key: cord-0803036-88t2kszg
authors: ZHOU, H.; Zhang, Q.; Cao, Z.; Huang, H.; Zeng, D.
title: Sustainable targeted interventions to mitigate the COVID-19 pandemic: A big data-driven modeling study in Hong Kong
date: 2021-02-02
journal: nan
DOI: 10.1101/2021.01.29.21250786
sha: eb7c0864d21aa158f3d3614f84baf77f1500fd96
doc_id: 803036
cord_uid: 88t2kszg

Background: The nonpharmaceutical interventions (NPIs) for contact suppression have been widely used worldwide, which impose harmful burdens on the population and the local economy. The evaluation of alternative NPIs is needed to confront the pandemic with less disruption. By harnessing human mobility data, we develop an agent-based model that can evaluate the efficacies of NPIs with individualized mobility simulations. Based on the model, we propose the data-driven targeted interventions to mitigate the COVID-19 pandemic in Hong Kong without city-wide NPIs. Methods: We develop a data-driven agent-based model for 7.55 million Hong Kong residents to evaluate the efficacies of various NPIs in the first 80 days of the initial outbreak. The entire territory of Hong Kong has split into 4,905 500mx500m grids. The model can simulate detailed agent interactions based on the demographics data, public facilities and functional buildings, transportation systems, and travel patterns. The general daily human mobility patterns are adopted from Google Community Mobility Report. The scenario without any NPIs is set as the baseline. By simulating the epidemic progression and human movement at the individual level, we proposed model-driven targeted interventions, which focus on the surgical testing and quarantine of only a small portion of regions instead of enforcing NPIs in the whole city. The efficacious of common NPIs and the proposed targeted interventions are evaluated by extensive Monte Carlo simulations. Findings: Without NPIs, we estimate that there are 128,711 total infections (IQR 23,511-70,310) by the end of the 80-day simulation. The proposed targeted intervention averts 95.85% and 94.13% of baseline infections with only 100 (2.04%) and 50 (1.02%) grids being quarantined, respectively. Mild social distancing without testing results in 16,503 total cases (87.18% infections averted), rapid implementation of full lockdown and testing measures (such as the control measure in Mainland China) performs the best, with only 805 infections (99.37% infections averted). Testing-and-quarantining 10%, 20%, 50% of all symptomatic cases with 24-hour/48-hour avert 89.92%/ 87.78%, 95.47%/ 92.42%, and 97.93%/ 95.61% infections, respectively. Interpretation: Big data-driven mobility modeling can inform targeted interventions, which are able to effectively contain the COVID-19 outbreak with much lower disruption of the city. It represents a promising approach to sustainable NPIs to help us revive the economy of the city and the world.

The global coronavirus disease 2019 (COVID- 19) pandemic has infected 62,844,837 patients, and claimed 1,465,144 lives as of 7:08pm CET, 1 December 2020 1 . Various nonpharmaceutical interventions (NPIs) were implemented in response to this public health crisis worldwide, ranging from case isolation and quarantine of contacts to the complete lockdown of the entire country [1] . These NPIs were demonstrated to be able to mitigate community transmission at the expense of economic disruption, harm to social and mental well-being, and costly administration required to ensure compliance [2] . However, a resurgence of COVID-19 has been seen in many parts of the world when the NPIs were lifted [3] . Population-wide vaccination program is not likely to cover a large-scale population in the near future considering the huge demand and difficulty in mass production and supply chains [4, 5] . It is therefore important to find the NPIs to contain outbreaks while minimizing the economic costs [6] .

To provide a reliable quantitative analysis of the efficacious of alternative NPIs, we explore a data-driven targeted intervention policy to precisely identify the subcommunities that need to be surgically tested and quarantine. In this study, we develop a data-driven mobility model (D2M2) to characterize the demographics, public facilities and functional buildings, transportation system, and travel patterns of the 7.55 million residents in Hong Kong, an international trade center and one of the most densely populated cities in the world. Given the mass amount and rich diversity of the data publicized by the authority and public sectors, Hong Kong represents an ideal opportunity to investigate the feasibility of using a detailed agent-based severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) transmission model to inform targeted interventions. Targeted interventions informed by the model can prioritize a set of grids with high risk for quarantine, so that the localized outbreak can be contained while preserving the normal state in the majority of the city. We examine the efficacy of the proposed targeted interventions in containing the outbreak of COVID-19 in Hong Kong.

Hong Kong was one of the first places outside Mainland China to report the COVID-19 confirmed cases owing to their strong connection [7] . From January to early March 2020, the COVID-19 infected individuals in Hong Kong mainly consisted of imported cases and secondary transmissions, which is considered the first wave of COVID-19 outbreak in Hong Kong. The number of infections in the first wave remained relatively low due to a series of rapid responses, including border control, manual contact tracing, voluntary community-wide use of personal protective equipment (PPE), and social distancing [8, 9] . The second wave, mainly consisted of the imported cases outside Asia and associated localized transmission, occurred from mid-March to May 2020. After implementing more stringent NPIs, only a few local 1 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint cases were reported in May and early June. On June 18, most social distancing restrictions were lifted. However, the third wave occurred in late June 2020 that pushed Hong Kong's coronavirus case tally to 5,009 by September 19, 2020 2 . The locally acquired cases appeared on 5 July, which is phylogenetically related to SARS-CoV-2 reported overseas [10] . There could be fourth and more waves to come, as humankind will likely to live with it until population-wide vaccination program takes place [11] . This study is mainly to simulate the third wave outbreak in Hong Kong from June 25, 2020 to September 24, 2020.

The contribution of this study is threefold. First, the proposed D2M2 is calibrated by the multi-level and multi-source data of the demographics, transportation, and human mobility. To the best of our knowledge, it is the most detailed characterization of the human mobility of Hong Kong. Second. Mobile phone data is difficult to obtain in most countries, thus D2M2, which is calibrated by open-source data, D2M2 can be easily extended to the modeling of other metropolises with various demographic and human mobility patterns. Given available mobile phone and contact tracing data, D2M2 can be further refined. Third, based on D2M2, we propose the targeted intervention, which can contain the outbreak with minimal disruption of the society. This is of particular importance to cities like Hong Kong, whose economy relies on international trade.

To quantitatively characterize the mobility patterns of the residents in Hong Kong, the proposed D2M2 is based on a recent human behavior model for daily movement [12] , and is calibrated with the real-world socio-demographic data of Hong Kong. A human-to-human contact and transmission network can be created by combining the human behavior pattern and a synthetic population (as shown in Fig 1) . Based on the demographic data from Hong Kong Population Projections 3 , D2M2 generates 7,550,000 synthetic agents at both household and individual levels and allocates them into 4,905 living grids. The generated households contained rich demographic characteristics, including house location (living grids), family size, age, gender, work/study locations, and nationality of household members obtained from the latest Hong Kong Population Census 4 . We also collected a comprehensive set of Points of Interest (POIs), including schools, restaurants, leisure and cultural services venues, shopping malls, pharmacies, groceries, various public entertainment places, etc. Agents with proper ages are also assigned to workplaces or educational facilities according to 2016 census data. In addition, the Foreign Domestic Helpers (FDHs) represent an integral part of Hong Kong society nowadays, and 5.29% of the 2 https://www.coronavirus.gov.hk/sim/index.html accessed by Dec 2, 2020. 3 https://data.gov.hk/en-data/dataset/hk-pland-pland1-projections-of-population-distribution-2019-to-2028 accessed by Dec 27, 2020. 4 https://www.bycensus2016.gov.hk/en/bc-dp-tpu.html accessed by Dec 27, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint population [13] . We set 39,9320 FDHs among the home working agents according to the statistics of Immigration Department in 2019 5 . We set FDHs gather on Sundays in public areas [14] , and thus represent the potential risk of localized outbreaks. The information of workplaces and educational facilities is obtained from the previous study and opened data from the public sector. The infection probability of an agent is related to the time that he or she spent where infected agents have visited [15] . Please refer to Supplementary Information for modeling details.

A recently introduced susceptible-latent-infected-removed (SLIR) model with some additional compartments is adapted to incorporate the unique characteristics of COVID-19 ( Fig. 1) [6, 15, 16] . In particular, at each time step (one day), the infectious asymptomatic ( ), infectious symptomatic ( ) and pre-symptomatic ( ) individuals can transmit the virus to susceptible (S) individuals with probability , and , respectively. If the transmission is successful, the susceptible agent will move to the latent asymptomatic state with probability or to the latent symptomatic state ( ) with probability (1 − ). A latent asymptomatic individual becomes infectious asymptomatic after a period ( ′) −1 , and some latent symptomatic individuals will transit to pre-symptomatic ( ) state after the same period. Then, during ( ) −1 period after being infected, the symptomatic individuals become onset cases and can either recover or removed after a period −1 .

The values for the basic reproduction number ( 0 ) were set as 2.5 according to the literature [17, 18] . Follow the previous study, we ran the models for 80 days to investigate the early stages of localized outbreak and randomly selected 50 agents as the initial infected on day 0 [15] . Refer to the Supplementary Information for more details.

To confront the pandemic with less disruption, we propose the targeted interventions, which identify a small subset of areas for quarantine while maintaining mobility in other areas. Specifically, we run 200 simulations of baseline scenario (i.e., no NPIs) with 50 randomly seeded agent. Then, following Koo's study [15] , based on the outcome of the baseline scenario simulations at day 80, the grids are ranked by the number of expected infections in the baseline simulation. Top 50 and top 100 grids are chosen as the High-Risk Grids (HRGs), while the others are non-HRGs. Compulsory quarantine is taking place at HRGs. Within HRGs, all essential materials are uniformly delivered by the authorities. Residents in non-HRGs are not affected except that they could not freely commute to HRGs.

As a comparison, we examine the efficacy of the state-at-home control, which involves the following scenario starting from day 15 (the actual delay in the response 5 https://data.gov.hk/en-data/dataset/hk-immd-set4-statistics-fdh accessed by 31/08/2020 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint in Hong Kong 6 ): (a) School closures. We transform all school students to the home working agents simultaneously. (b) Social distancing. We assume that all places are still open, but the volume of passengers is proportional to the Google Mobility Report on the same day.

Moreover, we examine the efficacy of the reactive control. Testing is a useful tool to detect, predict, and reduce the spread of disease [19, 20] . In this scenario, in addition to the mild social-distancing, patients with symptoms are quickly tested and quarantined since the authorities responds on day 15 [21] . We set that testing results will come out either 24 or 48 after being tested [6] . In the experiment, we evaluate the outcomes with different proportions of symptomatic patients who get tested (10%, 20%, 50%). The maximum testing capacity is set as 10,000 cases per day according to the governments' publication 7 . Finally, we investigate the fully lockdown control on the top of testing 50% of symptomatic cases, which represents the scenario that school closer the 70% reduction in human mobility is taken place on day 1, which follows the setting of strict lockdown in the study of US National Bureau of Economic Research [22] . Note that we run 100 simulations for non-baseline NPIs to account for the stochasticity and to calculate the confidence intervals (CIs). 6 https://www.bloomberg.com/news/articles/2020-07-10/hong-kong-said-to-shut-schools-again-aslocal-virus-cases-jump accessed on Dec 15, 2020 7 https://www.chp.gov.hk/files/pdf/statistics_on_covid_19_testing.pdf accessed on Sep 03, 2020 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

For the baseline scenario, a cumulative median of around 129,000 cases (IQR: 94,000-193,000) of the Hong Kong population were infected on day 80 ( Figure 2 , Table 1 ). At this level of transmission ( 0 = 2.5), mild social distancing resulted in a median of around 17,000 cases (IQR: 12,000-25,000), the full lockdown (city-wide quarantine) measure led to 805 cases (IQR: 636-10,60). Regrading reactive control measure with a 24-hour-delay, testing 10% of onset cases led to in a median of around 13,000 cases (IQR: 9,000 -21,000); testing 20% of onset cases led to 5,830 cases (IQR: 4,315-9,204); Testing 50% led to 2,667 cases (IQR: 2,187-3,618). With a 48-hour-delay, the infected cases number resulted by testing 10%, 20%, and 50% symptomatic cases would be 16,000 (IQR: 9,000-31,000), 9,761 (IQR: 7,584-17,675) and 5,649 (IQR: 3,691-9,433) respectively. Compared with the baseline scenario, the full lockdown led to the greatest reduction in cases (99.32, IQR: 95.83%-99.37%).

The location of the infection (school, workplace, domestic party, non-commuter place or home community) was set as the site where individuals acquired the virus from others. As shown in Table 1 , in most scenarios, home was the place with the highest number of infections. Workplace contributed the second highest number of infectious, indicating that reducing the mobility in workplace could reduce the risk of infections since the commuters interacted with agents in the same workplaces or functional buildings [15] . School closure could reduce the infectious among students, but contributed little to protecting other residents. This has also been confirmed by the previous study [23] . Moreover, the number of infections at FDHs' gathering of other POIs was relatively consistent and count small proportions in different NPIs. In particular, in the fully lockdown scenario the infections mainly occurred at the POIs that are neither school or workplace, including grocery, pharmacy, or retail store. This indicates the we may further reduce the number of infections by reducing the activities in POIs at the cost of wider social disruption.

The epidemic peak has not been reached by day 80 in all scenarios, because here we did not consider the contact tracing for model simplicity. As shown in Fig 2, the median number of newly infected cases in the baseline and mild social distancing scenarios was consistently increasing, whereas the number of new cases was relatively stable under the instantaneous response. From the spatial perspective, a considerable number of new infections were acquired from the urban areas in Kowloon and New Territories 8 (Fig  3) . The instantaneous response can effectively reduce the spillover effect of the epidemic outbreak from urban to the suburban area. The infection distributions in the baseline and mild social distancing scenarios were more widely distributed in less population-dense areas. Given the dense residential clusters mainly sited on the Kowloon, Hong Kong island, and the new towns of New Territories, the Large-scale and homogeneously distributed infections implied that the virus cannot be contained without massive testing or total lockdown strategy. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint

The targeted interventions were derived as follows. First, we ran simulations of the baseline scenario (Fig 4) . Second, the top 100 and 50 grids with the highest number of infections were chosen as the HRGs, which will be quarantined, whereas the people who live in other non-HRG grids were not affected. Therefore, the targeted interventions are proactive NPIs, which are based on the prediction of the epidemic patterns. Simulation results showed that the targeted interventions could effectively reduce the size of outbreak in the city, with 5,342 (IQR 4,578-6,607) and 7,550 (6,660-8,781) accumulative infected individuals after 80 days by only quarantining 100 (2.04% of the area and 24.46% of the population) or 50 (1.02% of the area and 13.05% of the population) grids in the city. This indicated that the COVID-19 outbreak was largely contained without much disruption of the society. The efficacy of quarantining 100 HRGs slightly outperformed the reactive control strategy with 50% symptomatic cases tested within 48 hours. Considering the high cost of contact tracing and mass testing, the targeted interventions stand a good potential to become reliable and cost-efficient NPIs to confront the long-term COVID-19 pandemic and even endemic.

Geospatially, the top 50 HRGs are mainly located in urban areas in Kowloon and New Territories, two of the mostly densely populated districts in the world. More grids in Hong Kong Island were identified as in the top 50-100 HRGs. The targeted interventions had a suppressive effect on new infections over time, though the number of cases continued to increase, and the cumulative number of cases are exponentially growing, implying a delayed intervention effect rather than a preventive effect here as observed in the fully lockdown scenario.

We examined the effect of testing by simulating three scenarios that 10%, 20%, and 50% percent of symptomatic cases are tested after the first 15 days (Fig. 5) . Moreover, to analyze the effect of testing speed, the testing results are set to come out 24-hour or 48hour after the testing. Compared with the baseline data, testing 10%, 20%, 50% of symptomatic cases within 24 hours reduced the infectious by 89.92% (IQR 88.88%-90.64%), 95.47% (95.23%-95.87%) and 97.93% (97.66%-98.13%), respectively. As contrast, testing 10%, 20%, 50% of symptomatic cases within 48 hours reduced in 87.78% (83.85% -90.44%), 92.42% (91.89%-92.84%) and 95.61% (95.11% -96.05%), respectively. Apparently, the speed of testing greatly affects the efficacy of NPIs. These results indicate that timely and prompt testing is essential to such reactive control measure [24] .

Regarding the location of infections, the home-community and workplace ranked 1st and 2nd in all testing scenarios. More importantly, due to the stay-at-home and testing strategies, the home-community was responsible for more than half of infections (Table  1 ). In the scenario of 24-hour testing delay, the home-community and workplace infections accounted for 89.76% (88.73%-90.81%), 88.78% (83.84%-88.93%) and 79.79% (74.89%-81.14%) at 10%, 20% and 50% testing scale level, respectively. The corresponding values were 88.62% (88.09%-88.93%), 89.04% (87.31%-90.30%) and is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint 88.52% (80.89%-88.66%) in the scenario of 48-hour testing delay. Compared to the fully lockdown scenario, the wide gap between the infection sizes implied that the community and workplace-based transmission is dramatically growing and became hard to control if no strict movement restrictions were implemented during the outbreak period. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint represent the MC simulation results. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

COVID-19 has resulted in enormous damage to the society and economy. Many countries have been suffering from a shortage of testing resources and economic stagnation due to the necessary lockdowns. The quantitative analyses with D2M2, which is calibrated with the real-world big data of human mobility, demonstrate that targeted interventions are feasible NPIs to contain the outbreak with only a small fraction of areas and population being affected.

Apparently, the best strategy to contain the COVID-19 outbreak is to have a stringent lockdown, which is also echoed by our results. However, city-wide quarantine and lockdown may lead to additional loss including the elevating unemployment rate, social isolation, and mental health crisis [25, 26] . Moreover, COVID-19 has the potential to become an endemic [11] , if long-term protection of vaccines is not satisfactory. Therefore, many countries would have to lift the NPIs with a certain level of COVID-19 risk for the near future [27] . In practice, the proposed surgically targeted intervention strategy can partially mitigate isolation fatigue and social disruption by proactively identifying only a small subset of the area and population for testing and quarantine, while preserving the mobility of most residents.

In practice, as an international financial center and a major supply chain hub, Hong Kong is likely to face the continuous risk of COVID-19 in the near future. The proposed D2M2 and associated targeted interventions have the potential to guide the sustainable NPIs that achieves a balance between lowering the risk and preserving the human mobility and economy of the city. This also applies to other major cities in the world, such as Beijing, New York, London, Tokyo, etc.

In the big data era, it is essential and promising to capitalize on the well of valuable data about human mobility and urban structure to benefit mankind in various contexts.

Our study is an effort in combining real-world data of population demographics, public facilities and functional buildings, transportation systems, and travel patterns to inform effective NPIs to combat emerging infectious diseases like COVID-19. The limitation of this study is the missing of real-time individualized mobility movement data such as the mobile phone data. Obtaining individualized mobile phone data is extremely difficult, if not impossible. However, the model can be applied to other cities where such individualized mobility data is available. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 2, 2021. ; https://doi.org/10.1101/2021.01.29.21250786 doi: medRxiv preprint

Estimating the effects of non-pharmaceutical interventions on COVID-19 in Europe

Modelling lockdown and exit strategies for COVID-19 in Singapore. The Lancet Regional Health-Western Pacific

Event-specific interventions to minimize COVID-19 transmission

The impact of COVID-19 and strategies for mitigation and suppression in low-and middle-income countries

Projecting the transmission dynamics of SARS-CoV-2 through the postpandemic period

Modelling the impact of testing, contact tracing and household quarantine on second waves of COVID-19

Impact assessment of non-pharmaceutical interventions against coronavirus disease 2019 and influenza in Hong Kong: an observational study. The Lancet Public Health

The epidemiology of COVID-19 cases and the successful containment strategy in Hong Kong

The role of community-wide wearing of face mask for control of coronavirus disease 2019 (COVID-19) epidemic due to SARS-CoV-2

Unique Clusters of Severe Acute Respiratory Syndrome Coronavirus 2 Causing a Large Coronavirus Disease 2019 Outbreak in Hong Kong

Will SARS-CoV-2 become endemic? Science, 2020

A human behavior integrated hierarchical model of airborne disease transmission in a large city

Foreign domestic helpers and evolving care duties in Hong Kong

Queer, brown, migrant: documenting the Hong Kong 'Helper'. Cultural Studies

Interventions to mitigate early spread of SARS-CoV-2 in Singapore: a modelling study

Reconstruction of the full transmission dynamics of COVID-19 in Wuhan

Nowcasting and forecasting the potential domestic and international spread of the 2019-nCoV outbreak originating in Wuhan, China: a modelling study. The Lancet

Projecting hospital utilization during the COVID-19 outbreaks in the United States

Population-scale longitudinal mapping of COVID-19 symptoms, behaviour and testing

Rapid Response to an Outbreak in Qingdao, China

Detection of 2019 novel coronavirus (2019-nCoV) by real