Steve McLaughlin: In 2020, COVID-19 entered the U.S. and changed our daily lives. Quarantining, face coverings, social distancing, and working from home have become the new normal, and many continue to face even greater challenges. The spread of COVID-19 is flattening out in some countries and areas of the U.S. and still increasing in others, and many experts warn of a second wave in the fall. So how do we move forward in the face of COVID-19 and its continued impact on our lives? And what should we do to plan for the future?

[steam whistle]

[applause]

[marching band music]

I'm Steve McLaughlin, dean of the Georgia Tech College of Engineering, and this is The Uncommon Engineer.

[music]

Man: [archival recording] We’re just absolutely pleased as punch to have you with us. Please say a few words.

[applause]

Steve: Our guest today is Professor Pinar Keskinocak. She's the William W. Chair from the H. Milton Stewart School of Industrial and Systems Engineering. She's also the co-founder and director of the Center for Health and Humanitarian Systems. Welcome to the show, Pinar.

Pinar Keskinocak: Thank you for having me. Great to be here.

Steve: You know,  just if it's OK, I just really want to jump straight into COVID-19. You know, we're recording this at a really difficult time for a whole bunch of reasons, and we are going to focus just on COVID-19 right now. You know, we've been working remotely what seems is probably over three months. And I know that your work is really, and is receiving lots and lots of attention around modeling— mathematical modeling— modeling that really predicts where the disease is, where it's headed, who's affected, and how those models get used for planners, for physicians, and so super, super important work. And so I think if it's OK, that's really where I would like to start is for you to share a little bit about the work that you're doing in modeling, and then we can just kind of go from there.

Pinar: Sure. So first of all, thanks again for having me here. It's  a pleasure. And let me first start by acknowledging our research team. I am one member of the research team working very closely with my colleagues Nicoleta Serban, John Asplund, and Ph.D. students Arden Baxter, Buse Eylul Oruc, April Yu, Daniel Kim and several others. I'm sure I'm forgetting some names now, but we have a pretty large team and it's been really an amazing group to work with.

So we focus on a couple of different directions. One of the things we try to do is to project the spread of the disease geographically and over time. And we do this by also considering the behavior of people, because how the disease spreads is going to depend in large part, how the public behaves. And so we call these nonmedical interventions, such as school closures, shelter in place. We looked at voluntary quarantine, which means that if a person in a household is experiencing symptoms, not only that person, but all the other members of the household also stay home until the entire household is symptom-free. So these are some of the interventions that we looked at. And in the absence of a vaccine and in the absence of a widespread available antiviral-type treatment, these nonmedical interventions are really our best options at the moment in terms of slowing down the spread of the disease. So that's basically what our disease spread modeling effort focuses on, looking at these different scenarios based on the combinations of interventions that may be in place at any given point in time, and of course, also people's compliance levels. Not everybody is going to comply with the interventions at the same level, so how does the disease spread the change, depending on the combinations of interventions as well as people's compliance levels. So that was one of the research focus areas for us.

And then we also look at the impact on healthcare capacity, specifically looking at the demand for hospital beds, ICU beds, and ventilators, so under these different scenarios, what kind of demand do we see specifically for COVID patients for these resources and also try to project where we might see shortages and when and at what levels.

And then finally, we also look at hotspots, looking at the trends and the changes, specifically in the data in terms of confirmed cases— can we say that certain areas are likely to become hotspots in the coming days, which gives us opportunities to better prepare for response in those areas. So those are some of the focus areas of our research team.

Steve: You know, there's so many aspects to being able to accurately predict, you know, what's happening. And you touched on, you know, your medical aspects. You touched on human behavior aspects. You touched on all kinds of various things. And I know many of our listeners are high school students, junior high students, engineers, engineering graduates, so I know, certainly I am, I think a lot of people are interested in how do you develop such a model? You know, what does a model look like and what are the kinds of tools? And how do you incorporate all this really varied information into something? Because, I mean, the other question is you're  an industrial engineer, and, you know, I think people don't necessarily think of engineers as developing a disease spread model, so I know I just threw a bunch of things that you there, but people are really, really curious about the model and how you develop the model and do that.

Pinar: So maybe I'll just give a very quick overview of the different types of models, because there are so many models out there, and broadly speaking, they fall into one of three categories. So we have what I would call the “curve-fitting models”. These models look at the data, the spread data [of] confirmed cases. And then, just like you will do, let's say in a forecasting model, they try to look at the trends and then from there try to project how things might look like, assuming more or less if similar trends follow. So these curve-fitting models tend to be quite good if you are looking forward, let's say, one or two weeks, maybe up to three. But beyond that, it's difficult for them to capture some of these other dynamics that we discussed earlier, such as human behavior or different interventions and so on. So that's one category.

And another category is what we call “SIR models”, standing for susceptible, infected, recovered. But these models basically put people into one of three categories. Because this is a pandemic, initially none of us have any immunity for the disease, so we are all susceptible. We all could get sick because we don't have any prior immunity in our body. And then at some point in time, a person may get infected and then eventually, you know, hopefully they recover. So you fall into one of these three categories, each individual at a point in time.

And then the recovered category basically assumes that once you recover, typically you will not get infected again. So we don't know if that's the case exactly for COVID yet, you know, hopefully or most likely that is the case. So these models can track these some of these different interventions to some degree, because we can say, OK, we have a school closures, then that means some of the interactions are going to be lower, the transmission is going to be lower. So they are able to incorporate some of these to some degree, but not usually at the high, high level at the population level, maybe at the subpopulations divided into age groups.

And then the third category is what we call “agent-based models.” That's the type of model that our team developed. And in the agent-based models, you have an agent and entity in a simulation representing each individual in the population. So we are essentially mimicking the entire population in a simulation model. So each of these individuals, what happens when you get sick? Some people might experience symptoms. Other people may not experience any symptoms. And then depending on whether you are symptomatic or asymptomatic, there's a chance you might pass the disease to other people. So there is kind of the individual, the progression of the disease within an individual, we capture this depending on age, and it's not the same for everyone. And then we also capture the interactions in the population where we model interactions in a household, interactions in a peer group which may be school or workplace, and interactions in a community such as a grocery store or a church or other public places that people get together. The agent-based models, in a way, capture a lot of the details that happen in the population, and they also make it very easy for us to implement and test different intervention strategies such as the ones that we discussed—school closures and volunteer quarantine and other disease— other interventions.

And the nice thing about the agent-based models is that even if an intervention has not been tried before— so we may not have data to fit for that particular type of intervention— we can still incorporate it in the model because we are explicitly modeling the specific human behavior. So they have a lot of flexibility. They are more complex than some of these other models and hence they take a bit longer to run, but they give us a lot of flexibility to do different things. So I mean, there may be other types of models, but broadly speaking, most of the models we see out there would fit into one of these three categories. And I think as a researcher, it's fantastic and wonderful to see these different types of models implemented and used by different research groups, because I think we learn something from each of them. And when we look at the results, you know, in an ensemble incorporating the projections from these different models, it gives us a lot of, you know, opportunities to observe what may happen in the in the best case, worst case, average case. And so I think it's nice to have these different options.

The CDC has on their website actually a collection of these models in ensemble, our model is also, you know, one of the models in the ensemble. And I would encourage everyone who is listening to take a look at it and see how these different models come up with different projections over time.

Steve: It feels to me like it exists at several scales, right? It sounds like you're able to model an individual and that individual's interactions with different groups within family. And then there's another scale around how those different groups might interact or how the disease spreads beyond those groups, and then at an even greater macro scale, how it transitions across states and country boundaries and things like that. Is that is that true? I mean, your model, even just the agent-based?

Pinar: That's exactly how it works. And so we assign each agent, each person, in our simulation to a household so they have a household. And then we use data specifically, you know, our model is focusing in terms of data on Georgia. We use Georgia data in calibrating the model, in populating the model, but it could be applied to any other geographic region, really. But we would use, for example, depending on your location, on the agent's location, what would be the household size. So all of that information we get from the census. So each person in the simulation belongs to the household, and then they belong to a peer group depending on their age— this may be a school group or a work group. Again, we look at the data in terms of school sizes, depending on the location, the school size or the work group size might differ. And then people travel. They travel for work quite a bit. We have the work travel data from Georgia. So you may be living in one county in one census this track, but traveling somewhere else, so then this is going to create interactions between different locations because people from different locations aggregate in a workplace and then go back home. And then as we discussed earlier, we also have the community interactions that, you know, you may be getting together with people from all over the state in some cases, probably more so from your own community.

And then in terms of transmission, the chances of transmission, you might imagine, is higher if you are in the same household with someone. It's a little bit lower if you are in the same workplace or peer group such as school. And it's even lower if you are interacting in the community because you are not spending so much time with them. So all of these different effects of transmission are incorporated into the model so that we can capture, you know, the likelihood of transmission depending on the interactions.

Steve: I mean, it just seems so incredibly complex, right? How you both build the underlying mathematics that describes all these things, but then more importantly, you want to be— the model needs to be informed and needs to have the right data. And whether the community— I don't know who I mean by “community”— but whether the community cooperates in providing that data, whether the community competes to share data, how easy is it for you to kind of pull those parameters broadly into your model? And I guess the last thing I'll say, I can now understand why your team is so big, right? That you know—

Pinar: Yeah. So I mean, the data part can become really tricky. So thankfully, a lot of the data comes from the census. So in terms of the, you know, age distribution based on location, the household sizes and so on and so forth, as well as the group travel, so that data is available. And, you know, we, of course, have to do quite a bit of work to take that available data, clean it, and then put it into the model, but it is out there.

What is more complicated is the behavior-related data. How do you know what percentage of the population will actually comply with a shelter in place order? Or after shelter in place ends,  anecdotally we know that some people are pretty much continuing voluntarily shelter in place, staying working from home and not interacting with others, whereas other people are not. So how do we know, you know, exactly what these percentages are? Of course, we don't know, so we use some estimates. For example, there is data available based on cell phone usage showing over time how far people travel and for how long, so that gives us some idea about the changes in behavior patterns, you know, how it was when you were in a shelter in place and how it's changing now over time. So we basically look at that kind of data and then that informs our model in terms of developing some scenarios. For example, we might say, let's say 80 percent of the people were following shelter in place, and then we fit the model and compared it to the actual data and see if that gives us a reasonable estimate. And then we might say, what if maybe 60 percent of the people continue sheltering in place one week after?  [It] maybe drops to 50 percent two weeks after shelter in place ends, but they are still voluntarily continuing. But will it this tapers off and then we reach maybe a plateau? Maybe five percent of the households are still continuing, and everybody else is going back to business as usual. But we can test a little of these different scenarios. And then as new data becomes available, we compare and contrast the model predictions with what we see in the data based on the actual confirmed cases and then make adjustments if needed. But because we are able to test a lot of different scenarios, for the most part we are able to capture what the range might be, you know, depending on the different behaviors and how it reflects in the actual number of infections we see.

The one thing that makes it really, really challenging is the fact that our testing capacity has been quite limited, which means that there may be many more infected people as well as infected and recovered people out there that we do not know about. So therefore, what we do is we take the reported numbers and, again, based on estimates, we say maybe the actual numbers might be about six or seven times what the confirmed cases are based on the testing availability and other information that we have. So there is a lot of— there are a lot of estimates in both calibrating the model parameters as well as how we interpret the number of confirmed cases that are out there.

So I guess despite the uncertainties, I think the models still give us very good information about what might happen under these different scenarios, under these different interventions, because otherwise we will be literally flying in the dark, you know, not even knowing the impact of, you know, a particular intervention or relaxing an intervention and, you know what we might see. And these are the kinds of situations where doing an actual experiment— oh, let's try and see what happens socially, economically, and from a public health perspective, very expensive experiments, right. I'm using the word “expensive” in a very general way, so I think that's where the models are quite helpful, you know, in terms of helping the decision makers to make more informed decisions.

Steve:  As most are familiar with, you know, hurricane models, right, we see all these pictures of, you know, the eight or 10 models that show the pathways of the hurricanes and so on, and I keep— one, I don't I don't want to oversimplify it and I don't want to offend, you know, you or others, you know, kind of trying, if there is or is not too much equivalence, but then, you know, when you talked about the models that the CDC has posted, I just have a whole bunch of different questions and thoughts. And I don't want to, like, stray into politics at all because it seems like that there's a wealth of— there's a wealth of models. Like you said, there's an ensemble of models that probably point to some concrete conclusions. And I'm curious about, you know, number one, how well the modelers, the different— I don’t know how many different models there are out there— that would be interesting to hear, you know, and then as someone assembles, you know, the general conclusions from those models that those feed into certain kinds of decision-making. And I'm really curious about— because it sounds like you're right in the middle of that— and I'm just curious about like what happens next when you have 20 or 30 or 50 or 100 models and how all that gets pulled together and who uses them, and do you collaborate? Do  people work well together on those?

Pinar: And so you are asking a fantastic question— How do we make sense out of these multiple models, and how do we incorporate them into our decision-making? So the models that I was referring to earlier, similar to our model, focus on projecting the disease spread, especially under different scenarios regarding intervention strategies. So in a way, our efforts, as well as some of the other efforts, take the perspective of public health. So we look at a number of infections, a number of infections with symptoms, number of hospitalizations, how many ICU beds might be needed. And we look at it over time so, of course, the cumulative numbers matter— how many people, will be infected through the course of the disease— but we also look at what might happen at the peak, which is the day or, you know, maybe a couple of days a week where this is really the most intense— we see the most number of infections, new infections, daily. So that's what we call the peak. And why does the peak matter? Because that's when our resources will be most stretched, right. So you have so many new infections, so many people who might need hospitalization or ventilators, and that's the resources are really, really stretched at that point, particularly from a public health perspective. So that's why over time we kept hearing this phrase, “flattening the curve,” right. So what is the curve? It starts slow with a small number of infections at the beginning, and then it picks up speed, and then it reaches a peak, and then after that, it will come down. This is kind of the most traditional way. If you did nothing, that's the kind of curve that you would see. But because we are doing interventions, we might actually see multiple peaks. So we are already observing, if you look at the data for Georgia on the Georgia Department of Public Health website, we already had one peak, and then because things slowed down during sheltered in place, we start to see things coming down, and then the number of new patients started to pick up again after shelter in place was released— another peak, right? So there may be multiple peaks because of these different kinds of public health interventions, businesses reopening, and so on.

And so, again, the focus of our model and several of the other models that project disease spread is on public health. So we basically look at what combination of interventions would reduce the number of infections or reduce the peak or delay the peak. So those are the kinds of things we are interested in just focusing on the on the health side. And of course, if you are a decision maker, that's not the only thing you would be considering. So we might find out that if you throw in five or six different interactions or just simply tell people all of you are going to stay home for two months, no going out, nothing; we’ll stop the whole country, yes, that might stop the disease, but that may not be feasible for various reasons. So there are all these other factors that come into play. The closing schools re-quantify public health impact of closing the schools, but then what about the social impact? There are already studies that show that being away from school and having to do schooling online disproportionately affects African-American and Hispanic communities and may increase the gap, and it may be more difficult for certain subpopulations to catch up with school if and when we go back to regular schooling. So those are some of these other considerations. Of course, there is the huge concern about the economy and so on. So I think the decisions are very complex. There are different models that capture different aspects of these decisions. Our model focuses on the public health aspects. There are other models that look at the impact on education, the impact on the economy, and then it's up to the decision makers how to pull the information from these different models, understand to the extent possible the tradeoffs, and then make, hopefully, informed decisions. I wish there was one big mega model that put all of these together and gave us the magic answer about what to do and when, but we are not there yet— maybe in the future.