OpenAI's o3 Cracks 18 Rare Disease Cases

OpenAI's o3 model helped researchers at Boston Children's Hospital clarify 18 diagnoses for children whose rare diseases had gone unsolved for years, according to findings published Thursday in NEJM AI, the AI-focused journal of the New England Journal of Medicine. The team at the hospital's Manton Center for Orphan Disease Research ran the genomes of 376 patients who had no diagnosis through the o3 system — the most capable model available when the work was done last year — and came away with new answers for almost 5% of them.

The setup was deliberately simple. For each case, the researchers fed o3 the clinicians' notes, a description of the patient's symptoms and a filtered list of genes that might be responsible, then asked the model to find the genetic culprit. Every output was reviewed by the human research team before any diagnosis was made final. The 18 new diagnoses spanned four disease areas: 10 children with rare neurodevelopmental diseases, four with neuromuscular disorders, two who had died suddenly without explanation, and two with early childhood psychosis.

Why a 5% hit rate is a big deal

Five percent sounds small until you remember these genomes had already been analyzed, often many times, without a result. The human genome holds roughly 20,000 protein-coding genes, and tying a patient's symptoms to the one variant that matters can take a skilled analyst days per case — time that hospitals simply don't have. 'There's pages upon pages of these genes that I have to get through for a case, while the LLM doesn't get tired,' said Catherine Brownstein, who helped lead the study. She called the result 'a total game changer,' adding that 'each one means an answer for a family.' An independent expert, Beth Israel's Adam Rodman, said a 5% diagnostic yield is 'truly meaningful' and could help clear large backlogs of unsolved cases.

Why this matters for you

For families living without a diagnosis, an answer is everything — even when there's no cure yet, it ends years of uncertainty and puts patients first in line when a treatment does arrive. One patient learned she has myofibrillar myopathy roughly 15 years after her symptoms began, because of this work. The broader point is that this was an off-the-shelf model, not a bespoke medical system, which means hospitals everywhere could use the same tools to speed up reanalysis. But the researchers were careful not to oversell it: humans reviewed every result, seven of the 18 were 'rediscoveries' that some clinic already knew but hadn't shared, and a diagnosis is only the first step toward treatment. OpenAI's own health lead, Ashley Alexander, put it plainly: 'We definitely don't want to overhype this.' I think that restraint is exactly why the result lands — this is AI doing unglamorous, high-value triage, not playing doctor.