AI cracks cold cases as OpenAI o3 diagnoses 18 kids with rare diseases
While tech bros use AI to generate pictures of the Pope in a puffer jacket, scientists just used a supposedly "outdated" model to solve medical mysteries that baffled human doctors for years. Here is how it actually worked.
Researchers from Boston Children's Hospital and Harvard University teamed up with OpenAI to re-examine 376 unsolved pediatric genetic cases. They threw these medical dead-ends at the o3 Deep Research model, which managed to crack 18 new diagnoses—about a five percent success rate for cases where human experts had completely run out of ideas.
The process wasn't just magic; they fed the AI anonymized patient profiles containing symptoms, doctor notes, and DNA sequences. Because medical data is notoriously messy, the AI acted like a digital Sherlock Holmes, comparing these files against public databases to figure out if genetic mutations were inherited or completely fresh anomalies.
Interestingly, the success rates varied wildly depending on the condition. While the AI diagnosed ten percent of neurodevelopmental disorders and over thirteen percent of early-onset psychosis cases, it struggled with sudden unexplained deaths, scoring only a single percent.
It turns out seven of the eighteen discoveries weren't even hidden mutations, but rather administrative failures. The crucial data already existed in public databases, but because hospital software is famously terrible at talking to other software, human doctors simply couldn't connect the dots.
In one bizarre case of early-onset psychosis, the AI noticed poorly read segments on chromosome 22 and correctly deduced that a whole chunk of the genome was missing. It diagnosed the child with DiGeorge syndrome, a prediction that was later confirmed by actual DNA re-sequencing.
It turns out the biggest barrier to saving lives isn't a lack of medical knowledge, but the nightmare of unorganized, siloed data that no human has the patience to sift through. If a glorified, outdated chatbot can outperform medical boards simply by reading the manuals, perhaps the entire healthcare system needs a hard reboot.
Source: OpenAI
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.