Rare diseases are typically thought to affect a tiny minority of the population, but the cofounder of Shoreditch-based healthtech startup, Mendelian, insists that this is a misconception.
"They're not rare at all," says Rudy Benfredj, "rare diseases are actually very common. Almost six percent of the population - 350 million people - are affected globally.”
Benfredj calls the treatment of rare diseases, “one of the biggest inefficiencies in healthcare", with patients waiting an average of seven years to be correctly diagnosed. “Patients call it the diagnostic odyssey,” he says.
During the several years it takes to receive a correct diagnosis, the patient is often shuffled from doctor to doctor, seeing an average of eight over the course of this time. “It's a journey that's really riddled with burden and turmoil,” says Benfredj.
However, he and his team saw the glimmer of possibility in this vexing area of healthcare. “To us, it looked a lot like a tech issue - like an information challenge.”
Mendelian, which is itself the name for a genetic trait which can lead to disease, says that the data on rare diseases is out there, it's just a case of making sense of it. “It's something that a machine could do very well, it's just very disorganised, very inconsistent, very contradictory, and comes from very different sources.”
Applying machine learning to rare disease
Since founding the company in 2015, the team have developed a potential solution: a machine learning algorithm which helps health specialists narrow down possible causes of symptoms. “It's the ability for us to start modelling some of the deep rules of biology that govern rare disease, that helps the patients get a diagnosis much faster,” says Benfredj. In a nutshell, the programme works by inputting symptoms, letting the relevant algorithms perform and produce a list of likely causes.
Most rare diseases are linked to genetic conditions, and this is an area that the team is captivated by, bolstered by technological advances such as cheaper sequencing techniques and emerging analytics.
The 100,000 Genomes Project involves the gene sequencing of NHS patients with rare diseases and their families to better understand the precursors to developing these diseases. The diagnostic yield is currently around 25%, but Mendelian say that they have opened up the data for researchers like them to help improve this rate.
Diagnosing Alstrom disease
There is a genetic root for between 75 and 90 percent of rare diseases, that include Wilson's Syndrome, SAPHO syndrome, Fabry disease and Alstrom disease. The prevalence of the latter is one in a million.
Early on, Mendelian linked up with a mentor, a woman called K. Parkinson, whose two children suffered with Alstrom disease and passed away aged 22 and 24 respectively. They weren't correctly diagnosed until the age of 18, meaning that their whole childhoods were characterised by confusion and mistreatment. Benfredj says that with the knowledge of the children's symptoms, Mendelian would have been able to produce a diagnosis far earlier.
In normal illness diagnosis, doctors and specialists rely on their experience to help them come up with a likely cause. However, in cases of rare diseases, the practitioner may not have encountered many, or any, similar cases over their lifetime.
“Doctors are good at recognising common patterns; we are at a different paradigm here,” says co-founder and director of Mendelian, Francisco Garcia, saying that the startup's aim is to prevent symptoms being mislabelled as different, more common health conditions.
However, this creates an issue with training an algorithm. “In rare disease, the whole training model is not going to work because we don't have enough examples per case,” explains Benfredj. “When you have a disease that is affecting 20 people in the UK, you're never going to have enough data in order to train your machine.” He says Mendelian instead employs a combination of a number of models.
“We use all sorts of different techniques in order to aggregate, curate, sort, store, and then retrieve this information from the big knowledge repository that we gather,” says Benfredj, adding that this provides the startup with a competitive advantage.
“In terms of the data, it's everything you can think of: signs, symptoms, proteins, disease. All the biological entities that can be linked and that allow us to get a better understanding of the underlying rules of the biology of rare diseases," says Benfredj.
The team’s aim is to “knock years off the diagnostic journey by flagging patients and raising suspicions, much much earlier,” says Garcia.
Benfredj adds: "We go to doctors, and we say, 'Let's look at the cases that you have diagnosed, and the cases that you have not diagnosed yet and let's look at how we can help, giving you some areas of investigation that can allow you to diagnose these patients much faster'."
Augmenting, not replacing, doctors
One of the tools currently offered by Mendelian is an app aimed at tertiary care specialists. “Not GPs, we're talking geneticists, cardiologists, paediatricians,” says Benfredj. The tool has been available for 18 months and is being used by 3000 doctors. They estimate that many patients have been helped by the tool and say that sometimes doctors will reference them in case reports on diagnoses.
They also have plans to roll out a tool in the NHS that will cater to primary care doctors. “Looking into GP practice, screening hundreds of thousands of patients and then being able to flag the patients that are most likely to have some of these very-hard-to-diagnose diseases,” says Garcia.
In terms of competitors, they say there is another tool that has been around for a while: Find Zebra. In terms of which is the more effective tool, Benfredj says, “The question is: how do you benchmark this? If you feed in a hundred patients, how many are being diagnosed by Find Zebra and how many are being diagnosed by Mendelian?”
He says the diagnostic yields are not comparable, with Mendelian achieving a far higher success rate. According to Benfredj, this is due in part to the different methods employed. Find Zebra, for example, uses a method called syntactic search, meaning it looks for words. “So if the patient has fever, it will be looking at F-E-V-E-R and checking if fever appears in some places,” says Benfredj. Instead, Mendelian is closer to a semantic search engine: “The machine understand what fever is, and realises it is some sort of headache and linked with other things."
“You always have to remember, in health, the amount of noise, the amount of imprecision that you have when you look at patient records, or when you get to describe a patient is absolutely huge,” says Benfredj. Finding truth in the noise is something that the group has been working on over the past few years.
Headlines on AI-powered health diagnoses have proliferated in recent years, is Mendelian another step towards AI driven healthcare? “When we do our research, we find out that we have a better accuracy than a group of experts. But then when you overlap their results with ours, the diagnostic yield explodes again,” says Benfredj, emphasising that the group is careful not be seen to be advocating for the replacement of doctors.
“The point for us is not to say that machine learning can do better than doctors at diagnosing, what's for sure is that doctors using machine learning will always be better than doctors without using it,” he adds.
Benfredj sees this kind of technology becoming more and more vital, as the incidence of diseases decreases, rare diseases will make up a larger proportion overall. "I wouldn't be surprised if in 20 years around 80 percent of the diseases that we know are rare," he says, adding that this has important consequences for the kinds of treatment required: "We want more personalised medicine; we want more attention to the patient. That's what precision medicine is all about."