Christoffer Nellaker, University of Oxford
Diagnosing illnesses isn’t always easy, especially when it comes to rarediseases. There may only be a handful of cases across the world meaning even the most experienced of doctors are unlikely to have seen them before. We hear from Helen Piper is the mother of 4 year old Alex, who has an undiagnosed development disorder and how the University of Oxford’s Christoffer Nellaker has harnessed the power of facial recognition technology to help diagnose conditions like Alex.
Christopher - So, a rare disease is, as the name implies, rare, but there are a lot of different types of rare diseases. So collectively, they're surprisingly common perhaps. One in 17 people might have a rare disease.
Ginny - That’s way more than I would’ve expected. How were you using facial recognition software to help diagnose these things?
Christopher - During the development of the face and head, very large portion of our genes in our genome are used in the development of the head and face. So, if there is a DNA change in one of these genes involved in this is very likely that it causes a change in the final form of the head or face. So, this is what clinicians had been using for the past 65 years to try and identify and help diagnose these rare diseases is to look if there are commonalities in characteristics between patients and thereby trying to group them into a potential diagnosis.
Ginny - You’ve trained a computer to do the doctor’s job for them so to speak.
Christopher - This research was only possible because we had a great collaboration with Professor Andrew Zisserman at the Department of Engineering Science who’s world leading at computer vision. And so together, we set out to try and see if we could use the latest computer vision and facial recognition type approaches to try and aid the job of clinicians in analysing faces and trying to cluster people by commonalities in the head and face.
Ginny - How good is the computer at picking up these differences?
Christopher - So, we used 8 rather well-known syndromes to train the computer to try and analyse these rare diseases. But then we applied this to a further 82 syndromes and the computer could then generalise and try and cluster patients even for these other syndromes that we didn’t explicitly use in training.
Ginny - How do you see this being used? Would it be the case of having a computer in every doctor’s surgery that could do this?
Christopher - This doesn’t require any fancy equipment. The algorithm uses any old photograph. If there's a 2D photograph in like a family album, in theory, if you could show that to a computer then this algorithm could be used to analyse them. Taking a photograph of patients with suspected rare diseases is standard clinical practice anyway. So, this would be seen as an expert tool for clinicians to try and help narrow down to potential diagnosis and sort of suggest which tests might be suitable to be done.
Ginny - So, you mentioned there that it’s a narrowing down procedure. It doesn’t mean that you could just tell exactly from a photograph what was wrong with the person. You'd still need to do other tests.
Christopher - This has to be used in the hands of an expert. It’s not sufficiently accurate to give 100%. Though it does do much better than expected. So, for these untrained syndromes, it narrows the search base for the right syndrome by 27 times better than random chance. This, again it was just a sort of proof of principle study to see if it was possible to do this. So, we’re hoping to expand on this.
Ginny - So, I guess that would save the taxpayer quite a lot of money if could run fewer tests rather than having to run all 27.
Christopher - Yes. I mean, this is also – since this algorithm just uses basic equipment, it could also be applicable in parts of the world where this type of expertise is not easily accessible. So, it would help narrow down the tests that might be done in countries and in healthcare systems where money is very much limited to the number of tests you can do.
Ginny - And what kind of tests are we talking about that would need to be done to get a sort of definite diagnosis?
Christopher - So ultimately, you want to identify a genetic cause if there is a genetic cause for a particular rare disease. That’s not always the case. But if there is a DNA change then that’s what you want to try and identify. And so, there are existing – a lot of specific tests to look for specific changes. As this new genome sequencing, as we heard earlier in the programme, is developing, of course, we’re hoping to make this – that this will make it a lot easier to identify the DNA changes in rare diseases.
Ginny - Just briefly to finish up, how long do you think it’ll be before something like this is in widespread use?
Christopher - So, we are trying to develop this as quickly as we can of course, but we do have to expand this and workout with clinician collaborations carefully. A number of years still, but we would be hoping to do it as a sort of ongoing research and development as it rolls out in collaboration with clinicians.