key: cord-0884717-7f9c4j5m
authors: Davis, E. L.; Lucas, T. C. D.; Borlase, A.; Pollington, T. M.; Abbott, S.; Ayabina, D.; Crellen, T.; Hellewell, J.; Pi, L.; Medley, G. F.; Hollingsworth, T. D.; Klepac, P.
title: An imperfect tool: COVID-19 'test & trace' success relies on minimising the impact of false negatives and continuation of physical distancing.
date: 2020-06-12
journal: nan
DOI: 10.1101/2020.06.09.20124008
sha: f7fee0f9a588d768daf6cf9a08e73eed0062d21f
doc_id: 884717
cord_uid: 7f9c4j5m

Background: Following a consistent decline in COVID-19-related deaths in the UK throughout May 2020, it is recognised that contact tracing will be vital to relaxing physical distancing measures. The increasingly evident role of asymptomatic and pre-symptomatic transmission means testing is central to control, but test sensitivity estimates are as low as 65%. Methods: We extend an existing UK-focused branching process model for contact tracing, adding diagnostic testing and refining parameter estimates to demonstrate the impact of poor test sensitivity and suggest mitigation methods. We also investigate the role of super-spreading events, providing estimates of the relationship between infections, cases detected and hospitalisations, and consider how tracing coverage and speed affects outbreak risk. Findings: Incorporating poor sensitivity testing into tracing protocols could reduce efficacy, due to false negative results impacting isolation duration. However, a 7-day isolation period for all negative-testing individuals could mitigate this effect. Similarly, reducing delays to testing following exposure has a negligible impact on the risk of future outbreaks, but could undermine control if negative-testing individuals immediately cease isolating. Even 100% tracing of contacts will miss cases, which could prompt large localised outbreaks if physical distancing measures are relaxed prematurely. Interpretation: It is imperative that test results are interpreted with caution due to high false-negative rates and that contact tracing is used in combination with physical distancing measures. If the risks associated with imperfect test sensitivity are mitigated, we find that contact tracing can facilitate control when the reproduction number with physical distancing, Rs, is less than 15.

isations, and consider how tracing coverage and speed affects outbreak risk. Findings: Incorporating poor sensitivity testing into tracing protocols could reduce efficacy, due to false negative results impacting isolation duration. However, a 7-day isolation period for all negative-testing individuals could mitigate this effect. Similarly, reducing delays to testing following exposure has a negligible impact on the risk of future outbreaks, but could undermine control if negative-testing individuals immediately cease isolating. Even 100% tracing of contacts will miss cases, which could prompt large localised outbreaks if physical distancing measures are relaxed prematurely. Interpretation: It is imperative that test results are interpreted with caution due to high false-negative rates and that contact tracing is used in combination with physical distancing measures. If the risks associated with imperfect test sensitivity are mitigated, we find that contact tracing can facilitate control when the reproduction number with physical distancing, R S , is less than 1·5.

Keywords: COVID-19, contact tracing, branching processes, SARS-CoV-2, testing strategy, case isolation, quarantine 1. Background 1 In December 2019, SARS-CoV-2, a novel coronavirus strain, was de-2 tected in Hubei Province, China. 1 By 31st January 2020 the first UK cases 3 of COVID-19, the disease caused by the SARS-CoV-2, were confirmed. 2 4 Initial modelling studies indicated that fast and effective contact tracing May. 10 Currently, testing of asymptomatic individuals is limited to workers and patients in NHS and social care facilities, 11 but from the 28th of May 20 the UK Government rolled out the initial stages of their 'test & trace' con-21 tact tracing programme to the general population. This new approach was 22 initiated with contact tracing of just over 2,000 confirmed cases. Crucially, 23 the current strategy only tests symptomatic contacts and notifies individuals 24 that they no longer need to isolate following a negative test. However, there 25 are critical limitations to the diagnostic test, with poor sensitivity (current 26 estimates imply close to 65% 12,13 ), especially in community-based settings, 27 leading to high false negative rates which are exacerbated by high variability 28 in symptom severity. 13 Infectious individuals who test falsely negative may 29 prematurely resume their normal activities, contributing to ongoing chains 30 of transmission.

Imperfect adherence and the innate difficulties in identifying contacts will 32 pose challenges for 'test & trace', particularly in crowded urban settings. 14 33 Therefore, evaluating both the limitations of contact tracing and how to 34 maximise its effectiveness could be crucial in preventing a second peak in initial study was that highly effective contact tracing would be sufficient to 42 control an initial outbreak of COVID-19 in the UK, however substantial new 43 evidence supports much higher pre-and asymptomatic transmission rates 44 than had initially been considered. 16, 17, 18 The focus on rapid testing in the 45 UK contact tracing programme also requires a detailed assessment of the 46 associated trade-offs through mechanistic modelling of the testing process.

Up-to-date modelling studies are needed to investigate the feasibility of con-48 tact tracing and the conditions under which it is effective. 49 We use improved incubation period and serial interval estimates, 19,20 im-50 perfect self-reporting and tracing rates, as well as simulating the use of diag-51 nostic tests both for detection and tracing of asymptomatic infection chains. 52 We also simulate decision-making regarding quarantine procedures for traced 53 individuals, and then explore the trade-offs introduced by poor test sensitiv-54 ity, particularly when negative test results are used to advise individuals to 55 cease self-isolation. 56 3 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 

In this extension of a previous COVID-19 branching process model, 3 the 58 number of potential secondary cases generated by an index case and the CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 12, 2020. period is fixed at 40%. 16, 17 The full infection profile is shown in Figure 2 . where an individual self-reports with a probability of 10% or 50%. The con-95 7 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 tacts of that individual are then traced with 40%-100% coverage. If a contact 96 is successfully traced they will always isolate. The time taken to trace and 97 isolate a contact is either one day or drawn from a Uniform distribution of 1- This is simulated by reducing the case reporting proportion to 0·06, reflecting 126 the hospitalisation rate in the UK. 24 Time from symptom onset to hospitali-127 sation is drawn from an Exponential distribution with mean 5·954 days (fitted 128 to published data. 24 ) We then defined the undetected outbreak size as the 129 number of cases that were exposed prior to the first hospitalisation, given an 130 8 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint initial seeding of 5 index cases at t = 0. We also consider a special case of 131 100 index cases to represent a large super-spreading event. 

We found that where a test sensitivity of 65% was assumed, the impact releasing false negative cases is mitigated by using a precautionary seven-day 151 quarantine period, which reduced the risk of a large outbreak from 27·2% to 152 15·3% for R S = 1·5, and from 12·6% to 2·7% for R S = 1·3, all with 80% 153 contact tracing ( Figure 3A ).

The negative consequences of early quarantine cessation for false negative 155 cases are further demonstrated by the fact that a two day delay in carrying 156 out the tests also led to a decrease in the probability of a large outbreak, from 157 27·2% to 20·4% for R S of 1·5 and 12·6% to 5·4% for R S of 1·3. Combining the 158 two-day delay in testing and the seven-day precautionary quarantine reduced 159 the risk of a large outbreak further. The risk of a large outbreak was reduced 160 from 27·2% to 13·1% for R S = 1·5 and from 12·6% to 1·9% for R S = 1·3, 161 both with 80% contact tracing coverage.

In the case of instant testing and an immediate end to quarantine if the 163 test is negative, there was a comparatively small benefit from scaling up 164 of contact tracing coverage from 40% to 100%, implying that much of the 165 9 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org/10.1101/2020.06.09.20124008 doi: medRxiv preprint

To assess at what point during an epidemic contact tracing would be un-180 able to control transmission, we looked at the probability of a large outbreak 181 (greater than 2,000 cases within 300 days) given the current outbreak size 182 ( Figure 4A ). Both the time taken to trace contacts and the proportion of is 24·1% and by 500 cases this increases to 36·8%. This is compared to the 188 initial probability of 2·3% for these parameter values given 5 initial cases.

With R S = 1·5 the risk of a large outbreak increased faster. At 250 cases 190 the risk of a large outbreak is already 78·2% and by 500 cases it is 88·5%, 191 compared to an initial risk of 15·5% when starting with 5 initial cases. Figure S1 ). With 80% contact tracing coverage, a four-day contact tracing 195 delay increased the probability of a large outbreak, relative to a one day delay, 196 from 13·1% to 17·3% for R S = 1·5 and from 1·9% to 4·0% for R S = 1·3. and improved test sensitivity can increase case detection: 95% sensitivity 202 and 100% self-reporting gives an increase from 30·5% to 73·9% compared to 203 65% sensitivity and 50% self-reporting (both for R S = 1.3). However, this 204 still results in 26·1% of cases being missed, hence detecting every case is 205 essentially infeasible.

Every missed case is a potential new chain of transmission and, given the 207 low value of k, there is a risk of super-spreading events. To demonstrate 208 this we consider a scenario where one missed case leads to a cluster of either 209 5 or 100 new cases in a population with poor adherence to self-reporting 210 guidelines ( Figure 4C and D respectively). We assume no self-reporting, so 211 the first observation of the outbreak occurs when the first case is hospitalised, 212 after which contact tracing may be initiated.

For a cluster of 5 new cases the median total outbreak size before the first 214 case is hospitalised is 13 cases for R S = 1·3 and 18 cases for R S = 1·5, which 215 translates to 4·1% and 30·1% probability of a large outbreak respectively if 216 11 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. Figure 4 : A) Comparing probability of outbreak by total number of cases so far. Sensitivity = 65%, self-reporting proportion = 0·5, individuals testing negative are isolated for a minimum of 7 days, time to test from isolation = 2 days. B) The proportion of cases detected with 100% contact trace and 50% or 100% self-reporting for 65% and 95% sensitivity tests. C) Total cases occurring before first hospitalisation in a population with no active tracing or case detection from one super-spreading event (5 new cases). D) Total cases occurring before first hospitalisation in a population with no active tracing or case detection from one super-spreading event (100 new cases).

80% contact tracing can be implemented ( Figure 4A ). For a cluster of 100 new 217 cases the median total unobserved outbreak size is 226 for R S = 1·3 and 249 218 for R S = 1·5, translating to 22·6% and 78·0% probability of a large outbreak 219 with 80% contact tracing. This emphasises the importance of maintaining 220 physical distancing measures that restrict the size of indoor social gatherings 221 12 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 to avoid extreme super-spreading events which could rapidly escalate.

For R S = 1·1 there is a 5·37% chance of seeing at least 200 cases in all 223 scenarios, even with slower tracing (up to four days' delay) and only 40% 224 of contacts traced. Comparatively, for R S = 1·3 there is a greater than 225 5% chance of seeing 800 or more cases unless 100% contact tracing, or 80% 226 contact tracing with a 1-day trace delay is achieved. For R S = 1·5 even 100% 227 tracing with a one-day delay won't bring the probability of a large outbreak 228 under 5%, but increasing tracing from 40% to 100% brings this probability 229 down from 22·5% to 6·8%. Figure 5 : Outbreak size, with risk of exceeding that number of cases i.e. seeing an outbreak of at least that size for contact tracing coverages of 40% to 100% (left to right) and one or four days maximum trace delay (top to bottom). Grey dashed lines represent 5% risk of seeing an outbreak of at least that size.

We also found that higher contact tracing coverage results in a lower 232 overall number of individuals which are traced, tested and quarantined, due 233 13 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 to the lower outbreak size (see Supplementary Figure S2 ). This means that 234 achieving greater efficacy in tracing will ultimately require fewer resources.

However, these resources are likely to be needed in a more condensed period CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 We demonstrated that small increases in the reproduction number under 270 physical distancing measures, R S , has a large impact on the feasibility of 271 contact tracing. We only consider values of R S up to 1·5, which is still 272 substantially lower than estimates of R 0 in the absence of interventions (R 0 ≈ 273 2·7 27 ) therefore, our estimates of R S reflect a decrease in social contacts of 274 almost 50% but even 80% coverage and a one day trace time still gives at 275 least a 15% probability of a large outbreak. This reiterates that physical 276 distancing is still vital, even with highly effective contact tracing, and that 277 contact tracing will likely be insufficient to allow a complete return to normal 278 life without additional measures, such as an effective vaccine.

In addition to general physical distancing, the risk posed by a single large 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . realistic, although symptomatic individuals will perhaps be more cautious.

However, this could also have repercussions on assuming that contact-traced 309 individuals will self-isolate when asked to do so, particularly asymptomatic CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 Contact tracing, incorporating diagnostic testing, is a well-established 376 method for controlling novel infectious disease outbreaks but has had vari- 

Added value of this study 388 We incorporate testing and updated parameter estimates into an existing 389 branching process model to assess how 'test & trace' programmes could be 390 used to help control outbreaks of COVID-19. We find that if recent test 391 sensitivity estimates (approx. 65%) are representative then using testing to 392 rule-out cases and immediately revoke isolation advice could substantially re-393 duce contact tracing efficacy. Additionally, even if these risks are mitigated, 394 e.g. by introducing a minimum isolation period for all traced contacts, con-395 tact tracing must be used in combination with physical distancing measures 396 to minimise risk of large outbreaks.

Implications of all the available evidence 398 Greater clarity in understanding of SARS-CoV-2 biology has allowed 399 more targeted analysis of contact tracing feasibility for COVID-19 control. 400 We find that success is highly dependent on targeting testing towards finding 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 12, 2020. 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org /10.1101 /10. /2020 

South China Morning Post, Coronavirus: China's first confirmed 408

Why did the UK need 100,000 tests a day?

Real-time tracking of self-reported symptoms to predict potential 446 COVID-19

Vari-448 ation in false-negative rate of reverse transcriptase polymerase chain 449 reaction-based SARS-CoV-2 tests by time since exposure

Quantifying SARS-CoV-453 2 transmission suggests epidemic control with digital contact tracing

How will country-based mitigation measures influence the course of the 457 COVID-19 epidemic?

Estimating the 459 asymptomatic proportion of coronavirus disease 2019 (COVID-19) cases 460 on board the Diamond Princess cruise ship

Suppression of COVID-19 outbreak in the municipality of Vo

Comparison of seven commercial RT-PCR diagnostic kits for COVID-484 19

Role 487 of testing in COVID-19 control

Transmissibility of 2019-506 nCoV

Just over half of adults strictly sticking to lockdown guide-511 lines as confidence in government falls

Interventions to mitigate early spread of sars-cov-2 in 516 singapore: a modelling study

Containing the coronavirus (COVID-19): Lessons

Methods: We extend an existing UK-focused branching process model for contact tracing, adding diagnostic testing and refining parameter estimates to demonstrate the impact of poor test sensitivity and suggest mitigation methods. We also investigate the role of super-spreading events, providing estimates of the relationship between infections, cases detected and hospital-