Boston Children’s Hospital researchers used OpenAI’s o3 model to help resolve 18 long-running pediatric rare-disease cases, according to a study in NEJM AI. The system analyzed 376 undiagnosed genomes alongside clinical notes and symptom profiles, yielding about a 5% diagnostic rate across neurodevelopmental, neuromuscular, sudden-death, and early-childhood psychosis cases. Investigators emphasized that all outputs received human review and that seven findings were “rediscoveries,” underscoring gaps in data sharing. Outside experts called the yield meaningful for clearing case backlogs but cautioned that trust, validation and oversight remain essential. OpenAI provided support for the project, and researchers stressed that consumer self-diagnosis with chatbots is inappropriate even as the tools help clinicians triage complex genomic data.
Related articles:
— Recognition of facial phenotypes of genetic disorders using deep learning (DeepGestalt)
— International evaluation of an AI system for breast cancer screening
— Highly accurate protein structure prediction with AlphaFold
— Large language models encode clinical knowledge
— Ethics and governance of artificial intelligence for health
