IP.com

AI cracks 18 medical mysteries that left specialists stumped: In a striking demonstration of what AI can do for medicine, researchers at Boston Children's and Harvard fed 376 unsolved pediatric genetic cases into OpenAI's o3 Deep Research model — every single one a dead end after exhaustive specialist review — and watched it surface leads that led doctors to confirm 18 brand-new diagnoses.

For each case, the team gave the model de-identified symptoms and a shortlist of suspect genes, then set it loose to weigh inheritance patterns, comb through public databases, and cross-reference the latest research. The result: 18 confirmed diagnoses across 376 cases — answers for nearly 5% of patients whose files had already exhausted conventional expertise.

What's especially telling is what surfaced in seven of those cases. The answer already existed somewhere — at another clinic or buried in a public database — it had simply never made its way into the patient's local records. The AI bridged a gap that human bandwidth alone couldn't.

Rare-disease medicine has a brutal backlog problem. Even after full genetic sequencing, roughly half of these cases go unsolved, and new files pile up far faster than clinicians can revisit old ones armed with fresh research. No doctor has the hours to continually re-examine every cold case or manually stitch together findings scattered across disconnected databases and clinics. But AI can — tirelessly, at scale, and across information silos that defeat even the most dedicated specialists. And remember: this was a year-old, general-purpose research model, not a bespoke medical system. It still gave 18 families answers they'd been told didn't exist. As these tools grow sharper and more specialized, the prospect of giving every "dead-end" case a genuine second chance moves from aspiration to reality. https://buff.ly/mSCHZeH