Rare genetic disorders: an AI that reads the face to guide diagnosis

Why can a face help spot a genetic disease?

Genetic syndromes, taken together, affect 8% of the population. Many produce recognisable facial features, and clinical geneticists use them as a clue.

But there are hundreds of syndromes, and the face-analysis tools built so far could recognise only a few. In clinic, the doctor has to consider hundreds.

Looking closely at the face is already part of a clinical geneticist's work: the authors describe it as a highly informative clue.

How was DeepGestalt built?

DeepGestalt uses computer vision and deep learning: techniques that let a program recognise shapes in images. Instead of saying "yes" or "no" for a single disease, it measures how much a face resembles hundreds of syndromes.

It was trained on over 17,000 images representing more than 200 syndromes. These images were gathered through a community-driven platform for describing patients' features.

Because it gives a degree of resemblance to many syndromes at once, it returns a ranked list rather than a single yes or no. That mirrors how a doctor thinks when several diagnoses are possible, and it leaves the final judgement to the clinician, which is where it belongs.

What results did they get?

In three initial experiments, DeepGestalt did better than clinicians: two to tell a target syndrome apart from other syndromes, and one to separate different genetic subtypes of Noonan syndrome.

The final experiment mirrored a real clinical situation: finding the right syndrome among all the possibilities. Across 502 different images, the correct answer was among the program's top 10 suggestions 91% of the time.

This last test matters most, because it resembles what a geneticist faces: one face and hundreds of possibilities to weigh. The authors recall that earlier technologies already matched clinicians, but only on a few diseases.

What does this change for patients?

For a family waiting for a diagnosis, a ranked list of leads can help the geneticist choose which test to order first. The authors see added value for assessing a patient's features, for genetic testing and for precision medicine.

The program does not replace genetic analysis: it saves time by guiding the search. Genetic testing remains the way to confirm which syndrome a person really has, and the clinician stays in charge of explaining the result to the family.

A list of syndromes ranked by resemblance can also help interpret a genetic test whose results are hard to read.

What are the limits?

The study does not show that patients are diagnosed faster or better in practice. It measures the program's ability to find the right syndrome on a set of images.

A photo is not enough to conclude: it is one clue among others. The program covers over 200 syndromes, while many more exist.