A study published June 18, 2026 in NEJM AI showed that researchers from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University and OpenAI used OpenAI’s o3 Deep Research reasoning model to re-analyze 376 previously unsolved pediatric rare disease cases. After expert review, additional testing and clinical confirmation, physicians established new diagnoses in 18 of those cases, an additional diagnostic yield of 4.8% on top of what specialists had already found through conventional analysis.
The Diagnostic Odyssey Problem
Families of children with rare genetic diseases often endure what specialists call a “diagnostic odyssey,” years of unexplained symptoms, repeated testing and specialist visits without a clear answer. The Manton Center for Orphan Disease Research at Boston Children’s Hospital specializes in exactly these unsolved cases, where standard genomic and clinical workflows have already been exhausted without identifying a cause. Even a modest additional diagnostic yield in this population represents a meaningful outcome for families who may have been searching for answers for years.
How the AI Model Was Used
The research team fed de-identified clinical and genomic information from the 376 unsolved cases into OpenAI’s o3 Deep Research model, a reasoning-focused AI system designed to work through complex, multi-step problems rather than simply retrieving information. The model’s role was to generate diagnostic hypotheses that human specialists could then investigate further through additional testing and clinical confirmation, rather than to issue diagnoses independently. All 18 new diagnoses required expert physician review and follow-up testing before being confirmed.
Part of a Broader Wave of AI Rare Disease Tools
The Boston Children’s Hospital and OpenAI collaboration arrives alongside other notable rare disease AI breakthroughs in 2026, including EvORanker, a new algorithm that compares genetic evolutionary patterns across more than 1,000 species to uncover previously unrecognized gene-disease relationships. Using EvORanker, researchers identified a gene not previously linked to a particular disease as a likely cause, opening a path toward better understanding and potential treatment. Together, these efforts reflect a broader push to apply large-scale computational tools to rare disease diagnosis, an area where the sheer rarity and diversity of conditions has historically made pattern recognition difficult even for experienced specialists.
The Limits of the Approach
Researchers involved in the study were careful to frame the AI model’s contribution as hypothesis generation rather than autonomous diagnosis, noting that every AI-suggested lead still required substantial expert validation before being confirmed. Genetics researchers writing more broadly about AI in pediatric rare disease diagnosis have flagged persistent challenges including data standardization across institutions, model interpretability, and the risk of algorithmic bias when training data does not adequately represent the full diversity of rare disease presentations across different populations.
Why an 18-Case Result Matters
Some observers might view an additional yield of under 5% as modest, but rare disease specialists note that these 376 cases had already been through exhaustive expert review without success, meaning any additional diagnoses represent genuinely difficult cases where conventional analysis had reached its limits. For the 18 families involved, a new diagnosis can open access to targeted treatments, clinical trials, and support communities that were previously out of reach without a named condition to search for.
What’s Next
The research team says it plans to explore applying similar reasoning-model approaches to larger cohorts of unsolved rare disease cases, potentially at other specialized centers beyond Boston Children’s Hospital. As reasoning-focused AI models continue to improve, researchers in the field expect this kind of AI-assisted second-look analysis to become a more standard part of the diagnostic workflow for the hardest, most persistent unsolved pediatric cases.
The Emotional Weight of a New Diagnosis
Genetic counselors who work with families in the Manton Center’s program describe the moment a diagnosis is finally confirmed, after what is often years of uncertainty, as profoundly significant even when the underlying condition has no cure, since a named diagnosis often opens access to patient advocacy communities, natural history studies and, in some cases, experimental treatment protocols that were previously inaccessible without a specific genetic answer. Researchers say this human dimension is part of why even a modest additional diagnostic yield from AI-assisted analysis is considered a meaningful success within the rare disease research community. The Manton Center team says it plans to publish its full methodology so that other rare disease research centers can replicate the approach on their own backlogs of unsolved cases, potentially multiplying the diagnostic yield demonstrated in this initial 376-case cohort.
