Dr. House, mean Dr. AI.
Online chatbots have been known to dole out some pretty questionable health advice — which, at its worst, has sent people to the ER. But in the right hands — that is, doctors’ hands — the tools have proven to be able to accomplish what doctors can’t on their own.
Now, new research from Boston Children’s Hospital has found that AI tools were able to diagnose 18 children with previously undiagnosed diseases.
With the help of these tools, children with rare illnesses and symptoms may now find answers in a much shorter time frame than before.
Published in NEJM AI, the research from the hospital’s center for rare diseases and OpenAI revealed that the tech company’s tools could identify errors in patient genomes.
Using OpenAI’s o3 Deep Research model, the team was able to analyze several hundred genomes of patients who hadn’t received a diagnosis for a rare disease — and found almost 5% of new diagnoses.
When done by humans, analyzing thousands of patient data to find the genetic cause of a disease can typically take several days, since there are around 20,000 protein-coding genes in the human genome.
The team ran the genomes of 376 patients who lacked diagnoses through the o3 system, along with clinicians’ notes about the case, a description of the patient’s symptoms, and a filtered list of certain genes that might be responsible.
New diagnoses were identified for 10 patients with rare neurodevelopmental diseases, four patients with neuromuscular disorders and two patients with early childhood psychosis illnesses.
The diseases of two children who had died suddenly without further specification were also identified.
“It’s a game changer,” Catherine Brownstein, lead study researcher, told NBC. “Considering how many times these had already been analyzed, that’s a huge number, and each one means an answer for a family.”
Searching the vast dataset of genomes is comparable to trying to find a needle in a haystack, with Brownstein noting that there are “pages upon pages of these genes that I have to get through for a case, while the LLM doesn’t get tired.”
Previous research has looked at similar topics of using large language models to search genomes for genes that have been sequenced in patients with rare diseases.
Several other health experts are applauding the findings of the most recent study, saying it will help doctors across the country find answers faster using commercial AI tools.
However, many also caution that the systems still need thorough human review.
While more people are turning to an AI model for health advice — despite bogus information — tech CEOs and experts warn it can be dangerous.
“Keeping humans in the loop isn’t optional — it’s the safeguard that protects lives,” Andy Kurtzig, CEO of the AI-powered search engine Pearl.com, previously told The Post.
While OpenAI expressed enthusiasm around the findings, the research team stressed that the findings are not a remedy for all unidentifiable diseases and that consumers shouldn’t use LLMs to self-diagnose.
Rather, the tools can be used by doctors to navigate complex medical information that would take a considerable amount of time to discern.
