Marcel Salathé: Tackle scientific challenges with transdisciplinarity
The ongoing Covid-19 pandemic has demonstrated the value of adopting transdisciplinary approaches to tackle scientific challenges. EPFL associate professor Marcel Salathé reflects on efforts to bridge computing and technology – particularly new tools like AI and machine learning – with the disciplines of epidemiology and health.
Artificial intelligence (AI) has been viewed by some as a panacea, an instant solution to many of the problems faced by humanity, especially health crises. But Salathé, associate professor at EPFL and head of the laboratory of digital epidemiology, is careful to rein in runaway enthusiasm with a healthy dose of reality. “I think by and large one can probably say that AI wasn’t a game-changer for this Covid-19 pandemic,” he tells Geneva Solutions.
He felt, however, that last month’s Applied Machine Learning Days event on AI and Covid-19 held at EPFL was a good moment to have a discussion on the role of AI in the context of Covid-19. It is important, he emphasises, to explore the range of possibilities that modern computational tools provide, with a focus on not just the present but also on the future.
One of the ways in which AI has been trialled in the context of Covid-19 is to detect the presence of the disease by looking at and listening to our lungs. “I’m heading this national research programme in Switzerland, where one of the projects is trying to figure out if machine learning can quickly assess whether a case is Covid-19 or not,” Salathé explains. “People have started to use machine learning on pulmonary images, which is almost classical medicine in that sense, to look for signs of Covid-19.”
But machine-learning algorithms must be trained on large volumes of data, in this case, images of lungs with and without Covid-19, in order to be able to extract the desired signal. At the start of the pandemic, this was impossible because there simply wasn’t enough data at hand. “Now the sad truth,” he continues, “is that this pandemic is far from being over. There will, of course, be many studies, and unfortunately, because the numbers are still going up there will be more and more data.” Salathé expects the studies to become more sophisticated in the months and years to come, and sees the lessons learnt being applied to future health crises.
Any techniques developed in the context of Covid-19 research will find immediate applications elsewhere, especially where internal imaging isn’t needed. “The general philosophy is that whatever a human can recognise, a computer should be able to recognise too,” Salathé explains. “It means that the limiting factor is not the human anymore, which in many cases of healthcare delivery it is.” Many people in the world aren’t fortunate enough to be around humans who can make smart decisions about their health during a crisis, he states. “If we can use AI – one day, it’s not going to be tomorrow – we can reduce that number, not because there is a human around but because there is a phone with an algorithm that’s just as accurate as a human.”
Another test-bed for AI in healthcare is in analysing coughs. “You’re trying to say, can you deduce from the sound of a cough, its particular audio characteristics, whether Covid-19 was present or not.” It sounds fantastic, in theory. After all, who wouldn’t prefer coughing into their phone’s microphones to find out if they have Covid-19 instead of providing uncomfortable nose swabs for the now all-too-familiar PCR tests. There’s just one problem: accuracy.
“We were having huge discussions about the accuracy of PCR tests,” Salathé remarks. “These are regular diagnostic tools that we have been using for years. And they are not perfect! But even if they are perfect to more than 99 per cent, that already caused quite some concern. So imagine an AI that gets it right 80 per cent of the time, which for an AI is pretty good; it's practically not usable.”
Of course, he points out, PCR is a technology that was developed decades ago. It was new and experimental at the time but has become cheap and widespread today. “So it will be with AI.”
Many of the projects testing the feasibility of using AI in the context of Covid-19 are ultimately research projects – “to figure out if we can use this technology in such a situation” – and the answer to open-ended research questions is never known in advance. While Salathé points out that today the answer to whether AI can help with a particular task will sometimes be yes and sometimes be no or not yet, he believes this will evolve.
The risk to any novel and interesting field such as AI research is that of hype. Expectations can never be met if they are impossible. “Let’s use AI where it’s useful,” Salathé adds, “but let’s not try to where it’s clearly not. There are always dangers with overselling a technology and if you’re talking about the health domain, those dangers can quickly have strong effects.”
We must not fall prey, he warns, to the well-known phenomenon of overestimating technology in the short term while underestimating it in the long term. “This is a consistent pattern where maybe the last one was the worldwide web,” he elaborates. “People thought the web would change the world immediately; they weren’t wrong about it changing the world, they were just wrong about the timescales. And I think it’s a similar thing with AI.” Salathé believes that AI is potentially one of the most profound technological developments for humanity but it won’t be that in the next five or ten years, but probably in the next 50 or 100 years.”
Another risk that Salathé warns of is “tech-solutionism”. “When we developed this technology for digital contact-tracing, which is very privacy-preserving,” he elaborates, “the response – accusing us of setting up a surveillance network – was quite astonishing.”
The root cause of this reaction, he explains, is that historically one domain, often technology, came to another domain claiming to have solutions to the latter’s problems without fully understanding them. “Sometimes this worked and sometimes it didn’t,” he adds. “Most of the time it turns out that the problem was somewhat hard and it would be good if people talked to experts in the field before. Covid-19 has been no exception there.”
“People have had their heart in the right place and everybody wanted to solve this problem,” he continues, “but sometimes there were approaches that were health-focused and dismissive of the technology, and sometimes there were solutions that were very technology-centred and just didn’t get key parameters about the health aspects right.” According to him, none of these methods worked. The pandemic has highlighted the value of collaborative approaches to problem-solving.
Salathé suggests that the way forward is to make transdisciplinarity across domains of knowledge and research the norm. “It’s something that’s so obvious. Everybody has been saying health should really work more with computer science and that this is the future of science. In reality, it’s still quite hard to do so.” He believes academia is structurally set up in such a way that it opposes this kind of work among disciplines and calls for cultural change.
“We don’t want to figure out in the middle of a crisis what’s the best way to work together,” he says. “We have to really be serious about building these bridges way before the flood comes.”