needhelp
← Back to blog

AI Found a Rare-Disease Clue in a Child's Face. The Diagnosis Still Needed Doctors

by needhelp
AI
Healthcare
Rare Disease
Genetics

Rachel Hinken spent years asking why her son Oliver was late to walk and speak and stayed near the bottom percentile for height. At age 10, he was under four feet tall. Doctors had told the New York mother that he would catch up.

An image search with Face2Gene gave her a lead they had missed: trichorhinophalangeal syndrome, or TRPS. Genetic counseling and testing later confirmed it, according to The Wall Street Journal’s August 17 report.

A match is not a diagnosis

Face2Gene analyzes facial phenotypes and ranks possible genetic syndromes. Its own terms call it a search and reference tool, not a replacement for clinical judgment.

That distinction matters. GeneReviews says TRPS can involve distinctive facial features, sparse slow-growing hair, short stature and skeletal changes. Molecular confirmation requires finding a pathogenic TRPS1 variant for TRPS I or a deletion spanning TRPS1, RAD21 and EXT1 for TRPS II.

The photo opened the right door; it did not complete the diagnostic work.

What the 13.3% result actually measured

A 2025 JAMA Network Open study tested ChatGPT-4o and Llama 3.1 8B on clinical summaries from 90 already-solved cases in the US Undiagnosed Diseases Network.

  • ChatGPT-4o included the exact final diagnosis in 13.3% of cases.
  • Llama did so in 10.0%.
  • Historical physician review had resolved 5.6%.
  • Only ChatGPT-4o’s difference was statistically significant.

These were retrospective, unusually hard cases. The models generated differential lists from summaries; they did not examine patients or autonomously confirm disease. Even the stronger model missed the exact answer almost seven times out of eight.

The work is moving from demos to research programs

In June 2026, Boston Children’s Hospital, Harvard and OpenAI reported a different workflow. OpenAI o3 Deep Research reanalyzed de-identified clinical and genomic data from 376 previously unsolved pediatric cases. It surfaced evidence-linked leads; specialists, extra testing and clinical laboratories ultimately confirmed 18 diagnoses, a 4.8% additional yield. OpenAI explicitly says the model diagnosed no patient.

Anthropic followed in July with a rare genetic disease grant call. Selected teams could receive up to $50,000 in Claude credits for six months, split between basic science and early-stage biotech work.

The useful role is a second set of eyes

Rare-disease diagnosis is a search problem with scattered records, changing literature and thousands of possible conditions. AI can rank a forgotten syndrome, connect a phenotype to a paper or make old genomic data worth reviewing again.

It can also be wrong, biased toward faces represented in training data, or expose highly sensitive health information. Face2Gene says uploaded photos and identifiable data stay in a private case library unless a user shares them, but its guidance still assumes proper clinical photo consent. A parent should not treat a consumer upload as a genetic test.

The breakthrough is narrower and more useful than “AI replaces the doctor”: it gives families and specialists another hypothesis to test.

References

Share this page