Abstract
A SEIR simulation model for the COVID-19 pandemic was developed (http://covidsim.eu) and applied to a hypothetical European country of 10 million population. Our results show which interventions potentially push the epidemic peak into the subsequent year (when vaccinations may be available) or which fail. Different levels of control (via contact reduction) resulted in 22% to 63% of the population sick, 0.2% to 0.6% hospitalised, and 0.07% to 0.28% dead (n=6,450 to 28,228).
There is pandemic spread of the new coronavirus “SARS-Cov-2”, causing the disease “COVID-19”, with the World Health Organization (WHO) reporting over 200,000 cases and over 8000 deaths on 19 March 2020 [1]. One approach to informing the potential health burden and relevant control measures for a new pandemic is to study its dynamics using mathematical models. Recently published mathematical modelling work on COVID-19 has reported that “in most scenarios, highly effective contact tracing and case isolation is enough to control a new outbreak of COVID-19 within 3 months” [2]. Another modelling study found that “combining all four interventions (social distancing of the entire population, case isolation, household quarantine and school and university closure) is predicted to have the largest impact, short of a complete lockdown which additionally prevents people going to work” [3]. Other such models have been used to estimate the impact of disease control measures in China [4, 5]. Given this background, we explore the potential health impact of the spread of the COVID-19 pandemic in a hypothetical European country of 10 million people to determine the potential impact of control measures, particularly to push the epidemic into a subsequent year, a time when a vaccine might become available.
Europe’s first cases of COVID-19 were reported to WHO on 25 January in France [6]. But 19 March there were 52 of these countries with reported cases, with 12 reporting 1000+ cases [1]. By this time major control interventions were in place in many of these countries [7].
In this modelling, we took the standard approach of using a deterministic SEIR model i.e., key compartments for: susceptible [S], exposed [E], infected [I], and recovered/removed [R]. We developed this model specifically for COVID-19. It is freely available online with a dashboard display to facilitate user interaction (http://covidsim.eu; version 1.0, 19 March). The Appendix details the parameters, derived variables and differential equations used in the CovidSIM model. Table A1 in the Appendix provides the input parameters used in the model, as based on available publications and best estimates used in the modelling work on COVID-19 to date (as known to us on 18 March 2020).
Our results suggest that pushing the peak of the epidemic into the next year, when assuming a low basic reproduction number (R0) of only 1.5, was achievable with “general contact reduction” at levels of over 11% to over 27% (for differing time periods) (Table 1). Similarly, it was achievable for probabilities of isolating symptomatic cases in hospital in the range of over 18% to over 41% (Table 1). Using instead a possibly more realistic value of 2.5 for R0, the only way of shifting the epidemic peak into the next year with interventions lasting for nine months, was if: (i) over 65% of contacts were reduced; or (ii) if the probability of cases being isolated in hospitals was extremely high at over 98%. Also for R0 = 3.5, pushing the epidemic peak into the subsequent year was generally not possible, except if contact reduction was over 61% for the rest of the simulated year.
Threshold analyses for pushing the peak of the COVID-19 epidemic in the hypothetical European country into the next year (i.e., pushing the peak to after day 365 of the simulation with the start of the simulation on 15 February 2020, the date we assumed that the infection spread began)
The feasibility of achieving high levels of contact reduction and case isolation is very uncertain – especially around sustaining these for long periods of time. At least in the short-term China has used intensive containment measures successfully as per the findings of the WHO-China Joint Mission Report [8]. This report stated that: “China has rolled out perhaps the most ambitious, agile, and aggressive disease containment effort in history.” While it is an open question around the generalisability of the Chinese approach to other jurisdictions [9], there is also evidence of containment success (as of late March 2020) outside mainland China, from Singapore, Hong Kong and Taiwan [10]. But for European countries currently with thousands of cases, it may be too late to adopt such intensive containment approaches and the best approach may simply be to apply less intense control measures that balance minimising health loss with minimising social and economic disruption (while also putting major resources into rapid vaccine development).
Figure 1 shows the scenarios at three different levels of R0 when combined with “25% general contact reduction”. At R0 = 3.5 the epidemic still peaks during the intervention period whereas for R0 = 2.5 the peak is pushed into the post-intervention period. For R0 = 1.5 the peak is nearly pushed into the subsequent year.
Numbers of prevalent symptomatic COVID-19 cases by day for the R0 = 1.5, 2.5, and 3.5 scenarios and all with 25% of “general contact reduction” as modelled using CovidSIM (unrecognised introduction of the infection occurs at time 0; control measures begin at day 5 and end six months later i.e., on day 189).
Figure 2 again demonstrates the importance of the timing of the intervention period. A higher level of contact reduction (at 50% vs 25%) more effectively suppresses the epidemic in the intervention period – but when the intervention period ends it causes a higher epidemic peak and a higher total health burden than for the 25% reduction (Table 2).
Estimated health impacts from COVID-19 in a hypothetical European country of population 10 million using CovidSIM for a range of basic reproduction number (R0) values and differing intensity of “general contact reduction” as the control measure (see Table A1 for input parameters)
Numbers of prevalent symptomatic COVID-19 cases by day for the R0 = 2.5 scenarios and different levels of general contact reduction as modelled using CovidSIM (unrecognised introduction of the infection occurs at time 0; control measures begin at day 5 and end six months later i.e., on day 189).
Considering the three levels of R0 and the two levels of “general contact reduction” (at 25% and 50%) resulted in: 22% to 63% of the population sick, 8.8% to 25.1% seeking a medical consultation, 0.2% to 0.6% needing to be hospitalised, 5,500 to 15,682 people needing critical care (in an ICU), 2,750 to 7,841 requiring ventilators, and 6,450 to 28,200 dying (0.07% to 0.28% of the population) (Table 2). The worst scenario with 0.28% of the population dying, compares to Europe in the 1918 influenza pandemic where 1.1% died (2.64 million excess deaths) [11].
Based on the recent age distribution data from China [12], these hospitalisations and deaths from COVID-19 would particularly occur amongst older age-groups. Overall, the levels of health service demand in the more severe scenarios would be completely unprecedented for a modern European country.
This is one of the first SEIR modelling studies of this new pandemic disease COVID-19 and the associated online simulation tool has advanced dashboard features and graphic visualisation of results that facilitate user engagement. Nevertheless, there are always limitations with modelling work. Some of these are discussed under issues with parameter uncertainty in Table A1, but here we note some of the more substantive limitations:
There is still a high degree of uncertainty around many aspects of COVID-19 epidemiology. For example, the value of R0 could conceivably be higher than the highest level in the scenarios we modelled (at R0 = 3.5). Also, the case fatality risk (CFR) could be underestimated, especially if there was health system overload.
The model was deterministic and not stochastic, though we largely offset this issue with modelling a wide range of scenarios. The lack of stochastic considerations mainly translates into increased uncertainty in the very early stages of epidemic spread, which then impacts on the timing of the peak.
The model neither considers any long-term health damage to survivors (especially among risk patients) nor does it consider the hard-to-estimate health loss arising from untreated other health conditions as a result of having an overburdened health system. Likewise we do not consider the additional health harm to the health workers involved e.g., adverse mental health impacts arising from working during a pandemic [13, 14].
Some of these issues can be addressed when improved data becomes available on the epidemiology of COVID-19. Nevertheless, public health workers can consider using this model in their own countries and adapt the parameters according to local settings and as the epidemiological characteristics of COVID-19 are better ascertained.
Data Availability
All the necessary data can be generated by the users themselves since the model is freely available online with a data file export capacity.
Parameters used in the modelling with CovidSIM
Input parameters for modelling the health impacts for a hypothetical European country using CovidSIM
Appendix: Description of the CovidSIM model and model input parameters
Model dynamics
Number of fully isolated cases at time t
Number of home isolated cases at time t
Initial values
![Embedded Image](https://www.medrxiv.org/sites/default/files/highwire/medrxiv/early/2020/03/23/2020.03.20.20039776/embed/graphic-8.gif)
Parameters
- N
- Population size
- X
- Number of initial infections
- tmax
- Day after introducti on of the infection when the transmissi on potiential is highest
- Qmax
- Maximum isolation capacity
- Time at which isolation measures start
- Time at which isolation measures end
- cHome
- Fraction of contacts which are prevented for cases who are in home isolation
- cCont
- Fraction of contacts which are prevented
- Time at which contact reduction starts
- Time at which contact reduction ends
- c(t)
- Fraction of contacts which are reduced at time t
- y
- Force of infection which originates fromoutside of the population (e.g. via travellers)
- R0
- Average value of the basic reproduction number
- a
- Amplitude of the seasonal fluctuatio n of R0
- DE
- Average duration of the latent period
- nE
- Number of stages for the latent period
- ε
- Stage transition rate in the latent period (ε = nE / DE)
- DP
- Average duration of the prodoromal period
- nP
- Number of stages for the prodromal period
- φ
- Stage transition rate in the prodromal period (φ = nP / DP)
- iP
- Relative infectiousness during prodromal period
- DI
- Average duration of the symptomatic period
- nI
- Number of stages for the symptomatic period
- γ
- Stage transition rate in the symptomatic period (γ = nI / DI)
- βI (t)
- Effective contact rate of individual s in the symptomatic period at time t βI (t) = R0/(iPDP + DI)× (1+ a cos(t / 365))
- βP (t)
- Effective contact rate of individual s in the prodromal period at time t (βP (t) = βI (t)iP)
- pSick
- Fraction of infected individual s who become sick
- pConsult
- Fraction of sick individual s who seek medical help
- pHosp
- Fraction of sick individual s who are hospitalized
- pICU
- Fraction of hospitalized individual s who are admitted to the ICU
- pDeath
- Fraction of sick individual s who die from the disease
Derived variables
![Embedded Image](https://www.medrxiv.org/sites/default/files/highwire/medrxiv/early/2020/03/23/2020.03.20.20039776/embed/graphic-9.gif)
Detection probability
SARS-CoV-2 infections which are brought into the country may not be detected and may spread without being noticed because the symptoms of COVID-19 may easily be confused with other influenza-like illnesses (ILI). Few practitioners may decide to order a SARS-CoV-2 test for what they regard a normal ILI patient while no community-transmitted cases in the population have been reported. If we assume that fraction of pTest ILI patients who (a) seek medical help or who (b) are hospitalized or who (c) die from the disease are tested for SARS-Cov-2, then the probability that not one single test has been performed on a COVID-19 patient by time t despite the ongoing transmission in the population is given by:
The probability that at least one test has been performed (and has returned a positive result) is then