Abstract
The spread of COVID-19 is expected to put a large strain on many hospital resources, including ICU bed space, and mortuary capacity. In this report we study the possible demands on ICU and mortuary capacity in Sydney, Australia, using an adapted SEIR epidemiological model.
1. Introduction
The novel coronavirus that causes COVID-19 has been spreading rapidly throughout the world, and has emerged in several hotspots in Sydney. Because the spread in other places has been so rapid, health systems have been overwhelmed. In response, authorities have instigated a number of measures to mitigate the spread, and attempt to ease this burden.
In this study we explore the consequences of various mitigation effects on demand for critical services, in particular intensive care units (ICU), and mortuary space. The key inputs into the model are the basic reproductive number of the virus, the R0, and the effect of mitigation strategies on this reproductive number. A number of estimates of R0 for COVID-19 have now been published (e.g., [3, 5, 10]), and likewise information is being gathered about the effect of certain mitigations elsewhere on the reproductive number [5]. We use these estimates as input into a published mathematical model, and apply the model to the Sydney environment.
The model we use is an extension of the well-established SEIR family of mathematical models (Susceptible, Exposed, Infectious, Removed), that has been adapted specifically for COVID-19 by the Neher lab at the University of Basel [6]. The extension involves including various stages of illness, from infectious, to severe (in hospital), and critical (in the ICU). This level of detail is important for predicting the demands on the ICU.
Aside from running the model for a range of values of the basic reproductive number R0 (consistent with published estimated ranges) and two different mitigation effects (moderate and strong), the model takes as input a number of other key parameters. These include age-specific severity data, which we obtain from a recent Imperial College, London report [3], data on the length of stay of COVID-19 patients in each stage of hospital care, also from [3], age distribution data from Sydney [7], and a seasonal forcing effect that takes into account the likely contribution of the winter months to transmissibility.
Key findings
For low values of the basic reproductive number, and strong effect of mitigation in Sydney, the model predicts an ICU capacity of 800 beds may be sufficient. However, for higher values of R0 or any “moderate” mitigation effect, it will not.
The range of possible outcomes for this outbreak is vast. For example if mitigation is moderate then any of the R0 values we tested leads to ICU capacity being exceeded. At the higher end, if R0 = 2.95 then the ICU capacity is predicted to be over-run by a factor of 7.
Demand for mortuary capacity is also very sensitive to the underlying R0 and strength of mitigation effect. For low values of R0 and strong mitigation, less than 50 spaces are predicted to be sufficient. However in any other scenario, capacity will need to exceed 150 spaces, and up to 1700 could be needed.
2. The extended SEIR model
We model the spread of COVID-19 in Sydney using an extended SEIR model developed by the Neher lab at the University of Basel [6]. In this model, Infectious are taken to be “in the community”, and there are several populations in the hospital system: Severe (H - in hospital), Critical (C - in ICU), and Dead (D). While normally R is “Removed”, which might include recovered and dead, we now use it for Recovered.
A schematic of the model is shown in Figure 1. The model can be expressed as a set of ordinary differential equations giving the rates of changes of the size of each sub-population. The mathematical details are given in Appendix A.
Like other models, ours has many assumptions including the choices of parameters. We provide these details in the Appendix. The key scenarios we have modelled are:
Basic reproductive numbers (R0) between 1.95 and 2.95.
Strong or moderate effects of mitigation strategies (scaling the R0 by a factor of 0.45 or 0.6 respectively).
These numbers were chosen using the modeling of the team at the London School of Hygiene and Tropical Medicine, published in The Lancet on March 11, 2020 [5]. Their estimate of the R0 in Wuhan before any intervention is R0 = 2.35, and one week after intervention began was R0 = 1.05 (that is, 45% of 2.35). The estimate of 2.35 pre-intervention is also commensurate with the estimate of the group at Imperial College London, who arrived at a value of R0 = 2.4 [3]. Further estimates in the literature are mentioned in Appendix A. We consider the intervention in Wuhan as “strong”, and also consider a “moderate” intervention with scaling factor These factors are also close to those considered in the Neher lab modelling, which considered mitigation factors of 0.4 and 0.6.
3. Results
The outputs of simulations relating to hospital demand (“Severe” cases, population H) are shown in Figure 2 (a) and (b), for ICU demand (“Critical” cases, population C) are shown in Figure 2 (c) and (d), and for mortuary demand (population D) in Figure 2 (e) and (f).
Key outputs of the model are summarized in Table 1.
4. Conclusions
The model predicts that hospital demand (excluding ICU) under the default scenario of R0 = 2.35 and strong mitigation will peak at about 570 beds by the end of October 2020. If mitigation is only of moderate effect, demand will swamp the system, with even a low R0 = 2.15 giving rise to demand for more than 5000 beds.
An ICU capacity of 800 beds is predicted to only be sufficient if there is a strong mitigation effect and for values of baseline R0 at 2.35 and below (Figure 2(c)). If the mitigation effect is only moderate (Figure 2(d)), they will not be sufficient, and there will be extreme overload beginning in May if the baseline R0 is high. Under the default R0 = 2.35 and strong mitigation, ICU demand will exceed 300 beds in November. Note: this models demand from COVID-19 alone and does not account for regular demands on the ICU capacity.
Mortuary requirements vary greatly depending on the R0 and strength of mitigation (Figure 2(e) and (f)). If R0≤2.35 and mitigation is strong, peak demand would be under 50 beds in Sydney. Any scenario with only moderate mitigation or a higher underlying R0 would result in the need for hundreds of mortuary beds.
Another output of the model is the number of fatalities over time, and this is shown in Figure 2(g) and (h). Here we see the major effect of the strength of mitigation: no scenario with only moderate mitigation predicts fewer than 6,600 fatalities. Under moderate mitigation effects, those large numbers of fatalities mostly stop increasing by late 2020, indicating the epidemic has washed through the population, albeit at massive cost to life. On the other hand, the effect of strong mitigation can be seen in the lower predicted numbers of fatalities, with some scenarios (R0 = 1.95 or 2.12) staying under 1000, but it also shows that the outbreak will persist longer, into 2021.
There are several key uncertainties in the modelling:
What is the effect on underlying R0 of the actual mitigation strategies that have been put in place and are being considered? While it would appear that the measures introduced in Sydney — physical distancing, and the closure of non-essential businesses — should have a “strong” effect on R0, this key question is beyond the scope of this study. What we can say from the results here is that mitigation strategies need to be strong enough to produce the desired effect on R0.
What is the baseline R0 for Sydney? The Imperial College study used 2.4 as default [3], and Kucharski et al. in Lancet estimated it at 2.35 [5]. These estimates have wide confidence intervals, and it could be lower in Sydney due to earlier awareness and also a more diffuse population. The Harvard group had a lower estimate of approx 1.9 but that assumed a quite high exposed-to-symptomatic probability of 0.5≥[10]. This assumption has the effect in their models of lowering estimates of R0.
5. Extensions, further questions
Some possible extensions that could be made within the same modelling framework include:
Focusing on specific region of Sydney, which could be relevant if each hospital receives patients from local catchment.
Including specific comorbidities may be valuable. For instance, Sydney may have higher rates of diabetes or obesity than Wuhan, which might affect death rates, while Wuhan has higher rates of smoking, especially in men. These variations could be implemented by changing the age-related parameters, assuming these conditions are aligned with different age profiles.
Data Availability
All data in the referred to in the manuscript is publicly available; links are given in the manuscript.
Appendix A. The mathematical model and implementation
A.1. The mathematical model
The basic model used in this study was introduced in [6], and is a generalisation of classical “SEIR” epidemiological models that have been in the literature since the SIR model was introduced in 1927 [4]. An introduction to such models can be found in the text [1].
In this implementation of the SEIR-type model shown in Figure 1, each compartment (sub-population) is further divided into age brackets, and these have different features based on what has been inferred about COVID-19 to date. Thus the model is really a set of models, one for each age bracket, denoted a (so that Sa refers to the susceptible population in age-bracket a). Age brackets and age-specific parameters are shown in Table 3.
The differential equations that govern the flow of people through the compartments are as follows, as in [6]:
For instance, the first equation says that the rate of flow of people out from Sa (susceptible in age category a), to Ea (exposed) at time t, is the total infectious population ∑ Ib times the infection rate βa at time t, times the proportion of the population still susceptible, Sa(t)/N. Patients leave the hospital on average after th days, so each day on average Ha(t)/th patients leave hospital (seen in the expression for dHa/dt). Of these, ca go to the Ca (the ICU), and (1 −ca) recover (go to Ra).
If a patient needs ICU but ICU is not available, we assume (following [6]), that leads to a fatality.
To summarize the parameters, we have:
Transmission rate βa(t). This is essentially a function of the basic reproductive number R0, but also accounts for mitigation effects and seasonal forcing:
The expression for seasonal forcing is standard (see eg [1, Sect. 1.8]), but includes the mitigation effect M (t) on R0.
Average times in each state of the model are given by
− tl: latent time exposed before becoming infectious,
− ti: time infectious before either recovering or becoming severely ill,
− th: time severe, in hospital, before either recovering or becoming critical,
− tc: time critical, in ICU, before either dying or leaving the ICU within the hospital.
Age-specific parameters:
− ma is the proportion of cases in Ia that stay mild (asymptomatic).
− ca is the proportion of cases in Ha that become severe.
− fa is the proportion of cases in Ca that are fatal.
A.2 Parameter settings
Table 2 gives a summary of key parameter settings in our implementation of the model.
The age related parameters given in Table 3 are from [3], and originate from studies of Wuhan cases in [9].
There are now many estimates of the value of R0 in the literature, ranging from a bit below 2 to above 3. Low R0 estimates of 1.8 −2.0 were found in [10], assuming a relatively high probability that an infection becomes symptomatic of Psym≥0.5. Lower values of Psym in that study give higher estimates of R0. An estimate of R0 = 2.2 was given in [8], using stochastic simulations and data from the early outbreak. A range of 2.0 to 2.6 was used in modelling in [3], with a default setting of 2.4. An agent-based modelling study, calibrated with similar data, estimated R0 = 2.27 [2]. Another study found pre-mitigation R0 in Wuhan to be 2.35 [5], dropping post-intervention to 1.05, which is approximately 45% of 2.35. And an estimate of R0 = 2.68 in [11], published 4th February, was based on the early spread within Wuhan and seeding of new outbreaks within China.
Taking these into account, in this study we take R0 = 2.35 as a central parameter setting (looking at several values to each side), and treat a mitigation scaling effect of 0.45 to be “strong” (since the Wuhan intervention was very strong), and a scaling of 0.6 for a moderate mitigation effect.
A.3 Implementations of the model
The publicly available code released by the Neher lab was the starting point for our modelling, and we adapted some of the javascript to produce the outputs required for this study. In parallel, we wrote an implementation of the model in the statistical package R to give us more flexibility in extending the model and changing assumptions. The two implementations were checked for agreement on identical input parameter settings. The R package will be made available for others to check and adapt for their own use.