By Chelsea Whyte
People with genetic syndromes sometimes have telltale facial features, but using them to make a quick and cheap diagnosis can be tricky given there are hundreds of possible conditions they may have. A new neural network that analyses photographs of faces can help doctors narrow down the possibilities.
Yaron Gurovich at biotechnology firm FDNA in Boston and his team built a neural network to look at the gestalt – or overall impression – of faces and return a list of the 10 genetic syndromes a person is most likely to have.
They trained the neural network, called DeepGestalt, on 17,000 images correctly labelled to correspond to more than 200 genetic syndromes. The team then asked the AI to identify potential genetic disorders from a further 502 photographs of people with such conditions. It included the correct answer among its list of 10 responses 91 per cent of the time.
Gurovich and his team also tested the AI’s ability to distinguish between different genetic mutations that that can lead to the same syndrome. They used images of people with Noonan syndrome, which can result from mutations in one of five genes. DeepGestalt accurately identified the genetic source of the physical appearance 64 per cent of the time.
“It’s clearly not perfect,” says Gurovich. “[But] it’s still much better than humans are at trying to do this.”
As the system makes its assessments, the facial regions that were most helpful in the determination are highlighted and made available for doctors to view. This helps them to understand the relationships between genetic make-up and physical appearance.
The fact that the diagnosis is based on a simple photograph raises questions of privacy. If faces can reveal details about genetics, then employers and insurance providers could, in principle, surreptitiously use such techniques to discriminate against people with a high probability of having certain disorders.
However, Gurovich says the tool will only be available to clinicians.
Christoffer Nellåker at the University of Oxford says this technique could bring significant benefits for those with genetic syndromes.
“This is not fundamentally different information than we’re sharing walking down the street, or we’re happy to share with Facebook or Google,” he says. “But interrogating the data in this way means you can extract information about health or disease status.
“The real value here is that for some of these ultra-rare diseases, the process of diagnosis can be many, many years. This kind of technology can help narrow down the search space and then be verified through checking genetic markers,” he says. “For some diseases, it will cut down the time to diagnosis drastically. For others, it could perhaps add a means of finding other people with the disease and, in turn, help find new treatments or cures.”
Journal reference: Nature Medicine, DOI: 10.1038/s41591-018-0279-0
More on these topics: