Reconstructing the first COVID-19 pandemic wave with minimal data in the UK

Introduction

The COVID-19 pandemic has inflicted devastating effects on global populations and economies [2, 3] and now still affecting countries in many different ways. Reviewing the challenges posted by the COVID-19 pandemic and evaluating previous responses is vital important for future pandemic preparedness [4-7]. Accurate estimation of exposure remains crucial for understanding the dynamics of disease transmission and assessing the impacts of interventions along different stages of pandemic. However, this is particularly challenging in the early phase since most of the characteristics of the pathogen are unknown and at the same time epidemiological data are sparse.

Confirmed COVID-19 cases was typically the first type of data to be collected and reported mostly due to the syndrome surveillance systems [8, 9]. However, it usually underestimates the true exposure in the population because of the limited capacity of diagnoses, the unsolid definition of cases, testing criteria and etc. Large-scale viral infection survey in the community can help to solve the testing issue. For example, the UK Office for National Statistics (ONS) conducted a national wide COVID-19 viral testing survey, namely Covid Infection Survey (CIS) [10] that successfully tracked the trajectories of COVID-19 infections in the community of UK since April of 2020. Because of its representative sampling across households in the general population this study is recognised to have a strong power to capture asymptomatic infections which might be missed out by symptomatic testing scheme in the early pandemic and can provide reliable estimates of prevalence over time [11]. However, this study started collecting samples from April of 2020 and then reporting the estimates of daily incidence from May of 2020 while the first death due to COVID-19 disease in the UK was documented in February 2020 [12]. This implies that the transmission of COVID-19 in the community began earlier than the survey, and the survey might not be able to recover the early epidemic curve.

Serologic studies that measure how many people have antibodies against the virus are a promising tool for pinning down the stage of the pandemic because of its ability of capturing past infections regardless of clinical symptoms [13]. If the antibody elicited by the virus lasts for lifetime, representative sampling in a population followed by the antibody testing will provide robust estimates of exposure. However, cohort studies following individuals over time after they’ve had a known COVID-19 infection were able to determine that antibodies are only measurable up to 6–9 months [14-16], on average, varying across testing assay [17] and antigen types [18]. The immediate implication is that serological studies will inevitably under-estimate the number of people exposed, since some will have a lower antibody count when the study is conducted and test negative. Linking multiple publicly available datasets, we proposed a method that have been published previously [1] to estimate the true level of exposure after considering the antibody decay. Here we further examined and evaluated the model in the context of reconstructing the first COVID-19 pandemic from three perspectives: validation from ONS Infection Survey, relationship between model performance and data abundance and time-varying case detection rate.

Result

Validation from ONS Infection Survey

Comparing the incidence of SARS-CoV-2 in England estimated by our model with those inferred by ONS Coronavirus Infection Survey (Figure 1), we found that our model could reveal the first but unobserved epidemic wave of COVID-19 in England from March 2020 to June 2020 additionally, with the second wave validated by the estimates from ONS Infection Survey. Further, we found our model results were highly consistent with those using SEIRS type compartmental models with time-varying force of infection [19, 20].

• Download figure
• Open in new tab

Figure 1.

Comparison of model predicted daily incidence of SARS-CoV-2 in England. The green lines show the predictions of median daily incidence by our model [1] based on Equation (1) and (2) in the Materials and Methods section. The orange lines show the predictions of median daily incidence from ONS Coronavirus Infection Survey while the orange shaded areas correspond to the 95% CrI.

Relationship between model performance and data abundance

We then examined the relationship between model performance and data abundance - how estimates of exposure from our model change with more serological data points being added into the fitting procedure one by one over time (Figure 2). We found a highly robust pattern of exposure across different regions of England was estimated in general.

• Download figure
• Open in new tab

Figure 2.

Comparison of estimates of exposure in seven regions of England as more serological measurements are given as model inputs (left to right). The green and orange lines show the model predictions of median exposure and seroprevalence, respectively, while the shaded areas correspond to the 95% CrI.

Specifically, the model could only start estimating the interested quantities: exposure and two parameters (infection fatality ratio and antibody decaying rate) when at least two serological measurements from April to June 2020 in each region were given as inputs. However, these estimates were already highly consistent with those when more serological measurements were added although the credible bands were wider. The wide credible bands suggested a bigger uncertainty around the estimates when little information was available. When three serological measurements in each of region were included the estimates of exposure level became largely consistent at the results when all serological measurements were used. This might be attributed to the timing of these third serological measurements since then the seroprevalence in most regions started decreasing. With more and more serological measurements being added, the credible bands of estimates of exposure gradually narrowed.

Time-varying case detection rate

While comparing the reported cases with the incidence estimated by our model (Figure 3), we found the UK government’s official COVID-9 online dashboard (https://coronavirus.data.gov.uk) reported COVID-19 cases in England only accounted for 9.1% (95%CrI (8.7%,9.8%)) of cumulative exposure by the end of October 2020. Further, the relative size of two infection waves in England in 2020 estimated by our model, Spring wave from February to June and Autumn wave from September to November, were reversed compared those reported by the confirmed cases. The case detection rate relative to the total exposure was also dramatically different in these two-epidemic waves. If separating the two waves from the first of August 2020, we found during January 2020 to August 2020 the case detection rate was only 4.3% (95%CrI (4.1%, 4.6%)) which increased to 43.7% (95%CrI (40.7%, 47.3%)) during August 2020 to October 2020, highlighting the dominate effect of testing effort in shaping the case curve in the early stage of a pandemic. The testing issue, e.g. the limited capacity of tests and symptom-based testing strategy posted a big challenge for understanding the early pandemic. Viral surveys in the general population can solve the sampling issue, but still have the problem of not sampling early on. Serological data even from some convenient samples, e.g., blood donors can help to pin down the progress of the pandemic when antibody decay is teased out.

• Download figure
• Open in new tab

Figure 3.

Comparison between estimates of daily incidence with reported cases of SARS-CoV-2 in England and case detection rate. Here, all serological measurements were used in the model fitting. In the top figure, the green lines show the predictions of median daily incidence by our model based on Equation (1) and (2) in the Materials and methods section while the shaded areas correspond to the 95% CrI. The red lines show the reported confirmed cases in England downloaded from GUV.UK dashboard. In the bottom figure, the blue lines show the estimates of median case reporting rate in England based on Equation (3) and (4) while the shared areas correspond to the 95% CrI.

Discussion

Accurate reconstruction of exposure time series is necessary to assess how policies influenced transmission over time, in particular when reporting is lagged, and multiple interventions may have been undertaken in succession. For example, [21] made use of the comparison of exposure between general population and pregnant women in New York City to conclude the effectiveness of shielding during pregnancy. Moreover, the prior exposure level in the population can be used to inform future intervention design, e.g., vaccination prioritisation. For example, in the early stage of the COVID-19 vaccination campaign, when dose supply and administrative capacity were initially limited worldwide, a modelling study [22] explored how uncertainty about previous exposure levels and about a vaccine’s characteristics affects the prioritization strategies for reducing deaths and transmission. This model showed use of individual-level serological tests to redirect doses to seronegative individuals improved the marginal impact of each dose while potentially reducing existing inequities in COVID-19 impact.

Here, we evaluated a simple dynamic model that we published previously and demonstrated its ability of reconstructing the first epidemic wave before large-scale survey sampling by providing robust estimates of exposure over time. One key element of the model was fitting model to serologic data that was generated from healthy adult blood donors supplied by the NHS Blood and Transplant (NHS BT collection) serum samples using the Euroimmun anti-spike IgG assay and reported in the Weekly national Influenza and COVID-19 surveillance report. This suggests that convenient samples, for example here serum samples from blood donors have the promising power to provide primary information of epidemic progress in a short timeframe especially during the emergency of a new outbreak from a novel pathogen.

Because of the rigorous sampling design and robust estimation power ONS Covid Infection Survey can almost be seen as a golden standard for estimating community prevalence. Our model does not take any results or estimates from the survey as inputs, so the comparison exercise that we conducted here between estimates of exposure from our model with ONS Covid Infection Survey provides a real-world validation. Moreover, we showed the modelling approach is a valuable early pandemic diagnostic tool and can clearly recover the first epidemic wave that the survey was unable to capture because of late starting time. Using the inferred daily incidence, we explicitly demonstrated the variation of case detection rates over two epidemic waves in England in 2020. It provides quantitative information for studying the association between the capacity, behaviour, strategy of testing with the epidemic evolution and further supported the argument that confirmed cases largely underestimate the extent of disease transmission.

Moreover, the simple structure of the presented model avoids unnecessary complexity and structure-based uncertainty in a full dynamic model where compartmental models simulating the disease spread in different groups of population including susceptible, expose, infected and recovered are developed. The exercise of studying the model performance against data abundance suggests the modelling results remain highly robust in data sparse setting that is particularly important, for example, in Low-or Middle-Income Country (LMIC).

Materials and methods

Method

We firstly calculated the incidence in England estimated by exposure model [1] by computing the difference of cumulative exposure in two successive days and adding together to the whole England as shown in Figure 1 and Figure 3: Here, E_i(t) is the daily exposure at region i estimated by exposure model [1], n is the total number of days from 1 January 2020 to 7 November 2020, i = 1, … 7 represents London, Southwest, Southeast, Northeast, Northwest, East, Midland. l_England(t) represents the daily incidence of England.

The 7-day average model predicted incidence can be calculated by Here, t = 4 refers to the fourth day of 2020, n is the end date of the comparison exercise, 7 November 2020.

The estimated reporting ratio as shown in Figure 3 was calculated by Here, C is the 7-day average reported cases in England from the UK government’s official COVID-9 online dashboard [12].

While testing the relationship between model performance and data abundance in Figure 2, we firstly obtained all the data and codes from paper [1] and rerun the model by adding the seroprevalence measurements one by one into the model.

Author contributions

L.J.W., S.C, J.A.F, and K.A.L conceived and designed the study. S.C. cleaned the data, S.C. and L.J.W. developed the methodology and conducted the formal analysis. S.C. and L.J.W. wrote the original manuscript. All authors reviewed and provided analytical input and approved the manuscript.

Competing interests

The authors have declared that no competing interests exist.

Data and materials availability

All codes and materials used in the analyses can be accessed at: https://github.com/SiyuChenOxf/Exposure_ONS-modelling. All parameter estimates and figures presented can be reproduced using the code provided. This work is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) license, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.