o3 from OpenAI solves 18 rare diseases!

Using OpenAI’s o3 model, Boston Children’s Hospital identified 18 novel diagnoses in 376 kids with previously undiagnosed rare diseases, according to a paper published Thursday in the New England Journal of Medicine’s AI journal. Researchers at Boston Children’s Hospital’s Manton Center for Orphan Disease Research used the o3 model, the most powerful at the time they conducted their study last year, on 376 patients with no diagnosis and got answers for nearly 5% of them.

It’s a straightforward process. They inputted the clinical notes, symptom information, and filtered set of genes into the o3 model, and asked it to identify the gene involved. All responses were checked by human researchers before a diagnosis was confirmed. The 18 novel diagnoses included 10 rare neurodevelopmental disorders, 4 neuromuscular conditions, 2 cases of sudden unexpected death, and 2 cases of early onset psychosis.

1Why a 5% hit rate is a big deal

5% might seem low, but keep in mind that these genomes had been examined multiple times before and yielded nothing. There are around 20,000 protein coding genes in the human genome, and pinpointing the single relevant variant can require days of work from an expert analyst, something that hospitals don’t usually have. “I have pages and pages of these genes that I need to go through for a particular patient, whereas the LLM won’t get tired,” said Catherine Brownstein, one of the lead authors of the study. “This is a total game changer. Each one of these means an answer for a family.” According to an outside expert, Beth Israel’s Adam Rodman, a 5% diagnostic rate is “truly meaningful” and could alleviate major backlogs of undiagnosed patients.

2Why this matters for you

A diagnosis can be life changing for families without one. Even if there is no treatment available, it takes away years of uncertainty and ensures patients receive treatment as soon as possible once it becomes available. One person discovered she had myofibrillar myopathy about 15 years after her symptoms started thanks to this research. More broadly, this is an out-of-the-box model being applied to medical problems, so any hospital can potentially use this method to accelerate reanalysis. However, the researchers were careful not to hype this too much, since they still had humans check all the results, 7 of the 18 diagnoses were “rediscovered” (i.e. known to a clinic but not reported), and a diagnosis is just the beginning of treatment. As OpenAI’s head of health Ashley Alexander noted, “We definitely don’t want to overhype this.” I think the modesty here makes this more believable — it’s AI performing grunt work and prioritizing high value tasks, not actually being a doctor.