Source: The Wall Street Journal — 2026-08-15
Summary
The Wall Street Journal profiles patients using AI to crack diagnoses that stumped their doctors: a mother who ran her son's photo through Face2Gene, a facial-analysis app, and got a strong match for trichorhinophalangeal syndrome that genetic testing later confirmed; and a 77-year-old Mayo Clinic patient whose ECG was flagged by AI at a 98% probability of cardiac amyloidosis, a diagnosis his care team hadn't reached. The piece cites a study where two AI chatbots correctly diagnosed 13% and 10% of 90 previously-solved complex rare-disease cases, versus 5.6% for doctors reviewing the same records.
Key Takeaways
- Face2Gene works by pattern-matching facial features against a database of known genetic-syndrome presentations — a task where AI's advantage is having effectively "seen" far more rare-syndrome cases than any individual clinician ever will in a career.
- The Mayo Clinic case shows AI applied to a routine test (an ECG) surfacing a specific, quantified risk (98% probability of cardiac amyloidosis) that the human care team's initial read didn't catch, in a patient whose symptoms had been attributed to something else.
- The cited study result — 13% and 10% correct diagnosis rates for two AI chatbots versus 5.6% for doctors on the same 90 previously-solved rare-disease cases — is a real head-to-head comparison, not a vendor claim, though all three numbers being low underscores how genuinely hard these cases are for anyone, human or AI.
- The pattern across both cases is patients or clinicians turning to AI specifically because rare diseases are, by definition, outside what any one doctor is likely to have personally encountered before — which is exactly the class of problem broad pattern-matching over large datasets is suited to.
Reel Script
Hook (18s, ~40 words): A mother uploaded one photo of her son's face to an app and got a rare genetic disease diagnosis doctors had missed. A 77-year-old's routine heart scan got flagged by AI with a 98% probability of a disease no one was looking for.
Core Concept (75s, ~170 words): Rare diseases are hard to diagnose for a structural reason, not a competence one: any individual doctor might see a given rare condition once or twice in an entire career, so there's no accumulated pattern recognition to draw on. That's precisely the kind of problem AI is well suited to, because it can be trained on far more rare-case examples than any one clinician will ever personally encounter. Face2Gene works by facial pattern analysis — think of it as running a photo against a massive reference library of what specific genetic syndromes tend to look like facially, something that takes a geneticist years of specialized training to do by eye, if they can do it at all for the rarest conditions. The Mayo Clinic case works differently: it's AI reading an ECG — a routine electrical heart-rhythm test — and picking up on a subtle pattern statistically associated with amyloid buildup in heart tissue, a disease that's easy to miss because its early symptoms mimic much more common heart conditions.
Hands-On (65s, ~150 words): The two real numbers here are worth sitting with. First: a 98% probability score from AI analysis of one patient's ECG, for cardiac amyloidosis — specific and quantified, not a vague "consider testing for" flag. Second, and more rigorous: a controlled study where two different AI chatbots were tested against 90 real, previously-solved complex rare-disease cases — meaning the correct answer was already known, so accuracy could be checked exactly. The chatbots correctly diagnosed 13% and 10% of those cases. Doctors reviewing the identical case files got 5.6% right. Notice all three numbers are low — these are genuinely hard cases by design — but the AI models still came out roughly twice as accurate as the physicians on the same test.
Takeaway (22s, ~50 words): This isn't "AI replaces doctors" — it's AI functioning as a second opinion generator for the specific case where a human's limited personal exposure to rare conditions is the actual bottleneck. If you or someone you know has an undiagnosed condition, these tools are a legitimate next step to try.