AI Cracked 18 Rare Disease Cases Doctors Couldn't Solve
Eighteen children have been diagnosed with rare diseases they might never have identified. Their cases had already been reviewed and filed away as unsolvable.
The diagnoses came from OpenAI's o3 model — the same off-the-shelf tool available to any developer through an API subscription. No custom clinical build. No proprietary hospital system. Just a general-purpose AI given access to 376 patients' genomic data at Boston Children's Hospital.
The 5% That Changes Everything for 18 Families
The research, published Thursday in NEJM AI — the New England Journal of Medicine's dedicated AI publication — found that o3 correctly identified disease-causing genetic variants in 18 of those 376 cases, a yield of just under 5%.
That number sounds modest. It is not.
Every one of those genomes had already been analyzed by specialist teams at one of the world's leading pediatric hospitals. These were not first-time referrals. They were the hardest cases — already reviewed, already unexplained, already waiting. According to NBC News, the cases were specifically chosen because they had stumped the hospital's own geneticists.
Catherine Brownstein, scientific director of the genetic investigations arm of the Manton Center for Orphan Disease Research at Boston Children's, said the team analyzed the genomes of patients who had received no diagnosis after standard review. A 5% yield on pre-screened, hardest-possible cases is a different number than 5% on a general patient population.
📰 Related: World's First AI-Designed Universal Vaccine Passes Human Trial at Cambridge
How o3 Found What Specialists Missed
The research team ran each genome through o3 in a structured workflow, asking the model to reason through connections between the patient's symptoms, genetic variants, and published medical literature — using only information publicly available to any researcher.
The AI did not discover new science. It connected existing science faster than human reading speed allows.
Rare diseases affect roughly 300 million people globally, according to the National Organization for Rare Disorders. Most involve variants so uncommon that no single clinician has encountered them before. A genome analysis might flag hundreds of candidates — ranking them accurately requires synthesizing thousands of papers. According to OpenAI's published case study, Boston Children's has now embedded AI across more than 50 operational workflows, saving over 60,000 staff hours and redirecting more than $7 million in labor costs toward direct patient care.
📰 Related: KPMG AI Report Exposes a Bigger Problem Inside Corporate AI
The Diagnostic Odyssey Has a Measurable Cost
Families of children with undiagnosed rare diseases face what researchers call the diagnostic odyssey — years of inconclusive tests and the exhaustion of watching a child remain unwell while medicine produces no answers.
The median time to a rare disease diagnosis exceeds five years. Roughly half of rare disease patients never receive a confirmed one in their lifetime.
The Boston Children's study does not propose replacing clinical geneticists. It demonstrates a tool that runs faster, reads wider, and does not fatigue — applied to cases that have already defeated human review.
John Brownstein, Chief Innovation Officer at Boston Children's, described the core problem as a cognitive limit rather than a knowledge gap. The data exists. The variants are in the literature. The constraint is processing speed.
📰 Related: Anthropic CEO Dario Amodei Calls for Government Power to Block Dangerous AI Models
What Comes Next for AI in Rare Disease Medicine
The NEJM AI study is a proof of concept, not a clinical protocol. The research team has been careful to frame the findings as a first demonstration that general-purpose AI can contribute meaningfully to genomic diagnosis — not a finished product ready for deployment.
The 376 genomes at Boston Children's represent a single institution with specific clinical infrastructure. Whether o3 performs comparably at hospitals without that depth of genomic data remains untested.
If a general-purpose AI can identify diagnoses that eluded expert teams at one of the world's best-resourced pediatric hospitals, what it means for rare disease patients seen at smaller centers — with fewer specialists and less data — is a question medicine has not yet answered.
Key Takeaways
- OpenAI's o3 model identified 18 new rare disease diagnoses from 376 previously unresolved cases at Boston Children's Hospital.
- The study was published June 18, 2026 in NEJM AI, the New England Journal of Medicine's AI-focused publication.
- All 376 genomes had already been reviewed and cleared by specialist teams — o3 succeeded where human experts had exhausted their options.
- Catherine Brownstein (Manton Center, Boston Children's) led the genomic research; John Brownstein (Chief Innovation Officer) oversaw the hospital's broader AI integration.
- Boston Children's AI deployment has saved over 60,000 staff hours and redirected more than $7 million in labor costs.
- The study is a proof of concept, not a clinical deployment — reproducibility at smaller, less-resourced hospitals remains unproven.
Sources
Also Read
Tech & AI Editor
David Park covers artificial intelligence, Big Tech, and the future of digital innovation. He translates complex tech developments into stories that matter for everyday readers.
