Researchers at Boston Children’s Hospital and Harvard used OpenAI’s o3 Deep Research model to identify 18 new diagnoses among children with rare genetic diseases that had previously gone unresolved, according to a study published in NEJM AI. The team reanalyzed 376 de-identified pediatric cases that had undergone genetic testing and expert review without yielding a diagnosis.
Using the o3 Deep Research model, the clinicians synthesized genetic data, clinical presentations, and medical literature. Diagnoses identified from this analysis included neurodevelopmental disorders, rare neuromuscular diseases, sudden unexpected death in pediatrics, and early-onset psychosis.
The tool is designed to assist clinicians and researchers in navigating complex medical information rather than serve as a direct consumer diagnostic product. John Brownstein, Chief Innovation Officer at Boston Children’s Hospital, noted the challenge in diagnosing complex cases is often due to cognitive limits. “We combine genetic information, phenotypic information, literature search, and the reasoning of AI to deliver diagnoses to families that were once left without any answers,” Brownstein said.
The NEJM AI publication is part of a broader initiative at the hospital, which announced in May that its AI efforts had led to over 40 rare disease diagnoses previously deemed unsolvable. OpenAI has committed $50 million to support Boston Children’s AI initiatives, which began in early 2025. The hospital reports that more than one-third of its employees now use AI tools in their daily work, resulting in approximately 60,000 hours saved, valued at over $7 million.
Rare diseases affect an estimated 300 million people globally, with many families navigating lengthy diagnostic processes. “This was unthinkable before, but is now providing hope to so many families,” Brownstein added. The researchers emphasized that each diagnosis still requires clinician verification and interpretation, demonstrating that AI serves to augment, rather than replace, human expertise.
