COVID-19 healthcare demand and mortality in Sweden in response to non-pharmaceutical (NPIs) mitigation and suppression scenarios

Introduction

The novel SARS-CoV-2 is highly transmissible(1), rapidly spread around the globe since it emerged in Wuhan, China(2), at a rate much faster than other emerging infectious diseases such as Ebola.(3) In response to the COVID-19 outbreak, China implemented extraordinary public health measures at great socio-economic cost, moving swiftly to ensure early identification of cases, prompt laboratory testing, facility-based isolation of all cases, contact tracing and quarantine.(4) In the community, physical distancing was implemented at a grand scale, all mobility put to an halt, and the city of Wuhan was in lock-down for about 9 weeks.(5) China’s tremendous efforts showed success.(6) Other Asian countries facing a major explosion such as South Korea also managed to curb the epidemic. South Korea employed very liberal testing, hospital-based isolation of all cases, combined with extensive contact tracing enhanced by mobile phone and digital technologies, but did not use a lock-down.(7, 8)

While the outbreak in China appears to be contained, since mid March 2020, the epicenter of the COVID-19 pandemic is in Europe, and since April in the United States. There is thus an urgent need to determine how best to reduce transmission rates, the height of the epidemic peak, the peak demand on healthcare services, and how to reduce fatalities.

In the absence of vaccines, a wide range of control measures can be considered to contain or mitigate COVID-19. These include active case finding with prompt isolation of cases, contact tracing with quarantine of contacts, school closures and closures of public places, mobility restrictions, physical distancing in the community, physical distancing only of the elderly, and a lock-down (also known as Cordon sanitaire).(4) There is currently no consensus about which measures should be considered, in which combination, and at which epidemiological threshold such measures should be implemented for maximum public health impact.(9)

Two strategies can be considered: (a) suppression which aims to rapidly reverse epidemic growth, thereby reducing case numbers to low levels, and (b) mitigation, which focuses on slowing but not necessarily immediately stopping epidemic spread – reducing peak healthcare demand while shielding those most at risk of severe disease from infection. Each policy has major challenges. Suppression aims to rapidly reduce the reproduction number, R₀, to below 1, thus causing case numbers to consistently decline. Mitigation aims to slow spread by reducing R₀ to a value close to but slightly above 1 for some time before the epidemic growth will cease gradually, partially in reponse to increasing levels of disease immunity in the population.

Public health measures need to be weighed up against economic repercussions and mental health of prolonged lock-downs. Although more strict community containment measures such as lock down will result in a shorter duration of the outbreak(10), the non-health sector negative consequences may be huge.

Sweden decided to implement public health interventions without a lockdown. Schools and universities were not closed, restaurants and bars remained open, instead Swedish citizens implemented "work from home" policies where possible, social distancing without police enforcement, and shielding of those older than 65 of age.

Here we aim to quantify the effects of the Swedish measures. We estimate the impact of COVID-19 on the Swedish population at the municipality level, considering demography and human mobility under various scenarios of mitigation and suppression. We estimate the time course of infections, health care needs, and the mortality in relation to the Swedish ICU capacity, as well as the costs of care, and compared alternative policies and counterfactual scenarios.

Methods

We developed a compartmental epidemiological model based on the SEIR formulation, and extended it to account for additional variables including compartments for health and ICU care. All these variables were age-structured (0-59, 60-79, and 80+ years). The model included age-structured compartments for susceptibles, exposed, infected, inpatient care, ICU care, dead and recovered populations based on Swedish population data at the municipality level (see Supplementary Information 1). Overall, the population of infected individuals were divided into two different groups, those that had sufficiently severe symptoms to potentially end up in hospital care (1 out of 6), and those who had mild or asymptomatic infections or were sick at home (5 out of 6).(11, 12) This parameter was calibrated to data. The model allowed for three different ways that deaths could occur: 1) after unsuccesful ICU treatment; 2) after triage and denial of ICU due to low chances of surviving or when ICU demand exceeds ICU capacity; and 3) outside of healthcare.

The model captured spatial demographic heterogeneities at the level of municipality in Sweden, and inter-municipality travelling based on a radiation model (see Supplementary Information 1). The radiation model was calibrated using a N1H1 Influenza A model and data for the period 2015-2018 in Sweden. Demographic data was obtained at the municipality level for the year 2018 from Statistics Sweden.

The parameterization of the model was achived in to ways: (1) some parameters were set to fixed values based on what is known through literature and data; (2) some parameters were given initial values from the international literature and then calibrated by fitting the model to current available outbreak-data on infection prevalence, deaths, ICU load and healthcare in Sweden. The full set of parameter values are given in the Supplementary Information Table S1.1. Age specific health-care need parameters from Ferguson et al. was initially used to represent the three age groups in our study (Supplementary Information 1, Table S1.2). These values where then calibrated to better reflect observed infection prevalence, in-patient care demand, ICU care demand and registered deaths due to COVID-19 from Stockholm region and Sweden (See Supplement S1.2).

The infectious period is likely to vary by the individual and range from days to weeks. Viral shedding is reported to occur from 7 to 22 days, including in mild cases of disease (13), and is a driver of disease transmission. Isolation of patients, or staying home if presenting with symptoms, will reduce transmission to contacts, and is a key strategy to contain COVID-19. It shortens the period that infected persons are able to infect others. Importantly, transmission from an infected but asymptomatic or pre-symptomatic individuals can still occur despite control measures.(14) We assumed that the average effective infectious period in the general population is 5 days, shortened from 7 days by natural isolation of symptomatic infected individuals. We assumed that individuals going into healthcare were admitted on average after 3 days of symptomatic infection, and were isolated from transmitting the virus to other individuals while in hospital care.

We used a measurement from the Swedish Public Health Agency of virus prevelence measured by nucleic acid detection in nasopharyngeal samples from the general population in the Stockholm region centered around the 1th of April. The measurement indicated 2.5% (95 % Cl = 1,4 – 4,2 %) of the population were infected.(15) To calibrate the infection prevalence in the model with the study result, we adjusted the model by allowing a latent compartment after infectiousness for the general infected population. Our model thus takes into account that virus ion can be detected for 7 days.

We calibrated the daily transmission rate, β, to 0.91c(t) to Swedish data on healthcare load, virus prevalence, and mortality (Supplementary Information 1), where c(t) is a time dependent contact-rate scaling parameter (Supplementary Information 1). In our model, R₀ is dependent on contact-rate and infectious period which parameters in turn are dependent on age-distribution (See Supplementary Information 1). Accordingly, we account for spatial and demographical heterogeneity in R₀. The country-level R₀ is the average over all municipality-local values for R₀. It is not with in the scope of the paper to derive exact values for R₀, yet we can see that the within-municipality R₀, for the no-countermeasures scenario (i.e., scenario a), is within the range of 2.73 to 4.55, and this range is consistent with the reported basic reproduction rates for COVID-19.(16) The country-level R₀ is given by an average over municipality-local R₀ values. Due to the many travelers infected with SARS-CoV-2 arriving from Italy, Austria and other parts of Europe, in the week of the 24^th of February, the model was seeded with 1 case per 120,000 individuals for all municipalities except for the municipalities within Region Stockholm that were seeded with 1 case per 30,000 individuals.

The model was set up to predicting the municipality transmission dynamics and inter-municipality spread across Sweden starting from 24^th of February and ending a little more than 6 months later, September 1, 2020. The model used scenarios to describe the countermeasures and counterfactual impacts. The mitigation and suppression scenario had onset the 20^th of March by a transient function with full effect by the end March (Supplementary Information 1, Table S1.1). The five different scenarios is summarized by:

1. no public health interventions (counterfactual scenario);

2. modest physical distancing in ages 0-59 years, moderate in ages 60+ years;

3. modest physical distancing in ages 0-59 years, moderately strong in ages 60+ years;

4. moderate physical distancing in ages 0-59 years, strong in ages 60+ years, and with an inceased degree of isolation of infectious individuals;

5. same as in scenario d), but with a slightly higher degree of isolation of infectious individuals.

A complete description of the scenarios is provided in the Supplementary Information 1.

In the scenarios the ICU capacity were compared against current baseline availability of 526 ICU beds in Sweden, which has been doubled in respons to the COVID-19 pandemic. The numbers of deaths and the infection fatality rate (IFR) associated with the different mitigation and supression scenarios were derived from our model. We also derived the direct healthcare cost for each of the scenarios (see Supplementary Information 4).

Additionally, we extracted total all-cause of deaths from Stockholm region and estimated a potential excess mortality beyond confirmed COVID-19 cases. This was done by comparing the excess mortality during the COVID-19 outbreak with the mortlity the weeks before the outbreak for the same and previous years using time series regression methods, while adjusting for time trends and the patterns of deaths previous months and years (Supplementary Information 3).

Results

The model showed a good fit against the reported COVID-19 related deaths in Sweden up to 20^th of April, 2020, in scenario d (Figure 1m-p). In Table 1 we present the R² and the mean square error of the observations to the predictions for scenario (a) to (e) in relation to: deaths for Sweden as a whole, deaths in Stockholm, ICU bed demand in Stockholm, and inpatient care in Stockholm. Scenario (d) further well describes the observation that 2.5% of the population in Stockholm were infected with the SARS-CoV-2 virus centered around the 1 of April (Figure 2d). Overall, the IFR for Sweden in scenario (a) to (e) is estimated to 0.40; 0.40; 0.37; 0.28 and 0.27, respectively (Table 1).

• Download figure
• Open in new tab

Figure 1.

Predicted number of total deaths from COVID-19 in the whole population in Sweden (first column), and for Stockholm region (second column); predicted demand of ICU beds in Stockholm (third column), and; inpatient care beds in Stockholm (fourth column). Actual observations in the early phase of the outbreak are illustrated as circles (O). The mitigation scenarios are organized in rows with panel a) no public health interventions (counterfactual scenario); b) modest physical distancing in ages 0–59 years, moderate in ages 60+ years; c) modest physical distancing in ages 0–59 years, moderately strong in ages 60+ years; d) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and improved isolation of infectious individuals; e) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and further improved isolation of infectious individuals. Mitigation giving rise to these predicted values had onset the 20^th of March.

In Figure 3 we present the scenarios of country level COVID-19 ICU bed demand over time in Sweden per the different age groups, and in total. According to scenario (a), the outbreak would peak in the middle of April and reach an ICU bed demand of around 10,000 patients (Figure 3; panel a). The age group below 60 years of age alone would take up more than the baseline ICU resources of 526 beds during a month at the peak. According to scenario (b), the ICU demand would peak at around 7,000 beds at the peak around the first of May (Figure 3; panel b). The demand would be flattened and continue for a longer period. The ICU demand for those below 60 years of age would again exceed the baseline ICU beds for 1 month. According to scenario (c), the ICU bed demand would peak at around 6,000 at the peak around the first of May (Figure 3; panel c). The demand would be flattened and continue for a longer period. The ICU bed demand for those aged below 60 years would almost take up the baseline ICU beds for a period slightly less than 1 month. According to scenario (d), the intensive care demand would not increase beyond 500, it would peak in June and decrease in August (Figure 3; panel d). According to scenario (e), the intensive care demand would be very small for the whole study period (Figure 3; panel e). We note, that the outbreak would likely resurge unless others means are there to control the transmission when the countermeasures are lifted in scenario (d) and (e). Corresponding estimates for Stockholm region are given in Figure 4 (panel a-e). Notewhorthy, the healthcare demand peaks earlier in Stockholm as compared with Sweden as a whole and decreases earlier in the summer.

• Download figure
• Open in new tab

Figure 3.

The predicted ICU bed demand per day from 24^th of February to the 1st of September, 2020, overall in Sweden in relation to different suppression & mitigation scenarios. Panel a) no public health interventions (counterfactual scenario); b) modest physical distancing in ages 0–59 years, moderate in ages 60+ years; c) modest physical distancing in ages 0–59 years, moderately strong in ages 60+ years; d) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and improved isolation of infectious individuals; and e) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and further improved isolation of infectious individuals. Mitigation giving rise to these predicted values had onset the 20^th of March.

• Download figure
• Open in new tab

Figure 4.

The predicted ICU bed demand per day from 24^th of February to the 1st of September, 2020, in the region of Stockholm in relation to different suppression & mitigation scenarios. Panel a) no public health interventions (counterfactual scenario); b) modest physical distancing in ages 0–59 years, moderate in ages 60+ years; c) modest physical distancing in ages 0–59 years, moderately strong in ages 60+ years; d) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and improved isolation of infectious individuals; and e) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and further improved isolation of infectious individuals. Mitigation changes giving rise to these predicted values had onset the 20^th of March.

In the Supplementary Information 2 we show how the timing of the increase in cases across the scenarios a-e presented in Figure 2 is sensitive to mobility between municipalities (Figure S2.1). There is a slightly earlier increase in ICU bed demand with higher inter-municipality mobility.

• Download figure
• Open in new tab

Figure 2.

The number of individuals in in the Stockholm region population predicted to carry the virus over time as determined by a virus detection assay. An actual observed result from the 1 April, 2020, using a population sample is illustrated by a circle (O) with 95% Cl (vertical bars). Panel a) no public health interventions (counterfactual scenario); b) modest physical distancing in ages 0–59 years, moderate in ages 60+ years; c) modest physical distancing in ages 0–59 years, moderately strong in ages 60+ years; d) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, and improved isolation of infectious individuals; and e) moderate physical distancing in ages 0–59 years, strong in ages 60+ years, with further improved isolation of infectious individuals. Mitigation giving rise to these predicted values had onset the 20^th of March.

In Table 2 we describe the predictions of the total number of individuals infected, the total person days of care, the total person days of ICU occupancy, the total number of deaths (assuming all ICU demands are satisfied), and the deaths from ICU capacity shortage (100% above baseline level). Following on this, we estimate the direct costs of the care and intensive care demands (Table 3). The number of infected individuals in Sweden is predicted very high by the model in scenario (a) to (c) with attach rates beyond 90% of the population already by the 1th of September, 2020 (Table 2). In the scenario (d) and (e), only 3.2 and 1.8 million people, respectively, would be infected by 1^th September.

Overall, the scenarios show the demand on inpatient care varies from just below 700,000 person-days to just above 100,000 person-days and the, while the demand on intensive care range from around 270,000 to around 30,000 person-days. The death rates, assuming no limits in ICU, varies from around 40,000 to 5,000 depending on the mitigation and suppression actions. Assuming instead a cap of the ICU bed capacity of 100% above baseline, the estimated number of additional excess deaths from lack of ICU capacity varies from 21,000 to 0 depending on the scenario. The total direct medical cost range between 20 billion to 4 billion SEK depending on the scenario (Table 3).

From the calculation of excess deaths associated with the COVID-19 outbreak in Stockholm (Supplementary Information 3), we observe a linear increase in all-cause deaths of 0.40 (95% Cl = 0.24, 0.57) for each registered COVID-19 death. The analyses excluded reported COVID-19 deaths from all-cause mortality, and thus indicate 40% additional deaths beyond reported COVID-19 cases.

Discussion

Our study shows an exponential growth of the number of COVID-19 infections, health care demands and deaths in Sweden which became apparent towards the end of March and the beginning of April, 2020. In April, it further suggests a strong effect of the physical distancing efforts put into place around the 20^th of March in Sweden. The epidemiological data from Sweden align best to our modelled secenario (d) which describe a moderate physical distancing in those below 60 years of age, a strong shielding of those above 60 years of age, and improved awareness and compliance of home isolation of symptomatic COVID-19 cases. The level of physical distancing and isolation in this scenario did not compromise the access to health care, and did not overwhelm the health care system. The extent of measures were less stringent and economically damaging compared to those introduced in other Scandinavian countries (Norway, Finland, Denmark), but the number of deaths by 5 May 2020 was much higher: according to the WHO situation report on 5 May 2020, there were 2,769 deaths with 22,721 confirmed infections in Sweden, versus 493 deaths with 9,670 reported cases in Denmark, 208 deaths with 7,847 cases in Norway, and 240 deaths with 5,327 cases in Finland.(17)

In the counterfactual scenario (eg no public health interventions), the intensive care unit demand was estimated to be almost 20 times higher than the intensive care capacity in Sweden and the number of deaths would be between 40,000 to 60,000

Our estimates using the Swedish model of physical distancing without lockdown show that by end August, 2020, about 30% of the Swedish population will be infected, resulting in about 9,000 deaths by that time point. Despite such high exposure, Sweden would remain far below the herd immunity needed to stop the outbreak, a threshold estimated to be around 60–70%). (18) Therefore, up to September 2020, the predicted impact is very dependent on ongoing adherence to physical distancing and shielding of vulnerable age groups. If such measures are not maintained, the consequences would be severe with demands substantially exceeding health care capacity and mortality rates rapidly increasing. Natural herd immunity (i.e., a situation where R₀ goes below 1 even if countermeasures were discontinued) appears not a viable objective to stop the virus circulation given the predicted death tolls. Further measures, including enhanced testing, prompt isolation of cases, more effective contact tracing and quarantining of contacts(19, 20), would result in further reducing transmission intensity and daily new cases, while avoiding lockdowns, until other control options such as vaccine and effective therapeutic options are readily available.

Of note, due to the strong triage in our model with only 15% of those aged 80+ being treated at the ICU care, the demand of this group is overall already estimated to be very low. In order to capture the Swedish mortality patterns, our model further estimates a substantial number of deaths occurred outside hospitals, mainly in care homes. The excess mortality estimates published by the European Centre for Disease Control indicate a substantial amount of additional deaths must have occurred due to COVID-19. (15) The total all-cause mortality in Stockholm indicate that a confirmed COVID-19 death is associated with an additional 0.40 (95% Cl: 0.24, 0.57) all-cause death eg 40% additional deaths beyond the reported COVID-19 cases.

In Sweden, we calibrated the rates of in-patient care and critical care to be lower than the ones predicted by Ferguson et al. for the UK (See Supplement Information, Table S1.2). Our results show that the demand on ICU beds can be reduced not only by a suppression strategy as successfully used in China, but also by a mitigation strategy. However, deaths cannot be effectively prevented in mitigation scenarios as many of those would occur independent of ICU demands. Of note, for the modes of death possible in our model (see Supplementary Information, Table S1.3), the ICU demand can be maintained at a lower level while deaths rise, which appears to align well to the Swedish reported data including a substantial number of deaths occurring outside hospitals.

Our analyses address the impacts from COVID-19 on the health-care demand, deaths, and direct healthcare costs in Sweden in relation to different public health interventions. As such it is in line with the assessment of the European Center for Disease Prevention and Control regarding the COVID-19 pandemic. We find that the direct health-care related costs are substantial ranging between 4 and 20 billion SEK dependent on the scenario. We thus showed that more stringent mitigation or suppression efforts yield larger direct health care cost reductions compard to less stringent mitigation or suppression, or the counterfactual scenario, with the maximum cost difference estimating 17 billion SEK. These cost estimates are likely underestimating the true costs because the estimates are only measuring the direct costs per patient per day during normal health-care demand in 2018. Further on, the costs of the health sector would need to be balanced against the cost of the economy as a whole. The estimates here do not capture impacts within the healthcare from other acute health-problems for which treatment is down prioritized or postponed due to the acute situation of the epidemic. It does further not consider economic impacts beyond the health sector.

We also did not consider the role of seasonality. Underreporting, or asymptomatic transmission, can be a driver of herd-immunity. Preliminary findings from Iceland found no more than 50% asymptomatic carriers.(21) Ferguson et al. estimated an CFR around 1.6% and and IFR of around 0.8% based on an 50% asymptomatic rate. Our study found higher rates of mild and asymptomatic transmission fitted better with the Swedish data and estimates slightly lower IFR’s. Preliminary reports from New York and Germany indicates this may be the case, but further investigation with more robust estimation of antibody prevalence and infection fatality rates are needed.

Our study supports that public health interventions such as social distancing combined with shielding of older persons, even without a strict lock-down, can protect the health care system by not exceeding the ICU capacities, but resulted in more cases including deaths compared to neighbouring countries with similar population densities that introduced more stringent lock-down measures.

Conflict of interest statement

None declared