Latest / Luke's English Podcast / AI Undermines Health Science

Transcript

AI Undermines Health Science · Luke's English Podcast

• 0:07
  Speaker 1: I have an interesting paper that's been published in a
• 0:09
  Speaker 1: journal called PLOS Biology and has been written up as
• 0:13
  Speaker 1: well in some of the science media, which looks at
• 0:17
  Speaker 1: the role of artificial intelligence in some papers looking at
• 0:22
  Speaker 1: questions to do with health and in particular, what factors
• 0:26
  Speaker 1: might predict whether someone is going to suffer from a
• 0:30
  Speaker 1: particular disease or health issue. Now, there's some ways you
• 0:36
  Speaker 1: can analyze these data really quite easily with artificial intelligence.
• 0:39
  Speaker 1: So you can download a big database of health outcomes
• 0:44
  Speaker 1: and predictors of those health outcomes like age and sex
• 0:47
  Speaker 1: and where you live and how much you earn. And
• 0:50
  Speaker 1: then you can do a big dredge through these data
• 0:53
  Speaker 1: with artificial intelligence and it spits out some results. And
• 0:56
  Speaker 1: then you look through those results and you decide to
• 0:58
  Speaker 1: report ones that you like the look of. The problem
• 1:02
  Speaker 1: is that's not necessarily the right statistical way in which
• 1:06
  Speaker 1: to analyze these data, because often health issues like much
• 1:10
  Speaker 1: of the things in the living world are really much
• 1:14
  Speaker 1: more complicated than single, relatively simple correlations. There's lots of
• 1:19
  Speaker 1: things that interact together in order to perhaps make you
• 1:22
  Speaker 1: depressed or make you obese or suffer from cardiovascular disease.
• 1:27
  Speaker 1: So but people are doing these analyses with artificial intelligence
• 1:31
  Speaker 1: and then they send them off to papers, to journals rather,
• 1:35
  Speaker 1: where those papers have been submitted, are peer reviewed and
• 1:39
  Speaker 1: then they come out and are published. And so what
• 1:42
  Speaker 1: this paper in PLOS One did is it did a
• 1:45
  Speaker 1: what's called a systematic review. So it did a big
• 1:48
  Speaker 1: search for this type of analysis that can be done
• 1:51
  Speaker 1: really easily with artificial intelligence to see whether the type
• 1:54
  Speaker 1: of analysis has become more popular since the tools became available.
• 1:59
  Speaker 1: And they have found a staggering increase, almost an exponential
• 2:03
  Speaker 1: growth in the number of these papers that are being
• 2:05
  Speaker 1: produced and are being submitted to journals and being published.
• 2:09
  Speaker 1: And the problem with this is we that the people
• 2:13
  Speaker 1: who did this, the expert medics who did this chose
• 2:17
  Speaker 1: diseases and chose and chose to focus in particular on.
• 2:22
  Speaker 1: on studies where we know there are lots of different
• 2:26
  Speaker 1: factors that can cause those diseases. And then they found
• 2:29
  Speaker 1: these papers that provided much simpler analyses that can be
• 2:32
  Speaker 1: done very simply with artificial intelligence. I should say more
• 2:36
  Speaker 1: complicated analyses can also be done with artificial intelligence, but
• 2:39
  Speaker 1: they're actually harder to do. So why does this matter? Well,
• 2:44
  Speaker 1: it matters for a number of reasons. First of all,
• 2:47
  Speaker 1: the way you progress in academia and get jobs is
• 2:50
  Speaker 1: through publishing papers, so there is a temptation, certainly in
• 2:53
  Speaker 1: some parts of the world, for scientists to try and
• 2:56
  Speaker 1: use AI to get papers quickly. And in fact, most
• 3:00
  Speaker 1: of these papers that this review identifies come out of China. Second,
• 3:05
  Speaker 1: it's flooding the scientific literature with fairly dubious and wrong conclusions.
• 3:13
  Speaker 1: And so that's kind of dangerous for the development of
• 3:16
  Speaker 1: drugs and diagnosis of health diseases if it's not picked up.
• 3:21
  Speaker 1: And so the authors of this paper sort of send
• 3:23
  Speaker 1: a warning, but also come up with some recommendations in
• 3:27
  Speaker 1: terms of how these AI approaches can be identified and
• 3:32
  Speaker 1: can be checked to make sure they don't get in there.
• 3:37
  Speaker 1: Another problem of this, of course, is it's now quite
• 3:39
  Speaker 1: easy to write a paper using AI and to do
• 3:42
  Speaker 1: these analyses. So you get some fairly formulaic writing. And
• 3:46
  Speaker 1: then this increases the burden on other scientists who ask
• 3:49
  Speaker 1: to peer review these papers. Now, I don't want to
• 3:52
  Speaker 1: have a big downer on AI because AI can be
• 3:55
  Speaker 1: an incredibly useful tool in science if used properly. Indeed,
• 3:59
  Speaker 1: we've used it in my group where we've linked artificial
• 4:02
  Speaker 1: neural networks to ordinary differential equations to help identify dynamical
• 4:07
  Speaker 1: patterns in systems. I mean, we wrote the code ourselves.
• 4:10
  Speaker 1: We developed the methods ourselves, mainly led by my postdoc,
• 4:14
  Speaker 1: Willem Bonnefeuille. So, you know, that's a useful tool that
• 4:18
  Speaker 1: other people can use. But, you know, like pretty much everything,
• 4:22
  Speaker 1: AI is creeping into our lives, sometimes in incredibly useful ways,
• 4:26
  Speaker 1: but sometimes in risky ways. And I thought that this
• 4:29
  Speaker 1: was an interesting paper, an interesting report, because it shows
• 4:32
  Speaker 1: that AI is even creeping into the production of scientific papers.
• 4:37
  Speaker 1: And in this case, in medicine, in not a particularly
• 4:41
  Speaker 1: helpful way. Anyway, I hope that's interesting. And I hope
• 4:45
  Speaker 1: if you want to read more, you can go and
• 4:47
  Speaker 1: look for the paper in PLOS One. And that's a
• 4:49
  Speaker 1: journal where all papers are open access so anyone can
• 4:52
  Speaker 1: download it and read it.