Could software reveal whether Abraham Lincoln had Marfan syndrome? Doctors faced with the tricky task of spotting rare genetic diseases in children may soon be asking parents to email their family photos. A computer program can now learn to identify rare conditions by analysing a face from an ordinary digital photograph. It should even be able to identify unknown genetic disorders if groups of photos in its database share specific facial features.
Rare genetic disorders are thought to affect 6 per cent of people. Genetic tests exist for the more common conditions such as Down's syndrome, but many people with the rarer disorders never get a proper clinical diagnosis. Genetic tests aren't available for many conditions because the gene variants that cause them haven't been identified. This means doctors often have to rely on the pronounced facial features that occur in between 30 and 40 per cent of rare disorders to make a diagnosis, but few people are trained to recognise them.
"Clinicians skilled in the use of facial features to support diagnosis are few and far between," says Alastair Kent, director of the charity Genetic Alliance UK. "As a result, families frequently experience long delays – years rather than months – before they receive a diagnosis for their child."
The software developed by Christoffer Nellåker and Andrew Zisserman of the University of Oxford and their colleagues should help family doctors or general paediatricians make a preliminary diagnosis. "The idea is to offer it to health systems right across the world because all you need is a computer and a digital photo," says Nellåker.
To train the system, Nellåker's team fed a computer vision algorithm 1363 publicly available pictures of people with eight genetic disorders, including Down's syndrome, fragile X syndrome and progeria (fourth, fifth and sixth in the graphic below). The computer learned to identify each condition from a pattern of 36 facial features in each shot, such as the shapes of eyes, brows, lips and noses.
"It automatically analyses the picture and annotates key feature points, producing from that a description of the face which expands the features that are important for distinctiveness," Nellåker says. These features are then compared with those from pictures of patients with confirmed disorders, allowing the system to suggest and rank predictions for new patients.
To show that it works, the team analysed photos of people with known genetic disorders. The accuracy of the software increases with the number of photos of a specific disorder it learns from. For the eight training diseases, for example, each disorder was represented by between 100 and 283 images. On average, this resulted in 93 per cent of the predictions being correct.