Prepared for the DELVE Initiative by Sylvia Richardson, Peter Diggle

Summary

Any TTI strategy that can only involve a fraction of the at-risk population must begin with a surveillance study to establish how incidence, and therefore potential disease transmission, varies between sub-populations and between individuals within sub-populations. This section describes how a surveillance study might be designed and analysed, taking account of patterns of heterogeneity of exposure and risk across the population, and potentially drawing information from multiple data-sources. Implementation of a surveillance system of this kind would require: 1) the design and regular (potentially weekly) implementation of a country-wide stratified quasi-random sampling scheme of antigenic diagnostic tests, to act as a “gold-standard”; 2) diagnostic testing of sampled individuals; 3) real-time analysis of the resulting data to produce stratum-specific predictive probability maps of incidence that can inform selection of individuals for TTI follow-up and adaptation of future sampling so as to maximise overall utility. Coordination with multiple sources of non-randomised self-reported symptom data on probable/possible case incidence would also be desirable, as this would enable gold-standard and self-reported sources of data to be analysed together to make best possible use of all relevant data-sources, thus creating an agile real-time surveillance system that can be exploited to increase the efficiency of the TTI process.

Citation

Sylvia Richardson, Peter Diggle (2020), Surveillance for Test, Trace, Isolate. DELVE Addendum TTI-TD2. Published 27 May 2020. Available from https://rs-delve.github.io/addenda/2020/05/27/surveillance-for-tti.html.

BibTeX
@Misc{addenda/2020/05/27/surveillance-for-tti, title = {Surveillance for Test, Trace, Isolate}, author = {Richardson, Sylvia and Diggle, Peter}, year = {2020}, publisher = {The Royal Society}, url = {https://rs-delve.github.io/addenda/2020/05/27/surveillance-for-tti.html}, abstract = {Any TTI strategy that can only involve a fraction of the at-risk population must begin with a surveillance study to establish how incidence, and therefore potential disease transmission, varies between sub-populations and between individuals within sub-populations. This section describes how a surveillance study might be designed and analysed, taking account of patterns of heterogeneity of exposure and risk across the population, and potentially drawing information from multiple data-sources. Implementation of a surveillance system of this kind would require: 1) the design and regular (potentially weekly) implementation of a country-wide stratified quasi-random sampling scheme of antigenic diagnostic tests, to act as a “gold-standard”; 2) diagnostic testing of sampled individuals; 3) real-time analysis of the resulting data to produce stratum-specific predictive probability maps of incidence that can inform selection of individuals for TTI follow-up and adaptation of future sampling so as to maximise overall utility. Coordination with multiple sources of non-randomised self-reported symptom data on probable/possible case incidence would also be desirable, as this would enable gold-standard and self-reported sources of data to be analysed together to make best possible use of all relevant data-sources, thus creating an agile real-time surveillance system that can be exploited to increase the efficiency of the TTI process.} }
Endnote
%0 Generic %T Surveillance for Test, Trace, Isolate %A Sylvia Richardson %A Peter Diggle %D 2020 %F addenda/2020/05/27/surveillance-for-tti %I The Royal Society %U https://rs-delve.github.io/addenda/2020/05/27/surveillance-for-tti.html %X Any TTI strategy that can only involve a fraction of the at-risk population must begin with a surveillance study to establish how incidence, and therefore potential disease transmission, varies between sub-populations and between individuals within sub-populations. This section describes how a surveillance study might be designed and analysed, taking account of patterns of heterogeneity of exposure and risk across the population, and potentially drawing information from multiple data-sources. Implementation of a surveillance system of this kind would require: 1) the design and regular (potentially weekly) implementation of a country-wide stratified quasi-random sampling scheme of antigenic diagnostic tests, to act as a “gold-standard”; 2) diagnostic testing of sampled individuals; 3) real-time analysis of the resulting data to produce stratum-specific predictive probability maps of incidence that can inform selection of individuals for TTI follow-up and adaptation of future sampling so as to maximise overall utility. Coordination with multiple sources of non-randomised self-reported symptom data on probable/possible case incidence would also be desirable, as this would enable gold-standard and self-reported sources of data to be analysed together to make best possible use of all relevant data-sources, thus creating an agile real-time surveillance system that can be exploited to increase the efficiency of the TTI process.
RIS
TY - GEN TI - Surveillance for Test, Trace, Isolate AU - Sylvia Richardson AU - Peter Diggle DA - 2020/05/27 ID - addenda/2020/05/27/surveillance-for-tti PB - The Royal Society UR - https://rs-delve.github.io/addenda/2020/05/27/surveillance-for-tti.html AB - Any TTI strategy that can only involve a fraction of the at-risk population must begin with a surveillance study to establish how incidence, and therefore potential disease transmission, varies between sub-populations and between individuals within sub-populations. This section describes how a surveillance study might be designed and analysed, taking account of patterns of heterogeneity of exposure and risk across the population, and potentially drawing information from multiple data-sources. Implementation of a surveillance system of this kind would require: 1) the design and regular (potentially weekly) implementation of a country-wide stratified quasi-random sampling scheme of antigenic diagnostic tests, to act as a “gold-standard”; 2) diagnostic testing of sampled individuals; 3) real-time analysis of the resulting data to produce stratum-specific predictive probability maps of incidence that can inform selection of individuals for TTI follow-up and adaptation of future sampling so as to maximise overall utility. Coordination with multiple sources of non-randomised self-reported symptom data on probable/possible case incidence would also be desirable, as this would enable gold-standard and self-reported sources of data to be analysed together to make best possible use of all relevant data-sources, thus creating an agile real-time surveillance system that can be exploited to increase the efficiency of the TTI process. ER -
APA
Richardson, S. & Diggle, P.. (2020). Surveillance for Test, Trace, Isolate. Available from https://rs-delve.github.io/addenda/2020/05/27/surveillance-for-tti.html.

What sort of epidemic surveillance is needed, and how it can aid TTI strategies?

Introduction

In the absence of unlimited resources, the necessary first step in a TTI strategy for COVID-19 is a system of real-time surveillance to provide an understanding of the evolving pattern of population-wide variation in disease incidence. Such a system can then inform a prioritisation of particular individuals or groups for TTI follow-up, by helping to identify individuals or groups of individuals who are most likely to transmit infection.

This task is particularly challenging when, as appears to be the case for COVID-19, a substantial (but unknown) proportion of cases are asymptomatic or pre-symptomatic but may nevertheless be capable of infecting others.

The remainder of this technical document (TD2) sets out a set of principles that could guide the design of a COVID-19 surveillance system and the analysis of the data that it provides. It then lists a range of potential sources of data and describes strategies for study-design and data-analysis, before discussing the subsequent progression from surveillance to TTI.

Principles of surveillance

In considering the design of a surveillance system, it will be important to take into account the following. These principles apply equally to surveillance systems aimed at monitoring the spatio-temporal evolution of cumulative prevalence determined, for example, by antibody tests, or of current incidence, determined by antigen tests.

  1. Set specific objective(s) For example, the objective may be to predict the pattern of variation in the current incidence of infectives for a COVID-19 across the UK population (sometimes called now-casting - using sampled data to understand the current state of a process that cannot be completely observed.

  2. Agree a minimal data-set In any resource-limited setting a balance needs to be struck between how much information is requested from each sampled individual and from how many compliant individuals data can be obtained. Individual-level information is useful for understanding individual risk-factors, but only information that is available country-wide is useful for predicting geographical variation in prevalence. For COVID-19, potentially useful covariates available from the most recent national census include LSOA-level deprivation scores and demographic summaries.

  3. Use a consistent definition of outcome. This requires either a single diagnostic to be used for case-ascertainment or (see below) a way of calibrating across different diagnostics.

  4. Identify major sources of heterogeneity in risk. In some sub-populations, notably care homes and hospitals, the patterns of exposure and risk are so different from those of the population at large that they are better treated as separate populations, each with their own surveillance system. Similarly, health workers need separate consideration. For the remainder of the population, it will be advantageous, both operationally and for statistical efficiency, to stratify by known sources of heterogeneity, for example by population density, deprivation and geographical region. Known individual level sources of heterogeneity, for which adjustment should be made at the analysis stage, include age, sex and ethnicity. Any residual variation in incidence can be considered as a proxy for all unknown, spatially or temporally structured sources of heterogeneity, allowance for which should also be made at the analysis stage.

  5. Sample as-if-at-random. Only a randomised sampling framework can guarantee unbiased predictions of incidence. A stratified random sample of the UK population, with strata defined by the major sub-population risk-groups and adjustment at analysis stage for individual-level risk-factors may, however, be an unattainable ideal. Whilst stratification does not eliminate the potential for bias in a non-randomised sampling scheme, it should reduce it. It may therefore be possible to construct a context-specific sampling instrument that can be regarded as “as-if-random’’ after stratification and adjustment.

  6. Choose spatial and temporal scales for analysis and for reporting. In principle, incidence is likely to vary continuously in time and space. For reporting purposes, it may be helpful to aggregate results to a temporal (weekly?) and spatial (regional?) scale on which operational decisions can be made. To avoid aggregation bias, surveillance data should be recorded and analysed at the finest substantively relevant temporal and spatial resolutions. In the current context, the finest relevant temporal scale could be daily or weekly, depending on the diagnostic used; for example, self-reported symptoms may be susceptible to weekday/weekend artefacts. The finest relevant spatial scale is likely to be a Lower Layer Super Output Area (LSOA); note that several of the known sources of heterogeneity (ethnicity, deprivation) can vary substantially between adjacent LSOA’s. Resolution to full post-code is technically possible but arguably too fine as, even in lock-down, a substantial proportion of the population will regularly move beyond their home post-code.

  7. Measure and report degree of uncertainty in predictions. A conventional measure of statistical precision is the standard error of an estimate. In disease surveillance, arguably, a more relevant measure is a predictive probability, i.e. the probability, given the observed data, that the underlying process is in a specified state; for example, that the incidence of COVID-19 amongst 70-year-old white males at a particular location is at least 10%.

  8. Take account of major public health interventions. In any statistical model, unexplained variation in the outcome is ascribed to stochastic variation, and the bigger the stochastic variation the less precise are the associated model-based predictions. Interventions that are likely to have a major effect on incidence should be included in the statistical model as change-points.

None of the above principles are specific to COVID-19, nor the statistical methods needed to analyse to resulting data. We suggest that the COVID-19 epidemic could be the stimulus for creation of a national, real-time spatial surveillance system for emerging infectious diseases that would be adaptable to any future public health concerns

Data sources

We make a distinction, admittedly not always sharp, amongst three broad types of data.

a. Designed studies will include measurement of a recommended diagnostic test of current COVID-19 infection with good sensitivity and specificity1, such as an RT-PCR swab-based test for COVID-19 infection. Two examples of such studies are the Imperial College REACT study2 and the Oxford-ONS surveillance study3.

b. Routinely recorded health outcome data potentially useful for surveillance can come from a variety of sources, including: the zoe app4, calls to NHS Direct; indicators monitored by PHE as part of its network of surveillance systems, such as FluSurvey, GP In Hours (GPIH) syndromic surveillance system, GP Out of Hours (GPOH) syndromic surveillance system, Emergency Department Syndromic Surveillance System (EDSSS)). Symptom-based self-reported indicators typically have lower sensitivity and specificity than diagnostic tests and may suffer from biases.

c. Other types of data, such as internet searches for specific symptoms, spikes in consumption of non-prescription medicines or sickness absences could potentially also be brought into the surveillance system to expand its detection of unusual events.

Data sources of types (b) and, more so, of type (c) should be used cautiously. At a minimum, they require careful calibration against high-quality data of type (a). Also, adding complexity to a predictive model risks losing, rather than gaining, precision if the signal content of the added data source is weak, and only weakly associated with the primary outcome of interest.

Design

Here, we assume that a method of locating an as-if-random sample of individuals has been agreed, the choice of diagnostic has been made and its performance characteristics are well understood. The remaining design considerations are: the selection of strata based on combinations of sub-population characteristics; the frequency of sampling; the individual-level characteristics to be recorded on sampled individuals; the sample size(s) to be taken in each stratum. With respect to the last of these, we need either to set a performance target and derive a set of sample sizes that will achieve this, or to set an achievable limit on total sample size, optimise their disposition across strata and evaluate the performance of the resulting surveillance system. Understanding the limitations of a range of affordable designs is more useful than setting a single performance target that is unattainable. Examples of suitable measures of performance might be: the maximum width of a 95% predictive interval for LSOA-level incidence within any stratum; or the ROC curve for predicting exceedance of a specified LSOA-level incidence.

As with any sample size calculation, an initial sampling design can only be constructed by assuming a specific statistical model for the underlying spatio-temporal incidence surface. However, data accruing in the early operation of the system can be used to assess the goodness-of-fit of modelling assumptions, estimate model parameters, and to adapt the sampling design accordingly. This will be particularly the case if resource constraints severely limit the proportion of probable or confirmed cases that can be followed up.

Analysis

In a surveillance system of the kind described here, taking account of spatial and/or temporal correlation in the underlying incidence process can materially improve predictive precision, sometimes by an order of magnitude5. Combining a relatively small sample of data from a designed study based on a recommended diagnostic test with a much larger sample of routinely collected data, typically based on self-reports, can also be a cost-effective way of making the best use of all available data.6 7

A suitable class of statistical models for problems of this kind is a generalized linear mixed model8 with a latent spatio-temporal Gaussian process included in the linear predictor9. The appendix gives an example of a model of this kind for the joint analysis of designed and routinely collected data.

The end product of the surveillance system is the availability of a set of stratum-specific space-time surfaces, adjustable for individual-level risk-factors, that quantify the variation in disease incidence, with associated measures of uncertainty.

In incidence surveillance system can provide a useful first stage in a TTI system.  TTI will be most cost-effective when the individuals at the start of a TTI chain are those most likely to be infective. Also, if surveillance and TTI use the same diagnostic, TTI-generated data can be fed back into the surveillance system to increases its local precision. Here we discuss two of the decision points in the TTI process where knowledge of the spatio-temporal incidence of disease that is provided by a surveillance system plays a role in informing the chosen course of action: the decision to test or not an individual who has either reported symptoms or contacted the primary health care system; and the decision to trace or not the contacts of this individual and when to initiate this.

Known routes of transmission of COVID-19 include transmission from asymptomatic, pre-symptomatic and symptomatic individuals through exhaled droplets, and environmental transmission through contaminated surfaces. Here we focus only on person-to-person transmission.

The surveillance system described thus far estimates the probability that an individual within any stratum is test positive (or is symptom positive if the data source is related to symptoms). We also need an understanding of the probability that an individual will have transmitted infection given that they are test positive, which will likely depend on the individual’s stratum membership and on their measured personal attributes.

In a recent paper, Ferretti et al (2020)10 quantify the relative contribution of different routes of transmission to the reproductive number, and estimate that the main transmission routes are from pre-symptomatic and symptomatic individuals, which altogether account for a large fraction of transmissions. This paper also shows that pre-symptomatic transmission is nearly sufficient for driving the epidemic, accounting for a proportion 0.9 of the total contribution to an R0 equal to 2.11 Both the incubation time and the generation time between source and transmission are estimated to be around 5 days. This shows: (i) that the primary focus of TTI is on “capturing” both pre-symptomatic and symptomatic individuals; (ii) that there is a time constraint on the effectiveness of the TTI process, which might require contact tracing to begin before confirmation through a diagnostic test.

Symptomatic individuals can be best captured rapidly through a range of symptom-related indicators (see details below), while pre-symptomatic and asymptomatic individuals might be captured only through the contact tracing process. Such individuals might also be captured through the designed surveillance system, but unless the designed testing strategy enrolled a substantial fraction of the whole population, focussing contact tracing on these would miss a large fraction of the infectives.

As is the case for surveillance, the definition of a minimum dataset needs care. However, as TTI will involve smaller numbers of individuals who are also more likely to be cases, the argument for recording individual characteristics is stronger. These could include age, sex, ethnicity and occupation. It is important to note that this approach considers only the general population, excluding special categories such as care-home residents, hospital in-patients and health workers, and assumes that resource constraints dictate a prioritisation of who should be tested as the first stage of TTI.

The testing strategy and the prioritisation of whom to test

Pragmatically, we want to maximise the capture probability, i.e. the probability for an individual of being found positive in the RT-PCR swab test if selected to be tested: Prob(RT-PCR+ | selected for testing). The steps to maximize the capture probability could be as follows:

  • Derive and agree with PHE and PH experts a list of symptom indicators, which encompass the most common symptoms of COVID-19 and are adapted to each of the routine data resources (111 calls, GPIH, GPOH, Zoe appl and NHSx app). Specific symptoms to COVID-19 such as loss of taste or smell are particularly useful as they will help to discriminate from influenza like symptoms;

  • For each of these sources, run a pilot study on a suitably stratified sample of people for whom both the set of symptoms indicators and the PCR test are measured.

  • Calibrate a predictive model of Prob (RT-PCR+ | indicators) using the pilot data for each data source. For example, Menni et al (2020)12 used the symptoms reported in the zoe app to derive a symptom prediction model using stepwise logistic regression. This gave a sensitivity of 0.65 [0.62; 0.67], a specificity of 0.78 [0.76; 0.80] and a ROC-AUC of 0.77 [0.72; 0.82]. The strongest predictor was loss of smell and taste. The overall Positive Predictive value was estimated at 0.69 [0.66; 0.71]. The sample size where both symptoms and PCR test were measured in the UK was 15,638, with 6,452 PCR+ tests. The authors did not find notable differences by age and sex.

  • Improve the positive predictive value of the symptom-based predictive model by adding external information from: (i) surveillance studies that predict the stratum-specific incidence surfaces of COVID-19; (ii) primary care EHR records of co-morbidities and other risk factors of sampled individuals.

  • Derive the final predictive probability of infection for each person that enters the system through any one of the data sources. Individuals with high predictive probability of infection (e.g. a person having specific symptoms in an area where current incidence is on the rise) are more likely to be COVID-19 positive than an individual with low predictive probability.

  • Embed the predictive probability into the first step of the TTI decision process so as to optimise allocation of testing resources to individuals most likely to be infected. The efficacy of different rules could be investigated through simulations.

The decision and timing of contact tracing

A predicted stratum-specific incidence surface provides a basis for deciding where and when to prioritise contact-tracing follow-up, with the ultimate goal of prioritising for follow-up those individuals most likely to have transmitted infection. This can be combined with the predictive probability of infection for the source individual to decide whether to immediately start tracing their contacts or to await confirmation through a positive test result. If the predictive probability of the source individual is high, it might be recommended that contact tracing is initiated immediately, since about 50% of transmission is in the pre-symptomatic period. Waiting for the result of the test would mean that the transmission chain is less likely to be broken, but this needs to be balanced against the opportunity-cost of contact-tracing individuals who subsequently test negative. Of course, if and when test results become available more quickly, this could be revisited. Timeliness of testing and manual contact tracing is also a key factor, as shown in the simulation scenarios considered in section

  1. See also Ferretti et al (2020)13, who discuss a quasi-instantaneous contact tracing scheme based on an app.

Discussion

Real time surveillance of incidence through surveys and TTI are two complementary components of public health action in the face of an emerging threat from an infectious disease. We have discussed how a designed surveillance study can be combined with self-reported symptom-based data to create an agile surveillance system that can track the spatial evolution of the disease in real time. If several self-reported symptom data sources are offered concurrently to the population, like zoe app and NHSx app, we recommend that they all contain an agreed standardised minimum set of information on the symptoms recorded, very much along the lines of the core outcome set which, by international agreement, is recorded in all clinical trials (COMET). This will allow information from diverse technologies to be meaningfully synthesised.

To be able to calibrate the link between designed diagnostic testing and self-reported symptom data, it is necessary to have records of both types of data on a subset of individuals. To prevent selection biases, it would be highly desirable that at least part of this subset is randomly chosen, rather than purely observational.

Steps in the TTI process can be informed by an agile, real-time surveillance system so as to prioritise testing for the individuals most likely to be infected, and to target the contact tracing accordingly. The algorithm used for this prioritisation can be refined and improved as the epidemic progresses, and more specific clinical characteristics of infected individuals are discovered. Finally, the TTI process can also feedback into the design of specific localised surveillance studies around hotspots.

Appendix 1 focuses on real time surveillance through a combination of designed testing studies and routine capture of self-reported symptoms or calls to the health system. As this framework develops, considering additional sources of data to detect unusual activity or behaviour might be beneficial to increase the agility of the surveillance system.

Appendix

An example of a surveillance system combining designed and routinely collected data to reconstruct the underlying space-time incidence data can be found in this docment.

  1. The sensitivity of a test is the proportion of true positives, the specificity is the proportion of true negatives. For RT-PCR based swab tests which are self-administered, a diagnostic sensitivity of 0.65 and specificity of 100% is used in the Oxford REACT study. ↩

  2. Laurance, J. and Alford, J. (2020) Home testing for coronavirus to track levels of infection in the community, Imperial College London News, 30 April 2020, available at: https://www.imperial.ac.uk/news/197217/home-testing-coronavirus-track-levels-infection/ ↩

  3. NIHR Oxford Biomedical Research Centre (2020) Large-scale COVID-19 infection and antibody test study launched, 23 April 2019. Available at: https://oxfordbrc.nihr.ac.uk/large-scale-covid-19-infection-and-antibody-test-study-launched/ ↩

  4. See: https://covid.joinzoe.com/ ↩

  5. Fronterre, C., Amoah, B., Giorgi, E., Stanton, M.C. and Diggle, P.J. (2020). Design and analysis of elimination surveys for neglected tropical diseases. Journal of Infectious Diseases (doi: 10.1093/infdis/jiz554) ↩

  6. Giorgi, E., Sesay, S.S., Terlouw, D.J. and Diggle, P.J. (2015). Combining data from multiple spatially referenced prevalence surveys using generalized linear geostatistical models. Journal of the Royal Statistical Society A 178, 445-464 (doi: 10.1111/rssa.12069) ↩

  7. Amoah, B., Giorgi, E. and Diggle, P.J. (2020). A geostatistical framework for combining spatially referenced disease prevalence data from multiple diagnostics. Biometrics, 76, 158-170(doi: 10.1111/biom.13142) ↩

  8. Breslow, N.E. and Clayton, D.G. (1993). Approximate inference in generalized linear mixed models. Journal of the American Statistical Association 88, 9-25.(doi: 10.2307/2290687) ↩

  9. Diggle, P.J., Rowlingson, B. and Su, T-L. (2005). Point process methodology for on-line spatio-temporal disease surveillance. Environmetrics, 16, 423-434. (doi: 10.1002/env.712) ↩

  10. Ferretti (n10) ↩

  11. Ferretti (n10), Figure 2 ↩

  12. Menni C et al (2020). Real-time tracking of self-reported symptoms to predict potential COVID-19. Nature Medicine (doi: 10.1038/s41591-020-0916-2) ↩

  13. Ferretti (n10) ↩