OpenAI o3 helps Boston Children's Hospital diagnose 18 previously unsolved rare genetic diseases in children

Researchers at Boston Children's Hospital have used OpenAI's o3 reasoning model to identify new diagnoses for 18 children whose rare genetic diseases had gone unresolved by standard clinical workup. The results were published June 18 in NEJM AI, the New England Journal of Medicine's AI-focused journal.

What's new

The research team analyzed 376 pediatric patients with previously undiagnosed rare diseases. Using OpenAI o3 alongside clinician notes, patient symptom data, and curated lists of gene candidates, the system surfaced diagnoses in 18 cases that clinicians had not resolved through conventional means:

• 10 children with neurodevelopmental conditions
• 4 with neuromuscular disorders
• 2 whose conditions were linked to sudden death
• 2 with early childhood psychosis

According to OpenAI, the Boston Children's partnership has now produced more than 40 previously unsolved rare disease diagnoses using AI tools, saved 60,000 hours of clinical work, and redirected more than $7 million in labor costs to other research priorities.

Context

Boston Children's is among the world's leading pediatric research hospitals and has been pursuing AI-assisted genetic diagnosis for several years. The diagnostic challenge in rare disease is particularly steep: many conditions affect fewer than a few hundred patients globally, making it nearly impossible for any individual clinician to maintain comprehensive knowledge. Rare diseases collectively affect an estimated 300 million people worldwide, and the majority are genetic in origin.

OpenAI's o3 model, which uses extended chain-of-thought reasoning, is well-suited to this type of structured medical reasoning: cross-referencing patient phenotype data against a large space of possible genetic etiologies, ranking candidates, and flagging diagnostic hypotheses for physician review. The model does not issue diagnoses autonomously — it surfaces candidates that the research team then evaluates.

Why it matters

The publication in NEJM AI gives this work significant scientific weight — NEJM is among the most selective journals in medicine, and placement there signals the research team met rigorous peer review standards for clinical AI evaluation.

Beyond the immediate diagnostic outcomes, the 18 new diagnoses represent a ceiling-lift, not a speedup. These were cases where the diagnostic answer was out of reach, not merely slow to arrive. That distinction matters: AI tools that find answers clinicians couldn't reach independently open qualitatively new capabilities in clinical medicine, rather than just accelerating existing workflows.

The combination of o3's extended reasoning with structured genomic analysis pipelines may serve as a template for other rare disease programs at academic medical centers. If the approach generalizes, it could substantially reduce the median time to diagnosis — currently measured in years for many rare genetic conditions — for patients and families who often exhaust conventional options before an answer arrives.