Close
Transcript: Episode 18
Can AI save us?
SUMMARY KEYWORDS
ai, people, pandemic, ucl, understand, question, modelling, diseases, animals, problem, human, data, doctors, predict, trace, mathematical models, species, technologies, system
SPEAKERS
Kate Jones, Michael Veale, Ali Parsa, Vivienne Parry
Vivienne Parry 00:03
Hello and welcome back to Coronavirus: The Whole Story. We hope you had a good time over the summer whether it was in 35 degree heat or lashing rain, there wasn't any middle ground was there? Speaking for ourselves, we're glad to be back just in time for the new term and to your academic year. My name is Vivian Perry. I'm a writer broadcaster, UCLA alumna and host of this award winning podcast, which surveys all things Coronavirus. Nothing has changed while we've been on holidays. This teeny microbe continues to have a giant global impact, every aspect of which is being assessed, tackled mitigated and illuminated by the incredible minds in all corners of UCL. If this is your first time with us, hello, you're so welcome! And once you heard one of our pods I'm warning you that our fans tell us that we're like a tube of Pringles, you can't resist snacking on all of them. So good know then that our entire back catalogue of past episodes is available from the UCL Minds website, or wherever you get your podcasts.
Now we like a big challenge and we're returning with one of our biggest yet - can AI save us? And in true UCL interdisciplinary style, I'm joined by an ecologist, a researcher of law, and a tech CEO to try and answer it in episode which has been yet produced by Professor Garrett Reese, Dean of the Faculty of Life Sciences. My first guest this week is Professor Kate Jones, the UCL Chair of Ecology and Biodiversity. In her research Kate uses statistical and mathematical modelling to understand the impact of climate change with a particular focus on emerging infectious diseases from animals. I'm also joined by Dr Michael Veal a lecturer in digital rights and regulation. Michael's work seeks to understand emerging digital technologies and how the law should be applied to them and the way they affect citizens both intentionally and unintentionally. And finally, my third guest is Ali Parsa, a UCL alumnus and the CEO and founder of Babylon health. Babylon health, as you all know, is a digital healthcare service that enables remote consultations, connecting patients and healthcare providers via text and video chat, something that's been evermore important during lockdown. Now, before we start, who's going to volunteer to tell us in 30 seconds? What exactly we mean by AI? Ali -
Ali Parsa 02:37
Thank you for volunteering me. I know it's a hard question to answer in 30 seconds because so many people interpret it and define it in so many different ways. But in my mind, it is attempts by machines to replicate some of what humans call intelligence. So it's the ability for machines to do some of our thinking. Now, that is what it should be. And I think the Turing test described it best that if another human being cannot know the difference between what the machine whether the other party is a machine or a human that kind of passed the test as artificial intelligence, and on but but but on that account, I think we are a long way away from having truly general intelligence, but we're kind of hacking our way through it little by little by being able to do the simplest stuff a little bit now and automate it, if that makes sense.
Vivienne Parry 03:43
Okay, very good. Let me start them with Kate. And you focus on emerging diseases which are principally what we call zoonotic. That is, they've jumped from an animal host into a human one with Korean or obviously we think of bats with pangolins maybe in the mix, too. The research you do in this area uses math and computer science, all kinds of tools, first of all do you use in your work? And how do you build them.
Kate Jones 04:08
So thank thanks for having me on. It's a pleasure to be here. I look at broad scale patterns across the globe to understand how, and when these kinds of diseases emerge into the human population. So it's an incredibly complex system. So you have to think about how climate change how anthropogenic activities human activities like land use change, building cities, agriculture, changes the patterns of animals present, and then the viruses or other pathogens present in the animals, and then how that changes the likelihood of spillover from animals to humans, and then how likely it is from a human that from a pathogen to get from a human to human and spread around the world. So I use a mixture of mathematical models. But also I do, I do use a lot of process based models, which are and I use AI for that. So I, I look for patterns in the data that I have to then extrapolate into areas I haven't measured. So I use AI in that way, combined with mathematical models to understand and predict what's going to happen next.
Vivienne Parry 05:25
And this is – is it a learning AI? In other words, you set it up to look for particular patterns and then develops and sees further patterns.
Kate Jones 05:34
Yes, so give you an example with Ebola. We've been producing predictive forecasts models of Ebola about where and when we would expect spill overs to happen. So, the first stage of that is where the pathogen is in the in the animal populations. So I would look at where these animals have been found in the past, records online and lots of different databases and correlate that with their habitats and where they live. And so for areas that have not been surveyed, I can then use a prediction an AI prediction of where those animals might be. The other way that I can do this is my experiments and I've been developing methods for automatic monitoring of populations of animals, so that I can use AI to recognise images from say, camera traps, the cameras you put out in the environment, you get all these images back, and I can recognise what species is present in there in in using an AI process. And I've also been using audio as well. So bats use echolocation and birds make sounds. You can use those sounds to to build recognizers, which can automatically go through the recordings that you make
Vivienne Parry 06:54
That’s fascinating. Of course we were all using AI on our phones all the time, which People perhaps don't realise, but you know, when our photos are sorted for us, that's an AI function, isn't it? So are you using AI to recognise particular kinds of animals that might be a special transmission risk?
Kate Jones 07:18
Yeah, we we've got several different projects across the world, which are looking at how human activities are promoting disassembly of ecosystems. So that means changing the animals that are present. And we're finding that if you change habitats into more human dominated places they become, but the animals present are more likely to give you and pass on these zoonotic diseases. So we've been setting up kind of grids of sensors in different countries in different areas along gradients of human pressure. So we would have camera traps and audio sensors and then we have like petabytes of data to go through, and we've been using AI to process that information to understand how ecosystems change when you have lots of human pressure and then what that means then for spillover
Vivienne Parry 08:15
So bats are always, perhaps rather unfairly for us lovers of that implicated but of course SARS came via civets. And MERS came by camels. I think. So. Are there any particular species that you are always on the alert for on what's it telling us now about emerging diseases,
Kate Jones 08:39
Bats do have a really unfair rap, to be honest, and there are lots of different species which are responsible for spill overs and rodents are also one of them. But I think it's I would be more comfortable and blaming certain taxa, by looking at our own activities, and how we're changing the landscape to promote species which can survive and those landscapes are very odd landscapes like London is a very, very odd, heterogeneous landscape full of concrete, but also parks, it's a very strange place to survive. And those kind of activities change the species which are then present in the area. And it turns out that but it's those kind of those city loving the agricultural loving species which are, are more risk to us in terms of, of, of the probability of spillover of a pathogen into us. So it's those kind of sinner back the city loving human loving animals which are more problems which which include passerine, birds, rodents, and some bats. So we do try and think about how to monitor those species using these remote AI in enabled sensors.
Vivienne Parry 10:01
Luckily, the one thing that Londoners know is not to hug pigeons. So what particular diseases are you most focused on and you’re most concerned about?
Kate Jones 10:12
I think the problem with zoonotic diseases is the unknown nature of the diversity out there. So I wouldn't like to pin my mouse to any particular pathogen. I think we need to set up proper surveillance systems and invest, in fact huge amounts of money and understanding what viruses and other pathogens are out there and in the animal populations, but crucially, we need to understand how our actions are changing ecological systems. And I really do think that the ecology part of the this, this puzzle has been missed out and I think a lot of money has been promoted as been pumped into vaccine development into understanding The symptoms and treatments, but so little money is actually spent on understanding how the environment changes the spillover probabilities from animals to people. And we really need a much more joined up approach in order to prevent the next pandemic.
Vivienne Parry 11:17
So does AI show you the hotspots globally, in other words, so are you using us in a predictive way or is in a way to understand what's already happening?
Kate Jones 11:28
I think I think that's a really, really interesting question and it comes down to what you think AI is useful for. And in my research, I found that relying on AI for prediction is is very difficult because you're usually forecasting into situations that the the pattern recognition system the AI has not seen before. And these underlying processes are very difficult to model in a pattern based analysis. So I've been using kind of both the AI part where I don't have the information that I need to fit, but I then feed that into a more mathematical model because then I can control the the actual processes, which I think are, are operating and then make a prediction and a forecast. So for example, our one of our latest analyses is on lassa fever. Lassa fever is a hemorrhagic fever like Ebola, and we've been showing that land use change and climate is a really good predictor of the number of Lassa cases and we can, we can use that to, to predict three months, three to four months ahead, what the size and shape of that epidemic is going to be like. And so that's a really novel way of trying to understand and using environmental data, AI and these mathematical processes to understand make informed decisions for the the governments involved.
Vivienne Parry
Here's a big question for you. Do you think I could have predicted this pandemic?
Kate Jones 13:09
AI? Perhaps not. But we certainly could have done this. And it's not a shock. This has happened. And we've been talking. We know myself and my colleagues have been predicting something like this would happen for at least 20 years. So it's not a shock. And this area of Wu Han where this diseases emerged has been flagged as risky for over 10 years, and then in 2019, at least three papers which pointed this out. So I don't think it's that hard. I think it's a policy problem. And I think people we need a much more joined up approach between public health ecology, agriculture and agricultural development and and try and vaccines element and forecasting. So I think, you know, I don't think it's a shock. I think it's a missed opportunity.
Vivienne Parry 14:07
Fascinating. So Kate has explained the great power of AI particularly to predict the shape and size of outbreaks, as with lassa fever, but it's your job, Michael, to understand how it can be used responsibly. And you get about evaluating new AI tools in that in terms of their potential to help or to harm society.
Michael Veale 14:31
That's a really good question. In many ways, we'd start by by looking at law, ethics, fundamental rights and human rights. The the core part I think with with AI is not to be dazzled by the technology. You have to very carefully and with care, look at the business models, how any use fits into what the public agency is trying to do, or to certain businesses trying to do in both the short term but also the medium and long term, we see that that there are different roles of AI that can, they can have immediate political effects. So we can look at the use of them. Admittedly, it was only linear regression and multi level modelling in the exam result algorithm from Ofqual in the UK, which caused an immediate political explosion. But we can also look at the use of AI to profile people online or in physical environments, and to create power for big platforms like Google and Facebook over the medium and long term. So it's important to not lose track of that, that aspect and how it interplays with the the broader systems.
Vivienne Parry 15:41
Now, can we talk about the track and trace scheme, which has been causing all sorts of issues by itself? Has it been helping to stop the spread of the virus?
Michael Veale 15:53
So I think it's certainly an idea of, I think if what you mean by that is that the technology side track and trace. So definitely contact tracing is a is a crucially important
Vivienne Parry 16:03
Yes. No, no, we it's the tech tech side. And I'm thinking of,
Michael Veale 16:07
yeah. So early on in the pandemic I was working with with colleagues, particularly in in Switzerland and Belgium. And we were responding to the notion that many countries, particularly Singapore, Hong Kong, and also China had taken very technological routes to dealing with a pandemic or an epidemic, then they were using large linked data sets. Many of them were trying to use telecoms data, or Bluetooth data from mobile devices to augment or even or take a large pay play a large part in contact tracing efforts. And the concern that we had there is that take that too many countries in the world that have probably greater respect for fundamental rights, and citizens are mistrustful. AI has been misused, and it has caused problems and anxieties and concerns around discrimination and transparency. So our role there was to say, could we build a kind of system that would would help contact traces, but which would would put privacy and human rights first. And we did this quite early on with with researchers, we effectively built something that public health authorities could use in Apps for, for Bluetooth connections between phones to see a phone and be near each other. But it would have the quality that no data would about you would be leaving your phone. So it can be done without creating a large central database of who saw who the worry with that central database, I think was exacerbated during COVID. Because because of the great uncertainty, nobody really knew if technical interventions will work or many of the qualities of this disease. And you could anticipate that if governments had access to really large sort of social graphs and networks of who saw who in society that could in many in the wrong hands become quite a coercive tool to allow a government to send certain people home or allow certain people out of their houses in a very orchestrated way that could really have effects on on, on on many groups persecution and and the like. So, we were very concerned about that.
Vivienne Parry 18:26
Is there any difference apart from in scale between the kind of track and trace systems that you know, shoe leather epidemiology has done, you know, since the time of John Snow, for for instance, very sensitive sexually transmitted disease and something like this covert tracking, is there a fundamental difference apart from that one of scale?
Michael Veale 18:56
Yes, there is so, so, the idea that it was never that the technological interventions like a Bluetooth tracing system would replace traditional contact tracing. It was very specifically aimed at times when individuals may have spent time with somebody they don't remember or they might not know the name of. And, and it was designed to do so to alert these people rapidly. So we were still at that time, we still are learning about the exact dynamics of the disease, how quickly it spreads to how long it's contagious, how long it can be incubated the effects of asymptomatic individuals on on the entire modelling of the system. And so there were there were specific things that you couldn't contract contact trace people before that an individual sitting behind on a bus, but hadn't you hadn't talked to or wasn't aware of the name of some places would do that with CCTV cameras and attempt to try that but that just really wouldn't scale. So there are some differences here. But I think time will still tell the the usefulness of any of these interventions because this is the first time that the technologies have had been deployed, but now they're deployed a really around the world and growing in adoption quite rapidly.
Vivienne Parry 20:12
We always talk about these kind of things in terms of risk versus benefits. And we perceive COVID-19 as a huge risk. Are we right to really put all our concerns to one side, because we face this? What seems to us to be such an enormous risk? Do you think track and trace for instance, would have been worth it? In the end?
Michael Veale 20:42
I think contact tracing will definitely be worth it. The good question that that we put forward early on was do you have to accept privacy or potential discrimination intrusions in order to have a technological intervention? So going back to the question of Can AI save us - one of the things that AI has done is it's really blinded a lot of people in computing to any alternative that isn't collecting a huge amount of data and putting it all in one place. And seeing later on what you want to do with it. I think we've seen it and Kate was was highlighting that AI has not is not necessarily very good at predicting dynamics we haven't seen before. It's a pattern recognition system, it can't go beyond what it's seen in training data. And what we'd found really, as an alternative is to say, do you need to collect all this data? Or can you focus on a particular purpose if you to make contact tracing happen effectively, and you want to do that technologically, you might not need to have data all centralised in one place, it can be kept on everybody's individual phones, that doesn't have to be a privacy or human rights trade off. We don't have to set aside human rights for the risk of COVID. We can actually have both. But that requires us to think about the problem we're facing really carefully. Say what are we trying to achieve with this technological intervention for contact tracing, and and how would we get there in a proportionate way and that's where we're law, human rights, Data Protection Law can really be a pretty guiding force.
Vivienne Parry 22:09
And hugely important. Thank you very much. Just to remind you all that you're listening to Coronavirus, The Whole Story, a podcast brought to you by UCL Minds. If there's a question about Coronavirus, you'd like our researchers to answer, email us at minds@ucl.ac.uk or tweet @UCL. Let me turn to Ali. You're the CEO of a company that enables remote health care at a time when everyone is looking for remote solutions to things we normally do in person. Tell us a bit about Babylon and how it works. Where does AI come into this?
Ali Parsa 22:49
I built a chain of hospitals and and what the insight we came up when we were delivering health care in hospitals was that actually if you give people a lot of healthcare upfront. You can help them avoid emergencies and crisis and emergencies and crisis is what you and I call sick care, which we're usually very good at in most healthcare systems, rather than keeping people healthy. Now, if you park the doctor inside your home, that person should probably monitor you and your family very well. You can ask any questions you want. And your chances of having emergency and crisis will reduce. There is obviously impossible to do because of the questions of accessibility and affordability of this. And if you think about accessibility, how do you make it accessible that's, that's easier problem to solve as long as you can deliver most of the healthcare most people need on devices Nowadays, most of them have that's how you accessible. We deliver health care to the population of Rwanda at scale, and most people don't even have a a smartphone. The issue is there is no accessibility without affordability. And if you look at where the cost in healthcare go, they about two thirds of all costs if you cut it by people going to salary, and if you cut it by diseases, about 70% of cost goes into predictable preventable diseases. And the role of AI is fundamentally to deal with this issue of how do we, on one hand automate as much as what expensive rare resources do in healthcare, to be able to help people to self-manage themselves better on one hand, and only let our doctors or nurses do what are more complex as the jargon in the industry says, trade at the top of their licence? And the second one is how do we how do we monitor people to be able to do, to see predictable preventable diseases and deal with them when they're a 10 pound problem before they become 1000 pounds solution.
Vivienne Parry 25:08
So what then did a Babylon do during the pandemic? You know, was it for you just business as usual but kind of on on steroids or have your services been adapted?
Ali Parsa 25:24
In a way it was business as usual. And in a way it was almost everything we've done up to now in the last five to six years was in in in waiting for this moment. Because if you look at what they describe as telemedicine companies, which is putting a doctor or a nurse behind the mobile phone, but a doctor or nurse has to take as much time and he's as real and he's as expensive behind the mobile phone as if they were inside their clinic. They're more accessible, they are more convenient, but they are no more affordable. So what you saw you is many of these telemedicine companies just have to go and rush and hire more and more doctors and nurses to deal with the excessive demand. What we saw was that we saw about, we quadrupled the amount of engagement we had, interactions we had the clients we had, but we barely had to increase our numbers of doctors or nurses, because what we managed to set up was to use technology to parse demand into the most appropriate way of solving it. So some, for instance, one of the things we did Vivian was we saw very early on people are calling or doctors and saying well everybody says wash your hands. How do I do that? Now when accessibility is immediate, you can imagine people very easily doing that. So we put produced a simple video there's nothing AI about it, that have over 100,000 people a day, were watching very early on, on how do you wash your hands clinically, and then and then obviously, when you go and present yourself with the symptom to a doctor, in early days, it was very simple basically rule base. If you have temperature if you have this, that the other, then you have to survive them. And so it's very easy. There's nothing more AI about this. It's a basic rule based thing to digitise that and give it to people. And then when they're at home, they need to be monitored. So our technologies monitor them. And if they needed to ask people something, what we found was most of the questions were highly repeatable and highly, almost non clinical. So we put in place a group of clinically supervised but non clinicians to look after people and their basic questions to a chat system. And, and long story short, what we saw was that almost 80% of our demand was dealt with without going to our doctors. And that was what a variety of technologies - AI has become really fashionable, but we need to be careful that we don't try to have a hammer. And then see every problem as a nail, sometimes much simpler, digitising technologies will do the same job.
Vivienne Parry 28:14
So there is, is almost doing a job of a kind of Harry Potter sorting hat. You know, it's, it's putting people into, you know, readily answered question buckets, then reserving doctors for perhaps what they're best at is the more complex diagnostics.
Ali Parsa 28:39
Yes, an AI can do some complex diagnostics too. We just recently published a paper our scientists published a paper in the nature communication, where he showed that on a series of tests and that's on the test environment, I want to emphasise that because for AI is so much easier to operate under control test environment than the real life. But under control test environment, it outperforms 73% of the doctors that it was that sat the same tests basically - and actually performed better when the situation was more complex, but it did so not by using basic machine learning techniques that are pattern recognition that was referred to earlier. They did so by using counterfactual simulations, which is which is a different technique in in AI, as so AI can do quite a lot. But yet it can do quite little when you compare it to human brain. And I think and I think there is a lot of hype and exaggeration that goes on. And what we find is exactly as you described, Vivienne, is that is has its best result when it is truly in the hands of the hands of human experts to get them to simplify and automate some of what they have to do.
Vivienne Parry 30:11
So a final question for you, Ali, how do you protect patients when collecting the data which you need your systems to train on and develop the service?
Ali Parsa 30:25
It's a really important question. And we try to the constant complaint and detriments of our scientists and our conditions to protect that data incredibly, incredibly carefully. And nowadays, obviously, it's much better because there's legislations in place that would force everybody to ask the patient or any user of either of any data, whether the data can be used for general research, but we try to make Those that data that we use is an identifiable as much as, as possible and collectivise. But it's a constant. It's a constant, if you wish, pursuit of excellence in this, there is no silver bullet. And I'm glad that more and more people are waking up. I mean, the other day, I used the tool test to see what Google has on me. I was shocked to see the amount of data they keep on me, which if I knew they had, I wouldn't want them to have. So it's it's an it's a very important work that Michael and others like him are doing to bring the ethic back into back into this field and ask the hard questions that policymakers need to ask because I think if you leave the market to its own devices, I am not sure it will end up with the right hands. And I think that's an area that we increasingly need to have clear policies on and impose a safeguard in solution.
Vivienne Parry 32:09
Thank you very much. Now I want to ask you all, because AI has, as we've heard been a very useful tool in understanding and in supporting the work of the health services during a pandemic, and adapting to the new normal. But I wonder how has AI itself been affected by the pandemic? And to finish up on episode this week, I want to ask each of you, how has the pandemic changed your thoughts about how we should use AI in the future? Let me turn to Michael, first of all on that.
Michael Veale 32:44
I think that that one thing that we've learned through the pandemic is, is how to deal not necessarily successfully with really, really strong uncertainty and going through and really being faced with something that is hugely disruptive to society but but which we Don't understand the dynamics of we don't understand how they're changing many, many uncertainties and having to get to grips with with also how the scientific method interacts with these uncertainties. And insofar as machine learning is a part of AI, machine learning is is really quite terrible at at predicting things that has not seen before, which changed quickly. And I think that the has really proven the value of a lot of more traditional modelling methods of understanding dynamics and putting them manually into models. And also the uncertainty that comes with those models. So we've seen a huge amount about the politics of modelling play out how diseases going to spread or not, and I think it's a it's created awareness and appreciation. And hopefully, it will take us beyond focusing on AI as a as a statistical silver bullet, and have us think more carefully about the role of computing statistics and modelling in society more generally.
Vivienne Parry 33:58
So for you, you know AI has been put firmly back back in its box. Kate, how about you?
Kate Jones 34:05
I think Allie and Michael have really eloquently said, you know, what I think about this fact that AI is is is a tool in human expert hands to help them. And I agree with what Michael and AI was saying completely. But But I also would like to add that I've been quite infused and encouraged by the creativity of a lot of people using AI for this pandemic. So for CT scans, so using AI to diagnose CT scans, and also thermal imaging of people's temperatures. So I think there's quite a lot on on testing and diagnosis, which is, is really interesting. And also kind of looking at Internet chatter to pick up clusters of symptoms, I think is is a really cool way of kind of an early warning system and an even like, just experimenting with different models and and combining different models together. So there was a, a group in Shanghai who have produced an AI model to kind of rival some of the process based models from Imperial, for example. Now, I think that's quite good, really, it's quite good to have different approaches and different thoughts on this problem.
Vivienne Parry 35:20
So Ali, have you been, you know, going through the pandemic, actually thinking about the future, you know, thinking about potential that's been thrown up by the pandemic for AI?
Ali Parsa 35:33
Absolutely. I think what the pandemic showed is that when the pressure comes, the existing systems will find it very difficult to cope with the extra demand. You must never forget Vivienne, that we spend $10 trillion annually on health care globally, and we only look after half of the world's population. Five out of 7 billion people in the planet have no access to secondary care, or surgery or a hospital. So we need to do something differently. And I think what AI has shown in this pandemic, but I think is true even before is that it will be a very important tool in enabling us to scale our ability to deliver health care to many more people. And I think the promises about AI is true about the promise about most technologies, which is at the beginning, it will do a lot less than people hype it to do, but eventually it will do a lot more than any of us can today, imagine. And so I am both encouraged, optimistic and scared and pessimistic about what it can bring because it's a Pandora's box we are opening and we better know what we are opening on how to control it, because I think it's going to do far beyond any of our imaginations today cannot typically.
Vivienne Parry 37:06
Well, that's a perfect place to end our absolutely fascinating discussion today. You've been listening to Coronavirus the whole story. This episode was presented by myself Vivienne Parry, produced by UCL with support from the UCL health of the public and UCL grand challenges and edited by the splendid Cerys Bradley. Our guests today were Professor Kate Jones, Dr. Michael veal and Allie Para. The episode was guest produced first for us by Professor Geraint Rees. If you'd like to hear more of these podcasts from UCL Minds. Subscribe wherever you download your podcasts, or visit ucl.ac.uk forward slash Coronavirus. This podcast is brought to you by UCL Minds bringing together UCL knowledge, insights and expertise through events, digital content and activities that are open to everyone. Looking forward to meeting you all again soon. Bye for now.