OpenAI's o3 Model Helps Boston Children's Hospital Diagnose 18 Children with Rare Diseases That Had Eluded Experts
核心洞察
Boston Children's Hospital researchers used OpenAI (搜索)'s o3 model to analyze genomes of 376 undiagnosed patients, yielding new diagnoses for 18 children across four disease areas.
The AI-assisted approach achieved a diagnostic yield of nearly 5%, which researchers called a "total game changer" given that these genomes had already been exhaustively analyzed by human experts.
Diagnoses included 10 neurodevelopmental disorders, 4 neuromuscular disorders, 2 sudden deaths, and 2 early childhood psychosis cases, with findings published in NEJM AI.
Researchers at Boston Children's Hospital, in collaboration with OpenAI (搜索), have demonstrated that off-the-shelf artificial intelligence tools can successfully identify elusive genetic diagnoses in children whose rare diseases had stumped specialists for years. The findings, published Thursday in the New England Journal of Medicine's AI-focused publication, NEJM AI, show that OpenAI's o3 model helped clarify 18 diagnoses among 376 patients who had long sought answers.
"It's a total game changer," said Catherine Brownstein, scientific director of the genetic investigations arm of the Manton Center for Orphan Disease Research (搜索) at Boston Children's Hospital and one of the study's lead researchers. "It got almost 5% new diagnoses, which doesn't sound like a lot, but considering how many times these had already been analyzed, that's a huge number, and each one means an answer for a family."
The Diagnostic Challenge of Rare Diseases
The Manton Center works with over 3,500 individuals globally and across all 50 states who are affected by rare diseases, partnering with hospitals and health centers around the world. Brownstein explained that the hospital routinely screens those patients' genomes against newly identified genes that might yield diagnoses, but the screenings often do not turn up any new answers.
Finding the genetic cause of a disease is inherently complex. There are approximately 20,000 protein-coding genes in the human genome, and while sequencing a patient's full genome is now straightforward, identifying clear cause-and-effect relationships in the messy data of a human genome is not always possible.
"A researcher can only spend so much time on a single case," said Suyash Shringarpure, a technical researcher at OpenAI (搜索) who focuses on health applications. "Maybe a case remained unsolved when it came to them first, but a year later a paper was published that clarifies the link between the gene and the disease."
How the AI-Assisted Analysis Worked
Conducting the research last year, Brownstein and the research team ran the genomes of 376 patients who lacked diagnoses through the o3 system, which was then the most powerful system available. To hunt for diagnostic clues in each genome, the researchers provided the o3 model with clinicians' notes about the case, a description of the patient's symptoms, and a filtered list of certain genes that might be responsible for the patient's symptoms. The human research team reviewed all of the system's outputs to make any final diagnosis.
Of the 376 cases spanning four different disease areas, the team identified new diagnoses for 10 patients with rare neurodevelopmental diseases, four patients with neuromuscular disorders, two children who had died suddenly without further specification, and two patients with early childhood psychosis illnesses.
Brownstein noted that there simply are not enough geneticists and analysts to find the biological needle in a genome-wide haystack that might cause a patient's symptoms. "There's pages upon pages of these genes that I have to get through for a case, while the LLM doesn't get tired," she said.
A Patient's Story: Kyra Benton
Kyra Benton was one of the individuals who received a diagnosis as a result of the new research. When Benton was nine, her mother became concerned that she seemed to move differently from her peers, starting to walk on her tiptoes and struggling to run with a normal gait. After years of worsening health, Benton visited Boston Children's Hospital, only to be told it also did not know the root of her disease. She soon faced severe heart problems and underwent a tracheotomy when she was only 13.
Benton said she came to terms with never knowing her diagnosis — until researchers from the Manton Center called her last year. "Last summer, about a week before my 20th birthday, we got a call from one of the researchers at the lab," Benton said. "She said, 'Hi, we know it's been about 15 years, but we have some news for you,' and it kind of just blossomed from there."
The AI-assisted analysis revealed Benton's condition to be myofibrillar myopathy (搜索), a progressive genetic neuromuscular disorder that causes muscle fibers to break down.
Expert Perspectives and Cautions
Adam Rodman, a doctor and expert on the use of AI in medicine at Beth Israel Deaconess Medical Center who was not involved in the research, called the new paper an exciting demonstration of AI systems' ability to diagnose diseases when used by doctors. "A diagnostic yield of 5% is truly meaningful and could serve as a significant screening tool to help speed up the reanalysis of significant backlogs of cases," he told NBC News.
Chunhua Weng, a professor of bioinformatics at Columbia University who was also not involved in the study, said the paper was a "wonderful" contribution to this area of research, though she — like the paper's research team — cautioned that LLM results still require rigorous human review. "The appropriate use of LLMs in diagnosis requires careful attention to trustworthiness," Weng said.
The paper also notes that seven of the Manton Center's identified diagnoses were actually "rediscoveries" — meaning a treatment team in one location had identified a patient's specific diagnosis but had not shared it with researchers around the world. Brownstein emphasized that even those rediscoveries were vital so that "when new treatments do come online, we can find the patients right away and make sure that they're first in line for any new technological development or any new therapy."
Broader Implications and Limitations
The research team was clear that its findings are not a panacea. Being diagnosed with a specific illness is often only an early step toward finding and then pursuing treatment options, and LLMs are not meant to be used by consumers to treat or diagnose diseases. Instead, the tools can help people and doctors navigate complex medical information.
"We definitely don't want to overhype this," said Ashley Alexander, head of health for OpenAI (搜索). "But I also want to make sure that people don't miss what's happening and what's possible with even just the version of ChatGPT that's in their pocket today."
For her part, Benton was surprised that AI was involved in the breakthrough. "Quite frankly, I'm the type of person that's not all that much favor of AI," she told NBC News. "On the other hand, I do acknowledge that it does have its advantages. Such as in this case, where it can lead to massive breakthroughs that can really change people's lives for the better."