key: cord-0691891-ealibraq authors: Rojas-Galeano, S.; Alvarez, L. title: Simulation of Non-Pharmaceutical Interventions on COVID-19 with an Agent-based Model of Zonal Restraint date: 2020-06-16 journal: nan DOI: 10.1101/2020.06.13.20130542 sha: 1ce4b4f580a5e4dcc098b2f935c60e817f2e2a6f doc_id: 691891 cord_uid: ealibraq Non-Pharmaceutical Interventions (NPI) are currently the only mechanism governments can use to mitigate the impact of the COVID-19 epidemic. Similarly to the actual spread of the disease, the dynamics of the contention patterns emerging from the application of NPIs are complex and depend on interactions between people within a specific region as well as other stochastic factors associated to demographic, geographic, political and economical conditions. Agent-based models simulate microscopic rules of simultaneous spatial interactions between multiple agents within a population, in an attempt to reproduce the complex dynamics of the effect of the contention measures. In this way, it is possible to design individual behaviours along with NPI scenarios, measuring how the simulation dynamics is affected and therefore, yielding rapid insights to perform a broad assessment of the potential of composite interventions at different stages of the epidemic. In this paper we describe a model and a tool to experiment with such kind of analysis applied to a conceptual city, considering a number of widely-applied NPIs such as social distancing, case isolation, home quarantine, total lockdown, sentinel testing, mask wearing and a distinctive "zonal" enforcement measure, requiring these interventions to be applied gradually to separated enclosed districts (zones). We find that the model is able to capture emerging dynamics associated to these NPIs; besides, the zonal contention strategy yields an improvement on the mitigation impact across all scenarios of combination with individual NPIs. The model and tool are open to extensions to account for omitted or newer factors affecting the planning and design of NPIs intended to counter the late stages or forthcoming waves of the COVID-19 crisis. In view of the absence of approved drugs or vaccines for COVID-19 (up to this day), Interventions (NPI) are the mechanisms that public health offices around the world are using in an 31 attempt to mitigate the impact of its epidemic (Lai et al., 2020; Ferguson et al., 2020) . Similarly 32 to the actual spread of the disease, the dynamics of the contention patterns obtained as a result 33 of application of NPIs are complex and may depend not only on interactions (contacts) between Figure 1 . Notice that not all of these conditions are mutually-exclusive, as they represent 141 incremental stages of the I state, evolving in a manner that is explained below. 142 Accordingly, we designed three types of agents: a healthy agent associated to the S and R 143 states (the latter marked with an immunity condition), a sick agent associated to the I extended 144 state tagged with the CARDS conditions, and a dead agent associated to the E state. In contrast to 145 population-dynamic simulation, instead of using transition rates between states, our model uses 150 Therefore, we consider the following transition events: is initially characterised as not confirmed, not tested, not severe, not deadly, not hospitalised 154 and not ICU-admitted. Besides, a recovery (or illness) period in days is assigned as a random 155 variate with a Gaussian distribution centered at the avg-duration model parameter. • I + Confirmed. This condition characterises an agent as a positive case. This may happen in any 157 of the following moments: when the patient is tested for the virus and the test results positive; 158 when the agent has not been previously diagnosed as positive at the moment of admission to 159 hospital, or admission to ICU bed or death; or when the patient feels sufficiently symptomatic 160 so as to self-isolate. On the other hand, if the Case Isolation NPI (see above) is lifted after 161 having been enforced during the course of the simulation, then the confirmed condition of 162 agents who have not been tested positive, is reversed (i.e those who self-isolated). The distinction between confirmed and non-confirmed cases is used to identify individuals 164 that need to stay isolated so as to prevent his capability to spread the virus. Besides, this 165 distinction is also useful to examine the discrepancy between "official" confirmed fatality rates CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 16, 2020 June 16, . . https://doi.org/10.1101 June 16, /2020 doi: medRxiv preprint a model parameter (%-asymptomatic). In the simulation tool, this condition is indicated with the colour of the agent (yellow for asymptomatic, red for symptomatic). • I + Risky. This condition characterises an agent as being in a high-risk population group. Cur-173 rently the model does not consider neither risk stratification of co-morbidities nor age struc-174 ture of the population. Therefore all these risk factors (obesity, diabetes, cardiovascular dis-175 ease or medium to old age) are encompassed in this single condition which is activated upon 176 acquiring the infection, with a probability defined as a model parameter (%-high risk). • I + Severity. This condition indicates progression of the disease to a severe state, which in turn 178 represents a higher threat of death, as it is defined in the I E transition. This event occurs 179 according to the following parameters (see Tuomisto et al. (2020) to hospital beds, on a daily basis but considering patients with a deadly condition instead. Here, ICU beds are released when agents recover or die. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. Since dead may occur in any day during the illness period, we modelled the chance of the day 208 of fatality as a triangular distribution with peak in the middle of such individual's particular 209 recovery period (half-recover-period). In other words, we assume: where CDF stands for cumulative distributed function. As a result, we compute the actual 211 probability of death in a given day as follows: 212 P(Extinct | day) = P(day) P(Extinct | Conditions). • I R (recovery). An infected individual gets cured when he reaches the end of his recovery • lifestyle: simulates the daily routine of agents, which currently encompasses moving the agent ahead towards its current destination, or head it backwards home if he has reached 246 his range limit. In any case, confirmed patients are not allowed to move around. If zonal enforcement is on, the movement is restricted only within the periphery of the zone 248 where the agent is resident. In the simulation view, residents of different zones are identified 249 with different shapes and colours (the latter coincides with the colour of their zone's ground). • epidemic: here the spread of the epidemic is simulated, with transmissions occurring in 251 proportion to proximity between infectious and susceptibles and probability of contagion. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10.1101/2020.06.13.20130542 doi: medRxiv preprint NetLogo Web" tab. NetLogo is a widely popular software platform for ABMs, which further to the 325 simulation language, also integrates a graphical view area and a test-bed for experimental design 326 (Wilensky and Rand, 2015) . 327 The developed tool implements the NPIs, epidemic SIRE+CARDS model and agent behaviour 328 rules described previously. A snapshot of the simulation view area is shown in Figure 3 . In there, 329 8 of 18 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. . Figure 3 . View area of the simulation tool. Agents are identified by a different shape and colour per zone; additionally, the epidemic state of agents is indicated by red (sysmptomatic), yellow (asymptomatic) or white (recovered, immune); deaths are shown by black cross marks (×). Besides, households and sentinels (ambulances) can also be seen. agents are represented with different shapes according to the zone where they reside. The colour 330 of the agent also represents its extended state (healthy: same colour as zone ground; immune: 331 white; sick: red, or yellow if asymptomatic; dead: black ×). Special-purpose agents designed to im-332 plement some of the NPIs such as households for home-quarantine and ambulances for sentinel-333 testing can also be seen. The control panel is organised in sections related to general, city and COVID-19 settings, mon-335 itors of epidemic indicators, parameters, action commands to execute the simulation, and a ded- In this section we describe a number of scenarios conceived to illustrate how to perform a rapid 342 assessment of the effect of NPIs, or combination of NPIs, to contain the spread of the epidemics 343 in the population of agents. We start with a baseline scenario where no measure is taken (Do 344 Nothing). Then we simulate scenarios where the individual NPIs are applied. Finally, we simulate 345 scenarios where these NPIs are combined with the zonal enforcement strategy in order to verify 346 its potential impact. The description of these scenarios is given in Table 1 . General, city and COVID-347 19 parameters used in all simulations are defined in Figure 2 . For each scenario, the epidemic is 348 assumed to begin with a "patient zero" seeded randomly in any zone of the city at 0d:12h after the 349 start of the simulation. The application of the configured NPI policies begins at 04d:00h. Table 3 and Table 4 . . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10.1101/2020.06.13.20130542 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. 28.4 ± 20.4 0.07 ± 0.05 Table 3 . Survivor, death and immune count in each scenario starting with a single infection. Statistical significant differences in counts of survivors and deaths between the single NPI scenarios compared to their corresponding combined NPI+ZE scenarios, are marked as * (p<0.01) and ** (p<0.001). Table 6 . Epidemic indicators obtained in each scenario starting with an outbreak of 5% population. The latter increase is caused of course, by the larger amount of carriers at the beginning of the 424 simulation (20 individuals or 5%) yielding a higher burden on the hospital facilities at the peak of The main finding however, in our opinion, is that simultaneous application of the Zone Enforc- . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 16, 2020. In addition, the proposed zonal enforcing NPI proved to be useful to boost the mitigation im-459 pact of combination of other individual NPIs. Although this seems an appealing finding, practical 460 application of said zonal enforcing may require logistic, residential and economical adjustments, 461 since people usually reside and work in different districts of a city. Therefore the feasibility of 462 limiting mobility of people within districts will depend on the urban planning and development 463 of sufficient decentralised infrastructure, such as industrial, residential, technological, commercial 464 and financial hubs distributed in a way that motivate people to reside and work within the same 465 district. This may not be feasible to achieve for the current COVID-19 pandemic, particularly in 466 cities in Latin America where proper urban zonal planning is unsatisfactory, but it is an alluring 467 idea to start exploring as a preventive measure to counter new epidemics that may come in the 468 forthcoming future. Finally, interesting gaps remain to be addressed in future work. Some of the ideas we are con-470 sidering are: expanding the model to include more realistic attributes regarding patients and epi-471 demic dynamics, such as differentiated infectiousness taking into account symptomatic structures 472 associated with age and gender; also considering incubation periods, co-morbidity risk stratifica-473 tion with age windows, inclusion of conglomeration centres (that is, mass transportation, schools, 474 cinemas, hospitals), as well as the estimation of indicators of economic impacts and the study of 475 the importance of educational aspects in the habits of collective social intelligence that may be 476 beneficial for the mitigation power of NPIs. Here we report plots of example executions of simulated scenarios with an outbreak infecting 5% 522 of the population. For each scenario in Table 1 , we include the individual NPIs outcome (left-hand 523 side) and the corresponding NPIs plus zonal enforcement outcome (right-hand side). . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10.1101/2020.06.13.20130542 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10.1101/2020.06.13.20130542 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10.1101/2020.06.13.20130542 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 16, 2020. . https://doi.org/10. 1101 /2020 Theory versus data: how to calculate R0? PLoS One