Background
The use of multiple outbreak control measures in the early phases of the COVID-19 pandemic was resource intensive and disruptive, but essential to minimise the loss of lives [
1,
2]. Measures such as case finding at the borders and healthcare touchpoints allow health authorities to assess the extent of disease importation and undetected spread in the community. Furthermore, contact tracing around notified cases can identify potential transmission routes and hence new cases [
3,
4]. When multiple control measures are implemented together, understanding the effectiveness of each measure enables public health authorities to focus on the most effective measures when resources are limited and to minimise interruption to economic and social activities. Studies typically evaluate the collective effectiveness of country or region-specific COVID-19 outbreak control measures by measuring changes to the reproduction number using overall observed case incidence [
1,
5‐
12] or only focus on the impact of specific interventions using outbreak data [
13,
14]. If analysis could disentangle the observed and unobserved transmission dynamics, it would therefore be possible to obtain higher resolution insights on the effects of each outbreak control measure.
Transmission chains from outbreak clusters have been used to characterise the reproduction number of infectious diseases other than COVID-19 and the relative contribution of different transmission routes (e.g. imported or environmental introduction vs community) to the overall spread [
15‐
18]. However, these studies typically do not account for the role of missed infections (e.g. asymptomatic or mildly symptomatic infections) in influencing the effectiveness of outbreak control measures. To our knowledge, the use of data on these transmission linkages to estimate the burden of infection for SARS-CoV-2 at the population level has yet to be documented. The extent of missed infections in the COVID-19 pandemic was commonly assessed via population-wide seroprevalence surveys [
19,
20], excess mortality studies [
21], random community testing [
22] or behavioural surveys [
23,
24]. However, during the initial phases of an outbreak of a novel pathogen, serological assays to measure the disease prevalence are generally not available. Moreover, these methods do not provide assessment on the extent of missed cases at the borders. Thus, methods to address these challenges and provide a more complete view of the outbreak are necessary.
With a population of 5.7 million inhabitants, Singapore was one of the first countries to report SARS-CoV-2 infections outside of mainland China at the beginning of the COVID-19 pandemic. The Ministry of Health monitored the daily incidence of imported, and linked and unlinked local COVID-19 cases and collected extensive information on the epidemiological events associated with each case (e.g. time of arrival, symptoms onset, notification, isolation or quarantine). In this study, we reconstructed the pandemic trajectory in Singapore and estimated the effectiveness of various outbreak control measures (Table
1) by combining the observed COVID-19 cases with a mathematical model. As countries redesign surveillance systems for future pandemics, this modelling framework has the potential to inform how the collection of different data fields can shape our understanding of disease transmission in the early phases of a pandemic.
Table 1
Outbreak control measures in Singapore. Observed case data were used to estimate the effectiveness of each measure. Cases are defined as infected individuals that tested positive and are notified, while infections include all notified and missed infected individuals
Border control (Minimise disease introduction into community)
| ■ Limiting the number of incoming travellers from countries with ongoing outbreaks ■ Quarantine or restricting movement of incoming travellers | ◆ Number of missed imported infections |
Case finding (Targeted testing at known or potential source(s) of infection)
| ■ Testing of symptomatic travellers upon arrival or when they developed symptoms during quarantine ■ Testing regime for non-symptomatic travellers | ● Imported case data |
■ Testing of suspect cases (e.g. persons with clinical signs and symptoms suggestive of pneumonia or severe respiratory infection, persons with acute respiratory infection and travel history to regions with ongoing outbreak) ■ Routine testing of high-risk populations (e.g. healthcare workers, nursing home residents) ■ Ad-hoc testing during cluster outbreak investigations | ● Local unlinked case data ◆ Effectiveness of case finding |
Contact tracing (Targeted testing at potential routes of infection)
| ■ Interviewing COVID-19 cases or use of Bluetooth contact tracing devices to identify close contacts ■ Testing of symptomatic contacts ■ Testing of contacts before the end of their quarantine | ● Local linked case data ◆ Effectiveness of contact tracing |
Use of other non-pharmaceutical interventions and vaccines (Untargeted community- or population-level preventive measures)
| ■ Physical distancing ■ School and venue closure ■ Large-scale population movement restrictions and the corresponding need to work-from-home ■ Population-wide face mask usage ■ Pre-event testing/vaccination ■ Accelerated development and roll-out of COVID-19 vaccines (primary doses and boosters) with priority given to frontline workers and the elderly before progressively offered to younger age groups | ◆ Average number of secondary cases generated by a single infectious individual over the course of the entire infectious period (i.e. R) |
Discussion
Using the growth patterns in the daily incidence of local linked and unlinked cases, and imported cases with community contact, we reconstructed the incidence of missed infections over time in Singapore. This enabled us to disentangle the effects of targeted measures such as case finding and contact tracing from other population-wide outbreak interventions. Our modelling framework was able to infer these missed infections without requiring large-scale serological surveys, which are typically challenging to conduct at the start of a pandemic. Such analysis can therefore provide early insights into the effectiveness of respective categories of outbreak control measures, and hence further inform the extent of measures required during different stages of an outbreak.
The changes in the estimated effectiveness of control measures largely coincide with the shifts in outbreak control policies, but there were other likely contributing factors. Changes in human behaviours such as a reduction in health-seeking behaviour coincided with a decline in the effectiveness of case finding,
ϵcf, from 42% in Mar 1 to Apr 6, 2020, to 10% during the lockdown from Apr 7 to Jun 18, 2020 [
23]. Furthermore, the interdependence of outbreak control measures can cause the effectiveness of measures to change in tandem. In particular, the contribution of contact tracing towards outbreak control hinges on the extent of case finding. Following the decline in
ϵcf during the lockdown, the effectiveness of contact tracing in identifying new cases declined from 78% in Mar 1 to Apr 6, 2020, to 63% during the lockdown. This observation is also supported by theory—when the effectiveness of isolating cases is low, a slight decrease in the effectiveness of contact tracing can result in a growing outbreak [
30]. Collectively, about 75% of the infections were estimated to be missed during the lockdown and this proportion was higher than other time periods due to the lowered effectiveness in both case finding and contact tracing. Thus, by identifying which outbreak control measures were contributing to the growth of an outbreak and the corresponding reasons for its lowered effectiveness, it is possible to address relevant aspects of human behaviour (e.g. promote use of telemedicine as patients feel more comfortable seeing their doctors online [
56]; discourage clinic hopping so the same doctor can better assess the need for follow up test [
57]).
In both wild-type SARS-CoV-2 and Delta variant outbreaks in Singapore, on average, there was less than 1 death per day. With prolonged periods of low death counts, we reconstructed the underlying outbreak dynamics using the incidence of linked and unlinked cases instead of using reported fatalities [
21,
58]. Prior to 2021, the Singapore population was largely unvaccinated and during the Delta variant outbreak about 60% of the population was vaccinated by Aug 2021. Our CFR estimates were less than 1% for the wild-type SARS-CoV-2 and Delta variant outbreak, which was less than the early CFR estimates of around 1.4% for wild-type SARS-CoV-2 and 3 times lower than the CFR estimates for the Delta variant in other studies [
59,
60]. The IFR estimates for both outbreaks in Singapore were also less than 0.5%, and in the lower range of IFR estimates as compared to other countries and regions [
58,
61,
62]. While the healthcare system was stretched in both outbreaks, ICU capacity was not exceeded and this helped to keep the number of deaths to a minimum. As deaths observed in small outbreak clusters would not be reflective of the number of deaths that could arise during a large epidemic wave, care is needed in the interpretation of underlying infection dynamics and how these influence measured disease outcomes.
We found that metrics derived from observed data alone do not always accurately reflect the underlying outbreak. Specifically, metrics such as the proportion of unlinked cases among all notified cases are not necessarily representative of the proportion of missed infections among all infections, and policy makers should therefore be careful when drawing conclusions of the latter from the former. This discrepancy is likely to occur because the missed infections have a much higher reproduction number as compared to notified cases, or when a single missed infection is the source of infection for multiple unlinked cases and the outbreak could be misinterpreted as growing or declining slowly in either scenario. In contrast, contact tracing data provides additional information on the source of infection of a case. The collection of such data expends minimal effort yet can help to improve our understanding of the underlying outbreak although misclassification could also affect the interpretation of the outbreak dynamics. Thus, the interpretation of common metrics should be done with a clear understanding of the data collection process. Previous studies have estimated the impact of measures such as border control by assessing correlations between the timing of interventions and national-level case incidence [
63], but our results suggest such analysis will not capture the complexity of interacting measures against a background of changing infection detection.
We also found that multiple independent notification datasets and informative priors helped to disentangle the model parameters and achieve more precision in estimates. Unlinked cases were generated by either missed imported or local infections with the former modelled as a factor of the notified imported cases, ρ. As such, the interaction of model parameters results in wide 95% credible intervals for ϵcf estimates. To improve these estimates, we could further stratify exposure histories of unlinked cases by their interactions with travellers from countries with ongoing outbreak for model fitting. Informative and uniform priors produced a similar set of parameter estimates when there were multiple independent notification data in the SARS-CoV-2 wild-type outbreak in 2020 for model fitting. However, the model output using a uniform prior was different from that of an informative prior for the Delta variant outbreak in Jul to Aug 2021. Unlike the wild-type SARS-CoV-2 outbreak, model fitting for the Delta variant was based on the time series of cases without accounting for the case linkage. As such, there was limited data to inform the extent of underreporting and hence the number of missed infections. The estimates of R from Jul 1 to 17, 2021 when using the informative prior falls within the lower range of the estimates derived from the uniform prior although both analyses suggest a growing outbreak.
There are some additional limitations to our study. One is that asymptomatic cases were assumed to have a similar distribution of delay from the time of infection to notification as symptomatic cases. To circumvent this, we can study the changes in the trajectory of the cycle threshold values (proxy for viral load) of cases that were tested multiple times over the course of the infection. The infection time of symptomatic and asymptomatic cases can be estimated from their respective viral growth trajectory [
64,
65] thereby informing the delay distribution for respective types of cases. Furthermore, we assumed that asymptomatic cases were as infectious as symptomatic individuals, and hence, no stratification of
R was modelled as there is no strong evidence to suggest that asymptomatic SARS-CoV-2 infections are less infectious than symptomatic individuals [
66,
67]. Our modelled outcomes for wild-type SARS-CoV-2 transmission were able to reproduce independent observations in a separate population-level serological survey and this lends support to our assumption of a homogeneous
R among most missed infections.
In addition, the burden of disease and infection estimates were averaged across all age groups, as there was insufficient data to estimate the transmissibility and susceptibility across different age groups in each time period. In our branching process model, we also assumed that each of the four parameters remains constant in a specified time period. As such, we are unable to provide a time-varying measure to characterise the impact of different outbreak detection and control measures that were progressively rolled out in the population at a granular level. Instead, time periods were chosen based on prior knowledge of major policies that would affect at least one of the four model parameters. In particular, from Jul 1 to 17, 2021, the outbreak of COVID-19 cases from a nightclub cluster and fishery port resulted in a reproduction number of more than 1. For cases at the end of this time period, the model assumes that their R is the same as the cases at the start of the same time period. However, as rapid and strict outbreak control measures were implemented around the period of Jul 18, 2021, the R of the cases around this transition period is expected to vary between the reproduction number estimated for Jul 1 to 17, 2021 and Jul 18 to Aug 18, 2021. With the potential for a larger reproduction number using a uniform prior, the exponential number of new infections generated by cases around the transition period causes the modelled peak outbreak to overshoot the observed peak in the subsequent time period. This further highlights the importance of having multiple independent data on case linkage to better inform the parameter estimates and to infer missed infections.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.