Human activity pattern implications for modeling SARS-CoV-2 transmission.

BACKGROUND AND OBJECTIVES: SARS-CoV-2 emerged in December 2019 and rapidly spread into a global pandemic. Designing optimal community responses (social distancing, vaccination) is dependent on the stage of the disease progression, discovery of asymptomatic individuals, changes in virulence of the pathogen, and current levels of herd immunity. Community strategies may have severe and undesirable social and economic side effects. Modeling is the only available scientific approach to develop effective strategies that can minimize these unwanted side effects while retaining the effectiveness of the interventions.
METHODS: We extended the agent-based model, SpatioTemporal Human Activity Model (STHAM), for simulating SARS-CoV-2 transmission dynamics.
RESULTS: Here we present preliminary STHAM simulation results that reproduce the overall trends observed in the Wasatch Front (Utah, United States of America) for the general population. The results presented here clearly indicate that human activity patterns are important in predicting the rate of infection for different demographic groups in the population.
CONCLUSIONS: Future work in pandemic simulations should use empirical human activity data for agent-based techniques.

Introduction

SARS-CoV-2 emerged in December 2019 and rapidly spread into a global pandemic [1], [2], [3]. Often considered strategies for reducing local SARS-CoV-2 spread are social distancing [4], [5], [6] and more aggressive lockdowns [7]. The impacts of social distancing and other public health measures including vaccination [8,9] and prophylactic medications (when these become available), for designing optimal community responses will be dependent on the stage of disease progression within the community [10], discovery of asymptomatic individuals [11,12], changes in virulence of the pathogen [13] and current levels of herd immunity. In some countries, there have been efforts to let the disease spread to develop a level of herd immunity, as well as calls for younger and low-risk individuals to join back into the workforce. The efficacy as well as the undesired effects of these approaches are largely not known [13].

As observed in previous pandemics, social determinants of health (SDOH) are likely to influence the ability of individuals and families to follow SARS-CoV-2 public health guidelines and may alter health outcomes [14,15]. Impacts of pandemics are amplified in lower socioeconomic, densely populated urban areas compared to distributed rural areas [14,16]. SDOH like education levels, household size, age-structure, population density, workplace – school – household – neighborhood structures, and area-level poverty, along with behaviors such as cigarette smoking that impact health, can directly and indirectly affect both immediate, and long term outcomes in pandemics [17,18]. The effects of selective movement of individuals across time and demographic groups in the evolution of disease are difficult to ascertain.

Modeling is the only viable scientific approach to develop effective strategies that can minimize these undesirable side effects and show the effectiveness of the interventions [19,20]. With the principal cause of SARS-CoV-2 spread being movement and interaction of individuals within a community [21], modeling should be informed by human activity patterns as different humans will come in contact with infected individuals depending upon their daily activities. Numerous simulation models have been reported, and most of them have been implemented in EpiModel [22]. Many papers have presented simulations of SARS-CoV-2 spread using a variety of mathematical and computational models; examples of these work can be found in Refs. [23], [24], [25], [26], [27], [28], [29], [30], [31], [32]. Of relevance to the work reported here, several agent based models have been used to predict SARS-CoV-2 spread [33,34]. Still, to the authors' knowledge, none of these models have incorporated empirical information about human activity patterns in their simulations.

We have shown that human activity patterns, as recorded by the American Time Use Survey (ATUS) [35], can be classified and used to simulate human activity in the context of estimating personal exposure to PM 2.5 [36,37]. Based on this classification of human activities, we have developed the SpatioTemporal Human Activity Model, or STHAM [37], to generate and characterize travel trajectories and activity patterns [38], [39], [40], and to integrate exposure-agent geographical distributions for estimating total human exposure in large populations. Here, we demonstrated that an infection model based on the STHAM [37] framework to simulate human movement in the Wasatch Front of the State of Utah, United States of America [US) [41] can reproduce the overall observed trends of SARS-CoV-2 transmission within this region's population. But more importantly we show that the rates of infection are highly dependent on the activity patterns of the agents considered.

Methods

The STHAM simulation method is described in detail in Ref. [37]. Here for completeness, we briefly describe the key steps to produce a STHAM simulation. The STHAM initialization process is detailed in Fig. 1 . At the core of the STHAM is the agent, which is represented as a single data record with demographic properties representing a hypothetical person. These properties (e.g., age, gender, census block location) are assigned, as constraints, to each agent randomly using the aggregate statistics from the 2010 United States Census at the census block level [42], [43], [44] so that the generated agents and the households to which they are assigned match the data provided in the census. Each household is then assigned a location in a round-robin fashion based on a list of physical postal addresses [45] that are matched to each census block. Next, each agent is assigned an employment status and, if relevant, an employment location, based on data from the Longitudinal Employee Household Dynamics Program (under the US Census Bureau) (LEHD) [46]. The LEHD also provides an approximation for employment location at the census block level, which allows the model to simulate commuter patterns accurately. Finally, school participation is assigned using age, and gender stratified enrollment rates from the American Community Survey (ACS) [43].

Fig. 1

STHAM agent creation process. Demographic properties from the Census Block tables are assigned to each agent, and then each agent is assigned to a household, which is given a home location. Additional locations for employment, school enrollment, and other regularly attended sites are then assigned. From Ref. [37].

Conceptually, for infectious disease modeling, individuals in STHAM are represented as agents belonging to households and work organizations, and residing in particular environments. Agents have properties representing individuals' characteristics of demographics, occupation, socioeconomic, and immune status, which are based on empirical observations. The environment has its own characteristics that could include weather, UV radiation, and air quality. Rules describe what happens when different classes of agents interact in their environment (e.g., the transmission of virus based on established rates), and the type of interactions possible under different public health scenarios.

The principal cause of SARS-CoV-2 spread is the movement and interaction of individuals within a community, and the colocation of agents is modeled as the primary cause of disease transmission. We executed the STHAM model for the Wasatch front using 2,364,637 agents. These agents were divided according to their demographics as Workers: 1,047,822, Students: 774,604, and OtherAgents: 717,141 based on the United States Census [44]. The categories of agents were defined by the following these rules: Workers are those agents that have a job (1 hour- 60 hours per week). Agents are considered as Students if the activities that they perform are consistent with attending any educational institution from preschool to college. Any other agents, i.e., those that do not have a job and are not Students are classified as OtherAgent. Since there is a significant intersection between workers and students, agents are classified as Worker and Student at the same time, and the sum of these three classes is larger than the total population. Based on the STHAM model, we calculated the daily activity patterns for all agents, including activity categories and activity locations. There are two kinds of daily life activities patterns in our model: one for weekdays and another for weekends. To simulate one week, we run the model with five weekdays and two weekends. We divided the 24 hours in each day into 288 five-minute segments and aggregated agents within a location at a five-minute interval. For any colocation, if there are one or more infected agents, that location is considered as an infected location, and any susceptible agents (never infected before) in these locations have a non-zero probability of being infected. This follows the current understanding that there have been no cases of SARS-CoV-2 reinfections, but could be modified as knowledge of SARS-CoV-2 immunity evolves [47]. A susceptible agent in an infected location has its activity contact index randomly sampled by a uniform distribution from 0 - 1. If the agent's contact index is lower than the transmission rate, this agent's status changes to infected. The simulation used a transmission rate defined by a Gaussian distribution and having a mean of 2.5/(288*14*30), and a standard error of 0.00001. We used a basic reproduction number (R0) of 2.5 over the span of 14 days in the model, which is based on current best estimates for R0 and the average length of disease as per the Centers for Disease Control and Prevention [48]. The STHAM model provides an average daily inter-agent contact rate of 30 [36]. Therefore, the mean value of the transmission rate is set to 2.5/(288*14*30). The infection process repeats every five-minutes, and after one-day simulation, the list of infected agents is updated by adding the new infected agents and excluding the recovered agents. Infected individuals remain infectious for 14 days. After this period, they are neither infective nor susceptible to reinfection. We arbitrary assigned five Workers, three Students, and two OtherAgents to be infected at the beginning of the simulations, and ignored all inbound and outbound travel from the region. We simulated the epidemic for 60 days.

The EpiModel [22] package utilizes a stochastic individual contact model (ICM), which simulates agent actions at an individual level for each time step. The ICM has a three-compartment model, where individuals in a population are assigned to compartments with labels – S, I, and R (Susceptible, Infected, and Recovered). The progression between compartments is dictated by user-specified percentages that are used to define random draws from a binomial distribution. Progression between compartments depends on actions taken by infected agents. In EpiModel, this is defined as the act_rate, or the number of both effective and ineffective contacts made by an infected agent. For the scenario reported here, the act_rate was set to 0.25 acts/day, which is consistent with the infecting probability used in the STHAM simulations. After an individual spends 14 days in the infected compartment, they are moved to the recovered compartment and treated as non-acting agents, which are neither infective nor susceptible to reinfection. A simulation of 10 weeks was run ten times, using randomly selected seeds, and the results for each simulation were aggregated and averaged to obtain the final results reported here. For starting parameters, there were 2,000,000 susceptible agents, ten infected agents, and 0 recovered agents assigned at the start of the simulation.

The number of infected individuals in the Wasatch Front, which is the most densely populated region of the State of Utah, United States of America, with a population of approximately 2,000,000 individuals, was obtained from the Utah Department of Health (UDOH). The first day in which the cumulative number of infected cases reported to UDOH was over ten individuals was March 15th 2020, with 18 cases. The raw values obtained from UDOH were scaled by a factor 1.8 to be consistent with the simulated results obtained with an initial population of 10 infected individuals at day zero, which was set to March 15th 2020.

Results

Fig. 2 shows the comparison of the evolution of cumulative number of cases in the population as a function of time between those reported by UDOH and those predicted by the STHAM model. The evolution of the pandemic predicted by the STHAM simulations for the total population is in good agreement with the observations for the first few weeks, but shows significant differences after 40 days. This is expected as the Wasatch Front established significant lockdown measures to control the epidemic, while the population of the STHAM model continued following their normal activity patterns. Another potential deviations between the simulations and the reported data is the number of unreported cases, which in a recent paper has been estimated in range between 52.1 % and 100 % [49].

Fig. 2

Comparison of the total cumulative number of STHAM predicted cases compared with the cumulative number of cases reported by the UDOH for the Utah Wasatch Front. Day Zero corresponds to March 15th, 2020, which is the day that the total count of infected individuals in the Wasatch Front exceeded ten individuals.

The STHAM predictions also show good agreement in the initial phase of the epidemic with the predictions of the EpiModel [22] (Fig. 3 ). The latest trend in cases continues to grow exponentially after the first 40 days. This is also expected because the STHAM model predicts a different rate of spread based on the agents' demographic, and a significant number of STHAM agents show substantially lower levels of infection (see below).

Fig. 3

Comparison of the total cumulative number of STHAM predicted cases with the cumulative number of cases predicted by the model simulations using EpiModel [22].

From Fig. 4 , it is apparent that since the start of the epidemic, the number of Workers infected grew much faster than any of the other categories, reaching an infection percentage that is approximately 30% higher than the population average. Conversely the agents classified as OtherAgents show a much lower infection rate leading to a reduced infection percentage approximately 30% lower than the population average. The agents classified as Students show infection percentages comparable to the OtherAgents in the early stages of the epidemic, but eventually, reach the same level of infection than the Workers. These results are consistent with the intuitive notion that those agents that have higher mobility, Workers and Students, are more likely to come in contact with other infected agents and therefore increase the total number of agents infected in their group. The results presented in Fig. 5 , further emphasize the importance of using empirical activity patters to model epidemic spread; it is noteworthy that the different demographic groups considered here show different spatial distributions since the early stages of epidemic.

Fig. 4

Temporal increase of predicted cases as a percentage of the population and the relative percentages for the different types of agents considered here. The agent simulation was performed using the STHAM modeling approach [36,37]. The left panel corresponds to the initial 21 days of the simulation, while the right panel extends up to 56 days (8 weeks).

Fig. 5

Spatial distribution of the increase of the cumulative number of predicted cases as a percentage of the population and the relative percentages for the different types of agents considered here. The agent simulation was performed using the STHAM modeling approach [36,37].

Discussion

While this study demonstrates the need of using empirical activity patterns to obtain higher accuracy in modeling epidemics, the study also has limitations that need to be recognized. The activities and mobility used here are those expected from people in the United States of America under normal times. These may be significantly different activity patterns than those expected during a pandemic, and better models based on data of human activity collected during an epidemic (e.g., COVID-19 Community Mobility Reports [50]) are necessary for developing more accurate predicted models. The demographic categories used here are quite broad, and future work should be the focus in developing more granular models for the demographics, these models should be grounded in empirical data, which can be transformed into probabilistic agent behaviors using the STHAM approach [36,37]. For simplicity, here we used a SIR compartment model (Susceptible-Infected-Recovered), but we acknowledge that using a SEIR (Susceptible-Exposed-Infectious-Recovered) model could provide more accurate results [51], [52], [53]. Our modeling software is able to accommodate any infection model including different rates of symptomatic/asymptomatic rates and scenarios that are potentially dependent on the age of the agent, load of exposure, immunization status, use of precautionary measures (e.g. agents wearing masks) among other factors.

As next steps, we will also evaluate the effects of the environment [54,55] and selective movement of individuals across time and demographic groups in the evolution of the disease for the State of Utah with different mitigation strategies and demonstrate that our model, taking into account the empirical evidence of human mobility, produces more detailed information than models homogeneously treating individual agents.

Conclusions

The principal cause of the spread of SARS-COV-2 is the movement and interaction of individuals within a community. The STHAM model [36,37] assigns different mobility patterns to an individual belonging to different demographic groups as defined by empirical activity sequences derived from the ATUS [35]. Here, we show that the STHAM modeling approach can reproduce the overall trends of the spread of SARS-COV-2 in the Utah Wasatch front, but also show that different demographic groups have significantly different infection rates. Future work should focus on using empirical data on human activity patterns for pandemic simulations using agent-based techniques and including more categories of agents to get higher specificity on the lead activates that contribute to the spread of the virus. Developing this kind of models will allow policymakers to make contention strategies than may be much more granular and leading to less undesirable effects.

Author's Contributions

JCF and RG conceived the research; YW and BL performed the calculations and all authors analyzed the data, developed the manuscript and approved the submission. All authors agree to be accountable for the contents of the paper.

Funding

This research was partially funded by a Special Emphasis: Emerging COV/0-19/SARS-CoV-2 Research grant funded by the Immunology, Inflammation, and Infectious Disease Initiative in partnership with the Office of the Vice President For Research, University of Utah; an Undergraduate Research Opportunities Program at the University of Utah Summer Research Experience fellowship awarded to BL; the Center of Excellence for Exposure Health Informatics supported by the Utah PRISMS Informatics Center (NIH/NIBIB U54EB021973), and the Utah Center for Clinical and Translational Science (NCATS UL1TR001067). Computational resources were provided by the Utah Center for High Performance Computing, which has been partially funded by the NIH Shared Instrumentation Grant 1S10OD021644-01A1. Content is the sole responsibility of the authors and does not necessarily represent official views of the NIH.

Declaration of Competing Interest

None