Coding COVID-19:  Computer Scientists Create Vaccination Strategies Using A.I. 

Rhys Waters 0:00

Welcome to Beyond Research. A podcast brought to you by Research Nova Scotia.

Stephanie Reid 0:08

Throughout the pandemic frontline health care workers and public health officials have been tasked with making difficult decisions about where to expand their finite resources, like beds, test kits, and eventually vaccines.

James Hughes 0:21

There might be a really good chance that when a vaccine becomes available, there might not be enough for everyone. You know, there’s logistical issues. There’s manufacturing time, we’re competing with every other jurisdiction on the planet to get their hands on those vaccines.

Stephanie Reid 0:34

This is Dr. James Hughes, an assistant professor in St. Francis Xavier University’s Department of Computer Science. Today, you will hear how his research in artificial intelligence aims to help public health officials answer the question, Who should we vaccinate first? Well, welcome to the podcast. James, we’re very happy you were able to join us today.

James Hughes 0:49

Very happy to join you as well.

Stephanie Reid 0:51

So obviously, your work has always focused in AI. How has that shifted during the pandemic this year?

James Hughes 1:07

Well, it’s it shifted significantly. I mean, if you asked me a year ago, even like, oh, would you be studying something like epidemic models, I, I probably wouldn’t have believed you if you said I would be before. A big part of my research program is focused on neuro-informatics, which is just a fancy word to talk about kind of like technology, computer science intersecting with neuroscience. But when COVID-19 hit it, I mean, like a lot of the researchers across the world, you know, it piques your interest, you think, you know, what could I do? Is there something along in my area of expertise, I could apply to this problem. And one of my colleagues that I hadn’t worked with in a while, but Professor Ashlock from Guelf, who was one of the collaborators on this project, I knew in the past, he’d done some epidemic stuff. So I reached out to him. And then we started to realize, really, we could probably start applying our our expertise for looking for ways to help the world with respect to COVID-19.

Stephanie Reid 2:04

So James, I’m hoping you can just give us a brief overview of your area of research, and your recent project involving AI and COVID-19.

James Hughes r 2:15

My area of research is the applied AI, applying AI to interesting new problems. AI is still a relatively new field. And as the technology becomes more and more effective, as we come up with new algorithms, as our computers become more powerful it becomes more practical to apply this technology and ideas to a number of different problems. In terms of what we’re doing now. We’re using the AI to to come up with vaccination strategies. So the reason this is important is, you know, we’ve got COVID-19 the pandemic, you know, there might be a really good chance that when a vaccine becomes available, there might not be enough for everyone on day one. We’re competing with, you know, there’s logistical issues, there’s manufacturing time, we’re competing with every other jurisdiction on the planet to get their hands on those vaccines. And we might expect to get some maybe shipments coming in every so often to have so many vaccines. So the question we had was, is there a clever way we could apply vaccines to a population based on what the shape of their social contact network looks like? And can we apply a very few and still really reduce the spread of the disease. So the World Health Organization, the Centers for Disease Control has a lot of important, very critical, very important guidelines that they have for how and when and who should be vaccinated first, and these are very important, nothing about what we’re doing is suggesting that these things should not be considered. All we’re looking for is a new way to provide new information to include with these considerations to hopefully help create more intelligent decision making.

Stephanie Reid 3:59

Right, you’re leveraging AI to discover and investigate new strategies in regards to covid-19 testing and treatment. So treatment in relation to who gets vaccinated. First, can you explain a little bit about what that looks like in practice?

James Hughes 4:18

So our idea was, you know, we can simulate the pandemic computationally. And we could also come up with vaccination strategies, like a simple strategy might be vaccinate someone who comes into contact with a lot of people, let’s say, and we can test these ideas on the simulation by kind of applying these vaccines during the simulation and seeing how those vaccines change the outcome of the pandemic in the simulation. So maybe we’re looking for things like how many people were infected in total, we want to minimize that amount, right. The problem with this is, I mean, we can sit down here and probably for another hour and a half on this podcast and talk about vaccination strategies, well, maybe we could do this, maybe we could do that. And these ideas, they’re probably fantastic. But there’s a lot of guesswork, and there’s gonna be a lot of testing involved. So our, our idea was, well, we can use AI to automate this process to look for these ideas, use ideas that seem to work and combine them with other ideas that might work in ways that we would never consider. So this is where we’re harnessing the AI. And we’re hoping that the end result, these little strategies that they produce, we can then go on and use and the policymakers, decision makers, they can take this idea and incorporate it into their decision making for prioritizing vaccine applications. This, of course, all does assume that there’s a limited supply. If we have, you know, a million vaccines, one for every Nova scotian. Tomorrow, well, fantastic. Give them to as many people as as they want. I mean, we’re in great shape, but we’re assuming less than ideal situation.

Stephanie Reid 5:56

Right. And obviously, the recent announcements from Pfizer and moderna have brought the subject of vaccine distribution to the forefront of the COVID-19 conversation. And you also have a lot of people discussing things like the ethical and political considerations around the acquisition of these vaccines and the advanced market commitments that have already been made. What are your thoughts on this? and How could your research help?

James Hughes 6:27

There are ethical considerations that came up when we were designing and using our system, but when talking about I mean, you look at Canada, I think Canada, as of right now has the most vaccines on order. And a lot of other developed countries in the world there, it’s the same thing. And you’ve got a lot of developing countries that don’t have the ability or the resources to get these vaccines. And I mean, we’re talking about a worldwide level of public health inequality, this is a big problem. If the technology we’re developing comes up with fantastic strategies that are effective, and people are following guidelines and rules, this, our technology, it won’t stop the spread of the disease, the strategies that we’re coming up with, well, it won’t stop the spread of the disease, but what it will do is slow the spread down. And if we can slow it down, it gives us that time to accumulate the vaccines and apply them such that overall the impact on people’s lives, the economy, can just be minimized. And that’s where this technology can come in place. In a more obvious thing we can talk about in the case of Nova Scotia, you know, we get the vaccines, we got the strategies, we incorporate it when we’re playing it, we’re in fantastic shape. But there might be areas where that limited supply, maybe we don’t have that much of a limited supply. But there are going to be places in this world that does have a very limited supply. And this is going to be very important for I mean, hopefully, it gives us results that are very important for the application of vaccines in those populations.

Stephanie Reid 8:01

How could… How is your modeling, addressing that? For someone who doesn’t understand what AI is, if we do have, you know, X number of vaccines in Nova Scotia, and we have a situation like that unfolding? How can we leverage what you’re working on? In the event, we did have a limited number of vaccines to stop or slow the spread.

James Hughes 8:27

In the mathematics world, we’ve got these structures, we call graphs, which is just a big word to describe a network. And you know, when you’re thinking of a network, we’d like to think of like nodes and the connections between the nodes, we call those edges. And you can imagine that, if you’re thinking of your household, every person in your family could be a node that could be represented by the little circle, you know, if you’re envisioning in your head, and all the people in your household, you’ve got a connection to you, we’ve draw those edge between them because they’re in close contact. And if heaven forbid, that someone in that household got the disease, it would very easily transfer from one person to another. But if we expand this, we take a step back, you might think, Well, you know, I go to work, I come into contact with a number of my colleagues try to distance as much as I can. So maybe I don’t have a connection to all of my colleagues. But there might be some, I get close enough that it’s possible that the disease could spread that way. And I maybe I go visit my neighbors, and so on. So you can kind of see how, depending on the scale that we can think of these graphs, whether it’s household community, whole province national worldwide, you can envision these graphs getting bigger and bigger and bigger, more connected more complex. And it’s on these graphs, these these networks based on social contact networks that we do our modeling on. So we look at how the disease spreads through these fancy structures, these graphs that hopefully represent our communities. So the idea is, our vaccination strategies are programs. They’re like a set of rules like it If degree greater than this and your neighbors infected is less than this, apply some vaccine, like it’s a program. So our a i, genetic programming is writing those programs to solve a problem. So we wrote a program that writes programs to solve our problems. And so that’s what we’re doing. But the one really good thing about this type of AI is the in the end, it gives us like a symbolic result. It gives us those programs that we can look at. So when someone comes along and says, like, you know, why did you tell me to vaccinate that person? I could tell you right there. I say like, Well, here’s the program. Here’s how it says, Here’s why. Here’s the intuition. Unfortunately, there are other types of AI out there, that they’re more black box where it says, oh, vaccinate this person. And if you ask the question, but why the answer is, because the black box told me to, and it’s usually right. And that’s the best You got it. Sometimes that works for you. Sometimes it does, if you’re a doctor, and your blackbox says, perform neurosurgery, if the physician asks you, but why? And your answer is because it’s usually right. They’re gonna say no, where if you have something that’s very explainable, you can explain why. So it was critical for us to use a type of AI that was very explainable. And so this program that writes programs, it’s very explainable. And when we take it to decision makers, we can say, here’s what it is, here’s why this makes sense. Here’s the intuition. And here’s how to apply it.

Stephanie Reid 11:29

Right. And for your COVID 19, modeling B vaccine distribution or testing, some of those algorithms or code would be telling the computer different things to do in the event, if someone was rural or urban, if they had children, no children, is that the idea?

James Hughes 11:48

Ultimately, as of now, it’s only it doesn’t incorporate things like Do they have children? Are they urban rural, the proxy for that could be things like, well, how connected they are rural people won’t be as connected and people who have children might be more highly connected. But one of the long term goals we have is to incorporate that information. We do have one of the research students working on this. We want to incorporate things like well, maybe certain groups of people are more susceptible? Or maybe you have like look like how does it interact with people who are older? Or how does it interact with someone who might be pregnant or something. So there’s a lot of complexity that we could add to our models, that might give us that much better results. So this is something that we are working towards. But that might be version like, not probably not version two, or three, but down the line.

Stephanie Reid 12:36

But going back to when you were talking about your research, and in the notion of the social networks, which we’ve heard talked about in different ways from, you know, public health officials, the media, Bubbles, clusters, community spread, community spread social networks, you’ve heard it, different variations of it, but in your opinion, and using AI, can a social network really be fully understood? And how do your clusters, nodes, social networks, work with all other non pharma measures right now? Like the social distancing, the isolating the mask wearing the hand washing, etc, etc? How does it all work together? And how is it accounted for?

James Hughes 13:26

Yeah, so there’s no question. And as you would expect, like, one of the best things we can do to stop the spread is like, minimize the number of people you come into contact with, minimize that minimize a number of possible ways that the disease could spread. This isn’t groundbreaking information. We know this, right? It’s intuitive. But one thing that we were able to show through our models, through our simulations with the AI was the two key things, and they’re intuitive there. You might even say it’s obvious, but it’s sometimes it’s really great when AI can confirm our intuition. And like empirically, so this is always really good news. So the one thing is, if you take our vaccination strategies, and you make them the contact networks, arbitrarily large, you make them huge. And you keep, like the connectedness level, the same. The strategies are similarly effective, they work just as good on the smaller versus larger networks. But, and this is where your intuition is going to come into it, if you keep them the same size, but increase their connectedness. So the number of people that come into contact with as you increase that these vaccination strategies become less and less effective. So although we’re coming up with strategies, they might only work if people are still following these social distance rules. If we’re in the world of a limited supply of vaccines, and some of the simulations we were doing on some of the better strategies we came up with. If you get connected to a certain level, it’s we end up Doing not that much better than doing literally nothing at all, in terms of stopping the spread of the disease. So it is critically important to really hit home, that the social distancing is critical, eliminating, like minimizing those extracurriculars where students can come into contact, those are your high degree nodes, if we can get rid of as many of those connections as possible, that’s gonna be really helpful for this reducing the spread of the disease. Wow. And there you have it. And But do you know what I mean? When I’m saying it’s intuitive, like what I said to you, I’m willing, like, if I asked you did that blow your mind? Probably not. But it’s great that the AI can like in our simulations, we empirically can confirm our intuition. This is important. This is good.

Stephanie Reid 15:43

Right. And are there differences between existing pandemic modeling and the modeling that your team like? Are there key differences between existing modeling and your modeling?

James Hughes 15:55

One of the key things in the design of our system is it’s modular. And we can kind of switch out the types of models we’re using for the pandemic for a specific disease. And we can change the the the disease parameters, like how likely it is to spread and whatnot. So all these things you can imagine, there’s just a bunch of knobs that we can tune and then run the simulation. So typically, when one of the most standard epidemic models, some of my use is the, they call it the the SI r model, where you have people in a population one, they’re either susceptible, they are infected, and they can spread it or they have been removed. So maybe they have passed away, or they’ve recovered from the disease. That’s, that’s your standard one. But that’s not great for COVID-19, because it doesn’t incorporate the really long incubation period, which is one of the things that makes this disease particularly difficult. So we’re using the S e. r. And this is actually like this is very standard. And what a lot of COVID-19 modeling is that this is what a lot of people are using. And it’s it’s basically you’re susceptible, and then you become exposed. We don’t know you’re exposed, we don’t know you have it, but you do have it. And it’s just a matter of time before you become infectious and you start showing symptoms.

Stephanie Reid 17:11

So the asymptomatic carriers.

James Hughes 17:14

Well, we need to be careful because asymptomatic can mean something different. In that you have it you are past the exposed period, you have it but you’re not showing symptoms, but you’re spreading it to people. So it’s funny that you say that because one thing we’re working on is we’re we’re you’re defining… Well, I suppose we’ve already defined it, we just haven’t included it yet, is a better resolution model that kind of tries to incorporate more information. So this is we call it the sea air model, because we had to call it you know, fancy and here we are on the Atlantic Ocean, right? So. So we have susceptible, and someone who’s susceptible, if they catch it, they become exposed. And when they’re exposed, we don’t know they have it. And they’re not spreading it yet. Because as you know, with the incubation period, there’s about a there’s quite a bit of time that you have it but you’re not spreading it. But then there is a little bit like about one two days where you have it, you don’t know you have it, you don’t have any symptoms, but you can spread it. So that is the E that’s the second he we put a little a little like apostrophe next to it. So it’s special. But then from there, you can become one of two things, you can become asymptomatic or infectious. If you become infectious, you’re spreading it, but we know you have it. If you become asymptomatic, you’re an asymptomatic carrier, you have it, you’re spreading it, but we’ve never we have not identified you to have it. This is where testing would be very, very important. But then after an amount of time, if you’re asymptomatic, or you’re infectious, you can then go to remote. So this is a bit of a higher resolution. So like maybe we’re kind of like overcome making our models a little over complex. So maybe once we start running these models, maybe it wasn’t that necessary to get this complex in this specific with our models, which would be good because if we can use a simpler model, that’s, that’s awesome. But maybe is required to get even better results. That’s more practical for COVID-19. But if tomorrow or the next day, a year from now another disease comes along. Well, we can use the same technology, tune those knobs, replace it, like plug and play what the what the standard model we’re using is for our simulations to apply to that disease as well. So it’s supposed to it should be very generally applicable.

Stephanie Reid 19:30

Do you think it’s reasonable to think that there’s one, like once we have a vaccine, that there’s one authority that can manage that distribution? And you had mentioned that you have preliminary results in your study? Like what is the timeline of the final development of this tool? What does that output look like? And is it something that Nova Scotia can latch on to immediately and actually deploy?

James Hughes 19:53

Well, so the first part of the question about the central authority if we’re talking about The level of the province, maybe we can get away with that. But we but you consider every community has its own nuances. And the decision makers in the specific communities are going to be more familiar with the nuances of their community. So perhaps the having those more local decision makers, they might have a, just a better idea of how to take the information we have that we give to them, and they can consider it and incorporate into their decision making. In terms of the timeline for our system, it’s like, for the most part, all the pieces are there a version, let’s say version one, we’ve had a lot of plans on what we want to incorporate into it what we want to make it better, but as of right now, it’s it’s pretty clunky and ugly looking code, you know, programs, it’s very ugly looking code and be very difficult to hand it off to someone who, who doesn’t have the expertise as like a programmer. But the goal over the next little while, and we’ve got a number of research students, both masters and undergraduate students and a PhD student working on the project, that’s goal is to make it a little bit more user friendly. And we are also we’re making a website where we can disseminate this information as well. The code is currently it’s publicly available. It’s on GitHub, which is this repository for code everyone uses. We’re making a website to disseminate this information, we got to get other people, other researchers interested and involved. Because if we can get this technology into the hands of other people, they’re going to have their own ideas. So you know, science is just a series of stepping stones, and I have my ideas. But if we can get other people involved, you know, we, as humans will just be that much more likely to have an even better solution. So I mean, there’s a lot of different things to consider when it comes to what makes a good strategy. And we’re a small team. So we’re hoping that our ideas, yeah, we can give them we can hand them off and just say, here’s the idea. It’s empirically tested, take it, use it, consider it when you’re making your decisions. But really, we’re hoping other researchers latch on to this project.

Stephanie Reid 22:04

So certainly, your team is working on some incredible work. And I’m excited to see how it’s applied to both the covid 19 pandemic and future infectious diseases. Hopefully, there are very few, but I can definitely see, as you mentioned, the general applications for this work as well long term for Nova Scotia in Canada, and beyond. So before we go, I just want to thank you so much, James, for taking the time to chat with us today.

James Hughes 22:31

Yeah, thank you.

Stephanie Reid 22:32

Wonderful. Well, thank you so much, James. You’ve been listening to Dr. James Hughes, an assistant professor in St. Francis Xavier University’s Department of Computer Science.

Rhys Waters 22:47

To find out more about this podcast and the research featured in this episode, visit researchnovascotia.ca. My name is Rhys Waters, and we will see you next time