Optimizing COVID-19 surveillance in long-term care facilities: a modelling study

We simulated nosocomial outbreaks of COVID-19 using a dynamic, stochastic, individual-based Susceptible Exposed Infectious Recovered (SEIR) model of SARS-CoV-2 transmission coded in C++ [39]. The model is described fully in Additional File 1 using the Overview, Design concepts, and Details (ODD) protocol for individual-based modelling [12, 18, 38‐57]. The goals of our model are to simulate (i) dynamic inter-individual contacts among patients and staff in an LTCF setting, (ii) transmission of SARS-CoV-2 along simulated contact networks, and (iii) clinical progression of COVID-19 among individuals infected with SARS-CoV-2. Throughout, COVID-19 refers to any case of SARS-CoV-2 infection, and not only symptomatic cases.

We used data from the i-Bird study to inform the population structure and dynamic inter-individual contact network used in our model. The i-Bird study has been described elsewhere [41, 58]; briefly, close-proximity interactions were measured every 30 s by sensors worn by all patients and staff over a 17-week period in 2009 in a rehabilitation hospital in northern France. There were 170 patient beds across the five wards of this LTCF, and staff were distributed across 13 categories of employment, grouped here as HCWs (caregiver, nurse, physiotherapist, occupational therapist, nurse trainee, physician, and hospital porter) or ancillary staff (hospital services, administration, other rehabilitation staff, management, logistical staff, and activity coordinator/hairdresser) (Fig. 1a).

This population structure was used in our model, but a novel contact network was simulated to account for missing data resulting from imperfect sensor compliance in the raw contact network (Fig. 1b; described in Additional File 1). Real patient admission and staff timesheet data were used to determine who was present in the LTCF: there were on average 170 patients and 240 staff present each week, stratified by ward and type of individual (e.g. patient, nurse, logistical staff) in Additional File 2: Table S1. The probability of coming into contact with another individual at each model time-step (30 s) was estimated from the raw contact data and stratified by hour of the day (e.g. 8:00:00–8:59:30, 9:00:00–9:59:30), day of the week (weekday vs. weekend), ward and type of individual. At each time-step, new contacts were simulated based on these probabilities, and contact durations were drawn stochastically from log-normal distributions stratified by the same variables. Contact behaviours were comparable between the raw and simulated networks, and fidelity of the simulated network has been validated previously by its ability to reproduce transmission dynamics from a real outbreak of methicillin-resistant Staphylococcus aureus in this LTCF [55, 58].

Parameter estimates from the literature were used to characterize SARS-CoV-2 transmission and clinical progression of COVID-19 (see Additional File 1: Table S2). We assumed that susceptible patients and staff could become infected with SARS-CoV-2 if in direct contact with an infectious individual, with the probability of transmission depending on the duration of contact. Owing to a lack of data from healthcare settings, SARS-CoV-2 transmissibility was estimated using data from the community (see Additional File 1). Assuming R₀ = 3 in France prior to lockdown [44], and using pre-pandemic inter-individual contact data [45], we derived a transmission probability per minute spent in contact with an infectious individual p = 0.14%. Under baseline assumptions in our simulated LTCF, this estimate for p resulted in a mean R₀ = 4.04 (Additional File 2: Fig. S1). This is consistent with the finding that, for any given value of p, R₀ is expected to vary between community and healthcare settings because of fundamental differences in inter-individual contact behaviour [43]. However, to reflect considerable uncertainty in the true value of p, we considered extreme values in sensitivity analysis (p = 0.07%, 0.28%).

Clinical progression of COVID-19 was characterized by a modified SEIR process, with (i) a non-infectious exposed period of 2–5 days, (ii) an infectious pre-symptomatic period of 1–3 days, (iii) an on-average 7-day infectious “symptomatic” period with three levels of symptom severity (severe, mild or asymptomatic), and (iv) eventual recovery with full immunity (Fig. 1c). Together, (i) and (ii) amount to an incubation period of 3–8 days including a 1–3-day window of pre-symptomatic transmission, consistent with estimates used elsewhere [18, 19, 59]. For (iii), we assumed that 70% of infected individuals develop clinical symptoms [51, 60, 61], 20% of which develop severe/critical symptoms [52]. Durations for each stage of infection were drawn probabilistically from their respective distributions for each infection. We assumed no difference in average time to symptom onset for mild symptomatic and severe symptomatic cases. As surveillance strategies were evaluated only for detection of outbreaks, death and potential long-term clinical outcomes were not explicitly simulated. Each simulation began by introducing one non-symptomatic index case (either exposed, pre-symptomatic or asymptomatic) into the facility on the first day of each simulation (t = 0). In the baseline scenario, both patients and staff introduced SARS-CoV-2 into the LTCF.

Simulated COVID-19 outbreaks were described using various outcome measures, including infection incidence, infection prevalence, case distribution (proportion of infections among patients, HCWs, and ancillary staff), outbreak size (cumulative number of cases after 12 weeks of unmitigated transmission, the full duration of each outbreak simulation), and outbreak size upon first presentation of COVID-19 symptoms.

We developed a stochastic surveillance algorithm to evaluate a range of surveillance strategies for their efficacy, timeliness, and resource-efficiency in detecting COVID-19 outbreaks. Strategies varied according to who received conventional nasopharyngeal swabs and RT-PCR tests, and with what priority. Demographic and clinical indications were used to administer swabs and tests, beginning on the first day of each outbreak simulation and assuming a daily maximum testing capacity ranging from 1 to 32 tests/day. We assumed a 24-h lag from swab to test result, perfect specificity (true negative rate) and imperfect sensitivity (true positive rate). Using data from a meta-analysis of RT-PCR for detection of SARS-CoV-2 in upper respiratory samples, sensitivity was modelled as a function of time since infection, ranging from 0% on the first 2 days after infection, to 33% after 4 days, peaking at 80% after 8 days, and gradually decreasing thereafter (Additional File 1: Fig. S1) [54]. Each test result was determined stochastically according to these parameters. The algorithm is described in further detail in Additional File 1.

Four types of surveillance were evaluated: (i) testing individuals with particular indications, (ii) random testing, (iii) testing cascades, and (iv) group testing (Table 1). Each was further divided into distinct surveillance strategies. For (i), three indications were considered: presentation of severe COVID-like symptoms (reference strategy), presentation of any COVID-like symptoms, or new admission to the LTCF. For (ii), tests were randomly distributed among patients, HCWs, or all patients and staff. In contrast to (i) and (ii), strategies (iii) and (iv) were conceived as hierarchical testing protocols, in which individuals presenting with severe COVID-like symptoms were always tested first to reflect their clinical priority. Remaining tests were subsequently allocated via cascades (iii) or as a single group test (iv).

For (iii), testing cascades were conceived as mixed testing strategies combining (i) and (ii), in which multiple indications were considered simultaneously but ordered according to their perceived clinical priority. If there were more tests available than individuals indicated for testing, remaining tests were distributed randomly among remaining patients, such that cascades always maximized daily testing capacity. For (iv), clinical specimens from individual swabs were pooled together and tested as one, up to a maximum of 32 swabs per test in the baseline analysis. Various studies have demonstrated the efficacy of this method for SARS-CoV-2 detection, with sufficient diagnostic accuracy to detect a single SARS-CoV-2-positive specimen pooled with 30+ negative specimens [38, 62, 63]. Yet diluting positive specimens nonetheless reduces the concentration of viral RNA in the sample, which should reduce sensitivity of a group test compared to an individual test of the same positive specimen [64]. We estimated a 0.7% reduction in test sensitivity per additional negative specimen added to a group sample when assuming a typical RT-PCR Cycle threshold cut-off (Ct = 40), and for sensitivity analysis estimated a faster rate (1.3%) corresponding to a stricter threshold (Ct = 35) (see Additional File 1). Various group testing procedures have been proposed elsewhere [33, 34, 37, 65, 66]; here we evaluated a simple two-stage “Dorfman” protocol that does not require additional investment or infrastructure, but which requires all individuals included in the initial group test to be re-swabbed and re-tested individually upon a positive group test result in order to determine which individual(s) is (are) infected [35, 67].

Each swab resulted in one RT-PCR test, except for group testing strategies, in which multiple swabs were combined per test. Admission-based tests were administered upon a patient’s arrival to the LTCF, and symptom-based tests were administered on the first day that symptoms appeared. We further assumed that no individuals refused swabbing/testing. Clinically, COVID-19 can resemble other common acute respiratory infections, such that individuals not infected with SARS-CoV-2 can nonetheless present with COVID-like symptoms and be indicated for symptom-based testing [68]. We used influenza-like illness as a proxy for COVID-like symptoms of aetiologies other than SARS-CoV-2. The daily incidence rate of influenza-like illness in our LTCF setting (1.1%) was calculated using data from 2008 to 2017 from French emergency departments (OSCOUR network) as the daily incidence rate of influenza-like illness among older adults (50–99 years) [53]. As with actual COVID-19, we assumed that 20% of these individuals also present with severe symptoms.

Surveillance strategies were evaluated for their ability to detect COVID-19 outbreaks using four primary outcome measures: first, the probability of detecting an outbreak (i) at any time t from the index case at t = 0, (ii) prior to any secondary cases (interpreted as the probability of detecting the index case before any nosocomial transmission), or (iii) prior to first presentation of COVID-19 symptoms. Second, detection lag, the number of days from the index case to outbreak detection (first positive test result). For group testing, this was taken as the date of the first positive group test result (first round of testing) and not the date of subsequent case identification (second round). We defined a maximum detection lag of 22 days, after which all outbreaks were assumed to be detected regardless of the surveillance strategy used. Third, outbreak size upon detection, the cumulative number of cases at first positive test result. Fourth, the total number of (i) nasopharyngeal swabs used and (ii) RT-PCR tests conducted until outbreak detection. Here, for group testing, this does include the second round of testing, i.e. resources required to individually re-swab and re-test all individuals included in the initial positive group test.

From a health-economic perspective, an efficient use of healthcare resources is one that yields better health outcomes than alternative uses of the same resources [69]. Efficiency can be measured using incremental analysis, in which the additional cost of a particular intervention compared to a reference baseline is scaled by its additional health benefit [70]. This is traditionally expressed as the incremental cost-effectiveness ratio using monetary costs and standardized units of health benefit (e.g. quality-adjusted life-years gained). To report on efficiency in terms of the surveillance cost and benefit outcomes measured in this study, we defined a similar metric, the incremental efficiency ratio (IER),

for each surveillance strategy S relative to the reference R. Efficiency results were calculated using the IER as the number of additional swabs and tests required per 1-case reduction in outbreak size upon detection (for simplicity, reported as mean additional swabs and tests used per case averted). Here, we took the perspective of an LTCF with a reference strategy of only testing individuals with severe COVID-like symptoms.

We ran a range of sensitivity analyses to account for uncertainty in (i) how SARS-CoV-2 was introduced into the LTCF, (ii) LTCF size and structure, (iii) transmissibility of SARS-CoV-2, and (iv) diagnostic sensitivity of RT-PCR for individual samples and (v) for group samples (see assumptions in Additional File 1). For each scenario, 100 epidemics were simulated using the transmission model. For each simulated epidemic, the surveillance algorithm was run 100 times across six testing capacities (1, 2, 4, 8, 16 or 32 tests/day), for a total 60,000 stochastic simulations for each surveillance strategy and each scenario. Across all scenarios, we also varied the maximum number of swabs potentially included per group test (2, 4, 8, 16, 32, or 64 swabs/test). Outcomes were only evaluated for epidemic simulations that resulted in nosocomial outbreaks, defined as simulations with ≥ 1 new case of COVID-19 within 21 days of the initial index case. Unless stated otherwise, outcome measures are reported as median values across all simulations, with uncertainties expressed as 95% uncertainty intervals, i.e. outcomes from the 2.5th and 97.5th percentiles. Supplementary results can be found in Additional File 2.

SARS-CoV-2 spread quickly, but with a great degree of stochasticity upon its random introduction to simulated LTCFs (Fig. 2, Additional File 2: Fig. S2). After 3 weeks of unmitigated transmission, a cumulative 86 (95% uncertainty interval 6–224) individuals were infected, predominantly other patients (median 72%), then HCWs (25%), and ancillary staff (3%) (Additional File 2: Table S2). Outbreaks were characterized by a median lag of 9 (2–24) days between the non-symptomatic index case entering the LTCF and first presentation of mild COVID-19 symptoms among any patient or staff in the facility (Additional File 2: Table S3). By the time symptoms emerged, an additional 5 (0–29) individuals had acquired SARS-CoV-2 but were not (yet) showing symptoms (Additional File 2: Table S4). Lags were longer for first presentation of severe COVID-19 symptoms (15 days from index case, 4–28), coinciding with a greater cumulative number of secondary infections (25, 0–101).

Surveillance strategies varied in their ability to detect emerging COVID-19 outbreaks. Surveillance efficacy depended on the stochastic nature of outbreaks, including how many, and which types of individuals became infected over time (Fig. 2c, d). Outbreaks grew exponentially at their outset, so delaying outbreak detection by just 1 or 2 days potentially coincided with tens more infections (Additional File 2: Fig. S3). Five days from the index case entering the LTCF, only 2 (1–12) individuals were infected; after 10 days, 9 (1–44) were infected; and after 15 days, 36 (2–124) patients and staff were infected.

Across all testing capacities, only testing individuals with severe COVID-like symptoms was among the least effective surveillance strategies considered (Fig. 3). This “reference” strategy took a median 16–17 days to detect outbreaks and had a 2–4% probability of detecting the initial index case prior to any secondary cases (Additional File 2: Fig. S4). Instead of only severe symptoms, testing individuals with any COVID-like symptoms was more effective, taking 9–15 days to detect outbreaks, with a 3–14% probability of detecting index cases prior to any secondary cases. Only testing patients at admission was overall ineffective by right of detecting neither staff index cases nor ongoing outbreaks already underway in the LTCF, resulting in long median delays to outbreak detection despite comparatively high probabilities of detecting COVID-19 prior to any secondary cases (10–33%). In the scenario where only new patients introduced SARS-CoV-2 into the LTCF, there was a 34% probability of detecting the index case when testing all patients upon admission, or 66% in a sensitivity analysis considering higher and more stable RT-PCR sensitivity. For random testing strategies, surveillance was highly ineffective when few tests were available, but increasingly effective at higher testing capacities. Conversely, for indication-based strategies, efficacy plateaued when capacity exceeded the number of individuals indicated for testing (Additional File 2: Fig. S4).

At high testing capacity (16–32 tests/day, ≈ 9–19 tests/100 beds/day), testing cascades were the most effective surveillance strategies. The four cascades considered here detected outbreaks within a median 7–10 days, coinciding with just 3–6 COVID-19 infections among all patients and staff. Cascades had a 19–36% chance of detecting outbreaks prior to any secondary cases, a 26–46% chance prior to the emergence of any COVID-19 symptoms, and a 64–85% chance prior to severe COVID-19 symptoms. Cascades that included both new patient admission and presentation of any COVID-like symptoms as indications for testing were most effective.

At low testing capacity (1 or 2 tests/day, ≈ 0.6–1.2 tests/100 beds/day), group testing was the most effective form of surveillance considered. Compared to the reference (16–17 days) and cascades (16–17 days), outbreaks were detected within 11–14 days (coinciding with a cumulative 9–16 infections) when pooling random patients, or 12–14 days (15–23 infections) when pooling random HCWs. At this low capacity, it was also more effective to pool symptomatic individuals in group tests (10–12 days, 9–17 infections) than to test them individually (12–15 days, 17–27 infections) because individuals with non-COVID but COVID-like symptoms were also “in competition” for limited tests. Compared to the baseline protocol, which assumed a maximum of 32 swabs/test, group testing was less effective given fewer swabs per test, despite potentially higher test sensitivity. For example, when pooling randomly selected patients in daily group tests, outbreaks were detected within 11–14 days at 32 swabs/test, 12–15 days at 8–16 swabs/test, and 14–17 days at 2–4 swabs/test. In a sensitivity analysis considering a stricter RT-PCR diagnostic threshold (Ct = 35 instead of 40), group testing was most effective at 16–32 swabs per test (12–15 days), but declined substantially for greater group sample sizes (16–17 days at 64 swabs/test) (Additional File 2: Fig. S5).

Surveillance strategies varied considerably in their use of testing resources (Additional File 2: Fig. S6) and in their efficiency for improving COVID-19 outbreak detection relative to the reference strategy (Fig. 4). The reference used the fewest swabs and tests, on average < 1/day regardless of the assumed daily testing capacity (owing to a low daily incidence of severe COVID-like symptoms). At high testing capacity (16–32 tests/day), the high incremental efficacy of cascades (outbreak detection a mean 5–8 days earlier than the reference, prior to 22–27 additional infections) resulted from extensive resource use (104–276 additional tests and swabs), for mean efficiencies of 4.0–11.2 additional swabs and tests per case averted. Although simply testing all patients and staff with any COVID-like symptoms was less effective than using testing cascades, it was a more efficient means to improve surveillance (mean 1.3 additional tests per case averted).

Group testing strategies were generally efficient with respect to tests, but used highly variable numbers of swabs to detect outbreaks. At high swabbing capacity (16–32 swabs/group test, ≈9–19 swabs/100 beds/day) and across all testing capacities, pooling randomly selected patients used a mean 11–38 excess tests to detect outbreaks 2–4 days earlier and prior to 14–21 additional infections (0.8–1.8 additional tests per case averted), but a median 94–384 additional swabs (6.7–18.7 additional swabs per case averted). Pooling the same number of randomly selected HCWs was less efficient than pooling patients (detection = 1–4 days earlier, prior to 6–14 infections; efficiency = 1.3–2.8 additional tests and 15.9–27.5 additional swabs per case averted). By contrast, for all scenarios and testing capacities considered, pooling individuals with any COVID-like symptoms was among the most efficient strategies in terms of both swabs and tests. In the most resource-limited scenarios (1–2 tests, ≈ 0.6–1.2 tests/100 beds/day; 2–8 swabs per group test, ≈ 1–5 swabs/100 beds/day), this was both the most effective means to detect COVID-19 outbreaks and the most efficient means to improve surveillance from the reference (detection = 1–6 days earlier, prior to 11–22 additional infections; efficiency = 0.3–0.6 additional tests and 1.0–1.3 additional swabs per case averted).