Validating and modeling the impact of high-frequency rapid antigen screening on COVID- 1 19 spread and outcomes 2

46 High frequency screening of populations has been proposed as a strategy in facilitating 47 control of the COVID-19 pandemic. We use computational modeling, coupled with clinical data 48 from rapid antigen tests, to predict the impact of frequent viral antigen rapid testing on COVID- 49 19 spread and outcomes. Using patient nasal or nasopharyngeal swab specimens, we demonstrate 50 that the sensitivity/specificity of two rapid antigen tests compared to quantitative real-time 51 polymerase chain reaction (qRT-PCR) are 82.0%/100% and 84.7%/85.7%, respectively; 52 moreover, sensitivity correlates directly with viral load. Based on COVID-19 data from three 53 regions in the United States and São José do Rio Preto, Brazil, we show that high frequency, 54 strategic population-wide rapid testing, even at varied accuracy levels, diminishes COVID-19 55 infections, hospitalizations, and deaths at a fraction of the cost of nucleic acid detection via qRT- 56 PCR. We propose large-scale antigen-based surveillance as a viable strategy to control SARS- 57 CoV-2 spread and to enable societal re-opening. 58 59


INTRODUCTION
compared to qRT-PCR -along with appropriate quarantine measures -can be more effective at 115 decreasing COVID-19 spread than less frequent molecular testing of symptomatic individuals. 116 Keeping in mind the realities of daily testing in resource-limited regions, we also hypothesize 117 that testing frequency can be adjusted according to the prevalence of the disease; that is, an 118 uptick in reported cases should be accompanied by more frequent testing. During the viral 119 incubation period, high infectivity correlates with a high viral load that can be detected by either 120 qRT-PCR or rapid antigen testing 18,21,26-28 . Rapid tests thus optimize diagnosis for the most 121 infectious individuals. Studies also point to the relatively small window of time during an 122 individual's incubation period in which the qRT-PCR assay is more sensitive than rapid tests 21 . 123 In this study we report the clinical validation of two direct antigen rapid tests for 124 detection of SARS-CoV-2 spike glycoprotein (S) or nucleocapsid protein (N) using 125 retrospectively collected nasopharyngeal or nasal swab specimens. Using the clinical 126 performance data, we develop a modeling system to evaluate the impact of frequent rapid testing 127 on COVID-19 spread and outcomes using a variation of a SIR model, which has been previously 128 used to model COVID-19 transmission [29][30][31][32][33][34][35] . We build on this model to incorporate quarantine 129 states and testing protocols to examine the effects of different testing regimes. This model 130 distinguishes between undetected and detected infections and separates severe cases, specifically 131 7 quarantine measures, and a previous simulation study has shown that frequent testing with 137 accuracies less than qRT-PCR, coupled with quarantine process and social distancing, are 138 predicted to significantly decrease infections 21,29,35-39 . Godio  By simulating the implementation of rapid testing strategies using parameters extracted 151 from data from realized outbreaks, we are able to expand on existing insights, including: 152 predicting the effectiveness of such schema on outbreaks with differing dynamics and at varying 153 intervention times, extracting parameters to train the comprehensive SIDHARTHE-Q model, and 154 demonstrating a method that is easily applied to fit parameters for any COVID-19 outbreak given 155 a data set including daily reports of confirmed cases, current hospitalizations, and deaths. Using 156 this method, we propose and test the effectiveness of a variety of testing strategies and analyze 8 evolving pandemic, and this is the first modeling system using publicly-available data to 160 simulate how potential public health strategies based on testing performance, frequency, and 161 geography impact the course of COVID-19 spread and outcomes. 162 Our findings suggest that a rapid test, even with sensitivities lower than molecular tests, 163 when strategically administered 2-3 times per week, will reduce COVID-19 spread, 164 hospitalizations, and deaths at a fraction of the cost of nucleic acid testing via qRT-PCR. Modern 165 surveillance systems should be well equipped with rapid testing tools to ensure that disease 166 tracking and control protocols are effective and well-tailored to national, regional, and 167 community needs. 168

Accuracy of Direct Antigen Rapid Tests Correlate with Viral Load Levels 171
Rapid antigen tests have recently been considered a viable source for first-line screening, 172 although concerns regarding the accuracy of these tests persist. We clinically validated two 173 different direct antigen rapid tests for the detection of either nucleocapsid protein (N) or spike 174 glycoprotein (S) from SARS-CoV-2 in retrospectively collected nasal or nasopharyngeal swab 175 specimens. Of the total number of nasal swab specimens evaluated by qRT-PCR for 176 amplification of SARS-CoV-2 N, S, and ORF1ab genes, 100 tested positive and 58 tested 177 negative ( Table 1). The overall sensitivity and specificity of the rapid antigen test for detection of 178 SARS-CoV-2 N, evaluated across the nasal swab specimens, was 82.0% and 100%, respectively. 179 interaction with an infected individual. Should an individual in S move into Q-U, they are 206 quarantined for 10 days before returning to S, a time period chosen based on current knowledge 207 of the infectious period of the disease and is consistent with CDC guidelines 43 . One could also 208 conceive of an effective strategy in which individuals exit quarantine after producing a certain 209 number of negative rapid tests in the days following their initial positive result or confirm their 210 negative result using qRT-PCR. Prolonged incubation beyond 10 days is assumed to be unlikely 211 -post-quarantine risk of transmission is estimated at 1% -and hence is not included in this 212 probabilistic model. 43 213 State I contains individuals who are infected but not diagnosed. Given that those 214 diagnosed are predominantly quarantined, the undiagnosed individuals in I -many of which are 215 pre-or asymptomatic -interact more with the S population than do those in Infected Detected 216 (D) and transmission due to this population is critically important to modeling outbreaks. 217 Therefore, the infectious rate for I is assumed to be significantly larger than for D. Furthermore, 218 a region's ability to control an outbreak is directly related to how quickly and effectively the 219 population in I tests into D, reducing transmission rates through quarantine. From both I and D 220 individuals may transition into Recovered (R), accounting for the many cases of infection that 221 are never detected. This study, in particular, highlights the critical role frequency of testing, 222 along with strict quarantine, has in mitigating the spread of the disease and provides specific 223 testing strategies based on rapid tests we predict to be highly effective. 224 In this model, we assume that individuals receive a positive diagnosis before developing 225 severe symptoms and that those with symptoms severe enough to be potentially fatal will go to 226 the hospital. If an individual develops symptoms, we assume they are tested daily until receiving 227 a positive result; hence, before severe symptoms develop, they will be diagnosed with high 228 . 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) These regions are also selected in our study due to the readily available surveillance data 261 provided by the local governments. We fit the model to the data from each region starting 1 April 262 2020. At this time point the disease reportedly is most advanced in NYC and least advanced in 263 SJRP, Brazil with estimated cumulative infection rates of 7.11% and 0.12%, respectively. 264 After calibrating the SIDHRE-Q model, the disease spread is observed with varying 265 validated rapid antigen test performances and frequencies (Fig. 2). Sensitivity (the ratio of true 266 positives to the total number of positives) and specificity (the ratio of true negatives to the total 267 number of negatives) compared to gold-standard qRT-PCR were used as measures of test 268 accuracy. 269 The rapid test frequency is varied while maintaining an accuracy of 80% sensitivity and 270 90% specificity, comparable to our clinical data collected in SJRP, Brazil. These testing 271 scenarios are then compared to symptomatic testing, in which individuals receive a rapid test 272 only when presenting symptoms, via either a rapid test or qRT-PCR. Since the primary testing 273 regimen deployed in MA, LA, NYC and SJRP, Brazil is qRT-PCR-based and focused on 274 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint symptomatic individuals, the symptomatic testing protocol via qRT-PCR is directly estimated 275 from the data to be the rate ߥ (Table 4). 276 The difference between the qRT-PCR and rapid test simulations (red and orange lines, 277 respectively) is therefore only sensitivity of testing (Fig. 2). Test outcome probability in this 278 model is a function only of whether an individual is infected and independent of other factors; 279 one can consider this a lower bound on effectiveness of a strategy, as sensitivity and infectivity 280 are often positively correlated with antigen testing. In this model with sensitivity s and frequency 281 of testing f, the probability an individual is diagnosed in a testing window is given by the 282 following: 283 To better understand the effect of rapid testing frequency and performance on healthcare 286 capacity and mortality rates, we simulate the testing strategy with 30%-90% sensitivity each with 287 80% or 90% specificity against the symptomatic testing strategy ( Supplementary Fig. 3). 288 As per our hypothesis, frequency and symptom-based testing dramatically reduced 289 infections, simultaneous hospitalizations, and total deaths when compared to the purely 290 symptom-based testing regimens, and infections, hospitalization, and death were reduced as 291 frequency increased. Although testing every day was clearly most effective, even testing every 292 fourteen days with an imperfect test gave an improvement over symptomatic testing with qRT-293 PCR. While the strategy works best when implemented at the very beginning of an outbreak, as 294 demonstrated by the results in SJRP, Brazil, it also works to curb an outbreak that is already 295 large, as demonstrated by the results in NYC. The difference between frequencies is more 296 noticeable when the testing strategy is applied to the outbreak in NYC, leading us to hypothesize 297 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint that smaller outbreaks require a lower testing frequency than larger ones; note the difference 298 between the dependence on frequency to curb a small initial outbreak in SJRP, Brazil versus a 299 large one in NYC (Fig. 3). 300 For test performance of 80% sensitivity and 90% specificity, the percent of the 301 population that has been infected in total from the beginning of the outbreak to mid-July drops 302 from 18% (MA), 11% (LA), 26% (NYC), and 11% (SJRP, Brazil) to 3%, 2%, 12%, and 0.26%, 303 respectively, using a weekly rapid testing and quarantine strategy (with regards to predictions of 304 overall infection rates, other studies based on seroprevalence and epidemiological predictions 305 have reached similar conclusions 49,50 ). If testing is increased to once every three days, these 306 numbers drop further to 1.6% (ΜΑ), 1.4% (LΑ), 9.5% (ΝΥC), and 0.19% (SJRP, Brazil) 307 (Supplementary Table 1). 308 To further examine the relationship between frequency and sensitivity, we model the 309 maximum number of individuals in a given state over the 105-day time period for four 310 geographic regions (Fig. 3, Supplementary Fig. 4). In all four geographic regions, as frequency 311 of testing increases, the total infections, maximum simultaneous hospitalizations, and total deaths 312 converge to small percentages regardless of the sensitivity at high frequencies. For example, the 313 predictions show that for the outbreak in LA, a testing strategy started on 1 April of every 10 314 days using a test of sensitivity 90% would have resulted in 2.5% of the population having been 315 infected, while using a test of sensitivity 30% would require a strategy of every 5 days to achieve 316 the same number. Thus, we conclude that frequency is more important than sensitivity to control 317 the outbreak using a test-based strategy, and a large range of sensitivities prove effective when 318 testing sufficiently often ( Supplementary Fig. 4-5) 29,51 . The following subsection contains a 319 discussion of a location-based method for varying the exact frequency of testing based on 320 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint evolving outbreaks. Frequency of testing can be significantly reduced to effectively contain the 321 disease once the initial outbreak has been controlled; it is clear that this takes only a matter of 322 weeks (Fig. 2). 323 On the other hand, according to the specificity of the rapid test and the quarantine 324 duration, larger testing frequency result in a larger percent of the population quarantined (Fig. 2). 325 Assuming a 90% rapid test specificity and 10-day quarantine duration, for the 1-, 3-and 7-day 326 frequencies almost 48%, 24% and 12% of the population, respectively, would be quarantined. 327 This figure may be reduced with additional rules for exiting quarantine early, such as after 328 complementary testing. An example of such a strategy is that individuals who test positive are 329 required to either quarantine for two weeks or produce two consecutive negative rapid tests in 330 the two days following their positive result. Assuming 80% sensitivity and 90% specificity, those 331 individuals will reenter the public while still infected with probability 0.04. If uninfected, that 332 individual will exit quarantine after two days with probability 0.81. However, a compromise 333 between the reduction of infections and the proportion of the population in quarantine would be 334 part of the planning for the appropriate testing protocol in each community or region. 335 While high frequency may be necessary to contain a large outbreak initially, relatively 336 infrequent testing, such as every one or two weeks, is sufficient to keep controlled outbreaks 337 small, while reducing the number of quarantined individuals to less than 10% of the population 338 using a two-week mandatory quarantine. 339 Additionally, quarantine adherence is of essential importance to the success of this 340 strategy, and we assume near-perfect quarantine compliance with a small transmission rate due 341 to diagnosed individuals (Table 4). Therefore, measures are needed to ensure quarantine is 342 widely adhered to ( Supplementary Fig. 8). Recent research has identified a number of ways to 343 . 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 13, 2021.
increase quarantine compliance, including compensating for wages lost, providing quarantine 344 facilities and effective handling of the health crisis. 52-55 345 Additionally, while high frequency may be necessary to contain a large outbreak initially, 346 relatively infrequent testing, such as every one or two weeks, is sufficient to keep controlled 347 outbreaks small, while reducing the number of quarantined individuals to less than 10% of the 348 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.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint of 1 will cause infections to rapidly drop. Both the threshold at which everyday testing begins 368 and the coefficient of log 2 T/D can be modified to produce a strategy that is more or less frequent 369 in testing or resource effective; a range of days between tests from 14 days to 1 day are used 370 ( Fig. 4a). 371 The purpose of this strategy is to tailor testing based on the specific characteristics of 372 unique outbreaks in different regions. A scan over different choices of T is shown in Fig. 4b; the 373 threshold we choose in Fig. 4a is 0.05% because it is successful in curbing the outbreak in 374 California within the time period we consider. Our analysis could be redone to select another 375 effective fine-grained strategy in other states or regions. The cost analysis is based on cost per 376 test -$7 per rapid test and $100 per PCR test -times number of tests used. Clearly that 377 calculation neglects the costs of storing, distributing, and administering tests, as well as 378 monitoring incoming results. The costs associated with these logistics would vary with differing 379 policies dictating the use of rapid tests; significantly, whether they would be administered at 380 home with self-reported results or in a testing facility or workplace for validation purposes. For 381 example, a company may choose to use the rapid tests to scan employees before allowing them 382 to enter the workplace, in a way similar to existing temperature checks. The cost of this 383 particular application would be minimal beyond that of the actual tests. Such costs would 384 inevitably be greater for PCR tests, which require a specialized testing facility, significant 385 equipment, and highly trained personnel. 386 Using a rapid test with a sensitivity of 80% and a specificity of 90%, the county-based 387 testing with threshold 0.05% reduces the active infections from 0.94% to 0.0005%, while the 388 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint uniform strategy with tests administered every 7 days results in double the number of active 389 infections (Fig. 4a). As the threshold is reduced, the total cost increases while the cumulative 390 infections, maximum percentage hospitalized, and cumulative deaths all decrease (Fig. 4b). 391 Appropriate choice of threshold is dependent on the severity of outbreaks in a specific region and 392 available resources, both logistically and fiscally. With regional data, such as that from 393 California used to produce Fig. 4b, this study can be reproduced to calculate an efficient testing 394 strategy that will effectively curb outbreaks of differing severities in any geographic entity. This 395 analysis does not include any delays in ramping testing up and down. If one were to reproduce 396 this analysis for a given testing strategy, a fixed-time delay could be introduced, depending on 397 the relevant logistical constraints. 398 Strategy B in Fig. 4 consists of qRT-PCR testing uniformly applied to the highlighted 399 population with a frequency of once weekly. The average cost per person per day is just under 400 $15. Despite this frequency and the accuracy of qRT-PCR, the strategy does not succeed in 401 curbing the spread as fast as strategy A, which uses a testing sensitivity and specificity of 80% 402 and 90%, respectively, and testing frequency that vary between counties depending on the 403 proportion of their population that is currently infected. The total cost for strategy A is estimated 404 at a fraction of the other at $1.53 per person per day. 405

DISCUSSION 407
In this study we examine the potential effects of a novel testing strategy to limit the 408 spread of SARS-CoV-2 utilizing rapid antigen test screening approaches. Our clinical data and 409 SIDHRE-Q modeling system demonstrate that 1) frequent rapid testing even at a range of 410 accuracies is effective at reducing COVID-19 spread, 2) rapid antigen tests are a viable source 411 . 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. qRT-PCR is considered the gold-standard diagnostic method because of its high sensitivity and 427 specificity, the logistical hurdles render it unrealistic for large-scale screening. 428 As qRT-PCR remains impractical for this strategy, and rapid tests are facing regulatory 429 challenges because they do not perform with qRT-PCR-like accuracy, rapid test screening is 430 either nonexistent in several countries or symptom-based. Even under best-case assumptions, 431 findings have shown that symptom and risk-based screening strategies miss more than half of the 432 infected individuals 60 . Some have argued that the need for widespread testing is overstated due 433 to the variability in test sensitivity and specificity 61 . Here, we present alternative large-scale 434 . 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) and clinical data into our modeling system has allowed us to incorporate regional differences -446 such as variances in healthcare access, state health policy and adherence, state GDP, and 447 environmental factors -under the same model. Significantly, our findings hold true across 448 Massachusetts, New York City, Los Angeles, and São José do Rio Preto, Brazil. We also present 449 the economic considerations of these testing regimes, showing that widespread rapid testing is 450 more cost efficient than less frequent qRT-PCR testing. In line with these economic 451 considerations, our model demonstrates the effectiveness of a geographic-based frequent testing 452 regime, in which high disease prevalence areas receive more frequent testing than low disease 453 prevalence areas. 454 Since COVID-19 is known to affect certain demographics differently, modeling would 455 benefit from incorporating demographic information correlated with disease progression and 456 spread to define sub-models and sets of parameters accordingly. Age, pre-existing conditions, 457 . 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. instructions regarding updated test frequency guidelines to enable adaptive testing strategies. 480 . 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 13, 2021. We emphasize that integral to the effectiveness of diagnostic schemes is 1) the proper 490 adherence to quarantine and public health measures and 2) the combined use of a variety of 491 diagnostic methods including nucleic acid, antigen, and antibody tests. According to these 492 models, rapid antigen tests are an ideal tool for first-line screening. Clinical molecular tests such 493 as qRT-PCR are vital to the diagnostic landscape, particularly to re-test suspected cases that were 494 negative on the rapid test. Because rapid tests present a higher rate of false negatives, methods 495 such as qRT-PCR remain integral to second-line screening. Antibody tests provide important 496 information for immunity and vaccination purposes as well as epidemiological surveillance. This 497 model also assumes that individuals will quarantine themselves before being tested and for 10 498 days following a positive diagnostic result and will not be infected while waiting for the qRT-499 PCR results. It is important to acknowledge the working definition of quarantine. The states 500 containing quarantined individuals (Q U , Q R and D) are defined as consisting of a population that 501 is meant to be quarantined, not a population that is necessarily in perfect compliance with the 502 mandate that they remain fully isolated from the population. Quarantine is assumed to be 503 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint imperfectly executed and the model accounts for a small, tunable interaction between 504 quarantined states and the general population, hence it is conservative. 505 There are important limitations to be considered in this model. Differences in disease 506 reporting between the geographical regions and the incomplete nature of COVID-19 surveillance 507 data, often due to the lack of testing or delays in reporting, are not considered in the model. It is 508 imperative that the testing results, hospitalization and death statistics, and changes in protocol are 509 reported in real-time to scientists and policy makers so that models can be accurately tuned as the 510 pandemic develops. Moreover, delays required to ramp testing strategies up or down are not 511 considered. Infectivity variations between individuals is also not applied to this model, and 512 future clinical studies should gather data on asymptomatic presenting COVID-19 cases. Non-513 compliant quarantine behaviors and possible infections during testing waiting times are also not 514 included in the calculations. The model also does not take into account infrastructural 515 limitations, such as hospital capacity and testing space, which depend on factors beyond the 516 scope of this analysis. Though the rapid antigen test offers several advantages such as 517 affordability, fast turnaround time, and ease of mass production, we are assuming that there are 518 systems in place to implement frequent and safe low-cost screening across different communities 519 and settings. 520 Our model underscores the need for a point-of-care or at-home test for frequent 521 screening, particularly as lockdown restrictions ease. Regulatory agencies can work towards 522 evaluating rapid tests to alternative standards other than comparison to high sensitivity molecular 523 diagnostics, as our model shows that frequency and scale of testing may overcome lower 524 sensitivities. Rather, we can refocus policy to implement first-line screening that optimizes 525 accuracy with efficiency and equitability. 526 . 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.

Development of Direct Antigen Rapid Tests for the Detection of SARS-CoV-2 530
We developed a direct antigen rapid test for the detection of the nucleocapsid protein or 531 spike glycoprotein from SARS-CoV-2 in nasal or nasopharyngeal swab specimens as previously 532

Validation of Direct Antigen Rapid Test for the Detection of SARS-CoV-2 544
In a retrospective study of nasal swab specimens form human patients, we compared the 545 accuracy of the rapid antigen test for detection of SARS-CoV-2 N to the viral loads of 546 individuals. All individuals were symptomatic between 1-10 days of fever. Nasal swab 547 specimens (n=158) were tested following approved human subjects use protocols. The nasal 548 swab specimens were banked frozen from suspected patients submitted to PATH for routine 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.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint were collected received ethical clearance from the PATH Research Ethics Committee, protocol 553 number 00004244; all participants provided written informed consent for the use of the samples. 554 The nasal swab specimens were de-identified, containing no demographic data, prior to analysis, 555 and the experiments were performed in accordance with relevant guidelines and regulations. 556 The nasal swabs were originally collected in 1 mL PBS, where 50 μ l was mixed with 50 557 μ l of Solution Buffer (0.9% NaCl and 0.1% Triton X-100). The 100 μ l mixture was then 558 pipetted onto the rapid antigen test for SARS-CoV-2 N detection and allowed to react for 15 559 minutes. After processing of the rapid antigen test, the visual positive or negative signal was 560 documented. 561 Additionally, in a retrospective study of nasopharyngeal swab specimens from human 562 patients, we compared the accuracy of the rapid antigen test to the viral load of individuals. 563 Nasopharyngeal swab specimens (n = 121) were tested in Brazil following approved human 564 subjects use protocols. The age of study participants ranged from 1 to 95 years with an overall 565 median of 37 years (interquartile range, 27-51 years), and 62% were female. All individuals 566 were symptomatic between 1-10 days of fever. The demographic summary of the patients are 567 included in Supplementary Table 2. The nasopharyngeal swab specimens were banked 568 refrigerated or frozen samples from suspected patients submitted to the lab for routine COVID 569 diagnosis. Prior to using the rapid test, the nasopharyngeal swab samples were validated by qRT-570 PCR using GeneFinder TM COVID-19 Plus RealAmp Kit (OSANGHealtcare, Anyang-si, 571 Gyeonggi-do, Republic of Korea I). The primary study under which the samples and data were 572 . 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. February 202) 1 . Although several affected US regions could have been modeled, we look at data 592 from Massachusetts, New York, and Los Angeles: these regions each contained "hotspots", or 593 areas of surging COVID-19 cases, at different points in time during the pandemic and have 594 days beginning on April 1 for Fig. 2 and Fig. 3, and 105 days beginning on April 10 for Fig. 4  596 (see "Modeling Parameters" in Methods). In order to understand the various testing proposals on 597 a global scale, we performed our clinical study in and expanded the modeling study to Brazil. 598 The specific data we use to fit our model are cumulative confirmed cases, total deaths, and 599 number of daily hospitalizations due to COVID-19. This surveillance data was retrieved from 600 government-provided online databases [66][67][68][69][70][71][72] . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint pandemic. The result of running the model with fixed infection and quarantine times as well as a 612 discussion of how that change is incorporated in Supplementary Fig. 9. The mean value method 613 of assigning values to parameters is standard in epidemiological modeling 29,73 . 614 In order to determine the numerical values of the parameters defining the flows between 615 states, we use a least squares regression to find fits for each seven day interval. All data points 616 within each interval and from each data set are fit collectively within each interval (the resulting 617 fits do not represent the mean of separately calculated fits). This procedure allows the model to 618 take into account the time dependent nature of the parameters, which rely on factors such as 619 social distancing regulations and changes in testing capacity. We also fit window sizes between 620 1 and 21 days and find that while the fit degrades with larger window size, the overall shape of 621 the curves do not change. We choose seven days assuming policy changes take a week to 622 become effective and that reasonably parameters can be expected to change within this time 623 period without causing problems with overfitting. Also, the seven day window size accounts for 624 the fact that often data is not reported as diligently over the weekend. Time series of the values 625 of the parameters for the geographic locations discussed in this paper are included in 626 Supplementary Fig. 6. 627 Given the restrictions on data available for the populations of various states, varying all 628 of the parameters results in an over parameterized system. Therefore, a subset of the model 629 parameters are fit while the others are either extracted from other sources; see Table 4. The 630 fitting procedure minimizes the sum of the squared residuals of the normalized total cases, 631 current daily hospitalizations, cumulative deaths, and percentage of total infected individuals 632 currently hospitalized. The first three are present in the data sets while the latter is derived from 633 the estimates of the ratio between infected undetected to infected detected individuals from the 634 . 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. one to see the difference in results between when a testing strategy is administered. 646 In order to determine the effectiveness of the county-based strategy when applied to the 647 state of California, we also fit all of the counties in California with a population greater than 648 1.5% of that of the entire state and with greater than zero deaths. The results do not depend on 649 these selections, but instead suggest a practical criteria to administer limited resources. The 650 fitting is done starting 10 April for these counties, as at this point the outbreak is sufficiently 651 well-documented in each to successfully model. For the county-level data we compute a seven 652 day running average of each of the data sets to which we then fit in order to smooth out 653 fluctuations in the data, likely due to reporting, which are more significant here than in the other 654 data sets considered, as the county populations are smaller and hence discrepancies impact the 655 smoothness of the data more. The fits for each of the counties can be found in Supplementary 656 Fig. 7. 657 . 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint inefficiencies in reporting. Additionally, it is difficult to gauge their consistency within the dates 659 provided or to compare between locations, as reporting mechanisms changed over time within 660 the same locations. Despite this lack of consistency, our model and fitting mechanism was 661 successful in reproducing the progress of the outbreak in each data set studied. 662

DATA AVAILABILITY 664
The authors confirm that the data supporting the findings of this study are available within the 665 article and/or its supplementary materials; any other data will be made available upon request. 666

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; 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 May 13, 2021. ; 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.

814
. 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 13, 2021. ; 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)

ߟ ߟ
is rate of transmission between D and S. It is defined as the probability that an interaction between an infected person and an uninfected person results in a new infection, multiplied by the average number of uninfected people a detected infected person comes into contact with in a given day. Units: 1/(days). accounts for a small but nonzero transmission due to the quarantined (detected) infected population. This value was chosen to be small, assuming an individual in a mandated quarantine will only interact with others with low probability, such as within a household, where complete isolation is difficult to maintain. . 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 13, 2021. ; 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 May 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint

Fig. 2. COVID-19 Outcomes in 3 US Regions and Brazil as a result of Frequent Rapid 943
Testing Protocol using the SIDHRE-Q Model. The Cumulative Detected Infected, 944 Hospitalized, Deceased, Active Infections, Recovered, and Quarantined are modeled over 105 945 days (top to bottom) using reported data from 4 global regions: Massachusetts, Los Angeles, 946 New York City, and São José do Rio Preto in Brazil (left to right). The COVID-19 population 947 spread and outcomes are modeled under a Rapid Testing Protocol (sensitivity 80%, specificity 948 90%) with variable testing frequencies ranging from 1-21 days between tests. This protocol is 949 compared to a symptom-based Rapid Testing protocol and a symptom-based PCR protocol. 950 . 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. 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 13, 2021. ; https://doi.org/10.1101/2020.09.01.20184713 doi: medRxiv preprint