Researchers at Boston Children’s Hospital’s Manton Center for Orphan Disease Research, working with OpenAI, reported in NEJM AI that OpenAI’s o3 model helped identify 18 new diagnoses among 376 patients whose rare diseases had gone unexplained despite prior genetic testing — a roughly 4.8% yield that the study’s lead researcher, Manton Center scientific director Catherine Brownstein, called “a total game changer” for a population of patients who have often already exhausted conventional diagnostic workups. One of those diagnoses went to Kyra Benton, who after roughly 15 years without an explanation for her symptoms was found to have myofibrillar myopathy, a progressive genetic neuromuscular disorder.
Why Rare Disease Diagnosis Takes So Long in the First Place
The human genome contains roughly 20,000 protein-coding genes, and linking a specific genetic variant to a patient’s specific set of symptoms is notoriously difficult even for experienced geneticists — a process that typically involves narrowing thousands of candidate variants down to a shortlist, then manually cross-referencing each against medical literature and clinical presentation. The Manton Center, which the hospital says works with more than 3,500 patients across all 50 states and internationally, exists specifically because so many of these patients cycle through specialists for years without answers; academic estimates have long put the average diagnostic odyssey for a rare genetic disease at several years from first symptom to confirmed diagnosis.
How the AI Model Was Actually Used
Researchers did not simply hand genomic data to o3 and accept its output. Instead, the model was given clinicians’ notes, a written description of each patient’s symptoms, and a pre-filtered list of candidate genes already narrowed down by conventional analysis — then asked to reason through which variant most plausibly explained the clinical picture. Every one of the model’s suggested diagnoses was reviewed by human researchers before being accepted; seven of the 18 total diagnoses turned out to be “rediscoveries,” meaning another medical team elsewhere had already identified the same diagnosis independently but had never shared that finding in a way the Manton Center’s own researchers could find it.
A Meaningful Yield, Not a Miracle
OpenAI’s head of health, Ashley Alexander, cautioned against overstating the results, saying the company “definitely” doesn’t “want to overhype this.” Outside researchers broadly agreed that the top-line number, while real, should be read in context: Adam Rodman of Beth Israel Deaconess Medical Center, who was not involved in the study, called a roughly 5% new-diagnosis yield meaningful precisely because these were cases that had already been analyzed by human experts and come up empty — suggesting the tool’s value may lie less in replacing genetic counselors than in cost-effectively re-screening the large backlog of unsolved cases that academic medical centers accumulate over years.
The Case for Caution
Chunhua Weng, a biomedical informatics researcher at Columbia University who was not involved in the study, said that using large language models for diagnosis “requires careful attention to trustworthiness” — a caveat the study’s own authors echoed by stressing that every AI-suggested diagnosis still required confirmatory human review before being delivered to a family. The researchers were also careful to note that a genetic diagnosis is frequently only the first step toward any treatment, not a cure in itself, and that OpenAI’s consumer-facing tools are explicitly not intended for patients to use to diagnose or treat disease on their own — a distinction the Manton Center’s controlled, clinician-supervised research process was specifically designed to preserve.
What’s Next
The Manton Center’s findings point toward a specific, near-term application rather than a sweeping one: using AI models to re-screen backlogs of previously unsolved genomic cases at academic rare-disease centers, where thousands of families are currently waiting with no diagnosis at all. Researchers say future work will need to test the approach on larger and more diverse patient cohorts, and to study how consistently the rediscovery problem — cases solved in one hospital system but never shared with another — might be reduced if AI-assisted re-analysis became a routine part of genetic medicine rather than a one-time research study.
Photo: Artem Podrez / PEXELS via Pexels
