key: cord-0792061-0ae5w8gx
authors: Nash, B.; Badea, A.; Reddy, A.; Bosch, M.; Salcedo, N.; Gomez, A. R.; Versiani, A.; Dutra, G. C.; Lopes dos Santos, T. M. I.; Milhim, B. H. G. A.; Moraes, M. M.; Campos, G. R. F.; Quieroz, F.; Reis, A. F. N.; Nogueira, M. L.; Naumova, E. N.; Bosch, I.; Herrera, B. B.
title: The impact of high frequency rapid viral antigen screening on COVID-19 spread and outcomes: a validation and modeling study
date: 2020-09-03
journal: nan
DOI: 10.1101/2020.09.01.20184713
sha: ac5e7070a6288ac6ac95f280621ab857fcde8ff7
doc_id: 792061
cord_uid: 0ae5w8gx

High frequency screening of populations has been proposed as a strategy in facilitating control of the COVID-19 pandemic. Here we develop a model to evaluate the impact of rapid testing on COVID-19 spread and outcomes, inspired by our clinically validated direct antigen rapid test (DART) for detection of SARS-CoV-2 spike glycoprotein. Using patient nasopharyngeal swab specimens we demonstrate that the DART sensitivity and specificity are 84.7% and 85.7%, respectively; moreover, sensitivity increases proportionally with higher viral loads. Based on surveillance data on COVID-19 from the United States and Sao Jose do Rio Preto, Brazil we show that frequent and strategic population-wide rapid testing, even at varied accuracy levels, is more effective than virus detection via polymerase chain reaction at reducing COVID-19 infections, hospitalizations, and deaths. While current policy emphasizes testing accuracy, we propose large-scale antigen-based surveillance as a vital strategy to control SARS-CoV-2 spread and to enable societal re-opening.

spread with rapid testing and model disease outcomes in three regions in the United 138 States and São José do Rio Preto, Brazil -the site of the clinical validation study -using 139 publicly available data. To date, COVID-19 modeling describes the course of disease 140 spread in response to social distancing and quarantine measures, and a previous 141 simulation study has shown that frequent testing with accuracies less than qRT-PCR, 142 coupled with quarantine process and social distancing, are predicted to significantly 143 decrease infections 12, 16, [22] [23] [24] [25] [26] . This is the first modeling system using publicly-available 144 data to simulate how potential public health strategies based on testing performance, 145 frequency, and geography impact the course of COVID-19 spread and outcomes. Our 146 findings suggest that a rapid test, even with sensitivities lower than molecular tests, when 147 strategically administered 2-3 times per week, will diminish COVID-19 spread, 148 hospitalizations, and deaths at a fraction of the cost of nucleic acid testing via qRT-PCR. The overall sensitivity and specificity of the rapid antigen test was 84.7% and 160 85.7%, respectively (Table 1 ). Our data demonstrate that the sensitivity of our test is 161 positively correlated to the viral load level (Fig. 1, Table 2 ). The rapid test result was 162 compared to the qRT-PCR cycle threshold (Ct) value measured across RdRp, N, and E 163 genes, and calculated as sensitivity and specificity. Table 1 . Clinical validation summary for the SARS-CoV-2 direct antigen rapid test (DART) evaluated using 121 retrospectively collected patient nasopharyngeal swab specimens.

. 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 September 3, 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 September 3, 2020. Positives ranks the samples in order of high qRT-PCR Ct value to low. In other words, 183 the higher the percentile, the more "positive" because fewer qRT-PCR cycles are required 184 for gene detection. Because the Ct value is a variable unit based upon qRT-PCR protocol 185 and instrumentation, we evaluated sensitivity against the percentile of cases across Ct 186 values.

As the percentile positive increases, the sensitivity between the rapid test and the 188 gold-standard qRT-PCR increases, reaching 100% sensitivity at 68.1% percentile Figure 1 . Performance of direct antigen rapid test (DART) for the detection of SARS-CoV-2 in 121 retrospectively collected patient nasopharyngeal swab specimens. Percentile Positive ranks the samples in order of high Cycle threshold (Ct) value to low Ct value. DART sensitivity is determined by calculating true positive agreement to qRT-PCR; the plot uses an ax b +c fit and 95% confidence intervals for the sensitivity. Shown are the percentile positive cases of the total positive population conditioned to qRT-PCR Ct for RdRp (RNA-dependent RNA polymerase), N (nucleocapsid), and E (envelope) genes, and the average of these genes.

. 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 September 3, 2020. (Fig. 2) . The differential equations governing the evolution of the Table 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. (which was not certified by peer review)

The copyright holder for this preprint this version posted September 3, 2020. 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint Given that those diagnosed are predominantly quarantined, individuals in I interact 218 more with the S population than do those in Infected Detected (D). Therefore, the 219 infectious rate for I is assumed to be significantly larger than for D. Furthermore, a region's 220 ability to control an outbreak is directly related to how quickly and effectively people in I 221 test into D, reducing their infectiousness through quarantine. This study, in particular, 222 highlights the critical role frequency of testing, along with strict quarantine, has in 223 mitigating the spread of the disease and provides specific testing strategies based on 224 rapid tests we predict to be highly effective.

In this model, we assume that individuals receive a positive diagnosis before 226 developing severe symptoms and that those with symptoms severe enough to be 227 potentially fatal will go to the hospital. If an individual develops symptoms, we assume 228 they are tested daily until receiving a positive result; hence, before severe symptoms 229 develop, they will be diagnosed with high probability. Those who do not develop 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint been a point of recent debate, it is assumed that the number of re-infected individuals is 241 small 27-31 . Therefore, individuals cannot transition from R to S, hence the separately 242 categorized quarantined populations. 243 We considered several variations and extensions of the SIDHRE-Q model. In 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint After calibrating the SIDHRE-Q model, the disease spread is observed with varying 263 validated rapid antigen test performances and frequencies (Fig. 3) . Sensitivity (the ratio 264 of true positives to the total number of positives) and specificity (the ratio of true negatives 265 to the total number of negatives) compared to gold-standard qRT-PCR were used as 266 measures of test accuracy.

The rapid test frequency is varied while maintaining an accuracy of 80% sensitivity 268 and 90% specificity, comparable to our clinical data collected in SJRP, Brazil. These PCR is directly estimated from the data to be the rate (Table 5 ).

The difference between the qRT-PCR and rapid test simulations (red and orange 275 lines, respectively) is therefore only sensitivity of testing (Fig. 3) . We assumed that test 

. 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint 280 . 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 September 3, 2020. To better understand the effect of rapid testing frequency and performance on 283 healthcare capacity and mortality rates, we simulate the testing strategy with 30%-90% 284 sensitivity each with 80% or 90% specificity against the symptomatic testing strategy. . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint As per our hypothesis, frequency and symptom-based testing dramatically 291 reduced infections, simultaneous hospitalizations, and total deaths when compared to the 292 purely symptom-based testing regiments, and infections, hospitalization, and death were 293 reduced as frequency increased. Although testing every day was clearly most effective, 294 even testing every fourteen days with an imperfect test gave an improvement over 295 symptomatic testing with qRT-PCR. While the strategy works best when implemented at 296 the very beginning of an outbreak, as demonstrated by the results in SJRP, Brazil, it also 297 works to curb an outbreak that is already large, as demonstrated by the results in NYC.

The difference between frequencies is more noticeable when the testing strategy is 299 applied to the outbreak in NYC, leading us to hypothesize that smaller outbreaks require 300 a lower testing frequency than larger ones; note the difference between the dependence 301 on frequency to curb a small initial outbreak in SJRP, Brazil versus a large one in NYC 

For test performance of 80% sensitivity and 90% specificity, the percent of the 304 population that has been infected in total from the beginning of the outbreak to mid-July 305 drops from 18% (MA), 11% (LA), 26% (NYC), and 11% (SJRP, Brazil) to 3%, 2%, 12%, 306 and 0.26%, respectively, using a weekly rapid testing and quarantine strategy (with 307 regards to predictions of overall infection rates, other studies based on seroprevalence 308 and epidemiological predictions have reached similar conclusions 32,33 ). If testing is 309 increased to once every three days, these numbers drop further to 1.6% (ΜΑ), 1.4% (LΑ), (Table 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 September 3, 2020. To further examine the relationship between frequency and sensitivity, we modeled 316 the maximum number of individuals in a given state over the 105-day time period for four 317 geographic regions (Fig. 4) . In all four geographic regions, as frequency of testing 318 increases, the total infections, maximum simultaneous hospitalizations, and total deaths 319 converge to small percentages regardless of the sensitivity at high frequencies. It is clear 320 that the difference in frequency required to achieve the same result using tests of differing 321 sensitivities is very small (Fig. 4) For example, we predict that for the outbreak in LA, a 322 testing strategy started on 1 April of every 10 days using a test of sensitivity 90% would 323 have resulted in 2.5% of the population having been infected, while using a test of 324 sensitivity 30% would require a strategy of every 5 days to achieve the same number. 325 Thus, we conclude that frequency is more important than sensitivity in curbing the spread, 326 and a large range of sensitivities prove effective when testing sufficiently often. How . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint testing can be significantly reduced to effectively contain the disease once the initial 330 outbreak has been controlled; it is clear that this takes only a matter of weeks (Fig. 3) . 331 On the other hand, according to the specificity of the rapid test and the quarantine 332 duration, larger testing frequency result in a larger percent of the population quarantined 333 (Fig. 3) . Assuming a 90% rapid test specificity and 14-day quarantine duration, for the 1- Additionally, while high frequency may be necessary to contain a large outbreak 345 initially, relatively infrequent testing, such as every one or two weeks, is sufficient to keep 346 controlled outbreaks small, while reducing the number of quarantined individuals to less 347 than 10% of the population using a two-week mandatory quarantine.

. 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 September 3, 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 September 3, 2020. (1) 369 The days between tests are chosen such that the detected active infections should 370 remain near to or below T. If the initial detected active infections are greater than T, then 371 the testing frequency of 1 will cause infections to rapidly drop. Both the threshold at which 372 everyday testing begins and the coefficient of log2T/D can be modified to produce a 373 strategy that is more or less frequent in testing or resource effective; a range of days 374 between tests from 14 days to 1 day are used (Fig. 5) . A scan over different choices of T 375 is shown in the supplements to Figure 5 ; the threshold we choose in Figure 5 Using a rapid test with a sensitivity of 80% and specificity of 90%, the county-based 384 testing with threshold 0.05% reduces the active infections from 0.94% to 0.0005%, while 385 the uniform strategy with tests administered every 7 days results in double the number of 386 active infections (Fig. 5) . As the threshold is reduced, the total cost increases while 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.

The copyright holder for this preprint this version posted September 3, 2020. In this study we examine the potential effects of a novel testing strategy to limit the 401 spread of SARS-CoV-2 utilizing rapid antigen test screening approaches. Our clinical data 402 and SIDHRE-Q modeling system demonstrate that 1) frequent rapid testing even at a 403 range of accuracies is effective at reducing COVID-19 spread, 2) rapid antigen tests are 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 September 3, 2020. 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 September 3, 2020. 440 We build upon these findings to show how in affected United States and Brazil regions, 441 population-wide frequent and rapid testing schemes, with sensitivities ranging from 30%-442 90%, can be more effective in curbing the pandemic than a PCR-based scheme.

Integrating real-world surveillance and clinical data into our modeling system has allowed 444 us to incorporate regional differences -such as variances in healthcare access, state 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint categories, each of which influence the risk of contracting and/or dying from COVID-19. 458 Further studies may benefit from incorporating these ideas should more information 459 become available.

Our findings also point to low-cost tools for implementation of this testing strategy, 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 September 3, 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint pandemic develops. The model also does not take into account infrastructural limitations 504 such as hospital capacity. Though the rapid antigen test offers several advantages such 505 as affordability, fast turnaround time, and ease of mass production, we are also assuming 506 that there are systems in place to implement frequent and safe low-cost screening across 507 different communities and settings. 

We developed a direct antigen rapid test for the detection of the spike glycoprotein 518 from SARS-CoV-2 in nasopharyngeal swab specimens as previously described 41 . Briefly, 519 the rapid antigen test is an immunochromatographic format with a visual readout using 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint at the test area reflects SARS-CoV-2 spike glycoprotein that is "sandwiched" between an 527 anti-spike glycoprotein antibody adsorbed to the nitrocellulose membrane and a second 528 anti-spike glycoprotein antibody covalently coupled to visible gold nanoparticles. 

A nasopharyngeal swab specimen (1 mL) was concentrated using a Vivaspin 500 546 centrifugal concentrator (Sartorius, Goettingen, Germany) at 12,000 x g for 10 minutes.

The concentrated nasopharyngeal swab specimen retentate was transferred to a 548 collection tube and the rapid antigen test was inserted into the tube with the retentate and 549 . 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 September 3, 2020. 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint

(2) 570 In order to determine the values of the parameters defining the flows between states, we 571 use a least squares regression performed at seven day intervals in the datasets to which 572 we fit. This allows the model to take into account the time dependent nature of the 573 parameters, which rely on factors such as social distancing regulations and changes in 574 testing capacity. We also fit window sizes between 1 and 21 days and find that while the 575 fit degrades with larger window size, the overall shape of the fits do not change. We 576 choose seven days assuming policy changes take a week to become effective and that 577 reasonable parameters can be expected to change within this time period. Also, the seven 578 day window size accounts for the fact that often data is not reported as diligently over the Table 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint is the probability that an interaction between an infected person and an uninfected person results in a new infection, divided by the average number of uninfected people a detected infected person comes into contact with on a given day. = 0.01 ⋅ The constant relating , accounts for a small but nonzero transmission due to the quarantined (detected) infected population. This value was chosen to be small, assuming a quarantined individual will only infect others with low probability.

is the probability that a symptomatic undetected individual is diagnosed on a given day. is estimated from the data. is multiplied by sensitivity (assume benchmark sensitivity 100% for PCR, as used when fitting). is the probability that an asymptomatic undetected infected individual is diagnosed on a given day. = 0 while fitting (during PCR symptomatic testing). =(sensitivity/days between tests) when the rapid testing strategy is activated.

is the probability that an undetected infected individual transitions to the recovered state on a given day.

= 1/14, or the inverse of average recovery time. 48 is the probability that an infected individual develops severe symptoms on a given day and transitions into the hospitalized state. The flow from to is assumed to be independent of the ratio / , but comes only from the detected infected population, hence why it is multiplied by ( + )/ .

is estimated from the data. is the probability that a detected infected individual transitions to the recovered state on a given day.

= 1/14, or the inverse of the average recovery time. 48 . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint is the probability that a hospitalized individual transitions to the recovered state on a given day. = 1/11, or the inverse of the average recovery time for a hospitalized individual. 48 is the probability that a hospitalized individual expires on a given day. is estimated from the data. is the probability of entering either of the quarantine states on a given day from either the Susceptible or Recovered populations. = 0 while fitting (during PCR symptomatic testing). = (1 −specificity) × (1/days between tests) when the rapid testing strategy is activated.

is the probability that an individual exits quarantine on a given day.

= 1/14, or the inverse of the quarantine period for fixed length quarantine.

The fitting procedure minimizes the sum of the squared residuals of the total cases, We consider the data sets for outbreaks in MA, NYC, LA, and SJRP, Brazil 42-47 .

While each location has testing and fatality information dating back to January, 599 hospitalization data was not included until late March (for NYC and SJRP) and April (for 600 MA and LA). Hence we begin our fitting procedure and testing strategy on 1 April for 601 each of the data sets; by this point, the outbreak is advanced in NYC, substantial in MA, 602 non-negligible, but far from its peak, in LA, and in early stages in SJRP, Brazil. Starting 603 . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint simulations at various stages of the outbreak allows one to see the difference in results 604 between when a testing strategy is administered. 605 In order to determine the effectiveness of the county-based strategy when applied 606 to the state of California, we also fit all of the counties in California with a population 607 greater than 1.5% of that of the entire state and with greater than zero deaths. The results 608 do not depend on these selections, but instead suggest a practical criteria to administer 609 limited resources. The fitting is done starting 10 April for these counties, as at this point The authors confirm that the data supporting the findings of this study are available within 624 the article and/or its supplementary materials; any other data will be made available upon 625 request.

Our code can be found on github:

626 . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint spread of the coronavirus disease 2019 (COVID-19) taking into account the undetected 650 . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 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. (which was not certified by peer review)

The copyright holder for this preprint this version posted September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint

The New York Times (2020). 728 . 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 September 3, 2020. . https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint

We thank Professor Lee Gehrke for critical reading of the manuscript. The study is 631 funded, in part, by a Bill and Melinda Gates Foundation Award (INV-017872) to E25Bio

EN is funded by Tufts University DISC Seed Grant. MLN is supported by a FAPESP 633 grant (#2020/04836-0) and is a CNPq Research Fellow. AFV is supported by a FAPESP 634 Fellow grant (#18/17647-0). GRFC is supported by a FAPESP Fellow grant

The funders had no 636 role in the design of the study

Inc. (www.e25bio.com), a company that develops diagnostics for epidemic viruses

Estimating the number of undetected COVID-19 cases exported 647 internationally from all of China. medRxiv

Mathematical modeling of the infections. The case of China

COVID-19 epidemic in Switzerland: on the importance of testing, contact 653 tracing and isolation

Evaluating the massive underreporting and undertesting of COVID-19 cases in 655 multiple global epicenters

Using influenza surveillance networks to 657 estimate state-specific prevalence of SARS-CoV-2 in the United States

Estimating the undetected infections in 660 the Covid-19 outbreak by harnessing capture-recapture methods

Development of a reverse transcription-loop-mediated isothermal 663 amplification as a rapid early-detection method for novel SARS-CoV-2

Mylab gets commercial approval from ICMR for Covid-19 antigen rapid testing kit

Rapid antigen tests now make up nearly half of daily checks | 669 India News -Times of India. The Times of India 670

Analytical sensitivity and efficiency comparisons of SARS-CoV-2 RT-673 qPCR primer-probe sets

Superspreading SARS events

Clinical progression and viral load in a community outbreak of 679 coronavirus-associated SARS pneumonia: a prospective study

Temporal dynamics in viral shedding and transmissibility of COVID-19

Modelling the COVID-19 epidemic and implementation of population-wide 684 interventions in Italy

Estimating the reproductive number and the outbreak size of COVID-19 in

Fitting and forecasting the trend of COVID-19 by SEIR(+CAQ) dynamic

Modified SEIR and AI prediction of the epidemics trend of COVID-19 in China 690 under public health interventions

Study on the epidemic development of COVID-19 in Hubei 692 province by a modified SEIR model

Spatial-temporal distribution of COVID-19 in China and its 695 prediction: A data-driven modeling analysis

SEIR Modeling of the Italian Epidemic of SARS-CoV-2

Using Computational Swarm Intelligence

Spread and dynamics of the COVID-19 epidemic in Italy: Effects of emergency 699 containment measures

The effectiveness of quarantine of Wuhan city against the Corona Virus Disease 701 2019 (COVID-19): A well-mixed SEIR model analysis

Preliminary prediction of the basic reproduction number of the Wuhan novel 703 coronavirus 2019-nCoV

Forecasting COVID-19-Associated Hospitalizations under Different Levels of

Results from an

Extended SEIR Compartmental Model

Risk of reactivation or reinfection of novel coronavirus (COVID-19)

Clinical recurrences of COVID-19 symptoms after recovery: viral relapse, 710 reinfection or inflammatory rebound?

Primary exposure to SARS-CoV-2 protects against reinfection in rhesus 712 macaques

Will we see protection or reinfection in COVID-19?

Estimation of the Probability of Reinfection With COVID-19 by the

Susceptible-Exposed-Infectious-Removed-Undetectable-Susceptible Model

COVID-19 Projections Using Machine Learning. COVID-19 Projections Using Machine 718 Learning

Seroconversion of a city: Longitudinal monitoring of SARS-CoV-2 720 seroprevalence in New York City

s Slow Start to Coronavirus Testing: A Timeline. The 723

The Lost Month: How a Failure to Test Blinded the U.S. to Covid-19. The 725

Despite Promises, Testing Delays Leave Americans 'Flying Blind

Epidemiological and clinical characteristics of the COVID-19 epidemic 729 in Brazil

It's Like Having No Testing': Coronavirus Test Results Are Still 731

The New York Times

Estimated 733 effectiveness of symptom and risk screening to prevent the spread of COVID-19

The Appropriate Use of Testing for COVID-19

Rapid antigen tests for dengue virus serotypes and Zika virus in patient serum

Massachusetts Department of Public Health. COVID-19 Response Reporting

California Department of Public Health. COVID-19 Cases -California Open Data

California Department of Public Health. COVID-19 Hospital Data -California Open Data

Department of Health and Human Hygiene. COVID-19 Daily Counts of Cases

Sao Jose do Rio Preto Public Health Office. COVID-19 Surveillance Data

New York State Government. Daily Hospitalization Summary by Region

COVID-19)

Non-Linear Least-Squares Minimization and Curve-Fitting for Python -Non-Linear Least

Squares Minimization and Curve-Fitting for Python

Matplotlib: Python plotting -Matplotlib 3.3.1 documentation

Plotly: The front-end for ML and data science models