key: cord-0848994-cr659fgk authors: van Wees, J.-D.; van der Kuip, M.; Osinga, S.; van Westerloo, D.; Tanck, M.; Hanegraaf, M.; Pluymaekers, M.; Leeuwenburgh, O.; van Bijsterveldt, L.; Verreijdt, P.; Brunner, L.; Tutu van Furth, M. title: Performance of progressive and adaptive COVID-19 exit strategies: a stress test analysis for managing intensive care unit rates date: 2020-05-20 journal: nan DOI: 10.1101/2020.05.16.20102947 sha: fafc36b3b9b500f105a0deb9a964434e0a45f2a2 doc_id: 848994 cord_uid: cr659fgk Background: In May 2020, many European countries have begun to introduce an exit strategy for the coronavirus disease 2019 (COVID-19) pandemic which involves relaxing social distancing measures. Predictive epidemiological modeling indicates that chances for resurgence are high. However, parametrization of the epidemiological nature of COVID-19 and the effect of relaxing social distancing is not well constrained, resulting in highly uncertain outcomes in view of managing future intensive care unit (ICU) needs. Methods and findings: For performance analysis of exit strategies we developed an open-source ensemble-based Susceptible-Exposed-Infectious-Removed (SEIR) model. It takes into account uncertainties for the COVID-19 parametrization and social distancing measures. The model is calibrated to data of the outbreak and lockdown phase. For the exit phase, the model includes the capability to activate an emergency brake, reinstating lockdown conditions. Alternatively, the model uses an adaptive COVID-19 cruise control (ACCC) capable to retain a targeted ICU level. The model is demonstrated for the Netherlands and we analyzed progressive and adaptive exit strategies through a stress test of managing ICU rates. The progressive strategy reflects the outcome of social and economic pressure to use one-way steering toward progressively relaxing measures at an early stage. It is marked by a high probability for the activation of the emergency brake due to an unsolicited growth of ICU needs in the following months. Alternatively, the two-way steering ACCC can flatten ICU needs in a more gradual way and avoids activation of the emergency brake. It also performs well for seasonal variation in the reproduction number of severe acute respiratory syndrome-coronavirus. Conclusions: The adaptive strategy (ACCC) is favored, as it avoids the use of the emergency brake at the expense of small steps of restrictive measures and allows the exploration of riskier and potentially rewarding measures in the future pathways of the exit strategy. One Sentence Summary: Progressive and adaptive COVID-19 exit strategies 23 . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint . CC-BY-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 20, 2020. The model is calibrated to data of the outbreak and lockdown phase. For the exit phase, the 34 model includes the capability to activate an emergency brake, reinstating lockdown 35 conditions. Alternatively, the model uses an adaptive COVID-19 cruise control (ACCC) 36 capable to retain a targeted ICU level. The model is demonstrated for the Netherlands and we 37 analyzed progressive and adaptive exit strategies through a stress test of managing ICU rates. The progressive strategy reflects the outcome of social and economic pressure to use one-way 39 steering toward progressively relaxing measures at an early stage. It is marked by a high 40 probability for the activation of the emergency brake due to an unsolicited growth of ICU 41 needs in the following months. Alternatively, the two-way steering ACCC can flatten ICU 42 needs in a more gradual way and avoids activation of the emergency brake. It also performs 43 well for seasonal variation in the reproduction number of severe acute respiratory syndrome- . CC-BY-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 20, 2020. . CC-BY-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. . CC-BY-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. Alternatively, the Adaptive COVID-19 Cruise Control (ACCC) is based on two-way steering 104 and can flatten ICU needs in a more gradual way to a given nominal capacity. We will show 105 that the ACCC strategy can avoid the need for an emergency brake, and also performs well if 106 . CC-BY-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 20, 2020. Mathematically, the SEIR compartments are modeled as fractions of the population, initially 119 susceptible to the virus. Subsequently, we solve the following differential equations starting 120 from initial values for S = 1-1/N, E = 1/N and setting I and R to 0. N is ratio of population size 121 and exposed persons introduced at the start of the simulation. The initial day of introduction 122 of the virus is 15 days prior to the first registered case, and the density of the virus per N is 123 varied to obtain a first order fit with the observed number of registered infections, 124 hospitalized, death. The differential model is formulated as follows: The R compartment is extended with the flow scheme for the hospitalization and ICU 130 treatment process as depicted in Fig. 3 . For the incorporation of the logistic spread of patients, the model adopts successive convolutions of the R prediction, with a Gamma distribution for 132 . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint 7 delay, recovery and treatment times dhos, dhosrec, dhosd, dicurec and dicud (Fig. 3 , Table 1 ). The . CC-BY-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 20, 2020. In the adaptive exit strategy, we consider the ACCC to automatically steer both 161 relaxing (mstep<0) and restrictive step-changes (mstep>0) toward reaching and retaining a 162 sustainable level of nominal ICU capacity and to avoid activation of the emergency brake and 163 renewed lockdown conditions. To this end, the need for a step-change and its sign is . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint 9 The daily admittance of ICU patients closely followed the trend of hospitalized 185 patients (Fig. 4) . Markedly, the fraction of hospitalized patients that required an ICU shows a 186 gradual decreasing trend toward a relatively constant of 20% from the end of March onwards. The case fatality of ICU patients is 30% (14) and the overall mortality of COVID-19 patients 188 (5,643 on 14 May 2020) in the Netherlands contributes to ~2% of the global death toll of the 189 pandemic ( Fig. 1) (2). The cumulative ICU mortality is approximately 10% of the registered CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint smoother to fit data on hospitalized (Fig. 5C ) and ICU occupancy (Fig. 5D) . The model is very 211 well capable of reproducing the observed hospitalization (Fig. 5C ) and ICU rates (Fig. 5D ) and 212 demonstrates the effectiveness of the social distancing in terms of reducing the hospitalization 213 inflow, resulting in a peak ICU occupancy in early April 2020 followed by a downward trend 214 to manageable levels on 30 April 2020 close to an occupancy of 700 ICU patients. It is expected 215 that ICU usage will further reduce well below 600 on 11 May 2020, the date the Dutch 216 government has been planning to relax the lockdown conditions. The posterior 1-a(t) The mortality beyond 7 April 2020 is significantly higher than expected from the 224 hospitalization flow model. One possible explanation for this strong deviation, is that a 225 growing share of the registered deceased patients are related to care centers dedicated to 226 elderly people (Fig. 5E ). Here COVID-19 has been marked by more active spreading than the 227 national trend in the last three weeks of April 2020. This may be related to a lack of testing 228 and protective measures for care-giving personnel in that setting of elderly care. It should also 229 be noted that overall many more casualties are related to COVID-19 than tested. It is estimated 230 from nationally recorded death rates that the real death toll of COVID-19 may be twice as high 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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint predicted number of cumulative infected is significantly higher in the model than actual 237 confirmed cases (Fig. 5B ). The SEIR model takes into account the effects of gradual build-up 238 of immunity through the gradual reduction of S in the mathematical formula (see methods). Such a reduction can potentially contribute to social relaxation and therefore the estimation 240 of the ratio of infected and hospitalized patients is important. However, with the adopted 241 parameters, the effect of immunity build-up is rather low and not of significant influence for 242 the presented results in this paper. In the second scenario (Fig. 6B) , the first scenario is complemented with an emergency 252 brake. This emergency brake is triggered either by reaching 700 ICU occupancy or if the daily 253 growth rate in ICU occupancy exceeds 20 (depicted with the colored dots in the right panels 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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint emergency brake was activated. This can indeed be observed in the ensemble results in Fig. 262 7. In the third scenario (Fig. 6C) , the step-change exit strategy adopts the relaxation in 264 five smaller step-changes with Da(t) = N(-0.04,0.01) , in intervals of three weeks' time, instead 265 of a single step-change of -0.2. The gradual relaxation results in a more gradual increase in 266 daily ICU needs for the entire ensemble when compared to the single step-change. Consequently, the emergency brake results in a lower peak of ICU admissions, marked by 268 most of the predicted peak values between 700 and 900, and no outliers higher than 1,000. In the fourth scenario (Fig. 6D) , we extend the third scenario with taking progressively In the fifth scenario (Fig. 6E) , we extend the previous scenario with seasonal variation 274 in the reproduction number of SARS-CoV-2 (7) and evaluate its effect on waxing and waning In summary, for the progressively relaxing exit scenario, a large step-change results in 284 a relatively early and possibly high peak, whereas if the change is implemented in more 285 gradual, smaller step-changes, the peak in ICU admissions is delayed and lowered. In both 286 scenarios, the emergency brake is capable of limiting forecasted ICU peaks in bounds of the 287 . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint 13 1,800-peak capacity in the Netherlands, but any scenario of progressively relaxing steps is at 288 some point likely to result in acceleration of growth and the triggering of the emergency brake. resulting in a(t) which on average is only a fraction lower (~0.03) than the low scenario, but 306 grows to ~0.05 in the last year due to slightly faster build-up of group immunity. From the 307 minor gain in social relaxation, it could be argued that the low scenario should be favored, 308 considering the more positive health effects. Within an alternative scenario, we consider seasonal variation in the reproduction as 310 for the progressive relaxation scenarios (conform Fig. 6E) . The results are shown for a period 311 of two years in Fig. 8C and 8D , causes an appropriate adaptive response of the ACCC, marked 312 by significant seasonal variation in a(t), while maintaining target ICU levels. 313 . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint In conclusion, the ACCC strategy can effectively steer toward the targeted nominal 314 ICU rates, and the adaptive capability allows it to closely follow seasonal variation in R0. In 315 both the low and high scenarios, the emergency brake is never activated proving ICU needs 316 remain below 700 and 1,400 ICU respectively at all times for all ensemble runs. We analyzed two dissimilar ways to exit the lockdown that most countries worldwide 320 are enduring. A progressively relaxing strategy can initially lead to a significant reduction of 321 distancing measures but is likely related to a high peak resurgence of COVID-19 cases for 322 which inevitably an emergency brake will be needed. Alternatively, our ACCC approach A key assumption in both strategies is that we can take limited and controlled step- . CC-BY-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 20, 2020. . CC-BY-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 20, 2020. . CC-BY-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 20, 2020. Table 1 ). The estimated ICU 484 mortality rate is 30%. Coronavirus disease (COVID-19) situation report-378 51 COVID-19 coronavirus pandemic World Health Organization (WHO) regional office for Europe Impact of non-pharmaceutical 387 interventions (NPIs) to reduce COVID-19 mortality and healthcare demand With COVID-19, modeling takes on life and death 392 importance Estimates of the 394 severity of coronavirus disease 2019: a model-based analysis Projecting the transmission 397 dynamics of SARS-CoV-2 through the postpandemic period Call for 402 transparency of COVID-19 models Forecasting hospitalization and ICU rates of the COVID-19 outbreak: an efficient SEIR 405 model Nowcasting and forecasting the potential domestic and 408 international spread of the 2019-nCoV outbreak originating in Wuhan, China: a 409 modelling study Ensemble Smoother With Multiple Data Assimilation Countries by density COVID-19 infecties op de IC's Epidemiologische situatie 420 COVID-19 in Nederland 6 mei Rijksinstituut voor Volksgezondheid en Milieu (RIVM) door Jaap van Dissel [Technical briefing to parliament We thank Anke Zindler for helpful comments. . CC-BY-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 20, 2020. . CC-BY-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. . CC-BY-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 20, 2020. . CC-BY-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 20, 2020. . https://doi.org/10.1101/2020.05.16.20102947 doi: medRxiv preprint Same as panel A but with emergency brake (C) Small step-changes with emergency brake.