Controlling SARS-CoV-2 in schools using repetitive testing strategies
Abstract
SARS-CoV-2 remains a worldwide emergency. While vaccines have been approved and are widely administered, there is an ongoing debate whether children should be vaccinated or prioritized for vaccination. Therefore, in order to mitigate the spread of more transmissible SARS-CoV-2 variants among children, the use of non-pharmaceutical interventions is still warranted. We investigate the impact of different testing strategies on the SARS-CoV-2 infection dynamics in a primary school environment, using an individual-based modelling approach. Specifically, we consider three testing strategies: (1) symptomatic isolation, where we test symptomatic individuals and isolate them when they test positive, (2) reactive screening, where a class is screened once one symptomatic individual was identified, and (3) repetitive screening, where the school in its entirety is screened on regular time intervals. Through this analysis, we demonstrate that repetitive testing strategies can significantly reduce the attack rate in schools, contrary to a reactive screening or a symptomatic isolation approach. However, when a repetitive testing strategy is in place, more cases will be detected and class and school closures are more easily triggered, leading to a higher number of school days lost per child. While maintaining the epidemic under control with a repetitive testing strategy, we show that absenteeism can be reduced by relaxing class and school closure thresholds.
Editor's evaluation
This paper evaluates different testing strategies on the SARS-CoV-2 transmission dynamics in a primary school environment and shows that repetitive testing significantly reduces the infection attack rates in the schools. It provides insights into policy design to keep schools open as much as possible in the era of transition from COVID pandemic to endemic.
https://doi.org/10.7554/eLife.75593.sa0Introduction
The SARS-CoV-2 pandemic has caused over 200 million COVID-19 cases and over 4 million deaths around the world up to September 2021 (World Health Organization, 2022). Although vaccines have been approved, even for young children, there is an ongoing debate whether such age classes should be vaccinated or prioritized for vaccination (Wong et al., 2021). While the contribution of children in the COVID-19 epidemic is still subject to discussion (Gaythorpe et al., 2021), there is a consensus that more infectious variants can cause significant outbreaks among children (Milne et al., 2022). Furthermore, recent work indicates that children, who typically undergo an infection with little or no symptoms, might still be highly contagious and as such generate new infections in the community (Meuris et al., 2021). As alternative to a vaccination-based strategy, the only means to mitigate outbreaks of SARS-CoV-2 in primary schools, is through non-pharmaceutical interventions, including the use of masks, social distancing, hygienic precautions, and diagnostic testing. Here, the aim of diagnostic testing is to detect and subsequently isolate infected individuals. Therefore, it is important to advance our understanding on how different testing strategies impact primary schools, considering the evolution of SARS-CoV-2 contagiousness through different variants of concerns (VOCs). When defining intervention policies in a school setting, attention needs to be devoted to limiting the number of school days lost. In fact, Engzell et al. evaluated the impact of school closures on students’ learning performance, finding that students of age 8–11 years made less progress while learning from home (Engzell et al., 2021). Several scientific investigation discussed the use of testing strategies in school settings, for example (Colosi et al., 2022; Leng et al., 2022; GOV UK, 2021a; GOV UK, 2021b; Paltiel and Schwartz, 2021; Chang et al., 2020a; Hamer et al., 2021), suggesting that a repetitive testing strategy reduces transmissions in a school context but increases absenteeism. In this work, we explore the effectiveness of testing strategies in a primary school setting by varying factors related to the considerate strategies and school environment, and by testing viral and immunological characteristics representing different SARS-CoV-2 VOCs. To do so, we construct an individual-based model that explicitly represents a set of primary school pupils. These pupils are allocated to a fixed set of classes and are taught by a fixed set of teachers. Through this micro-model, we perform a fine-grained evaluation of testing strategies, keeping track of both the attack rate and the number of school days lost. We conduct experiments considering different R0 values to reflect the increase in infectiousness exhibited by the Delta VoC and we vary the incubation period and the proportion of immune individuals to represent the surge of the Omicron VoC. In addition, we investigate the impact of class and school closure thresholds, incubation period, proportion of symptomatic infections, school size and seeding frequency.
Results
To investigate the efficacy of the testing strategies, we consider both the attack rate (i.e., proportion of the infections generated in the school population, excluding seeded cases) and the average number of school days lost per child (NSDL). In order to differentiate between different phases of the epidemic, we first compare two scenarios that represent the Wuhan strain and the Delta VoC, characterized by a different transmission potential given a contact. Further, we present the case of the Omicron VoC, where we reduce the proportion of immune individuals and we consider a shorter incubation period.
On the one hand, our modelling experiments (details on the simulation model in the Methods section) show (Figure 1) that symptomatic isolation results in the infection of a significant proportion of the school population (Wuhan strain, median: 0.03, 95% quantile interval [0.01,0.04]; Delta VoC, median: 0.14, 95% quantile interval [0.10,0.18]). This comes as no surprise, as in our model we assume that 80% of the pupils will go through the infection asymptomatically, and by following this testing policy, we are only able to pick up infections that make up the tip of the iceberg. On the other hand, the attack rate consistently decreases when a testing policy is used that performs a wider screening of the school population, such as reactive testing and repetitive testing. Such policies enable the detection of both symptomatic and asymptomatic cases, that can be subsequently isolated, thereby limiting further transmissions. Among the two screening options, we observe that repetitive approach is the strategy that most reduce the attack rate (Wuhan strain, median: 0.02, 95% quantile interval [0.01,0.03]; Delta VoC, median: 0.05, 95% quantile interval [0.04,0.07]). However, contrary to intuition, our experiments indicate that the reactive screening strategy performs only slightly better than symptomatic isolation (Wuhan strain, median: 0.03, 95% quantile interval [0.01,0.04]; Delta VoC, median: 0.12, 95% quantile interval [0.10,0.17]). This can be explained by the low probability that pupils will be symptomatic when infected, hence a low probability to trigger the reactive screening. When we assume that 80% of infections in children progress asymptomatically, we can expect (by assuming a geometric distribution) that four asymptomatic generations take place, on average, before a symptomatic infection is observed. Therefore, when a reactive screening procedure is triggered by a symptomatic individual, the infected individuals that share a class with this individual might already be recovered or in the end phase of their infectious period. To confirm this reasoning, we simulated a multiple class screening strategy that is triggered when a pupil tests positive. We notice similar attack rates when the screening procedure is repeated (Appendix 1—figure 4). Hence, on average, only a limited number of generations can be avoided by employing a reactive screening strategy, when the infection is predominantly driven by asymptomatic carriers. Note that we also assume that only a limited percentage of symptomatic children is detected (30%), due to the fact that many children exhibit only minor symptoms.

We show the base scenario for the Wuhan strain (left panel) and Delta VoC (right panel) for a moderate seeding of 5 seeds per week.
In each panel, we consider three testing strategies: symptomatic testing (SI), symptomatic testing in combination with reactive screening (ReaS) and repetitive screening (RepS). For each of the testing strategies we show a boxplot of the attack rate (green boxplot) and NSDL (orange boxplot) together with their mean values (respectively, yellow and blue dots). The epidemic is simulated for 100 days.
To interpret the experimental results with respect to the average number of school days lost per child (NSDL), we need to recognize that children can miss school due to isolation when infected or due to quarantine due to a high risk contact. For the symptomatic isolation strategy, only children with symptoms are isolated, resulting in an average NSDL per child that is directly proportional to the fraction symptomatic cases (Wuhan strain, median: 0.08, 95% quantile interval [0.03,0.3]; Delta VoC, median: 0.3, 95% quantile interval [0.07,0.63]). For reactive screening, additional asymptomatic pupils might be identified, thereby quickly reaching the class or school closure thresholds, with a higher NSDL as a result (Wuhan strain, median: 0.59, 95% quantile interval [0.11,1.34]; Delta VoC, median: 1.99, 95% quantile interval [0.68,11.58]). This effect is most pronounced when we apply repetitive testing, where we effectively detect a high proportion of the infections, thereby rapidly meeting the class and/or school closure thresholds, with a very high NSDL as a consequence (Wuhan strain, median: 13.85, 95% quantile interval [3.50,31.93]; Delta VoC, median: 42.18, 95% quantile interval [32.97,51.64]).
We note that the high NSDL associated with repetitive testing, renders this testing policy impractical. We argue that, by using repetitive testing, more lenient thresholds could be applied, as we are able to detect a larger proportion of cases. We investigate this in Figure 2 where we remove the school closure threshold, and investigate a set of class closure thresholds while considering a repetitive testing strategy. This experiment confirms that a larger class threshold can be used, with only a limited impact on the attack rate, and that such thresholds result in a more acceptable NSDL. A more stringent school threshold shows a positive effect on controlling the attack rate, but drastically increases the NSDL (Appendix 1—figure 9). While the overall trends of the reported measures are similar for the Wuhan and Delta scenarios, the difference between the testing strategies is most pronounced in case of the more infectious virus strain (i.e., the Delta VoC). In our Omicron scenario with low immunity levels and a shorter incubation period, we observed high attack rates that are more difficult to control (Appendix 1—figure 20). To further reduce the number of infections, a twice weekly testing could be considered (Appendix 1—figure 21).

We show the repetitive testing strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different class closure thresholds, and no school closure threshold.
The epidemic is simulated for 100 days. This experiment shows that when repetitive testing is in place, a higher class closure threshold has little effect on the attack rate, yet it significantly reduces the NSDL (Table 1).
Median and 95% Quantile Interval - Class closure threshold scenario.
Class Closure Threshold | Summary Measure | Median | 95% Quantile Interval |
---|---|---|---|
2 | Attack Rate | 0.08 | [0.07,0.11] |
4 | Attack Rate | 0.09 | [0.07,0.12] |
6 | Attack Rate | 0.09 | [0.07,0.12] |
8 | Attack Rate | 0.09 | [0.07,0.12] |
No Threshold | Attack Rate | 0.09 | [0.07,0.11] |
2 | NSDL | 9.5 | [7.86, 11.16] |
4 | NSDL | 3.6 | [2.50, 4.81] |
6 | NSDL | 2.01 | [1.46, 2.83] |
8 | NSDL | 1.71 | [1.36, 2.17] |
No Threshold | NSDL | 1.61 | [1.39, 1.91] |
In order to challenge some of the assumptions of this study, we conduct a series of sensitivity analyses. We show these results in the Supplementary Information and briefly report the main findings here. We investigate the impact of the amount of weekly introductions, by seeding 1 and 10 cases on a weekly basis, next to the baseline scenario of 5 cases. We note that the impact of additional seeding cases amplifies the attack rate, but overall repetitive testing proved to be robust in regard of this parameter (Appendix 1—figure 1). Futher, we notice that a higher attack rate is detected in school with smaller sizes when the seeding number is kept the same (Appendix 1—figure 17). When introducing a number of seeds proportional to the school size, the attack rate and NSDL are similar, but a higher stochasticity is observed for the smaller school size (Appendix 1—figure 18). Furthermore, the high efficacy of repetitive testing is also observed when varying the level of contact reduction between classes, when considering different levels of immunity in children and adults, in a low prevalence setting and when a high probability of symptomatic infections is considered (Appendix 1—figures 5–7, 12 and 13). Next, we assume that asymptomatic individuals are as infectious as symptomatic individuals, as recently observed by Meuris et al., 2021. Also in this case, the trends of attack rate and number of school days lost among the different testing strategies are consistent with the baseline scenarios reported above (Appendix 1—figure 8). Repetitive testing and reactive screening are less effective when the turnaround time is increased (Appendix 1—figure 15). While reactive screening performs similar to symptomatic isolation for a turnaround time of 3 days, repetitive screening is still the strategy that is most successful to reduce the number of infections. In addition, testing strategies show a lower performance when a shorter incubation period is considered (Appendix 1—figure 14). We also consider a repetitive testing scenario where we test the entire school population twice per week, which shows that such a strategy squashes a highly contagious epidemic such as driven by the Delta VoC (Appendix 1—figures 2 and 3) and reduces the attack rate of an immune-evasive VoC with shorter incubation period, as the Omicron VoC (Appendix 1—figure 21). In addition, we noticed a decrease in the NSDL when assumptions on school and class closure are relaxed for the Delta VoC (Appendix 1—figure 3) compared to a single repetitive testing strategy. However, when school and class thresholds are present and the Delta Voc considered (Appendix 1—figure 2), or when the thresholds are relaxed and the Omicron VoC assumed (Appendix 1—figure 21), the NSDL increases if testing twice per week. Considering a repetitive testing strategy, we also tested the compliance to testing, showing that attack rate decreases and NSDL increases when compliance is increased (Appendix 1—figure 16). Interestingly, in our experimental setting, a compliance of 60% leads to a similar attack rate than a compliance of 100%.
Discussion
This simulation study compares the efficacy of testing strategies for mitigating COVID-19 outbreaks in a school setting. We evaluated such strategies computing both the attack rate and the number of school days lost. The former quantity is related to the risks of importations into households and communities, and of complications from infection, e.g. long COVID and Multysystem Inflammatory Syndrome, while the latter to educational disruption.
Throughout all simulated scenarios, a repetitive testing procedure is shown to be most efficient to reduce the attack rate. Simulations indicate that such a testing strategy limits the number of transmission events even when no class and school closures are in place. The low efficacy of the symptomatic testing and reactive screening procedures is explained by the asymptomatic nature of SARS-CoV-2 infections, especially for children. In fact, when surveillance is based just on the onset of symptoms, asymptomatic carriers avoid detection and intervention, sustaining the spread of the virus.
Class and school closures affect the number of school days lost of healthy children. To limit the learning loss caused by such closures, a control strategy in which only infected cases are isolated would be optimal. This is the aim of the a repetitive testing strategy for which no school or class thresholds are considered. Within our experimental settings, we observe that repetitive testing can keep transmission under control and limit the number of school days lost.
In our experiments, we consider PCR tests as gold standard, as we argue that the available testing infrastructure is most appropriate for performing reactive and repetitive screening procedures. To make this procedure more efficient, a class pooling approach could be used to reduce the number of samples to be analyzed (Libin et al., 2021). To further reduce the number of required PCR tests, the use of a repetitive testing strategy can be targeted to areas where prevalence is particularly high.
The viral input parameters chosen in the simulation study were set to describe the spreading of COVID-19. However, other infectious diseases can easily be represented by incorporating the specific transmission characteristics of the respective pathogens in the simulator. Especially in the case of emerging epidemics or pandemics with higher contagiousness in child-to-child interactions and/or a higher severity for children, an appropriate testing strategy in a school setting is pivotal to dampen epidemic spread. By using the simulation model presented in this paper, ad-hoc testing strategies can be easily simulated offering valuable insights in controlling epidemics.
We assume that teachers are allocated to specific classes and are assumed to interact only with individuals with whom they share the same class. This means that the interaction between teachers in the school environment is limited. We argue that this is a reasonable assumption at this stage of the epidemic, where a large proportion of teachers is either immune or vaccinated. In order to add such functionality to the model, an additional contact structure could be added to the model in which teachers meet, that is, a teacher room, to be informed by the contact frequencies between adults in a school environment (Verelst et al., 2021).
In the baseline scenario, we assume perfect compliance by school individuals for both the reactive and repetitive screening. We argue that this is a reasonable assumption, as the threshold for participating in saliva sampling is low, and societal awareness and support for this policies can be achieved, via prompt governmental communication. Nonetheless, we investigate the effect of compliance to testing for the repetitive testing strategy. Interestingly, in our experimental setting a similar attack rate is observed for a compliance level of 60% and 100%.
Methods
Individual-based primary school model
We construct an individual-based model to describe COVID-19 outbreaks in a primary school setting, which we briefly introduce in this section We refer to the Supplementary Information for a full description of the model. Children are assigned to classes and we simulate interactions among children both within and between classes. Teachers are allocated to specific classes and are assumed to interact only with individuals with whom they share the same class. We assume that symptomatic individuals develop symptoms at the peak of their infectiousness, at which they can be detected and placed in isolation for 10 days. We implement three testing policies aimed at mitigating school outbreaks:
Symptomatic Isolation (SI). Symptomatic individuals are detected with probability and tested. Individuals that test positive are put in isolation.
Reactive Screening (ReaS). Symptomatic individuals are detected with probability and tested. Individuals that test positive are put in isolation. In addition, all members of the class where this case originates from are also tested. When any additional cases are detected, these individuals are also put in isolation.
Repetitive Screening (RepS). All of the school’s members are tested on a repetitive basis once per week. All individuals that test positive are put in isolation.
All testing policies will close a class when the number of infections in this class exceeds two cases. Analogously, all testing policies will close the school when the number of infections over all classes exceeds 20 cases. When the class, or school, threshold is triggered, the respective class, or the entire school, is closed for 10 days. The length of isolation and of class/school closure is set according to viral clearance observations (Chang et al., 2020b), and in line with isolation policies in place in European countries in the first half of 2021. Infection counts are recorded in a 14-day time window to determine class and school closures. We assumed a weekly screening as the baseline scenario because a strategy based on a single test can be more easily applied at a national level when a high amount of tests need to be quickly analyzed. However, we also consider a repetitive screening strategy based on twice weekly testing. The assumptions on class and school thresholds, and on the frequency of weekly testing are challenged in a sensitivity analysis, which we discuss in the Results section.
Experimental framework
Model parameters are set to describe COVID-19 spreading. We represent both the Wuhan strain of SARS-CoV-2 and the Delta variant by setting different transmission potentials given a contact, informing such values from the literature (Li et al., 2020; Burki, 2021). We consider a distinct detection probability of symptoms for children () and adults (), as children typically exhibit mild symptoms that are easily overlooked (Sinha et al., 2020). Children are set to be halve as susceptible as adults (Davies et al., 2020). We assume that 30% of school children are immune due to prior infection, and that 90% of the teachers are immune, due to their vaccination status or due to prior infection (Sciensano, 2022).
The simulated testing procedure accounts for the use of PCR tests on saliva or throat washing samples. The sensitivity of such tests is set to 86%, and there is a one day delay in reporting the result (Butler-Laporte et al., 2021). Recent reports show that the performance of saliva sampling in combination with PCR testing is on par with nasopharyngeal swab sampling in combination with PCR testing (Wyllie et al., 2020). We assume full compliance to testing, that could potentially be reached since saliva sampling is less invasive compared to other specimen collection procedures. Infectious individuals become PCR detectable 2 days after infection, as previously assumed (Torneri et al., 2020; Torneri et al., 2021). For the reactive screening testing policy, we assume that there is a one day screening delay.
Every simulated week, five susceptible children are assumed to acquire infection outside the school environment, accounting for disease importation or seeding. The epidemic is simulated for 100 days and we consider an ensemble of 100 simulation runs to present our final results. The number of simulations was selected allowing for producing clear and stable results, and we show the full distribution of the different statistics, such that the reader can directly interpret the full scope of the simulation results. For each simulated outbreak, we compute two summary measures that account, for the number of transmissions at school and absenteeism, respectively. The former is defined as the total number of cases (minus the index cases) divided by the the number of pupils in the school, and we refer to this quantity as the attack rate. The latter is defined as the sum of the school day lost divided by the school size, and we refer to this quantity as number of school days lost (NSDL).
Appendix 1
Individual-based school model
Transmission model
We extend the SARS-CoV-2 transmission model presented by Torneri et al., 2021 to investigate school settings. In this model, the infection dynamic is described as follows. Individuals are initially susceptible and once infected, they enter the exposed stage. The infection can be asymptomatic or symptomatic. Symptomatic individuals develop symptoms after a pre-symptomatic period. Each symptomatic individual is assumed to show symptoms at the peak of their infectiousness, as indicated by literature findings (Sun et al., 2021; He et al., 2020). After infection, individuals will eventually recover, after which they are assumed to be immune to reinfection.
Infection events are simulated with a counting process approach. First, contacts between individuals are generated. Contacts are effectives, i.e., lead to the transmission of the virus, according to a Bernoulli trial, based on the time since infection. Effective contacts that take place between susceptible and infectious individuals result in infection events. The probability that a contact is effective is composed of two factors: the infectivity measure, and the transmission potential , which accounts for the transmissibility of the pathogen and the susceptibility of the exposed individual. In this context, the basic reproduction number of an infectious disease is approximated with the mean number of effective contacts infectious individuals generates in a fully susceptible population throughout their infectious periods (Torneri et al., 2021).
The infectivity measure is defined over the exposed and infectious period of the infected individual and is set to represent the shape of the viral load curve for a SARS-CoV-2 infection, under the assumption that a higher amount of virus corresponds to a higher transmission probability (Buonanno et al., 2020). Based on literature findings, we define an infectivity measure that peaks at symptom onset and lasts 10 days, on average (Zhou et al., 2020b; Zhou et al., 2020a; Kim et al., 2020; Long et al., 2020; Liu et al., 2020; He et al., 2020; Cevik et al., 2021). In addition, has an initial plateau with value zero that accounts for the exposed phase.
Asymptomatic and symptomatic individuals are assumed to have the same viral progression, as argued in Zou et al., 2020; Zhou et al., 2020a, but we introduce a different level of infectiousness between infectious individuals based on the clinical outcome. Precisely, the relative infectiousness of asymptomatic compared to symptomatic is 0.5 (Davies et al., 2020).
School classes, pupils, and teachers
We consider a population of children in primary school (6–12 years old), where each child is randomly allocated to a class. To this end, we construct a set of classes of which the size is sampled from a probability mass function informed by a survey on Belgian primary schools (https://www.agodi.be/nieuwe-omkadering-basisonderwijs), up until at least 1000 pupils are allocated to these classes.
In the school, we consider teaching and supportive staff, to which we will refer as teachers from this point forward for brevity. The number of teachers is proportional to the number of pupils (ratio ) (https://www.vlaanderen.be/publicaties/vlaams-onderwijs-in-cijfers). We consider different contact ratios within and between classes, and assume that teachers have contacts only with children and other teachers of their allocated classes. Instead, children can have contact with children of the same class and children of different classes. The within and between classes contact rates for children are set accordingly to a contact data survey that took place in Belgium (Hoang et al., 2019). The within class contact rate is given by number of contacts that take place in primary schools and last more than 1 hr (). The between class contact rate is computed as the the number of contacts that last less than 1 hr (). However, we assumed that in a pandemic setting the number of between class contacts is reduced. In the baseline scenario, we assumed that the between contact rate in a COVID-19 pandemic scenario is 30% of . We test such assumption in the sensitivity analysis by varying this proportion among 20%, 50% and 90%.
Sensitivity analysis
Seeding number

We compare the testing strategies in the context of the Delta VoC for a seeding of 1 seed per week (left panel) and 10 seeds per week (right panel).
School and class thresholds are set, respectively, to 20 and 2 detected cases. The epidemic is simulated for 100 days. This experiment shows that increasing the number of seeds leads to an increase in both the attack rate and NSDL.
Number of tests per week

We show the repetitive testing strategy in the context of the Wuhan strain and the Delta VoC for a moderate seeding of 5 seeds per week, where we consider a repetitive testing strategy for which the entire school population is tested either once or twice per week.
We consider class closure threshold of 2 and school closure threshold of 20 cases. The epidemic is simulated for 100 days. This experiment demonstrates that for a highly infectious virus strain, repetitive testing can further reduce the number of transmissions at school while increasing the NSDL.

We show the repetitive testing strategy in the context of the Wuhan strain and the Delta VoC for a moderate seeding of 5 seeds per week, where we consider a repetitive testing strategy for which the entire school population is tested either once or twice per week.
We consider class closure threshold of 8 and no school closure threshold. The epidemic is simulated for 100 days. This experiment demonstrates that twice testing can reduce the number of transmissions at school and the NSDL when assumptions on class and school thresholds are relaxed.
Multiple screenings

We compare the reactive screening strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the number of screening.
The class closure threshold is of eight detected cases, and there is no school closure threshold. The epidemic is simulated for 100 days. This experiments shows that increasing the number of screening has a small effect on the attack rate and NSDL.
Between-classes contact rate

We compare testing strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the proportion of between classes contacts compared to a pre-pandemic scenario.
The class closure threshold is of eight detected cases, and there is no school closure threshold. The epidemic is simulated for 100 days. This experiments shows that increasing between classes contact rate increases both the attack rate (left panel) and the NSDL (right panel).
Immune population proportion

We compare the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the proportion of immune children.
The class closure threshold is of eight detected cases, and there is no school closure threshold. The epidemic is simulated for 100 days. This experiments shows that increasing the proportion of immune children decrease both the attack rate (left panel) and the NSDL (right panel) for all the testing strategies.

We compare the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the proportion of immune adults.
The class closure threshold is of eight detected cases, and there is no school closure threshold. The epidemic is simulated for 100 days. This experiments shows that increasing the proportion of immune adults decrease both the attack rate (left panel) and the NSDL (right panel) for all the testing strategies.
Increased infectivity asymptomatic carriers

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider asymptomatic carriers to be as infectious as the symptomatic ones.
The class closure threshold is of eight detected cases, and there is no school closure threshold. The epidemic is simulated for 100 days. The repetitive screening strategy is shown to reduce the attack rate (left panel), while leading to a higher NSDL (right panel).
School closure threshold

We show the repetitive testing strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different school closure thresholds.
The class closure threshold is set to eight detected cases. The epidemic is simulated for 100 days. This experiments shows that a low school closure threshold decreases the attack rate but it increases the NSDL.

We show the reactive screening strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different school closure thresholds.
The class closure threshold is set to eight detected cases. The epidemic is simulated for 100 days. This experiments shows that a low school closure threshold decreases the attack rate but it increases the NSDL.

We show the symptomatic isolation strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different school closure thresholds.
The class closure threshold is set to 8 detected cases. The epidemic is simulated for 100 days. This experiments shows that a low school closure threshold decreases the attack rate but it increases the NSDL.
Low seeding scenario

We show the testing strategies in the context of the Delta VoC for a seeding of 1 seeds per month, where we consider different school closure thresholds.
The class closure threshold is set to eight detected cases. The epidemic is simulated for 100 days. The repetitive screening strategy is shown to decrease the attack rate (left panel) compared to reactive screening and symptomatic isolation, while increasing the NSDL (right panel).
Probability of symptomatic infections

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different probability of having a symptomatic infection.
The class closure threshold is set to 8 detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. This experiment shows that to an increase in the probability of having a symptomatic infection it corresponds an increase in the attack rate (left panel) and the NSDL (right panel).
Incubation period

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different incubation period.
The class closure threshold is set to eight detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. The repetitive screening strategy is shown to decrease the attack rate (left panel) compared to reactive screening and symptomatic isolation, while increasing the NSDL (left panel). Reactive screening and repetitive screening are less effective in reducing the attack rate when the incubation period is shorter.
Test result delay

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different turnaround time for the test result.
The class closure threshold is set to 8 detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. An increase in the attack rate is shown for repetitive testing and reactive screening with an increase in the turnaround time (left panel). The NSDL increases with an increase in the turnaround time, while a similar trend is shown for the other strategies (right panel).
Compliance

We show the repetitive testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the compliance to testing.
The class closure threshold is set to eight detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. For an increase in the compliance the attack rate decreases and the NSDL increases.
School size

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seeds per week, when varying the school size.
The class closure threshold is set to eight detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. For an increase in the school size, the attack rate increases and the NSDL decreases.

We show the testing strategies in the context of the Delta VoC for a seeding of 1 case per week for a school of size 200, 5 seeds per week for a school of size 1000 and 10 seeds per week for a school of size 2000.
The class closure threshold is set to eight detected cases and there is no school closure threshold. The epidemic is simulated for 100 days. This experiment shows that a high stochasticity for a small school size, but similar attack rates and NSDL when the proportion of seeds over the school size is the same.
No school and class closures

We show the testing strategies in the context of the Delta VoC for a moderate seeding of 5 seed per week when no class and school closures are considered.
The epidemic is simulated for 100 days. This experiment shows that repetitive testing decreases the attack rate while increasing the NSDL compared to the other testing strategies.
Omicron scenario

We show the testing strategies in the context of the Delta and Omicron variants for a moderate seeding of 5 seeds per week, when no school closure is considered and the class threshold is set to value 8.
The epidemic is simulated for 100 days. Omicron is implemented by lowering the immunity proportion to 0.1 for children and 0.5 for adults, and by considering a shorter incubation period (mean 3.3 days and standard deviation of 2.2 days). This experiment shows that it is more difficult to contain the Omicron strain in a school setting. The repetitive screening strategy is shown to decrease the attack rate compared to reactive screening and symptomatic isolation while increasing the NSDL.

We show the repetitive testing strategy in the context of the Omicron VoC for a moderate seeding of 5 seeds per week, where we consider a repetitive testing strategy where the entire school population is tested either once or twice per week.
We consider class closure threshold of 8 and no school closure threshold. The epidemic is simulated for 100 days. This experiment demonstrates that in the contest of Omicron, repetitive testing can further reduce the number of transmissions at school when twice testing per week is considered.
No testing scenario

We show the testing strategies compared to a no-testing scenario (No Testing) in the context of the Delta variants for a moderate seeding of 5 seeds per week.
when school and class closures are set, respectively, to 20 and 2 detected cases. The epidemic is simulated for 100 days. This experiment shows that symptomatic isolation and the no testing scenario have a similar performance.
Class closure threshold: symptomatic isolation

We show the symptomatic isolation strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different class closure thresholds, and no school closure threshold.
The epidemic is simulated for 100 days. We note that in this experiment symptomatic isolation performs similarly for different class closure thresholds.
Class closure threshold: reactive Screening

We show the reactive screening strategy in the context of the Delta VoC for a moderate seeding of 5 seeds per week, where we consider different class closure thresholds, and no school closure threshold.
The epidemic is simulated for 100 days. We note that in this experiment symptomatic isolation performs similarly for different class closure thresholds.
Data availability
The current manuscript is a computational study, so no data have been generated for this manuscript. Source code of the individual-based model was implemented in R (version: R/3.6.0-foss- 2018a- bare) and is freely available in a Zenodo repository at the following DOI: 10.5281/zenodo.6488473. Modelling code is also uploaded as source code on a publicly available Github repository (https://github.com/AndreaTorneri/TestingStrategies, copy archived at swh:1:rev:5b6845b34f9d9a98d7f9438c2b9ffdac00db0a6b).
References
-
Lifting of COVID-19 restrictions in the UK and the Delta variantThe Lancet. Respiratory Medicine 9:e85.https://doi.org/10.1016/S2213-2600(21)00328-3
-
Repeat SARS-CoV-2 testing models for residential college populationsHealth Care Management Science 24:305–318.https://doi.org/10.1007/s10729-020-09526-0
-
Time Kinetics of Viral Clearance and Resolution of Symptoms in Novel Coronavirus InfectionAmerican Journal of Respiratory and Critical Care Medicine 201:1150–1152.https://doi.org/10.1164/rccm.202003-0524LE
-
Screening and vaccination against COVID-19 to minimise school closure: a modelling studyThe Lancet Infectious Diseases 22:977–989.https://doi.org/10.1016/S1473-3099(22)00138-4
-
A Systematic Review of Social Contact Surveys to Inform Transmission Models of Close-contact InfectionsEpidemiology (Cambridge, Mass.) 30:723–736.https://doi.org/10.1097/EDE.0000000000001047
-
Viral kinetics of SARS-CoV-2 in asymptomatic carriers and presymptomatic patientsInternational Journal of Infectious Diseases 95:441–443.https://doi.org/10.1016/j.ijid.2020.04.083
-
Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus-Infected PneumoniaThe New England Journal of Medicine 382:1199–1207.https://doi.org/10.1056/NEJMoa2001316
-
Viral dynamics in mild and severe cases of COVID-19The Lancet Infectious Diseases 20:656–657.https://doi.org/10.1016/S1473-3099(20)30232-2
-
Assessing COVID-19 Prevention Strategies to Permit the Safe Opening of Residential Colleges in Fall 2021Annals of Internal Medicine 174:1563–1571.https://doi.org/10.7326/M21-2965
-
COVID-19 infection in childrenThe Lancet. Respiratory Medicine 8:446–447.https://doi.org/10.1016/S2213-2600(20)30152-1
-
Transmission heterogeneities, kinetics, and controllability of SARS-CoV-2Science (New York, N.Y.) 371:eabe2424.https://doi.org/10.1126/science.abe2424
-
On realized serial and generation intervals given control measures: The COVID-19 pandemic casePLOS Computational Biology 17:e1008892.https://doi.org/10.1371/journal.pcbi.1008892
-
Should children be vaccinated against COVID-19 now?Archives of Disease in Childhood 106:1147–1148.https://doi.org/10.1136/archdischild-2020-321225
-
Saliva or Nasopharyngeal Swab Specimens for Detection of SARS-CoV-2The New England Journal of Medicine 383:1283–1286.https://doi.org/10.1056/NEJMc2016359
-
Viral dynamics in asymptomatic patients with COVID-19International Journal of Infectious Diseases 96:288–290.https://doi.org/10.1016/j.ijid.2020.05.030
-
SARS-CoV-2 Viral Load in Upper Respiratory Specimens of Infected PatientsNew England Journal of Medicine 382:1177–1179.https://doi.org/10.1056/NEJMc2001737
Decision letter
-
Kathy LeungReviewing Editor; The University of Hong Kong, Hong Kong
-
Jos W van der MeerSenior Editor; Radboud University Medical Centre, Netherlands
-
Bernadette C YoungReviewer; University of Oxford, United Kingdom
In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.
Decision letter after peer review:
Thank you for submitting your article "Controlling SARS-CoV-2 in schools using repetitive testing strategies" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by David Serwadda as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Bernadette C Young (Reviewer #2).
The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.
Essential revisions:
Reviewer #1's points in brief:
1. Frequency of tests in the strategies.
2. Thresholds of class and school closure, and comparison across strategies.
3. Comparison of low, medium, and high prevalence and infection attack rates.
4. Proportion of symptomatic infections among children.
5. Parameters about the virus such as: (a) Δ vs. Omicron and (b) shorter incubation or latent period.
Reviewer #2's major points in brief:
1. Thresholds of class and school closure, and comparison across strategies.
2. Parameters about the virus such as: (a) Δ vs. Omicron and (b) reinfection.
3. Test sensitivity and turnaround time.
4. Compliance of taking tests.
5. Infection attack rates and Number of School Days Lost.
6. Frequency of tests in the Repetitive Screening strategy.
Reviewer #3's major points in brief:
1. Control baseline and infection attack rates.
2. Parameters about the virus such as: (a) Δ vs. Omicron.
3. Class or school size.
4. Consideration of severe diseases and long COVID.
5. Duration of class/school closure.
Reviewer #1 (Recommendations for the authors):
1. For the reactive screening strategy, it is not clear why it was assumed that the whole class would be screened only once, if one symptomatic individual was identified. It is not surprising that the reactive screening strategy was only slightly better than symptomatic isolation because every child was only screened once. Could you please explain in more detail why repeat testing was not considered when the reactive screening strategy was triggered? This strategy is still different from repetitive screening.
2. It is not clear how the thresholds of class or school closure were selected (i.e., 2 for class closure and 20 for school closure). While I understand that a higher threshold (to trigger class or school closure) would have little effect on the attack rate (Figure 2), a lower threshold should also be assessed in the sensitivity analysis, such as 2 for school closure. Lowering the closure threshold would have more substantial effects on attack rates and NSDL, and thus an "optimal" threshold could be potentially identified to minimize the attack rates or NSDL.
3. It is not clear why "continuous" weekly introduction events to the school was assumed, by seeding 1, 5, or 10 cases weekly. Since continuous introduction was assumed, it is quite intuitive that the repetitive screening strategy is the most effective, because none of the three strategies would be able to "eliminate" infections for more than a week. In a low prevalence setting such as seeding 1 case every month, is repetitive screening still more effective than reactive screening or symptomatic screening?
4. How sensitive are the results to the assumption about the proportion of symptomatic infections among children? This also matters more in a low prevalence school setting.
5. How sensitive are the results to a shorter incubation period and/or a shorter latent period? Incorporating this assumption in the sensitivity analysis would be potentially useful to guide public health practice.
Examples of a few similar studies to be included in the literature review in the introduction:
a) Chang JT, Crawford FW, Kaplan EH. Repeat SARS-CoV-2 testing models for residential college populations. Health care management science. 2021 Jun;24(2):305-18.
b) Hamer DH, White LF, Jenkins HE, Gill CJ, Landsberg HE, Klapperich C, Bulekova K, Platt J, Decarie L, Gilmore W, Pilkington M. Assessment of a covid-19 control plan on an urban university campus during a second wave of the pandemic. JAMA Network Open. 2021 Jun 1;4(6):e2116425-.
c) Paltiel AD, Schwartz JL. Assessing COVID-19 prevention strategies to permit the safe opening of residential colleges in Fall 2021. Annals of internal medicine. 2021 Nov;174(11):1563-71.
Reviewer #2 (Recommendations for the authors):
This study uses mathematical models to assess the likely outcome of 3 models of testing on two outcomes in a primary school like setting (where students are in a single class and teachers teach a single class).
– Number people infected by a case in school (attack rate).
– Number of school days missed.
Three strategies are compared
– Testing individuals who develop symptoms.
– Testing individuals who develop symptoms, and if positive, their class.
– Testing all individuals repetitively. The paper does not specify what this interval is, but later results suggest the starting assumption is weekly.
A strength of this approach is the ability to predict how strategies might work based on what we know about viral transmission. The models are clearly defined, and allows others to repeat the work. However as with all modelling, the study is based on a number of assumptions, some of which are widely accepted based on available evidence.
– Not all cases are detected, and children less likely to be detected than adults.
– That individuals are most infectious when they develop symptoms.
Some of these assumptions are specific to public health measures that do not necessarily apply across other settings.
– That 2 cases in a class within 14 days would lead to closing the class, and 20 cases in a rolling 14 day window would close the school.
– 10 days isolation of cases (now lower in many parts of the world).
One of the primary outcomes (days of school missed) would be very sensitive to these policies, especially class closure. This policy is an important part of the intervention, and should be highlighted in abstract and title to aid with interpreting the results. Sensitivity analysis has been performed around these assumptions, however I have concerns with the clarity and accuracy of the presentation of the results of this analysis, which I will discuss in the findings.
Additional assumptions, which have not been tested by sensitivity analysis in this paper, do give me some concern about the applicability of the model.
The viral parameters are modelled for original pandemic and Δ strains, and the model makes assumptions about high immune rates in adults, lower in children, reasonable in settings of high vaccination rates. However it also assumes reinfection does not occur, which is reasonable over the 100 days of the model but not with a newly introduced VOC or over longer time periods.
The model further assumes that testing in clinical practice has imperfect sensitivity (83%) and turn- around time of 1 day. This latter is plausible at some times but not in all settings, and often not at times of high case rates. The sensitivity analysis should include delays to turn around times.
Finally, the model assumes all invited to testing take part, which is unlikely to reflect clinical reality, particularly for asymptomatic testing in the repetitive testing strategy, and evidence is needed to support this assumption.
In the model 5 cases in a school acquired infection outside school are imported, and cases and absence over the next 100 days are computed. The model runs 100 times to determine cases and absences.
The authors show a large reduction in the attack rate is lower if screening is pursued, and greater impact with reactive screening. Reactive (ReaS) screening has a limited effect, likely due to incomplete case ascertainment (high number cases not being detected due to high number asymptomatic). Repetitive screening (RepS) leads to average 4 as the Number of School Days Lost (NSDL) in 100 days for Wuhan and over 40 /100 in Δ. This means students would miss almost half days in education.
The effect of changing these thresholds is reviewed, and notably removing school closure threshold and reducing class thresholds does reduce NSDL. However Figure 2 shows results only for RepS, with Δ, and does not show a direct comparison with other testing strategies. However for all the results presented, the upper bound of the estimated attack rate are now within the confidence interval for the AR in ReaS. The paper does not present numbers for the results, but from graphical representation the Δ AR in ReaS has a median around 13% with an observed range of around 8-17%. The Δ AR for RepS has an average 5% with the original closure rules, but median increases to around 7.5 -8-5 with the changes in closure rules and upper bound extends to about 12%, so well within the interquartile range of the RepS strategy. I therefore disagree that author's statement at line 198 that "a higher class closure threshold has little effect on the attack rate, yet it significantly reduces the NSDL " is supported by the presented data.
Twice weekly screening was more effective than weekly in the model and with permissive rules on class closure did not increase school days missed, which is an important result. Given that other published modelling has assumed twice weekly screening (eg https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1013533/S1355_SPI-M-O_Consensus_Statement.pdf), it is unclear why weekly screening would be the assumed parameter.
The authors report that RepS is 'the most efficient to reduce attack rate' – and while it has the greatest impact on attack rate, the high cost of school days missed in the core model leads me to question the 'efficiency' of the strategy. Further the qualifying statement "Simulations indicate that such a testing strategy limits the number of transmission events even when no class and school closures are in place." Is not supported by the results presented, as discussed above. Likewise, I do not think their statement "keeping transmissions under control, a limited number of school days lost is computed when no thresholds are in place" is currently supported. However if the model were investigated with twice weekly testing, this may be the case.
The authors highlight that this paper demonstrates the limitation of symptomatic based testing and isolation strategies in a population likely to have low rates of symptoms and immunity (as is currently the case in many current primary school settings).
I think this is an important piece of work and will help guide public health decision making, but some further analysis is needed to see if the findings support the authors' claims, or if the impact of repetitive screening in primary school could only be achieved with very stringent school closure rules, which would be unacceptable to many.
Abstract
– Suggest editing abstract to allow reader to determine the findings (and effect size) in abstract.
– Since isolation and closure policies are important in the model and for the primary outcome, these should be included or indicated in the abstract.
Introduction
– The piece should be placed in context of other published work modelling impact of testing strategies on school attack rate and absence, for eg (not exhaustive).
– UK SPI-MO reports around mass testing, and secondary school based testing (both mass screening) and reactive
– https://www.gov.uk/government/publications/spi-m-mass-testing-of-the-whole-population-25-november-2020
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1013533/S1355_SPI-M-O_Consensus_Statement.pdf
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/976324/S1146_SPI-M-O_Daily_contact_testing.pdf
Leng et al., have presented extensive modelling evidence in secondary schools
https://www.medrxiv.org/content/10.1101/2021.07.09.21260271v1.full
https://www.epicx-lab.com/uploads/9/6/9/4/9694133/ms-testing_school-rev.pdf
– Suggest remove results from introduction section Lines 52-64, introduction should not contain results.
Methods
– Number days of school lost is not defined, nor the method for calculating defined.
– Attack rate should be defined in the methods. From results I infer it is number of cases in the school?
– Testing frequency in RepS should be made clear.
Results
– Results section should outline results prior to drawing conclusions about them (eg line 131-134) makes a comparative statement about results not yet presented.
– Results should be presented quantitatively as well as graphically, so that the reader is not left trying to deduce them.
– Figure 2 should allow comparison to other models performance, as I believe that with relaxed school/class closure rules ReaS and RepS will not be strongly different.
Discussion
– Line 219 "keeping transmissions under control, a limited number of school days lost is computed when no thresholds are in place" is currently supported, because the median attack rate was close to 10% in these simulations. However if the model were investigated with twice weekly testing, this may be the case. I encourage the authors to apply twice weekly testing to the base model of RepS and see if RepS provides a clearer benefit.
– Paragraph at line 221 is unclear. Appears the authors recounting recommendations from another paper, but this may be a matter of linguistic confusion.
– Paragraph at line 242 is unclear, I do not know which part of the paper this refers to.
Supplement
– Suggest insertion of a qualifying comment on reinfection given observation of high rate of reinfection with Omicron variant.
Reviewer #3 (Recommendations for the authors):
In this manuscript the authors set out to investigate the effectiveness of different COVID-19 testing strategies within school settings to reduce infections, while also reducing the loss of in-person teaching time. The authors show that (reactive) testing based on symptoms is generally ineffective (due to large proportion of asymptomatic infections) compared to a repetitive testing strategy – but also that such a strategy should be utilised with appropriate thresholds for closing in-person education given the greater number of tests. Overall, the article is well written and the authors offer several sensitivity analyses showing the robustness of their core and timely result – that frequent testing can be used to reduce infections while keeping schools open. There are four main areas I think warrant consideration in the text: (a) a control baseline – do symptomatic isolation and reactive testing strategies exhibit lower attack rates than in scenarios with no testing – it is currently unclear how (in)effective these strategies are? (b) how well do the results hold up in the context of new variants e.g. Omicron? (c) how does school size impact findings? (d) while infection may be mild for most children, neither non-mild infection and/or persistent symptoms (long COVID) are built into the modeling framework – both are additional reasons for wanting to minimize the attack rate (as is reducing community transmission), while additionally minimizing disruption to education and should be discussed.
One key detail I was unable to find in the text, how long are school/class closures implemented for when triggered?
I would suggest that the authors include the length of epidemic simulations (100 days) for figures where NSDL is reported – or maybe consider reporting this result as a proportion of school days? Also, knowing that 5 infections are seeded per a week into a population of 1000 is useful for interpretation.
The authors note that the attack rate does not include seeded infections; could the authors clarify how they calculate this?
I found it unclear in the Supplemental description what the operational differences are between pupils and teachers in the model, could the authors clarify?
I thank the authors for making their simulation code available online. I would encourage the authors to also archive their code (e.g. via Zenodo/Figshare) to create a version of record with a persistent identifier (https://docs.github.com/en/repositories/archiving-a-github-repository/referencing-and-citing-content). This would create a resource that is accessible even if GitHub goes away, changes its domain name etc. For the purposes of replication I would also encourage the authors to list which version of R and its packages were used in the code.
I would suggest the authors change "his or her" to "their" in the Supplement (line: 318)
With respect to school size – I wonder if 1000 pupils is a typical size for a primary school? In my experience many primary schools are much smaller than this, is there a particular reason this number was chosen? I think school size influences the manuscript in two ways: (a) school closure threshold (just Figure 1?) (b) appropriate importation rate of infections. So I do expect the key finding of the value of repetitive testing to remain. Are other results affected? If so, I think it would be useful to communicate some of these nuances, especially with respect to those who wish to translate these research results into actionable policy.
https://doi.org/10.7554/eLife.75593.sa1Author response
Essential revisions:
Reviewer #1's points in brief:
1. Frequency of tests in the strategies.
2. Thresholds of class and school closure, and comparison across strategies.
3. Comparison of low, medium, and high prevalence and infection attack rates.
4. Proportion of symptomatic infections among children.
5. Parameters about the virus such as: (a) Δ vs. Omicron and (b) shorter incubation or latent period.
Reviewer #2's major points in brief:
1. Thresholds of class and school closure, and comparison across strategies.
2. Parameters about the virus such as: (a) Δ vs. Omicron and (b) reinfection.
3. Test sensitivity and turnaround time.
4. Compliance of taking tests.
5. Infection attack rates and Number of School Days Lost.
6. Frequency of tests in the Repetitive Screening strategy.
Reviewer #3's major points in brief:
1. Control baseline and infection attack rates.
2. Parameters about the virus such as: (a) Δ vs. Omicron.
3. Class or school size.
4. Consideration of severe diseases and long COVID.
5. Duration of class/school closure.
All the essential revision points listed above are addressed in the present revision. For a clearer response to the reviewers comments, we report our responses directly below each comment made by the reviewers.
Reviewer #1 (Recommendations for the authors):
1. For the reactive screening strategy, it is not clear why it was assumed that the whole class would be screened only once, if one symptomatic individual was identified. It is not surprising that the reactive screening strategy was only slightly better than symptomatic isolation because every child was only screened once. Could you please explain in more detail why repeat testing was not considered when the reactive screening strategy was triggered? This strategy is still different from repetitive screening.
We selected a strategy based on a single screening of the class in which a case is detected to represent a strategy that is used in practice in different European countries. In the first version of the paper, we did not consider additional screenings because the majority of cases was supposedly detected with the first screening. To get more insights on the effect of multiple screening rounds, we extended the simulation model and we ran a scenario in which classes of detected cases are screened 2 or 3 times. Results show that multiple screenings have a low impact on attack rate and number of school days lost (Figure 6). A sentence describing this experiment has been added to the Result section (lines: 174-176).
2. It is not clear how the thresholds of class or school closure were selected (i.e., 2 for class closure and 20 for school closure). While I understand that a higher threshold (to trigger class or school closure) would have little effect on the attack rate (Figure 2), a lower threshold should also be assessed in the sensitivity analysis, such as 2 for school closure. Lowering the closure threshold would have more substantial effects on attack rates and NSDL, and thus an "optimal" threshold could be potentially identified to minimize the attack rates or NSDL.
There was no commonly defined protocol in the literature, nor among different European countries that we could use to inform the threshold values. Therefore, we had to make an initial ad-hoc choice, and we assumed a threshold of 2 for class closure and 20 for school closure. Acknowledging the uncertainty around this choice, we challenged both class and school closure thresholds in a sensitivity analysis, of the initial revision. As suggested by the reviewer, we extended this sensitivity analysis on the school closure threshold by considering a wider range of thresholds. The additional results demonstrate the effect of a school closure threshold on attack rate and NSDL. This shows that a low school closure threshold induces a low attack rate in combination with a high NSDL (Figures 11, 12 and 13).
3. It is not clear why "continuous" weekly introduction events to the school was assumed, by seeding 1, 5, or 10 cases weekly. Since continuous introduction was assumed, it is quite intuitive that the repetitive screening strategy is the most effective, because none of the three strategies would be able to "eliminate" infections for more than a week. In a low prevalence setting such as seeding 1 case every month, is repetitive screening still more effective than reactive screening or symptomatic screening?
We assumed a continuous weekly introduction to describe a context of a high COVID19 prevalence. Testing strategies are expected to be more relevant when the virus circulates widely. However, following the comment of the reviewer we simulate a scenario with a single introduction per month, and results are added to the sensitivity analysis (Figure 14). In such a scenario, the attack rate is low for all the testing strategies. Nevertheless, repetitive screening still performs better in decreasing the attack rate. We added a sentence discussing this analysis in the result section (lines 228-232).
4. How sensitive are the results to the assumption about the proportion of symptomatic infections among children? This also matters more in a low prevalence school setting.
Thank you for this question. We added a simulation scenario in which the assumption about the proportion of symptomatic individuals was varied (Figure 15). Overall, both the attack rate and the NSDL increase, when the probability of children being symptomatic is higher. Reactive screening has a larger impact on the attack rate compared to symptomatic isolation when the probability of symptomatic infection is larger. However, even when the probability of symptomatic infection increases, repetitive screening remains the strategy that most reduces the number of infections in all of the tested scenarios. We added a sentence discussing this analysis in the result section (lines 228-232).
5. How sensitive are the results to a shorter incubation period and/or a shorter latent period? Incorporating this assumption in the sensitivity analysis would be potentially useful to guide public health practice.
We extended the simulation model to enable a shorter incubation period, that we set in line with observation for the BA.1 Omicron VoC (Backer et al., 2022), and we ran a simulation comparing such incubation period with the baseline (Figure 16). Results indicate that testing strategies are less effective in decreasing the attack rate when the incubation period is shorter. In this scenario, repetitive screening remains the strategy with most impact on the number of infections. We added this scenario to the sensitivity analysis (Figure 16), and we report the findings in the result section (lines: 241-242).
Examples of a few similar studies to be included in the literature review in the introduction:
a) Chang JT, Crawford FW, Kaplan EH. Repeat SARS-CoV-2 testing models for residential college populations. Health care management science. 2021 Jun;24(2):305-18.
b) Hamer DH, White LF, Jenkins HE, Gill CJ, Landsberg HE, Klapperich C, Bulekova K, Platt J, Decarie L, Gilmore W, Pilkington M. Assessment of a covid-19 control plan on an urban university campus during a second wave of the pandemic. JAMA Network Open. 2021 Jun 1;4(6):e2116425-.
c) Paltiel AD, Schwartz JL. Assessing COVID-19 prevention strategies to permit the safe opening of residential colleges in Fall 2021. Annals of internal medicine. 2021 Nov;174(11):1563-71.
We thank the reviewer for pointing out studies on testing strategies in school settings. We acknowledge them in the introduction to highlight the current state of the art (lines 46-49).
Reviewer #2 (Recommendations for the authors):
This study uses mathematical models to assess the likely outcome of 3 models of testing on two outcomes in a primary school like setting (where students are in a single class and teachers teach a single class).
– Number people infected by a case in school (attack rate).
– Number of school days missed.
Three strategies are compared
– Testing individuals who develop symptoms.
– Testing individuals who develop symptoms, and if positive, their class.
– Testing all individuals repetitively. The paper does not specify what this interval is, but later results suggest the starting assumption is weekly.
A strength of this approach is the ability to predict how strategies might work based on what we know about viral transmission. The models are clearly defined, and allows others to repeat the work. However as with all modelling, the study is based on a number of assumptions, some of which are widely accepted based on available evidence.
– Not all cases are detected, and children less likely to be detected than adults.
– That individuals are most infectious when they develop symptoms.
Some of these assumptions are specific to public health measures that do not necessarily apply across other settings.
– That 2 cases in a class within 14 days would lead to closing the class, and 20 cases in a rolling 14 day window would close the school.
– 10 days isolation of cases (now lower in many parts of the world).
One of the primary outcomes (days of school missed) would be very sensitive to these policies, especially class closure. This policy is an important part of the intervention, and should be highlighted in abstract and title to aid with interpreting the results. Sensitivity analysis has been performed around these assumptions, however I have concerns with the clarity and accuracy of the presentation of the results of this analysis, which I will discuss in the findings.
We revised the manuscript by highlighting the role of the number of school days lost, as suggested by the reviewer. We included the findings and main assumptions in the abstract, and we added several sensitivity analyses to better investigate how this quantity is affected by the testing strategies in different scenarios. We also improved the presentation of the results, increasing the clarity of the investigation and better supporting the statements made.
Additional assumptions, which have not been tested by sensitivity analysis in this paper, do give me some concern about the applicability of the model.
The goal of the developed simulation model is to obtain general insights on specific aspects of the transmission dynamics, and we agree with the reviewer that to do so the assumptions made need to be challenged in a sensitivity analysis. Following the comments of the reviewer we extended our investigation, by further challenging the assumptions made.
The viral parameters are modelled for original pandemic and Δ strains, and the model makes assumptions about high immune rates in adults, lower in children, reasonable in settings of high vaccination rates. However it also assumes reinfection does not occur, which is reasonable over the 100 days of the model but not with a newly introduced VOC or over longer time periods.
We thank the reviewer for the comment. To investigate the effect of a possible immune evasive VoC we challenged the assumption on immunity rates (Figures 8 and 9). Results show that lower immunity corresponds to higher attack rates and higher NSDL. We included these results in the result section (lines 228-232). It would be interesting to include reinfections and simulate testing strategies in longer time periods, but this requires additional model developments. Furthermore, as pointed out by the reviewer, it is reasonable not to consider reinfection in our simulation scenarios, since we look at a time interval of 100 days.
The model further assumes that testing in clinical practice has imperfect sensitivity (83%) and turn- around time of 1 day. This latter is plausible at some times but not in all settings, and often not at times of high case rates. The sensitivity analysis should include delays to turn around times.
We added a sensitivity analysis investigating the turn around time of 1,2,3 days. We observed an increase in the attack rate for the reactive screening and the repetitive testing strategies with an increase in the turn around time, while a similar trend is observed for the symptomatic isolation strategy (Figure 17). We report these findings in the result section (lines 236-241).
Finally, the model assumes all invited to testing take part, which is unlikely to reflect clinical reality, particularly for asymptomatic testing in the repetitive testing strategy, and evidence is needed to support this assumption.
We agree that compliance to testing is an important assumption, especially for asymptomatic testing or in low-prevalence settings. Therefore, we extended the simulation model to include compliance to testing and we ran a sensitivity analysis in which we vary the compliance to testing for the repetitive screening strategy. Overall, for a decrease in compliance the attack rate increases and the NSDL decreases (Figure 18). Interestingly, in our experiment we observe a similar attack rate between a 60% and 100% compliance. We added a sentence about compliance in the result section (lines 252-256). In addition, we rephrased the paragraph about compliance in the Discussion section according to the result obtained in the sensitivity analysis (lines 310-317).
In the model 5 cases in a school acquired infection outside school are imported, and cases and absence over the next 100 days are computed. The model runs 100 times to determine cases and absences.
The authors show a large reduction in the attack rate is lower if screening is pursued, and greater impact with reactive screening. Reactive (ReaS) screening has a limited effect, likely due to incomplete case ascertainment (high number cases not being detected due to high number asymptomatic). Repetitive screening (RepS) leads to average 4 as the Number of School Days Lost (NSDL) in 100 days for Wuhan and over 40 /100 in Δ. This means students would miss almost half days in education.
The effect of changing these thresholds is reviewed, and notably removing chooll closure threshold and reducing class thresholds does reduce NSDL. However Figure 2 shows results only for RepS, with Δ, and does not show a direct comparison with other testing strategies. However for all the results presented, the upper bound of the estimated attack rate are now within the confidence interval for the AR in ReaS. The paper does not present numbers for the results, but from graphical representation the Δ AR in ReaS has a median around 13% with an observed range of around 8-17%. The Δ AR for RepS has an average 5% with the original closure rules, but median increases to around 7.5 -8-5 with the changes in closure rules and upper bound extends to about 12%, so well within the interquartile range of the RepS strategy. I therefore disagree that author’s statement at line 198 that “a higher class closure threshold has little effect on the attack rate, yet it significantly reduces the NSDL “ is supported by the presented data.
We apologise for the confusion. To facilitate the comparison among testing strategies when different class closure thresholds are considered, we added Figures 25 and 26 presenting the results of a sensitivity analysis in which symptomatic isolation and reactive screening are considered.
Furthermore, we report both the median and the 95% quantile interval in the result section (Table 2), to allow for a clearer interpretation. By specifying these values, we support the statement that the NSDL significantly reduces.
We changed the sentence:
“This experiment shows that a higher class closure threshold has little effect on the attack rate, yet it significantly reduces the NSDL” into
“This experiment shows that when repetitive testing is in place, a higher class closure threshold has little effect on the attack rate, yet it significantly reduces the mean NSDL (Table 2). (line 259)”.
Twice weekly screening was more effective than weekly in the model and with permissive rules on class closure did not increase school days missed, which is an important result. Given that other published modelling has assumed twice weekly screening (eg https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1013533/S1355_SPI-M-O_Consensus_Statement.pdf), it is unclear why weekly screening would be the assumed parameter.
We assumed a weekly screening as the baseline scenario because a strategy based on a single test can be more easily applied at national level when a large number of tests need to be quickly analyzed. We added a sentence in the manuscript to clarify our choice (lines 96-100). Furthermore, by keeping this approach we stress even more that even a single test per week is sufficient for a repetitive testing strategy to perform better than symptomatic isolation and reactive screening in reducing transmission. However, we acknowledge that twice weekly testing is suitable to further reduce the attack rate and we show the effect of such a strategy in two scenarios in which we consider, respectively, the Δ VoC (Figure 4,5), and the Omicron VoC (Figure 23). For a twice weekly testing strategy in the setting of the Δ Voc, we noticed a decrease in the NSDL when the assumption on school and class closure are relaxed compared to a single repetitive testing strategy (Figure 5). However, the opposite trend is observed when school and class thresholds are present (Figure 4), or when the Omicron VoC is considered (Figure 23). We describe this in the result section (lines 242-252).
The authors report that RepS is 'the most efficient to reduce attack rate' – and while it has the greatest impact on attack rate, the high cost of school days missed in the core model leads me to question the ‘efficiency’ of the strategy. Further the qualifying statement “Simulations indicate that such a testing strategy limits the number of transmission events even when no class and school closures are in place.” Is not supported by the results presented, as discussed above.
We agree with the reviewer and included a sensitivity analysis in which the three testing strategies are considered when no school or class closure thresholds are in place (Figure 21). Results show that repetitive testing is the most successful intervention under study in reducing the attack rate.
Likewise, I do not think their statement “keeping transmissions under control, a limited number of school days lost is computed when no thresholds are in place” is currently supported. However if the model were investigated with twice weekly testing, this may be the case.
We included the sensitivity analysis in which no school and class closures are considered (Figure 21) to support the claim. Furthermore we added a scenario in which school threshold is varied for the reactive screening and symptomatic isolation strategies (Figures 11,12 and 13), showing that such strategies can reach a similar control to repetitive screening only with a low school closure thresholds, leading to a high NSDL. We also included two additional scenarios in which twice weekly testing is considered (Figures 4 and 22). We noticed that while in one case the NSDL are further reduced compared to testing on a weekly basis (Figure 5), in the other scenario we observe an increase in the NSDL (Figure 4, and 22).
The authors highlight that this paper demonstrates the limitation of symptomatic based testing and isolation strategies in a population likely to have low rates of symptoms and immunity (as is currently the case in many current primary school settings).
I think this is an important piece of work and will help guide public health decision making, but some further analysis is needed to see if the findings support the authors’ claims, or if the impact of repetitive screening in primary school could only be achieved with very stringent school closure rules, which would be unacceptable to many.
We thank the reviewer for the appreciation of our work, and for the comments made which helped to improve the manuscript. We believe that the revised manuscript addresses the comments made, further clarifying and supporting the claims made.
Abstract
– Suggest editing abstract to allow reader to determine the findings (and effect size) in abstract.
– Since isolation and closure policies are important in the model and for the primary outcome, these should be included or indicated in the abstract.
We changed the following sentence of the abstract:
“Through this analysis, we demonstrate that repetitive testing strategies can significantly reduce the attack rate in schools, contrary to a reactive screening approach. Furthermore, we investigate the impact of these testing strategies on the average number of school days lost per child.” Into:
“Through this analysis, we demonstrate that repetitive testing strategies can significantly reduce the attack rate in schools, contrary to a reactive screening or a symptomatic isolation approach. However, when a repetitive testing strategy is in place, more cases will be detected and class and school closures are more easily triggered, leading to a high number of school days lost per child. While maintaining the epidemic under control with a repetitive testing strategy, absenteeism can be reduced by relaxing class and school closure thresholds.” (lines: 14-22)
Introduction
– The piece should be placed in context of other published work modelling impact of testing strategies on school attack rate and absence, for eg (not exhaustive).
– UK SPI-MO reports around mass testing, and secondary school based testing (both mass screening) and reactive
– https://www.gov.uk/government/publications/spi-m-mass-testing-of-the-whole-population-25-november-2020
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1013533/S1355_SPI-M-O_Consensus_Statement.pdf
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/976324/S1146_SPI-M-O_Daily_contact_testing.pdf
Leng et al., have presented extensive modelling evidence in secondary schools
https://www.medrxiv.org/content/10.1101/2021.07.09.21260271v1.full
https://www.epicx-lab.com/uploads/9/6/9/4/9694133/ms-testing_school-rev.pdf
– Suggest remove results from introduction section Lines 52-64, introduction should not contain results.
The introduction has been adapted including the suggested investigations concerning testing in schools, and removing results from the introduction (lines 46-49).
Methods
– Number days of school lost is not defined, nor the method for calculating defined.
– Attack rate should be defined in the methods. From results I infer it is number of cases in the school?
– Testing frequency in RepS should be made clear.
We added the definition of both the number of school days lost and the attack rate as follows:
"For each simulated outbreak we compute two summary measures that account, re-
spectively, for the number of transmissions at school and absenteeism. The former is defined as the total number of cases (minus the index cases) divided by the number of pupils in the school, and we refer to this quantity as the attack rate. The latter is defined as the total number of school days lost divided by the school size, and we refer to this quantity as number of school days lost (NSDL)." (lines 131-137)
In addition we specified that repetitive testing is considered to have a frequency of once per week in the baseline scenario. (lines 85-87)
Results
– Results section should outline results prior to drawing conclusions about them (eg line 131-134) makes a comparative statement about results not yet presented.
– Results should be presented quantitatively as well as graphically, so that the reader is not left trying to deduce them.
– Figure 2 should allow comparison to other models performance, as I believe that with relaxed school/class closure rules ReaS and RepS will not be strongly different.
We restructured the Result section according to the reviewer’s comments, and we added the median estimate together with the 95% quantile interval for a clearer presentation of the results.
We also added a scenario in the sensitivity analysis simulating the effect of class closure thresholds for the symptomatic isolation (Figure 25) and the reactive screening (Figure 26), which can be compared to Figure 2. We observed that the summary measures computed for the repetitive screening and reactive screening are different. In particular, repetitive screening performs better in reducing the attack rate when class thresholds vary, compared to reactive screening.
Discussion
– Line 219 "keeping transmissions under control, a limited number of school days lost is computed when no thresholds are in place" is currently supported, because the median attack rate was close to 10% in these simulations. However if the model were investigated with twice weekly testing, this may be the case. I encourage the authors to apply twice weekly testing to the base model of RepS and see if RepS provides a clearer benefit.
We added the requested simulation to the sensitivity analysis (Figure 4). We noticed that testing twice per week further reduces the attack rate also in this case, compared to testing once per week. However, twice weekly testing increases the NSDL (Figure 4). In addition, we also investigated the use of twice weekly for the Omicron VoC (Figure 23.)
– Paragraph at line 221 is unclear. Appears the authors recounting recommendations from another paper, but this may be a matter of linguistic confusion.
We restructured the paragraph to present more clearly the point we wanted to make.
(lines 284-290)
– Paragraph at line 242 is unclear, I do not know which part of the paper this refers to.
We apologize for the confusion, and we decided to remove this paragraph, in favour of a more clear and structured Discussion section.
Supplement
– Suggest insertion of a qualifying comment on reinfection given observation of high rate of reinfection with Omicron variant.
We added sensitivity analyses in which we consider the Omicron VoC and we varied the proportion of immune individuals in the sensitivity analysis.
Reviewer #3 (Recommendations for the authors):
In this manuscript the authors set out to investigate the effectiveness of different COVID-19 testing strategies within school settings to reduce infections, while also reducing the loss of in-person teaching time. The authors show that (reactive) testing based on symptoms is generally ineffective (due to large proportion of asymptomatic infections) compared to a repetitive testing strategy – but also that such a strategy should be utilised with appropriate thresholds for closing in-person education given the greater number of tests. Overall, the article is well written and the authors offer several sensitivity analyses showing the robustness of their core and timely result – that frequent testing can be used to reduce infections while keeping schools open. There are four main areas I think warrant consideration in the text:
a) a control baseline – do symptomatic isolation and reactive testing strategies exhibit lower attack rates than in scenarios with no testing – it is currently unclear how (in)effective these strategies are?
We included the sensitivity analysis in which we compare the testing strategies with a baseline scenario, in which individuals are not tested. Results show that symptomatic isolation performs similarly well to a no testing strategy, while reactive screening performs slightly better (Figure 24).
b) how well do the results hold up in the context of new variants e.g. Omicron?
We extended the sensitivity analysis by including a scenario which represents the Omicron variant. Omicron has been simulated by reducing the immunity for children and adults to, respectively, 0.1 and 0.5, and by including a shorter incubation period (Mean 3.3 days, sd 2.2 days). Results show that all testing strategies are less effective against the Omicron variant (Figure 22). Twice weekly testing improves the performance of repetitive testing (Figure 23). We included these results in the result section (lines 212-215).
c) how does school size impact findings?
We investigated simulation scenarios to further investigate the impact of the school size (Figures 19 and 20). Results show that the size can affect the attack rate when the same number of seeds are added among schools of different sizes. Precisely, to a lower school size, we observe a higher attack rate (Figure 19). This is possibly caused by the number of seeds, which accounts for a different proportion of the school population when the school size varies. We noticed a similar effect in Figure 3, where we tested the impact of the seeding number. To get more insights into this “seeding effect”, we simulated a scenario in which the number of seeds are weighted according to the population size. Results show that results are similar for different school sizes when we considered the weighted number of seeds (Figure 20). We describe this in the result section indicating the effect of the school size (lines 223-227).
d) while infection may be mild for most children, neither non-mild infection and/or persistent symptoms (long COVID) are built into the modeling framework – both are additional reasons for wanting to minimize the attack rate (as is reducing community transmission), while additionally minimizing disruption to education and should be discussed.
We thank the reviewer for the point made and we mention this in the discussion (lines 264-267).
One key detail I was unable to find in the text, how long are school/class closures implemented for when triggered?
We thank the reviewer for pointing this out. School/class closure, as well as isolation, are implemented for 10 days, according to the duration of viral clearance (Chang et al., 2020) and in line with the isolation policy in place in Belgium in 2021. We specified the length of the school/class closure in the methods section. (lines 91-95).
I would suggest that the authors include the length of epidemic simulations (100 days) for figures where NSDL is reported – or maybe consider reporting this result as a proportion of school days? Also, knowing that 5 infections are seeded per a week into a population of 1000 is useful for interpretation.
We added the following sentence in each caption: "The epidemic is simulated for 100 days". The number of seeds is also reported in each caption.
The authors note that the attack rate does not include seeded infections; could the authors clarify how they calculate this?
The attack rate is calculated as the total number of cases excluding the number of seeds divided by the number of pupils in the school. We explicitly add this definition in the text (lines 133-135).
I found it unclear in the Supplemental description what the operational differences are between pupils and teachers in the model, could the authors clarify?
We clarified that teachers are assumed to be linked to a single class, while pupils are linked with both children of the same class and children of other classes (lines 391-395).
I thank the authors for making their simulation code available online. I would encourage the authors to also archive their code (e.g. via Zenodo/Figshare) to create a version of record with a persistent identifier (https://docs.github.com/en/repositories/archiving-a-github-repository/referencing-and-citing-content). This would create a resource that is accessible even if GitHub goes away, changes its domain name etc. For the purposes of replication I would also encourage the authors to list which version of R and its packages were used in the code.
As suggested, we uploaded the code in a Zenodo repository (DOI:10.5281/zenodo.6488473).
We also listed the R version used, both in the repository and in the manuscript (lines 314-316.). The repository will be made freely available after acceptance of the manuscript by the journal.
I would suggest the authors change "his or her" to "their" in the Supplement (line: 318)
We made the suggested change (line 369).
With respect to school size – I wonder if 1000 pupils is a typical size for a primary school? In my experience many primary schools are much smaller than this, is there a particular reason this number was chosen? I think school size influences the manuscript in two ways: (a) school closure threshold (just Figure 1?) (b) appropriate importation rate of infections. So I do expect the key finding of the value of repetitive testing to remain. Are other results affected? If so, I think it would be useful to communicate some of these nuances, especially with respect to those who wish to translate these research results into actionable policy.
School size depends on national or regional school system, and on whether these are in metropolitan or rural areas. We assumed a size of 1000, but we challenged this assumption in the sensitivity analysis, following the suggestion made. By investigating the point raised by the reviewer we noticed that the size can affect the attack rate when the same number of seeds are added among schools with different sizes. As discussed in the reply to comment(c), this is because of the different proportion of seeds over the population, and when weighted seeds are considered results are in line among schools with different sizes. Similarly, the school size could affect the number of school closures, when thresholds are kept the same among school with different sizes. However, repetitive testing is always expected to be the strategy that most reduces the number of transmissions at school.
https://doi.org/10.7554/eLife.75593.sa2Article and author information
Author details
Funding
Fonds Wetenschappelijk Onderzoek (234620N)
- Lander Willem
Fonds Wetenschappelijk Onderzoek (1242021N)
- Pieter JK Libin
Horizon 2020 - Research and Innovation Framework Programme (101003688)
- Niel Hens
European Research Council (68254)
- Niel Hens
Antwerp Study Centre for Infectious Diseases (ASCID)
- Niel Hens
CHU Liege
- Christelle Meuris
F.R.S. - FNRS
- Gilles Darcis
European Union (VERDI project (101045989))
- Niel Hens
- Andrea Torneri
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
LW and PJKL acknowledge support from the Research Foundation Flanders (FWO, fwo.be) (postdoctoral fellowships 1234620 N and 1242021 N). NH and PJKL acknowledge support from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant number 101003688—EpiPose project). NH and AT received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant number 682540—TransMID project). NH acknowledge funding from the Antwerp Study Centre for Infectious Diseases (ASCID) and the chair in evidence-based vaccinology at the Methusalem-Centre of Excellence consortium VAX-IDEA. CM received funding from the ”Fondation Léon Fredericq” and the ”Fond d’investissement de recherche scientifique” from the CHU of Liège. GD received ”Post-doctorate Clinical Master Specialists” funding from the Fund for Scientific Research (F.R.S.–FNRS, frs-fnrs.be). We used computational resources and services provided by the Flemish Supercomputer Centre (VSC), funded by the FWO and the Flemish Government. This project was supported by the VERDI project (101045989), funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the Health and Digital Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.The funding agencies had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Senior Editor
- Jos W van der Meer, Radboud University Medical Centre, Netherlands
Reviewing Editor
- Kathy Leung, The University of Hong Kong, Hong Kong
Reviewer
- Bernadette C Young, University of Oxford, United Kingdom
Version history
- Preprint posted: November 15, 2021 (view preprint)
- Received: November 15, 2021
- Accepted: June 15, 2022
- Version of Record published: July 5, 2022 (version 1)
Copyright
© 2022, Torneri et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 647
- Page views
-
- 140
- Downloads
-
- 5
- Citations
Article citation count generated by polling the highest count across the following sources: PubMed Central, Crossref, Scopus.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Epidemiology and Global Health
Background:
Machine learning (ML) techniques improve disease prediction by identifying the most relevant features in multidimensional data. We compared the accuracy of ML algorithms for predicting incident diabetic kidney disease (DKD).
Methods:
We utilized longitudinal data from 1365 Chinese, Malay, and Indian participants aged 40–80 y with diabetes but free of DKD who participated in the baseline and 6-year follow-up visit of the Singapore Epidemiology of Eye Diseases Study (2004–2017). Incident DKD (11.9%) was defined as an estimated glomerular filtration rate (eGFR) <60 mL/min/1.73 m2 with at least 25% decrease in eGFR at follow-up from baseline. A total of 339 features, including participant characteristics, retinal imaging, and genetic and blood metabolites, were used as predictors. Performances of several ML models were compared to each other and to logistic regression (LR) model based on established features of DKD (age, sex, ethnicity, duration of diabetes, systolic blood pressure, HbA1c, and body mass index) using area under the receiver operating characteristic curve (AUC).
Results:
ML model Elastic Net (EN) had the best AUC (95% CI) of 0.851 (0.847–0.856), which was 7.0% relatively higher than by LR 0.795 (0.790–0.801). Sensitivity and specificity of EN were 88.2 and 65.9% vs. 73.0 and 72.8% by LR. The top 15 predictors included age, ethnicity, antidiabetic medication, hypertension, diabetic retinopathy, systolic blood pressure, HbA1c, eGFR, and metabolites related to lipids, lipoproteins, fatty acids, and ketone bodies.
Conclusions:
Our results showed that ML, together with feature selection, improves prediction accuracy of DKD risk in an asymptomatic stable population and identifies novel risk factors, including metabolites.
Funding:
This study was supported by the National Medical Research Council, NMRC/OFLCG/001/2017 and NMRC/HCSAINV/MOH-001019-00. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
-
- Epidemiology and Global Health
The NIH-funded RECOVER study is collecting clinical data on patients who experience a SARS-CoV-2 infection. As patient representatives of the RECOVER Initiative’s Mechanistic Pathways task force, we offer our perspectives on patient motivations for partnering with researchers to obtain results from mechanistic studies. We emphasize the challenges of balancing urgency with scientific rigor. We recognize the importance of such partnerships in addressing post-acute sequelae of SARS-CoV-2 infection (PASC), which includes ‘long COVID,’ through contrasting objective and subjective narratives. Long COVID’s prevalence served as a call to action for patients like us to become actively involved in efforts to understand our condition. Patient-centered and patient-partnered research informs the balance between urgency and robust mechanistic research. Results from collaborating on protocol design, diverse patient inclusion, and awareness of community concerns establish a new precedent in biomedical research study design. With a public health matter as pressing as the long-term complications that can emerge after SARS-CoV-2 infection, considerate and equitable stakeholder involvement is essential to guiding seminal research. Discussions in the RECOVER Mechanistic Pathways task force gave rise to this commentary as well as other review articles on the current scientific understanding of PASC mechanisms.