key: cord-0868032-dwr0gwqu authors: Keeling, M. J.; Hill, E.; Gorsich, E.; Penman, B.; Guyver-Fletcher, G.; Holmes, A.; Leng, T.; McKimm, H.; Tamborrino, M.; Dyson, L.; Tildesley, M. title: Predictions of COVID-19 dynamics in the UK: short-term forecasting and analysis of potential exit strategies date: 2020-05-11 journal: nan DOI: 10.1101/2020.05.10.20083683 sha: 7967e33e39803ee2526e9741d8877351bc7ca18c doc_id: 868032 cord_uid: dwr0gwqu Background: Efforts to suppress transmission of SARS-CoV-2 in the UK have seen non-pharmaceutical interventions being invoked. The most severe measures to date include all restaurants, pubs and cafes being ordered to close on 20th March, followed by a "stay at home" order on the 23rd March and the closure of all non-essential retail outlets for an indefinite period. Government agencies are presently analysing how best to develop an exit strategy from these measures and to determine how the epidemic may progress once measures are lifted. Mathematical models are currently providing short and long term forecasts regarding the future course of the COVID-19 outbreak in the UK to support evidence-based policymaking. Methods: We present a deterministic, age-structured transmission model that uses real-time data on confirmed cases requiring hospital care and mortality to provide up-to-date predictions on epidemic spread in ten regions of the UK. We simulated a suite of scenarios to assess the impact of differing approaches to relaxing social distancing measures from 7th May 2020, on the estimated number of patients requiring inpatient and critical care treatment, and deaths. With regard to future epidemic outcomes, we investigated the impact of reducing compliance, ongoing shielding of elder age groups, reapplying stringent social distancing measures using region based triggers and the role of asymptomatic transmission. Findings: We find that significant relaxation of social distancing measures on 7th May can lead to a rapid resurgence of COVID-19 disease and the health system being quickly overwhelmed by a sizeable, second epidemic wave. In all considered age-shielding based strategies, we projected serious demand on critical care resources during the course of the pandemic. The reintroduction and release of strict measures on a regional basis, based on ICU bed occupancy, results in a long epidemic tail, until the second half of 2021, but ensures that the health service is protected by reintroducing social distancing measures for all individuals in a region when required. Conclusions: Our work supports the decision to apply stringent non-pharmaceutical measures in March 2020 to suppress the epidemic. We provide strong evidence to support the need for a cautious, measured approach to relaxation of lockdown measures, to protect the most vulnerable members of society and support the health service through subduing demand on hospital beds, in particular bed occupancy in intensive care units. In late 2019, accounts emerged from Wuhan city in China of a virus of unknown origin that was leading 2 to a cluster of pneumonia cases [1] . The virus was identified as a novel strain of coronavirus on 7th 3 January 2020 [2] and the first known death as a result of the disease occurred two days later [1] . Over 4 the next few days, cases were reported in several other cities in China and in other countries around 5 the world including South Korea, Japan and the United States of America. On 23rd January, the 6 Chinese government issued an order for Wuhan city to enter "lockdown", whereby all public transport 7 was suspended and residents were not allowed to leave the city. Over the next 24 hours, these measures 8 were extended to all the major cities in Hubei province in an attempt to prevent further spread of 9 disease. 10 Whilst the introduction of these severe social distancing measures began to have an effect upon re-11 ducing the growth rate of cases in Wuhan [3] [4] [5] , reported cases outside China continued to grow and 12 by late February the virus, now designated by the World Health Organisation as SARS-CoV-2, and 13 the disease it causes as coronavirus disease 2019 , had spread to Europe, with a growing 14 number of cases being reported in northern Italy [6] . As more countries in Europe and around the 15 world started to experience a dramatic rise in cases, similar measures were put in place in an effort to 16 protect the most vulnerable members of society and to ensure that health services capacities were not 17 exceeded [6, 7] . 18 In the UK, the first cases of COVID-19 were reported on 31st January 2020, in the city of York in 19 the north of England. In the early stages of the UK outbreak, the government focused on a strategy 20 of containment, to reduce the likelihood of large-scale within-country transmission occurring. This 21 strategy involved rapid identification and isolation of infected individuals, achieved through contact 22 tracing and testing of suspect cases. However, by early March it was evident that sustained community 23 transmission was occurring and there was a growing concern that a large epidemic could rapidly 24 overwhelm the health service, resulting in a significant number of deaths. This led to the government 25 considering the introduction of a range of social distancing measures in order to slow the growth of the 26 outbreak, thus delaying and flattening the epidemic peak and reducing the risk of exceeding hospital 27 capacities owing to an influx of COVID-19 patients. On 12th March, the UK officially entered the 28 "delay" phase, with the government declaring that all individuals with a cough or fever should self-29 isolate for a period of seven days. Over the following days, several major sporting events were cancelled 30 and the Prime Minister advised against all non-essential travel and contact with others. With daily 31 cases and deaths continuing to rise, the government introduced its most severe measures, with firstly 32 all restaurants, pubs and cafes being ordered to close on 20th March, followed by a "stay at home" 33 order on the evening of 23rd March and the closure of all non-essential retail outlets for an indefinite 34 period. By this time the reported number of deaths in the UK had reached 335. regarding the future spread of disease and predictions regarding the potential impact of a range of 38 intervention measures. This evidence, together with behavioural considerations from SPI-B (Inde-39 pendent Scientific Pandemic Influenza Group on Behaviours) and advice from NERVTAG (New and 40 Emerging Respiratory Virus Threats Advisory Group), is collated and coordinated by SAGE (Scien-41 tific Advisory Group for Emergencies) to support UK cross-government decisions in the Cabinet Office 42 Briefing Room (COBR). As the UK lockdown continues, modelling groups, including those of SPI-M, 43 are analysing potential "exit strategies", which could allow some relaxation of social distancing mea-44 sures, whilst minimising the future impact of the disease on the health service. Epidemiologists are 45 critially aware that, should measures be relaxed too rapidly when there are still sufficient susceptible 46 individuals in the population, there is a high risk of a second infection wave that may once again 47 threaten to overwhelm health services. 48 In this paper, we present a novel mathematical model of SARS-CoV-2 transmission that is currently 49 being utilised to provide short and long term forecasts regarding the future course of the COVID-19 50 outbreak in the UK. Our model uses real-time data on confirmed cases requiring hospital care and 51 mortality to provide up-to-date predictions on epidemic spread in ten regions of the UK. We investigate 52 how compliance with social distancing affects future epidemic outcomes. We compare and contrast 53 different exit strategies, namely: relaxing social distancing by age group, or the regional lifting and 54 imposition of restrictions according to healthcare system capacity. Finally, we explore the sensitiv-55 ity of our conclusions to a key biological aspect of SARS-CoV2 which remains unknown: whether 56 different age groups differ in their core susceptibility to infection, or their likelihood of displaying 57 symptoms. Transmission model 60 Here, we describe a compartmental model that has been developed to simulate the spread of SARS-61 CoV-2 virus (resulting in cases of in the UK population. In the ongoing outbreak in the 62 UK, cases of COVID-19 are confirmed based upon testing, with priority for testing given to patients 63 requiring critical care in hospitals [8] -generating biases and under-reporting. There is evidence to 64 suggest that a significant proportion of individuals who are infected may be asymptomatic or have only 65 mild symptoms [9, 10] . These asymptomatic individuals are still able to transmit infection [11] , though 66 it remains unclear whether they do so at a reduced level. Our modelling approach has consequently 67 been designed to consider the interplay between symptoms (and hence detection) and transmission 68 of COVID-19. We developed a deterministic, age-structured compartmental model, stratified into 69 five-year age bands. Transmission was governed through age-dependent mixing matrices based on 70 UK social mixing patterns [12, 13] . The population was further stratified according to current disease 71 status, following a susceptible-exposed-infectious-recovered (SEIR) paradigm, as well as differentiating 72 by symptoms, quarantining and household status (Fig. 1 ). Susceptibles (S) infected by SARS-CoV-2 73 entered a latent state (E) before becoming infectious. Given that only a proportion of individuals who 74 are infected are tested and subsequently identified, the infectious class in our model was partitioned 75 into symptomatic (and hence potentially detectable), D, and asymptomatic (and likely to remain 76 undetected) infections, U . We assumed both susceptibility and disease detection were dependent upon 77 age, although the partitioning between these two components is largely indeterminable (additional 78 details are given in Table 1 and Supporting Text S1). We modelled the UK population aggregated to 79 ten regions (Wales, Scotland, Northern Ireland, East of England, London, Midlands, North East and 80 Yorkshire, North West England, South East England, South West England). A drawback of the standard SEIR ordinary differential equation (ODE) formation in which all individ-82 uals mix randomly in the population is that it cannot readily account for the isolation of households. 83 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 May 11, 2020. For example, if all transmission outside the household is set to zero in a standard ODE model, then 84 an outbreak can still occur as within-household transmission allows infection between age-groups and 85 does not account for local depletion of susceptibles within the household evironment. We addressed 86 this limitation by extending the standard SEIR models such that first infections without a household 87 (E F , D F , U F ) are treated differently from subsequent infections (E S , D S , U S ). To account for the 88 depletion of susceptibles in the household, we assumed that all within household transmission is gen-89 erated by the first infection within the household (for further details, see Supporting Text S1). We stratified the population into susceptible, exposed, detectable infectious, undetectable infectious, and removed states. Solid lines correspond to disease state transitions, with dashed lines representing mapping from detectable cases to severe clinical cases that require hospital treatment, critical care (ICU), or result in death. The model was partitioned into five-year age bands. See Table 1 for a listing of model parameters. Note, we have not included quarantining and household status on this depiction of the system. The model is concerned with epidemiological processes and so predicts the number of symptomatic 92 and asymptomatic infections on each day. However, in order to provide evidence regarding the future 93 impact of the outbreak in the UK, it is crucial to be able to predict the number of severe cases that may 94 require hospital or critical care. We utilised two processes in order to estimate hospitalisation rates: 95 (i) we estimated the proportion of clinical cases in each age group that would require hospitalisation 96 by comparing the age distribution of hospital admission to the age structure of early detected cases -97 4 . 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 May 11, 2020 . . https://doi.org/10.1101 Population size of a given age By region ONS assuming these detected cases were an unbiased sample of symptomatic individuals. (ii) we used age 98 independent distributions to determine the time between onset of symptoms and hospitalisation. A 99 similar process was repeated for admission into intensive care units. Both of these distributions were 100 drawn from the COVID-19 Hospitalisation in England Surveillance System (CHESS) data set that 101 collects detailed data on patients infected with COVID-19 [14] . Information on the distributions of length of stay in both intensive care units (ICUs) and hospital was 103 used to translate admissions into bed occupancy -which adds a further delay between the epidemio-104 logical dynamics and quantities of interest. In terms of matching the available data and quantities of interest, we also use the prediction of 106 symptomatic infections to drive the estimated daily number of deaths within hospitals. The risk of 107 death is again captured with an age-dependent probability, while the distribution of delays between 108 hospital admission and death is assumed to be age-independent. These two quantities are determined 109 from the Public Health England (PHE) death records. Model fitting 111 We fit on a region-by-region basis to four timeseries: (i) new hospitalisations; (ii) hospital bed occu-112 pancy; (iii) ICU bed occupancy; (iv) daily deaths (using data on the recorded date of death, where-ever 113 possible). The relative transmission rate from asymptomatic cases (τ ) and the scaling of whether age-114 structure case reports are based on age-dependent susceptibility or age-dependent symptoms (α) were 115 treated as free parameters. 116 We performed parameter inference using the Metropolis-Hastings algorithm, computing likelihoods 117 5 . 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 May 11, 2020. . assuming the daily count data to be drawn from a Poisson distribution. At the outset of each iteration, 118 based on the α and τ instances at that moment, we first set the recovery rate γ, age-dependent 119 susceptibilites (σ a ) and age-dependent probabilities of displaying symptoms (d a ). We achieved this 120 through calibration to early age-stratified UK data from 1st February 2020 to 1st April 2020; initial 121 assumptions were an R 0 of 2.7 and doubling time of 3.3 days, but these are allowed to vary between 122 regions. After a burn-in of 250,000 particles, the algorithm was run for a further 250,000 iterations. We thinned 124 the generated parameter sets by a factor of 100, giving 2,500 parameter sets representing samples from 125 the parameter posterior distributions. In order to capture the impact of social distancing measures that were introduced in the UK (on 128 23rd March) to reduce transmission, we scaled down the mixing matrices associated with schools, 129 work and other activities while increasing the within household transmission matrix (see Supporting 130 Text S2). This approach allowed us to flexibly vary the effectiveness of different social distancing 131 measures and investigate the impact of compliance with social distancing (φ) upon the future spread 132 of disease. We considered a range of compliance levels, scaling from zero (no-compliance) to one 133 (maximal compliance), as well as inferring the compliance parameter from the available data (φ = 134 0.53(0.36 − 0.70) across all regions). Another prominent intervention measure that was put in place in the UK to reduce the spread of infec-136 tion has been household quarantining, whereby an entire household was instructed to self-quarantine 137 when any member of that household started to show symptoms of infection. To incorporate house-138 hold quarantining measures into the ODE formulation, we added a quarantined class into our model, 139 whereby a fraction (H) of the first detectable infection in any household (and therefore by definition a 140 symptomatic case) is quarantined as are all their subsequent household infections. Accounting for the 141 effect of household saturation also ensures that subsequent household infections do not contribute to 142 further transmission. For a complete description of the model equations, see Supporting Text S1. 143 We used this model framework to perform a series of analyses assessing the impact of social distancing 144 strategies on the future spread of infection. Unless otherwise stated, all intervention shown represent 145 the mean dynamics from the posterior parameters inferred by a Monte Carlo Markov Chain (MCMC) 146 fitting scheme; where practical credible intervals are also shown. To provide a baseline for comparison of our intervention scenarios, we initially simulated our model 149 to investigate the impact of the current intervention policies, continuing from their introduction on 150 23rd March 2020. We simulated the model from 1st March 2020 to 30th April 2020 and compared 151 the results to a scenario where no lockdown measures were ever introduced. To quantify prediction 152 uncertainty, a total of 200 simulations were run for each scenario (lockdown activated or no lockdown 153 imposed) using distinct parameter sets produced by the MCMC procedure, representing samples from 154 the posterior parameter distributions. We focused our attention on estimates of deaths as well as 155 hospitalisation and ICU bed occupancy, as key public-health considerations. To investigate the longer term impact of the epidemic, we explored several scenarios in which control 158 measures are relaxed on 7th May. The first scenario investigated a policy whereby social distancing 159 measures were relaxed on 7th May for all individuals, regardless of age. To reflect the uncertainty 160 6 . 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 May 11, 2020. . in the degree of relaxation of the lockdown at this point, we varied our social distancing compliance 161 parameter (φ = 0, 0.25, 0.5, 0.75, 1), which allowed us to consider how the epidemic trajectory may be 162 affected for a range of relaxation policies. In these simulations we assumed that any remaining social 163 distancing measures were fully removed at the end of 2020. Age-dependent relaxation of lockdown measures 165 We next investigated policies imposing age-dependence upon the relaxation criteria. Specifically, we 166 allowed all social distancing measures to be lifted from 7th May for all individuals below a certain age 167 (from 45 to 75 year old). For those above the age threshold, we assumed that social distancing measures 168 remained in place until the end of 2020. Simulations were then run to the end of 2021, to capture 169 any subsequent waves of infection. For each age threshold under consideration, we again consider the 170 cumulative deaths, as well as cumulative hospital and ICU bed occupancy. We differentiate between 171 these health impacts that occur when age-specific restrictions are in place and when all restrictions are 172 lifted. We also focus of the number of days in which ICU bed occupancy exceeds 4,000, as a measure 173 of the immediate severity of the outbreak and the pressure on the health services. Full relaxation of lockdown measures with region-based reintroduction 175 Our penultimate set of simulations considered an adaptive intervention strategy, whereby lockdown 176 measures were fully relaxed on 7th May, but then reintroduced when occupancy of intensive care 177 units exceeded a given capacity and relaxed again when ICU occupancy declines. To account for 178 regional variation in the outbreak and local hospital capacities, we assumed that control measures 179 would operate locally, using the ten regions. We therefore used a pro-rata threshold, which equated 180 to 3,000 occupied beds on a nationwide scale, as a trigger for reintroducing or relaxing controls (see 181 Table S2 ). Given the sizeable delay between the implementation of controls and their effects on ICU 182 occupancy, the dynamics only predicted a low number of switches between control and relaxation. We 183 gathered regional predictions of daily deaths, ICU bed occupancy and hospital bed occupancy, with 184 simulations run to the end of 2023. When evaluating the impact of lockdown measures, we are reliant upon recorded data on confirmed 187 cases, hospital admissions and ICU occupancy in order to infer parameters of our model. However, 188 there is still ongoing uncertainty in the relative level of transmission from asymptomatic individuals 189 (τ ) and the mechanisms driving age-specific detection rates (α). A range of α and τ parameter values 190 are all able to generate predictions that closely match the available data. We therefore carried out a 191 sensitivity analysis to these two parameters, investigating the impact of applying lockdown measures 192 for specific age groups, as these parameters vary. We allowed τ , the relative level of transmission from 193 asymptomatic individuals, to vary between 0 and 0.5; while α varied between 0 and 1. For large α, 194 higher proportions of confirmed cases in a particular age group is as a result of greater susceptibility; 195 whereas low vales of α indicate that a higher proportions of confirmed cases is due to greater severity of 196 symptoms. This key parameter interacts with the relative transmission from asymptomatic infection 197 (τ ), although τ plays a minimal role when α is small. To assess the impact of these parameters on the 198 effectiveness of lockdown measures, we computed the early epidemic growth rate under restrictions 199 that target four specific age-groupings: (i) pre-school children under 5 (PS), (ii) school-aged children 200 and young adults, 5-20 (S), (iii) adults between the age of 20 and 70 (A) and (iv) the elderly over 70 201 (E). 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 May 11, 2020 . . https://doi.org/10.1101 Our model predicts that, should the current lockdown policies be continued, the number of daily deaths 206 would peak in April across all regions before starting to decline (Fig. 2) . England and Wales are found 207 to be most severely affected, with the highest number of predicted deaths per capita, whilst we predict 208 a lower number of deaths per capita in Scotland and Northern Ireland (noting that though our regional 209 model fits generally had strong correspondence with the data, the fit to Scotland was weaker). All 210 English regions show similar behaviour, other than the South East and South West, where we predict 211 a lower number of deaths (Fig. 2) . We observe similar behaviour in the levels of hospital and intensive 212 care unit occupancy throughout this period ( Fig. S1 and Fig. S2 ). Our model predicts that, under 213 continued total lockdown, the average total deaths would be approximately 39,000 ( Table 2) . If the epidemic in the UK had been allowed to progress with no introduction of lockdown measures, 215 our model predicts that the epidemic would have continued to grow throughout April, with deaths 216 exceeding 200,000 by the end of 2021 (Table 2 ). This provides strong evidence to support the necessity 217 of the social distancing measures that were introduced in order to reduce the growth rate of the 218 epidemic and ensure that the health service is not overwhelmed with admissions. In each panel: filled markers correspond to observed data (squares are for reported deaths, circles are for date of death), solid lines correspond to the mean outbreak over a sample of posterior parameters; shaded regions depict prediction intervals, with darker shading representing stricter confidence (dark shading -50%, moderate shading -90%, light shading -99%); red dashed lines illustrate the mean projected trajectory had no lockdown measures being introduced. (Prediction were produced on 23rd April, using data until 21st April). . 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 May 11, 2020 . . https://doi.org/10.1101 Measured age-independent relaxation protocols to reduce health system burden 220 Evaluating a policy whereby social distancing measures were relaxed on 7th May for all individuals, 221 we found that for a significant relaxation of lockdown, the epidemic rapidly resurges with a peak in 222 daily deaths of over 4,000 occurring in late June (Fig. 3, top panel) . We project intensive care unit 223 occupancy to near 10,000 by the end of June (Fig. 3, second panel) , implying that significant release 224 of lockdown measures would not be advisable. For more measured relaxation protocols, we found 225 that, whilst there may be a slight resurgence in cases in the short term, hospital and ICU occupancy 226 remained within capacity. Whilst the short term forecasts from these simulations suggest that keeping 227 some lockdown measures in place can have a positive impact upon reducing cases and deaths, we note 228 that, for simulations in which more severe lockdown remains in place after 7th May, a second infection 229 wave is predicted in 2021, when all social distancing measures are removed. This implies that, should 230 relaxation of measures occur, there may need to be a reintroduction of social distancing interventions 231 at a later stage to reduce the impact of a potential second wave occurring later in 2020 or early in 232 2021. Assessment of age-based shielding strategies 234 We next analysed the lifting of social distancing measures on 7th May for all individuals below a 235 certain age, with social distancing measures remaining in place for the remainder of the population 236 until the end of 2020. We observe that continuing lockdown for anyone over the age of 45 for the 237 duration of 2020 results in the lowest number of deaths and number of admissions into hospital and 238 ICU wards during 2020 (Fig. 4, first column) . However, upon release of these lockdown measures, we 239 observed a significant second wave in 2021 when isolation includes these younger age groups (Fig. 4 , 240 second column). When isolation is only in place for older age groups, a large initial wave of infection 241 occurs during 2020, but a subsequent secondary wave is not observed. If we consider the overall 242 impact from 2020-2021, we find that a strategy of continuing lockdown measures for anyone over the 243 age of 65 minimises the total number of deaths, and continuing these measures for anyone over the 244 age of 60 minimises hoapital and ICU occupancy, though the overall effect of this when compared 245 with other age-related lockdown policies is marginal (Fig. 4 , third column). Continuing lockdown for 246 the over 60s throughout 2020 whilst relaxing measures of the remainder of the population results in, 247 on average, 138,000 deaths by the end of 2021 (Table 2 ). Finally, we note that, as the age-threshold 248 at which shielding is implemented increases, the total number of days for which ICU bed occupancy 249 exceeds 4,000 increases, implying that only shielding older age groups may put severe demands upon 250 the health service (Fig. 4 , third row). Utility of reintroducing lockdown measures regionally with ICU occupancy triggers 252 Performing subsequently occurs, but an increase in ICU occupancy triggers the reintroduction of social 253 distancing measures on a region by region basis (Fig. 5) . This results in a second, smaller peak in late 254 May, with ICU and hospital occupancy remaining at manageable levels (Fig. 5 , second and third pan-255 els). As regional lockdowns start to be released yet again, we see a slight resurgence for a third time. 256 The consequence of this adaptive strategy is that the number of deaths and confirmed cases gradually 257 reduce over a long period of time (Fig. 5, top and bottom panels) , with the epidemic reaching low 258 levels in late 2020. In addition, the result of such a policy is that, throughout the epidemic, ICU and 259 hospital occupancy stabilises and gradually decreases, thus providing a necessary level of protection 260 for the health service. We assume social distancing measures were relaxed on 7th May for all individuals. The paler lines correspond to the dynamics using differing levels of relaxation (φ = 0, 0.25, 0.5, 0.75, 1), with φ = 0 (lightest line weight) corresponding to a total removal of social distancing measures, and φ = 1 (heaviest line weight) representing a continuation of lockdown measures. Shaded regions represent the 95% posterior prediction intervals. We display statistics on daily counts of (Row one) deaths; (Row two) ICU occupancy; (Row three) hospital occupancy. At the start of 2021, all remaining social distancing measures are removed (the "no control" phase). . 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 May 11, 2020 . . https://doi.org/10.1101 Fig. 4 : Impact of age-based shielding strategies on outbreak burden. In these simulations, social distancing measures were lifted on 7th May for all individuals below an age threshold, with social distancing measures remaining in place for the remainder of the population until the end of 2020. No interventions were applied post-lockdown release, with simulations continued until the end of 2022. Box plots per considered age threshold convey median values for each statistic and extent of spread. Solid lines depict the profile of median estimates across age threshold space. The following statistics were computed for the period 23rd March 2020 to the end of 2021: (Row one) cumulative deaths; (Row two) cumulative ICU bed occupancy; (Row three) amount of days ICU occupancy exceeded 4000; (Row four) cumulative hospital bed occupancy. We stratify the outputs occurring across the considered time horizon in three ways: (Column one) during lockdown; (Column two) after lockdown; (Column three) combined (entire time horizon). . 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 May 11, 2020 May 11, . . https://doi.org/10.1101 May 11, /2020 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 May 11, 2020 . . https://doi.org/10.1101 Role of asymptomatics crucial in determining the effect of age-based lockdown relaxation 263 measures 264 Finally, we investigate the impact of applying lockdown measures for specific age groups, whilst varying 265 τ , the relative level of transmission from asymptomatic individuals, and α, the scaling determining 266 whether the age-dependence in cases comes from susceptibility (α = 1) or symptoms (α = 0). We 267 observe that, regardless of the values of τ and α, applying control on only a single age group (PS, 268 S, A or E) results in large-scale epidemics (Fig. 6) . Similarly ineffective strategies are observed when 269 combining PS control with one of S, A or E. However, control of school aged children, adults and the 270 elderly, results in epidemics that are under control for all values of τ and α. Should we exempt the elderly from lockdown we find that, for high levels of α, large epidemics are 272 observed, whereas if the true value of τ is high and α is small, applying control on the younger age 273 groups and releasing lockdown on the elderly can result in epidemics that will rapidly die out. In 274 contrast, if we relax lockdown on school children but keep it in place for other age groups, we note 275 that this only has a positive effect upon the epidemic if the true value of α is high, or the true value of τ 276 is low. If α is low and τ is high, then this implies that the age-dependence of reported cases is primarily 277 as a result of clinical symptoms rather than susceptibility and the transmission rate of asymptomatic 278 cases is high. Therefore, school children will play a much larger role in transmission, implying that a 279 policy of re-opening school would cause a much larger epidemic. These results reinforce the need to 280 resolve uncertainty regarding the role of asymptomatic individuals in the infection process in order to 281 establish the optimal intervention strategy. Potential exit strategies comparison 283 Our findings are summarised in Table 2 , where we focus on deaths (and the associated QALY losses), 284 hospital occupancy and the scale of the lockdown as a measure of potential economic burden. QALYs 285 (Quality Adjusted Life Years) are a standard measure in health economics which accounts for the 286 number of life years lost due to an illness or disease, which also takes into account quality of life. 287 Hence, under the QALY framework deaths in younger individuals have greater impact than deaths of 288 older individuals due to the additional years of life lost (for further details, see the Supporting Text 289 S3). Our lockdown scale measures the pro rata number of days the population is under lockdown; so 290 if 50% of the population is under lockdown for 200 days, we report a value of 100 (50% × 200). A completely uncontrolled outbreak is predicted to lead to around 200,000 deaths, 4.5 million QALY 292 losses but no lockdown impacts. If the current controls are maintained until the end of 2020, then 293 we predict only 39,000 cases this year, but a further 159,000 if controls are then completely removed. 294 Regional switching and age-dependent strategies provide alternative exit strategies in the absence 295 of pharmaceutical interventions. Of these, the lockdown of those age 60 or over generates the lowest 296 mortality and also the lowest lockdown scale, thereby minimising socio-economic disruption. However, 297 it is unclear if a protracted lockdown of this age-group would be practical, ethical or politically 298 acceptable. 299 . 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 May 11, 2020 May 11, . . https://doi.org/10.1101 May 11, /2020 doi: medRxiv preprint . 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 May 11, 2020 . . https://doi.org/10.1101 Discussion 300 In this paper, we have developed an age-structured compartmental SARS-CoV-2 transmission model 301 that has been used to make short-term predictions and analyse the effectiveness of the current social 302 distancing measures that have been implemented in the UK. We have also extended our model to 303 investigate a suite of alternate policies to establish whether some relaxation of intervention measures 304 may be considered as part of potential exit strategies. Our model shows that, without the introduction 305 of the large scale social distancing measures that were introduced on 23rd March, the epidemic would 306 have continued to grow exponentially and hospital and ICU occupancy would have rapidly exceeded 307 capacity. However, under current policies, the epidemic is expected to peak in April for all regions of 308 the UK, before starting to decline. One of the most important questions in the current epidemic is when and how social distancing 310 measures might be relaxed. We consistently find that any relaxation of control measures in the 311 short term leads to a rapid resurgence of COVID-19 disease with the health system potential being 312 overwhelmed by a sizeable second epidemic wave (Fig. 3) . In contrast, moderate or no adjustments to 313 current social distancing measures allows hospital and ICU occupancy to remain within capacity over 314 the duration of the outbreak. It is apparent from the data on confirmed cases and deaths as a result 315 of COVID-19 disease that the risks associated with infection increase with age [15] [16] [17] . We therefore 316 also investigated the impact of age-specific control policies, whereby lockdown measures remained in 317 place for all individuals over a certain age until the end of 2020. We found that, whilst some marginal 318 gains can be made should everyone over the age of 60 be put under isolation measures, extending this 319 policy to include younger age groups increases the risk of a second wave occurring when measures are 320 relaxed. Furthermore, we projected critical care to be stretched and ICU bed occupancy to exceed 321 4,000 during the course of the pandemic in all but the most wide-ranging age-specific lockdown policies 322 (Fig. 4) . As a note of caution, our sensitivity analysis shows that the effectiveness of any age-specific intervention 324 policy is critically dependent upon the precise role of asymptomatic individuals in the epidemic. 325 Undocumented infection has been inferred to have facilitated the spread of SARS-CoV-2 in China [18] . 326 We note that government advice upon self-quarantining focuses upon individuals who show symptoms 327 of COVID-19 (primarily a fever and a dry persistent cough) and therefore, should asymptomatic (or 328 pre-symptomatic) infections play a significant role in the transmission process, such a policy may 329 not be as effective as thought due to the potential for ongoing transmission through asymptomatic 330 cases. In practice, to minimise the risk of a second large epidemic wave occurring in the UK, adaptive 332 policies may need to be considered. To that end, we examined a more bespoke intervention policy 333 whereby measures were relaxed and re-introduced on a regional basis, with a defined trigger for 334 the reintroduction of interventions when ICU occupancy exceeded a certain level. This results in 335 a longer epidemic tail, until the second half of 2021, but ensures that the health service is protected 336 by reintroducing social distancing measures for all individuals in a region when required. Several countries around the world have now seen significant epidemics of COVID-19 and many have 338 implemented severe lockdown policies in an effort to contain the disease. In China and other countries 339 in East Asia, once the epidemic was regarded to be under control, in seeking to prevent the occurrence 340 of a large second infection wave the relaxation of isolation measures has been implemented in a gradual 341 fashion. Our model findings support the need for this form of relaxation policy. We recognise that 342 there is a need for certain measures to be lifted as soon as is feasible, for a range of practical, social and 343 economic reasons. However, government agencies should be prepared to resume lockdown if needed, 344 based upon the the progression of the epidemic following relaxation. Identifying triggers, such as ICU 345 occupancy exceeding a certain threshold, may be beneficial in allowing decision makers to follow a 346 16 . 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 May 11, 2020 . . https://doi.org/10.1101 clear set of guidelines for controls to be reintroduced. The identification of such triggers will be based 347 upon the objective of an intervention measure and the ability to resolve epidemiological uncertainty as 348 the outbreak progresses. To this end, formal adaptive management approaches may help to facilitate 349 the establishment of state dependent intervention strategies [19] . The model described is necessarily a simplified representation of reality based on several assumptions 351 and has various limitations. Data informing contact structure for the UK were measured histori-352 cally [12] . Were contact patterns in early 2020 (pre-lockdown) to substantially differ from the preex-353 isting data, the influence of projected intervention effects may be impacted. Similarly, while we can 354 infer the compliance to the currently imposed rules, we have limited understand of how people will 355 behave if the controls are released -will they remain wary of potentially infectious situations, or 356 will they compensate for the time in lockdown. Throughout we have assumed that mixing patterns 357 would return to their pre-pandemic norm. Heterogeneities in compliance and in infection patterns, 358 such as increased transmission in hospitals and institutions may affect the outcome of the measures 359 considered. We note that the estimates of deaths resulting from an individual strategy does not take 360 into account the potential for increased deaths due to exceeding hospital or ICU capacities, and so may 361 underestimate deaths from strategies resulting in high occupancies. In addition, though there have 362 been recorded instances of superspreading events for COVID-19 [20] , our model does not explicitly 363 account for such dynamics. However, beyond the early stages of the outbreak the dynamics at the 364 population-level are generally driven by the average pattern of social mixing, rather than individual 365 level variation. All the strategies we have considered here assume that an exit strategy will have to rely on non-367 pharmaceutical intervention. In which case, a second (or subsequent) wave of infections follows any 368 return to normality while there is sufficient susceptibility in the population. We are therefore faced 369 with three potential exit solutions: 1) Seek a measured reduction in restrictions that minimises the 370 impact of the unfolding outbreak, but acknowledging that the majority of the population will become 371 infected (although not symptomatic); 2) Accept a substantial and long-term change to our social 372 interactions (practising far better prevention of transmission), such that the reproductive ratio of the 373 virus is constantly held below one -electronic and traditional methods of tracing and isolation [21] 374 fall into this category; or 3) rely on the development of an effective vaccine, in which case the best 375 approach may be to extend the lockdown, reducing infection until mass vaccination can occur. In conclusion, the current pandemic of COVID-19 has resulted in the introduction of strict social 377 distancing measures in the UK and many other countries around the world. Government agencies are 378 currently analysing how best to develop an exit strategy from these measures and to determine how 379 the epidemic may progress once measures are lifted. Our work provides strong evidence to support 380 the need for a cautious, measured approach to relaxation, in order to provide necessary support for 381 the health service and to protect the most vulnerable members of society. 382 . 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 May 11, 2020 . . https://doi.org/10.1101 . 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 May 11, 2020 . . https://doi.org/10.1101 Supporting information items Supporting Text S1 Description of the complete system of model equations. Details on the mechanisms underpinning social distancing measures within the model framework. Explanation of the QALY losses computation. EQ-5D index population norms for England. UK population aggregated to ten regions (rounded to nearest 10,000). With regard to our intervention scenario in which regional ICU occupancy triggered the reintroduction and relaxation of social distancing measures within that region, the final column lists each of the regional ICU bed occupancy thresholds (equating to 45 occupied ICU beds per one million population). Regional projections for hospitalisations per 100,000 with and without imposition of lockdown. In each panel: filled markers correspond to observed data (squares are for reported deaths, circles are for death of death), solid lines correspond to the mean outbreak over a sample of posterior parameters; shaded regions depict prediction intervals, with darker shading representing stricter confidence (dark shading -50%, moderate shading -90%, light shading -99%); dashed lines illustrate the mean projected trajectory had no lockdown measures being introduced. Regional projections for ICU bed occupancy per 100,000 with and without imposition of lockdown. In each panel: filled markers correspond to observed data (squares are for reported deaths, circles are for death of death), solid lines correspond to the mean outbreak over a sample of posterior parameters; shaded regions depict prediction intervals, with darker shading representing stricter confidence (dark shading -50%, moderate shading -90%, light shading -99%); dashed lines illustrate the mean projected trajectory had no lockdown measures being introduced. . 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 May 11, 2020 . . https://doi.org/10.1101 Massimiliano Tambor-388 rino Writing -original draft Writing -review & editing: Matt J. Keeling A Novel Coronavirus from Patients with Pneumonia in China World Health Organization. Novel Coronavirus (2019-nCoV) Situation Report -1 Early dynamics of transmission and control of COVID-19: a mathematical modelling study Evolving Epidemiology and Impact of Nonpharmaceutical Interventions on the Outbreak of Coronavirus Disease The effect of control strategies to reduce social mixing on outcomes of the COVID-19 epidemic in Wuhan, China: a modelling study The early phase of the COVID-19 outbreak in Impact assessment of nonpharmaceutical interventions against COVID-19 and influenza in Hong Kong: an observational study Priorities for SARS-CoV-2 (COVID-19) testing Estimating the asymptomatic proportion of coronavirus disease 2019 (COVID-19) cases on board the Diamond Princess cruise ship SARS-CoV-2 Infection in Children Presumed Asymptomatic Carrier Transmission of COVID-19 Social Contacts and Mixing Patterns Relevant to the Spread of Infectious Diseases Projecting social contact matrices in 152 countries using contact surveys and demographic data COVID-19 Hospitalisation in England Surveillance System (CHESS) -daily reporting Clinical Characteristics of Coronavirus Disease 2019 in China Estimates of the severity of COVID-19 disease Monitoring trends and differences in COVID-19 case fatality rates using decomposition methods: Contributions of age structure and age-specific fatality Substantial undocumented infection facilitates the rapid dissemination of novel coronavirus (SARS-CoV2). Science (80-. ) Adaptive Management and the Value of Information: Learning Via Intervention in Epidemiology Secondary attack rate and superspreading events for SARS-CoV-2 Quantifying SARS-CoV-2 transmission suggests epidemic control with digital contact tracing. Science (80-. ) The authors would like to thank the RAMP Rapid Review Group for their useful comments on this 384 manuscript.