Researchers claim AI can sort cause from coincidence
Healthcare company spruiks new algorithm, but will it work?
By Barry Keily
As scientists and school children know, correlation doesn’t necessarily mean causation. Just because something looks like it’s having a certain effect, it doesn’t mean it is. If this, then not necessarily that.
Teasing actual causation out of big collections of data is a continuing challenge for researchers, but now a British company claims it has come up with an artificial intelligence application that will do exactly that.
Babylon Health, a business that specialises in remote location telehealth services, says it has engineered a system capable of analysing results from disparate but overlapping datasets in a way that identifies causal relationships that would otherwise remain hidden.
The system, says company chief science officer Saurabh Johri, can reveal information that might not be discoverable without additional, expensive research.
"Until now, we have been limited to piecing together answers from studies that needed to capture all the data really neatly,” he says.
“But when we've seen a correlation between obesity and low vitamin D in one study, and obesity and heart failure in another, we have not been able to say whether vitamin D has a causal role in heart failure without doing another, hugely expensive clinical trial. Now we can put the pieces of the jigsaw together."
A draft of the research underpinning the AI – written by employees Anish Dhir and Ciarán Lee, and inspired, says the company, by quantum cryptography – is available on the preprint site arxiv.
Babylon says a fully scrutinised version also will be presented this week at the Association for Advancement of Artificial Intelligence (AAAI) conference in New York.
The algorithm takes its start from the concept of entropy. In the physical universe, any system becomes more disordered over time. The same principle holds in the realm of information, the researchers maintain.
It holds therefore that data describing a causative force should be more ordered than the data describing its effect. By working backwards, it should be possible to distinguish genuine cause from mere correlation.
The researchers admit that the system is far from infallible.
“This obviously isn't a magic wand that will give us all the answers but there are so many studies with missing data, where researchers wish they had tested for something else and could combine it with a study someone else had done, or had thought to ask their questions in a different way,” says Lee. “Now they can.”
When it is unveiled at the AAAI conference this week, the system is likely to be met with a high degree of scepticism – and not without good reason.
Babylon’s reputation for providing AI-driven telehealth services suffered an embarrassing setback in 2019 after The Sunday Times reported that its diagnostic app delivered starkly different findings in two almost identical cases.
Babylon, however, disputes the story, which stated that a 60-year-old male smoker reporting chest pains was told he was likely having a heart attack. A woman of the same age, with same smoking habits and similar chest pains was told she was having a panic attack.