AI Diagnosed 18 Rare Pediatric Diseases That Stumped Doctors for Years

For families of children with rare genetic diseases, the path to a diagnosis is often measured in years, not days. Multiple specialists, repeated tests, and endless dead ends define what clinicians call the "diagnostic odyssey."

A new peer-reviewed study published in NEJM AI offers a glimpse of a faster path. Researchers at Boston Children's Hospital's Manton Center for Orphan Disease Research, Harvard University, and OpenAI used the o3 Deep Research reasoning model to reanalyze 376 de-identifed pediatric cases that had already gone through standard genetic testing and expert review without yielding a diagnosis . The result: 18 new confirmed diagnoses for children with previously unsolved rare genetic conditions .

How the NEJM AI Study Worked

Rare disease diagnosis is extraordinarily difficult. Genomic sequencing can surface millions of genetic variants, and medical knowledge about these variants changes constantly . The traditional workflow — a human expert manually reviewing candidate variants against the latest literature — is slow, painstaking, and limited by how much one person can hold in working memory.

To address this, the research team fed 376 de-identifed cases into OpenAI's o3 Deep Research model. The AI was tasked with connecting clinical features, inheritance patterns, variant evidence, and the most up-to-date scientific literature into evidence-linked candidate explanations. Every lead the model surfaced then went through human adjudication: expert review, additional testing, and clinical confirmation .

The conditions identified spanned neurodevelopmental disorders, rare neuromuscular disease, sudden unexpected death in pediatrics, and early-onset psychosis . Many of these cases had eluded years of expert analysis.

Chief Innovation Officer John Brownstein acknowledged the modest yield. "Achieving nearly 5% new diagnoses may not sound substantial," he told NBC News, "but considering the frequency of prior analysis and expertise applied, it's actually quite significant" .

Beyond the Study: AI as Core Infrastructure at Boston Children's Hospital

This NEJM AI study is not an isolated experiment. It is part of a much larger enterprise-wide AI transformation at Boston Children's Hospital that the institution describes as treating AI as "core infrastructure" .

Key results from the broader initiative alongside OpenAI:

• Over 40 rare conditions diagnosed that had previously gone unresolved .
• $7 million in cost savings and 60,000 hours of staff time saved across clinical, research, and operational teams .
• More than a third of employees now use AI tools integrated into daily workflows, spanning a secure internal ChatGPT environment used across research, clinical care, and operational areas such as supply chain and surgical scheduling .

The hospital reports that AI has reduced administrative burden, cut costs, expanded clinical capacity, and made it possible to tackle diagnostic cases once thought impossible .

The $50 Million Commitment from OpenAI

The $50 million figure that frequently appears alongside this news has two components.

First, in March 2025, OpenAI launched the NextGenAI consortium with 15 research institutions (including Harvard) and $50 million in funding for research grants, compute access, and API resources . Boston Children's Hospital has been a direct beneficiary. Chief Innovation Officer John Brownstein publicly credited the $50 million commitment for advancing their rare disease AI work in a March 2025 LinkedIn post .

Second, in early 2026, the OpenAI Foundation committed an additional $50 million to its People-First AI Fund, part of a broader commitment to invest at least $1 billion over the next year across life sciences and disease research .

What This Means for the Global Rare Disease Community

An estimated 300 million people worldwide suffer from rare diseases . Many endure a years-long diagnostic odyssey: visiting multiple specialists, undergoing repeated testing, and often never receiving a definitive answer. The emotional and financial toll on families is immense.

The NEJM AI study demonstrates that AI reasoning models can surface diagnoses from existing genomic data that human experts missed — potentially collapsing years of searching into weeks or days . The approach is scalable, reproducible, and improves as medical knowledge grows. The implication is that AI-assisted genomic reanalysis could become a standard second-line tool, dramatically reducing the diagnostic burden for millions of families.

A 5% diagnostic yield from previously exhausted cases may seem small. But in the world of rare disease, where each percentage point represents real children who finally get answers and treatment plans that can actually begin, the impact is immediate — and the trajectory is only accelerating.