI’s o3 model as part of an expert‑led workflow.
- The model generated evidence‑linked hypotheses that human experts reviewed and then validated by follow‑up testing, producing 18 confirmed diagnoses for an additional diagnostic yield of about 4.8%.
- Diagnoses spanned four cohorts — 10 neurodevelopmental, four neuromuscular, two early psychosis and two sudden unexpected pediatric death cases — and included a confirmed chromosome 22 deletion the model had flagged.
- The team found seven of the 18 were rediscoveries that were not in the records the researchers reviewed, highlighting gaps in record integration and the need to reconcile fragmented clinical data.
- Authors and outside experts warned the study is retrospective, did not measure time, cost or patient outcomes, and stressed that LLM outputs can be plausible but wrong, so prospective multicenter trials and rigorous human oversight are needed before clinical rollout.