WEBVTT Kind: captions Language: en 00:00:00.789 --> 00:00:06.700 This is a project about improving computational epidemiology with higher fidelity models of 00:00:06.700 --> 00:00:07.740 human behavior. 00:00:07.740 --> 00:00:12.540 This is a project that I'm doing with my Co-PIs, Christian Lebiere and Mark Orr,]. 00:00:12.540 --> 00:00:16.270 Christian’s at CMU [Carnegie Mellon University] and Mark is at University of Virginia. 00:00:16.270 --> 00:00:22.980 And we have a larger cast of contributor - contributors that are working on all varieties of things 00:00:22.980 --> 00:00:30.860 ranging from natural language processing to epidemiology. 00:00:30.860 --> 00:00:37.700 This project was motivated by the realization that last year, we were in the midst of, historically, 00:00:37.700 --> 00:00:41.870 the most massive attempt ever to change human behavior. 00:00:41.870 --> 00:00:47.390 And I'm talking about specifically non-pharmaceutical interventions, such as social distancing, 00:00:47.390 --> 00:00:52.540 hand washing, mask wearing and now, vaccination. 00:00:52.540 --> 00:01:02.810 And decision makers and people who are trying to manage public health, rely on epidemiological 00:01:02.810 --> 00:01:08.180 models to forecast rates of infections and deaths, and to try to understand what the 00:01:08.180 --> 00:01:13.100 possible effects are of these NPIs - non-pharmaceutical interventions. 00:01:13.100 --> 00:01:19.841 Unfortunately, a lot of these models are not very detailed and have a huge abundance of 00:01:19.841 --> 00:01:20.841 uncertainty. 00:01:20.841 --> 00:01:25.950 And to some degree, we believe that that's partly because they do not really have accurate 00:01:25.950 --> 00:01:32.619 models of how people respond, psychologically, and behaviorally to NPIs into what is going 00:01:32.619 --> 00:01:34.840 on in the environment around them. 00:01:34.840 --> 00:01:39.729 And just to give you a concrete example, this is one of the many models that we've seen 00:01:39.729 --> 00:01:45.480 on the news, or on the web over the past year. 00:01:45.480 --> 00:01:48.950 It's a snapshot that I took in October. 00:01:48.950 --> 00:01:54.620 And on the right hand side of that graph is a projection for the fall - the month following 00:01:54.620 --> 00:01:56.670 October, when this was presented. 00:01:56.670 --> 00:02:02.870 And what you can see is this huge pink error bar around the prediction and that the difference 00:02:02.870 --> 00:02:08.070 between the top and the bottom of that confidence interval is an order of magnitude. 00:02:08.070 --> 00:02:12.340 And within that pink area, you can reasonably say: things might go up, they might stay the 00:02:12.340 --> 00:02:14.519 same or they might go down. 00:02:14.519 --> 00:02:17.329 So these models have a large degree of uncertainty. 00:02:17.329 --> 00:02:27.019 And we are making the bet that by understanding and modeling more specifically an individual's 00:02:27.019 --> 00:02:30.030 psychology, that we will be able to do better. 00:02:30.030 --> 00:02:34.970 And this is partly because I think we all believe that people's beliefs and attitudes 00:02:34.970 --> 00:02:39.870 and intentions and self-efficacy all have an impact on how they respond. 00:02:39.870 --> 00:02:45.019 And there's certainly evidence out there that that is the case. 00:02:45.019 --> 00:02:49.530 And it is also the case that these responses seem to change over time. 00:02:49.530 --> 00:02:55.579 So we've heard a lot about COVID fatigue, and how people's attitudes change over time. 00:02:55.579 --> 00:02:58.890 And these things also seem to vary across regions. 00:02:58.890 --> 00:03:03.099 So some regions seem to respond differently than others. 00:03:03.099 --> 00:03:11.169 So our aim was to build computational predictive models that are based on a variety of things 00:03:11.169 --> 00:03:13.609 that we had already been working on. 00:03:13.609 --> 00:03:19.269 So one set of things was around theories of individual health psychology that some of 00:03:19.269 --> 00:03:20.280 us had worked on. 00:03:20.280 --> 00:03:28.269 Another is a large amount of experience with a particular theory and computational modeling 00:03:28.269 --> 00:03:33.540 system called ACT-R, which allows us to build computational models of behavior change and 00:03:33.540 --> 00:03:36.919 to develop agent-based simulations. 00:03:36.919 --> 00:03:41.829 And so out of this, our goal was to develop what we call psychologically valid agents 00:03:41.829 --> 00:03:48.200 that we can build into agent-based models that will allow us to accurately predict the 00:03:48.200 --> 00:03:54.989 dynamics of changing behavior over time and how those dynamics are impacted by these NPIs 00:03:54.989 --> 00:04:02.229 by government messaging, mass media, social media, disinformation, campaigns, etc. 00:04:02.229 --> 00:04:08.849 The theory itself, that's the core of our work ACT-R is a constrained, very principled 00:04:08.849 --> 00:04:11.629 framework for modeling human behavior. 00:04:11.629 --> 00:04:14.430 It's a theory of the structure of the brain and the functioning of the mind. 00:04:14.430 --> 00:04:16.030 It's also a simulation environment. 00:04:16.030 --> 00:04:22.850 And it essentially says, how the modules of the brain that carry out goals and memory 00:04:22.850 --> 00:04:28.120 and perception - how they operate together dynamically over time to produce behavior, 00:04:28.120 --> 00:04:33.170 that allows us to model both the symbolic knowledge that people have as well as their 00:04:33.170 --> 00:04:36.180 statistical adaptivity to the things that are going on around them. 00:04:36.180 --> 00:04:43.450 And it includes about 45 years of research, based in the laboratory as well as real world 00:04:43.450 --> 00:04:49.190 applications, a lot of fMRI and EEG imaging data. 00:04:49.190 --> 00:04:56.440 And so one way to think of it is - we're trying to build these individual level agents that 00:04:56.440 --> 00:05:02.810 can simulate the, what we call the response profiles of people. 00:05:02.810 --> 00:05:07.790 That is, you know, whether they will, in fact wash their hands or wear masks or social distance, 00:05:07.790 --> 00:05:13.020 or are they going to go out and party or go to - go out to restaurants. 00:05:13.020 --> 00:05:19.200 And these agents are going to be seated with representations of individual level attitudes 00:05:19.200 --> 00:05:21.060 and beliefs and intentions. 00:05:21.060 --> 00:05:26.290 And then those agents will be embedded in an agent-based simulation of given regions 00:05:26.290 --> 00:05:27.570 and periods. 00:05:27.570 --> 00:05:32.770 And from that, we want to be able to predict actual behavior that we will compare against 00:05:32.770 --> 00:05:39.090 some proxy measures that we have of behavior, including mobility data from Unicast, and 00:05:39.090 --> 00:05:43.550 mass scoring data that's collected daily from COVIDcast. 00:05:43.550 --> 00:05:49.310 And the way that we are seating these models is using a variety of data that are out there 00:05:49.310 --> 00:05:53.430 already, including these daily polling data sites. 00:05:53.430 --> 00:05:58.580 We're also doing a lot of analysis of mass media and online information. 00:05:58.580 --> 00:06:05.780 And using that to, to get representations that we think characterize individuals in 00:06:05.780 --> 00:06:07.970 these different regions over time. 00:06:07.970 --> 00:06:15.340 So just to give you some examples, we're, we're ingesting a dataset called Third Eye 00:06:15.340 --> 00:06:23.360 Chyron dataset, which is basically a texturized version of CNN, MSNBC, and all the other major 00:06:23.360 --> 00:06:26.260 news networks. 00:06:26.260 --> 00:06:35.260 We've got access to a variety of datasets of Twitter, including geo-tagged COVID dataset 00:06:35.260 --> 00:06:36.530 for the world. 00:06:36.530 --> 00:06:42.940 And our CMU partners have a system called CASOS, which analyzes data from the United 00:06:42.940 --> 00:06:46.080 States in great detail and great volumes. 00:06:46.080 --> 00:06:52.500 And so here are some plots of pro versus con tweet volumes in a variety of cities in California 00:06:52.500 --> 00:06:53.500 that we've collected. 00:06:53.500 --> 00:07:03.720 And we use GPS mobility tracking provided by Unicast, as well as a day by day polling 00:07:03.720 --> 00:07:08.780 data about behaviors from the CMU COVIDcast folks at Delphi. 00:07:08.780 --> 00:07:17.210 So just to give you one thread of analysis that we're doing, a bunch of folks who are 00:07:17.210 --> 00:07:23.690 doing natural language processing, and machine learning over Twitter are inducing what we 00:07:23.690 --> 00:07:28.340 call stances, which are representations of attitudes, beliefs, and intentions from individual 00:07:28.340 --> 00:07:33.960 level tweets, which are then aggregated up to the users who are making those tweets. 00:07:33.960 --> 00:07:40.970 And these are stances or attitudes, beliefs towards certain things like mask wearing, 00:07:40.970 --> 00:07:42.860 or social distancing. 00:07:42.860 --> 00:07:47.580 We do that at large scale, and then use the representations that come out of that natural 00:07:47.580 --> 00:07:54.180 language processing to see the representations inside of this psychological valid agent, 00:07:54.180 --> 00:07:58.370 computational agent, that we're using in our sim[ulations]. 00:07:58.370 --> 00:08:02.200 Just to give you some ideas of the kinds of behaviors or phenomena that we're trying to 00:08:02.200 --> 00:08:10.280 model - using our psychologically valid agents, we are modeling a variety of phenomena, these 00:08:10.280 --> 00:08:13.550 are just a couple on the left here. 00:08:13.550 --> 00:08:20.590 One phenomena that one sees over and over again, across the world is that as the pandemic 00:08:20.590 --> 00:08:30.419 hit, there was a great decrease in the effective transmission rates down to around one and 00:08:30.419 --> 00:08:37.950 then this kind of dampened oscillation around, around a transmission rate of one that seemed 00:08:37.950 --> 00:08:43.370 to indicate that people were adjusting their behavior to modulate that transmission rate. 00:08:43.370 --> 00:08:49.779 And our psychological models can in fact, model mask wearing in relation to what's going 00:08:49.779 --> 00:08:54.459 on in the ambient environment in that kind of dampened oscillating pattern, which is 00:08:54.459 --> 00:08:56.350 in that lower quadrant there. 00:08:56.350 --> 00:09:02.740 On the right here are just showing how, how, at the aggregate level, this is for four states 00:09:02.740 --> 00:09:09.740 of polling data about mask wearing, we can - our models can predict pretty well what 00:09:09.740 --> 00:09:14.490 the actual mask learning probabilities will be in those four states and we can get down 00:09:14.490 --> 00:09:17.199 into finer grained regional areas. 00:09:17.199 --> 00:09:21.601 So if you want to find out more, please contact me [ppirolli@ihmc.org]. 00:09:21.601 --> 00:09:26.180 And I just want to mention that our research is funded by NSF and IARPA, and I want to 00:09:26.180 --> 00:09:29.430 thank the various folks at the bottom here for providing us with data. 00:09:29.430 --> 00:09:29.829 Thank you!