AI Finds Rare Disease Causes
AI Tools Diagnose 18 Children Whose Rare Diseases Stumped Doctors for Years
A collaboration between Boston Children's Hospital and OpenAI has cracked 18 cases that had stumped doctors for years. Using off-the-shelf AI tools — not specialized medical software — researchers identified the genetic errors causing rare illnesses in children who had no diagnosis. The findings were published in NEJM AI, the New England Journal of Medicine's AI-focused journal.
The 18 newly diagnosed children include 10 with rare neurodevelopmental diseases and 4 with neuromuscular disorders. Two cases involved children who had died suddenly — the AI identified the genetic cause for their grieving families. Every AI-suggested diagnosis was confirmed by human geneticists, according to NBC News.
Years of Searching Ended by AI in One Pass
For rare disease patients, the wait for answers is brutal. On average, families spend 5 to 7 years bouncing between specialists, collecting 2 to 3 wrong diagnoses along the way. The problem is not finding DNA errors — modern sequencing can scan an entire genome. The problem is knowing which of the millions of variants actually causes the illness.
That interpretation bottleneck is exactly what the AI tackled. The Manton Center for Orphan Disease Research at Boston Children's Hospital works with more than 3,500 patients across all 50 states. It used GPT-4-based tools to rank overlooked genetic variants by reading clinical notes alongside obscure medical literature — work that would take a human specialist far longer, according to NBC Connecticut.
Off-the-Shelf AI, Not a Custom Medical Tool
One of the study's most striking findings is what the AI was not. It was not a purpose-built medical system trained on patient records. OpenAI's team used general-purpose large language models with no special fine-tuning for genetics. The AI simply read doctors' unstructured notes and matched them against patterns in medical literature it had already learned.
Dr. Catherine Brownstein, co-director of the Manton Center's Gene Discovery Core, stressed that the AI surfaced leads — humans made the final calls. Every diagnosis the AI flagged was verified by board-certified geneticists using Sanger sequencing, a standard DNA confirmation method. The AI acted, in one description, as a "tireless, ultra-well-read intern," according to NBC Boston.
A Low Hit Rate That Still Carries Huge Weight
Critics note that 18 diagnoses from a pool of 3,500 patients is a yield of roughly 0.5%. Some medical journals, including The Lancet, have argued the AI helps only in the most extraordinary cases and may not move the needle for most undiagnosed patients. That is a fair point — but for those 18 families, the number is everything.
For rare disease advocates, the framing is about closure. Organizations like NORD — the National Organization for Rare Disorders — point out that a diagnosis is not just a label. It can unlock access to targeted clinical trials and end years of emotional uncertainty. For the families of the two children who died, the AI gave them an answer no doctor ever could, according to NBC Philadelphia.
Privacy Concerns Shadow an Otherwise Hopeful Result
Not everyone is celebrating without reservation. Bioethicists have raised the liability question: if an AI points to a diagnosis that leads to a risky treatment and something goes wrong, who is responsible — the doctor or the AI company? Groups like the Electronic Frontier Foundation have also flagged the risk of sending sensitive genomic data to private AI systems, warning it could sidestep HIPAA protections if not managed carefully.
The economic case for moving forward is powerful. Shortening the rare disease diagnostic process by 50% could save the U.S. healthcare system more than $10 billion a year by cutting redundant tests and unnecessary hospital stays. The research is already fueling calls for legislation that would require insurers to cover AI-assisted genomic reviews, according to NBC Chicago.
Sentiment
NeutralPublishers
11
Articles
11
Reach
11
