Effect of vaccinating health care workers to control Ebola virus disease: a modelling analysis of outbreak data

Background Health care workers (HCW) are at risk of infection during Ebola virus disease outbreaks and therefore may be targeted for vaccination before or during outbreaks. The effect of these strategies depends on the role of HCW in transmission which is understudied. Methods To evaluate the effect of HCW-targeted or community vaccination strategies, we used a transmission model to explore the relative contribution of HCW and the community to transmission. We calibrated the model to data from multiple Ebola outbreaks. We quantified the impact of ahead-of-time HCW-targeted strategies, and reactive HCW and community vaccination. Results We found that for some outbreaks (we call “type 1”) HCW amplified transmission both to other HCW and the community, and in these outbreaks prophylactic vaccination of HCW decreased outbreak size. Reactive vaccination strategies had little effect because type 1 outbreaks ended quickly. However, in outbreaks with longer time courses (“type 2 outbreaks”), reactive community vaccination decreased the number of cases, with or without prophylactic HCW-targeted vaccination. For both outbreak types, we found that ahead-of-time HCW-targeted strategies had an impact at coverage of 30%. Conclusions The optimal vaccine strategy depends on the dynamics of the outbreak and the impact of other interventions on transmission. Although we will not know the characteristics of a new outbreak, ahead-of-time HCW-targeted vaccination can decrease the total outbreak size, even at low vaccine coverage. summary Targeting health care workers for Ebola virus disease vaccination can decrease the size of outbreaks, and the number of health care workers infected. The impact of these strategies decrease depends on timing, coverage, and the dynamics of the outbreak.


Outbreak Data
Information on the occupation of cases is rarely available [16], although this information is critical to determining the role of HCW in transmission. We used data for outbreaks where we could find occupation (HCW or not) of cases. This resulted in twelve timeseries drawn from local outbreaks during the West African epidemic, and from the large 1995 Kikwit outbreak in the Democratic Republic of Congo (DRC) (Supplementary Section 1). All twelve timeseries are provided in the supplement, and five (for brevity) are shown in Figure 1. We noted that the number, timing, and dynamics of HCW infections during these outbreaks were not consistent, and we therefore used the dynamics of HCW and community infections to classify the outbreaks into types.

Classification of EVD outbreaks into two types
We compared the HCW infection dynamics of the outbreaks and determined that they fell loosely into two types. The distinguishing characteristics we used were grouped into four categories: i) the proportion of HCW and community infected through time; ii) the shape and timing of the cumulative distribution of HCW infections; iii) the weekly proportion of HCW infected; and iv) the total size of the outbreak in both number of cases and duration (Supplementary section S5). Data from Kikwit (1995, DRC) were uniquely detailed and therefore provided an opportunity to quantify the role of HCW in transmission by fitting a mechanistic model. According to our classification, Kikwit was a "type 1" outbreak (described in Results), and therefore we also calibrated the transmission model to the other observed outbreak pattern: "type 2".

Kikwit outbreak data
These data contained epidemiologically-inferred links between cases, the occupations of both infectors and infectees, and daily granularity of symptom onset (Supplementary Section 2).
The outbreak started with infrequent cases in rural areas before introduction to Kikwit General Hospital on April 7 th (Figure 1a) [6]. On May 2 nd a haemorrhagic fever was diagnosed, on May 8 th this was confirmed as EVD, and on May 10 th international assistance was initialised. Further control measures started on May 12 th (Supplementary Section S2). The final case died on July 16 th , resulting in 317 cases reported, 248 deaths, and a case-fatality ratio of 78% [17].
For each case we used their occupation; the occupation of their likely infector (obtained by real-time epidemiological investigation); date of onset; and date of recovery or death. There were some missing data in each field (Supplementary Section S2). We censored cases with date of onset before April 7 th , when the first case was admitted to Kikwit General Hospital, which gave 284 cases, of whom 73 were HCW (26%). A likely infector was available for 191 cases.

Transmission Model
We developed a deterministic compartmental model of EVD transmission stratified by occupation, where individuals were either HCW (h) or community (c) (Figure 3

Calibration of type 2 outbreak scenario
Data from available type 2 outbreaks did not include links between cases, and therefore were too incomplete to fit the same model framework. Observed type 2 outbreaks were characterised by a longer time period of HCW infections with no early rapid increase, and the epidemics were longer and larger in both occupation groups (see Results and Figure 2). We used published evidence to calibrate the type 2 scenario. We used initial HCW and community reproduction numbers from a large study of transmission in Guinea [21], which found low HCW-related transmission, and higher community-related transmission.
Contemporaneous analyses of the West African outbreak suggested a pattern of initially sustained transmission in the community, followed by a slow decline in transmission [22][23][24][25].
Therefore, we calibrated the parameters of the sigmoid functions to give a slow decrease in ",$% . We computed the four reproduction numbers between each occupation group using these published estimates for the overall reproduction numbers. To fully quantify the uncertainty, we used the parameter uncertainty from fitting the Kikwit data. The modifications we made to the transmission parameters ( Table 2) resulted in simulated outbreaks with higher community reproduction number and slower transmission decrease in type 2 compared with type 1 outbreaks. We kept the same parameters for population size, number of HCW, and reporting fractions as in Kikwit, which allowed direct comparison of type 1 and 2 scenarios, and therefore the impact of vaccination.

Simulation of vaccination
We extended the model to include vaccination of HCW and community and compared the impact of eight vaccination strategies (Table 3). We sampled 600 parameter sets from the joint posterior distribution and generated 15 stochastic simulations for each. We compared the number of cases and the time to extinction (0 individuals in E or I) to the baseline scenario without vaccination for each parameter set and random number seed. We used the parameter uncertainty inferred from the Kikwit data for both type 1 and type 2 outbreaks, and report 95% credible intervals in the text.
We simulated ahead-of-time HCW vaccine coverage values of 50%, 30%, and 10%. These values reflect the high turnover of HCW in recently affected countries [26,27], and the possibility that protection could wane. Vaccine coverage should be interpreted as effective levels of coverage: 30% coverage is equivalent to 100% vaccination of HCW and waning to 30% protection, or as 30% vaccination and 100% protection.
Vaccination reduced susceptibility to infection ( Figure 3). For single-dose vaccine, efficacy was 90%, and protection was reached after one week. For prime-boost vaccine, efficacy was 90, where 80% was reached one week after prime, and boost was 28 days after prime. We simulated vaccination of 1500 people per day, which was an operational maximum suggested by field teams. Reactive vaccination started with unvaccinated HCW and continued until all HCW and 70% of community members were vaccinated, at the same rate for single dose and prime-boost vaccination.
In the type 1 scenario, reactive strategies began on day 20 (April 27 th ), which is when health authorities were alerted to an outbreak of bloody diarrhoea in Kikwit [17]. This was earlier than detection of EVD, but we assumed that there would be improved surveillance and quicker EVD confirmation compared with 1995. For type 2 simulations, early transmission was slower. We therefore started reactive vaccination when the number of cases was similar to the type 1 outbreak on day 20, which was day 40 (median=53 cases in type 1, 38 in type 2).
At 40 days, no simulation of the type 2 scenario had more than 100 cases. This is similar to the number of reported cases at commencement of vaccination in the recent outbreak in Nord In ahead-of-time vaccination strategies of type 1 outbreaks the number of exposed and infected HCW at the start of the epidemic simulation were drawn from independent Poisson distributions with means from the joint posterior. For type 2 outbreaks, epidemics were seeded with 5 infected and 5 exposed community members (Supplementary Section S4.5).

Sensitivity analysis
The type 2 scenario from published parameter estimates implies that HCW have lower onward transmission than community members. To explore this assumption, we conducted a sensitivity analysis where HCW transmission mirrors the transmission characteristics of community members, where both have moderate transmission and a later time of decrease in transmission (Supplementary Section S7).

Classification of outbreaks into two types
Using key characteristics of the HCW and community transmission dynamics, we classified twelve localised EVD outbreaks into two broad types ( Figure 2, Table 1, and Supplement S5).
In both outbreak types, HCW were at high risk of infection, however, in type 1 outbreaks there was an early rapid increase in HCW incidence and in the cumulative proportion of HCW infected. Type 1 outbreaks also had a higher total proportion of HCW infected, shorter duration of HCW infections, and a smaller total outbreak size. In contrast, type 2 outbreaks exhibited a lower overall proportion of HCW infected, and a less obvious time period of high HCW incidence, with longer period of HCW infections. These outbreaks also showed a lower overall proportion of HCW infected, and larger total outbreak size.
The outbreak types are broad classifications, based on a combination of features of the dynamics of HCW infections. By classifying outbreaks in this way, we were able to determine the effect of HCW-targeted vaccination strategies under the range of observed transmission scenarios.  Figure 4). In contrast, the within-community reproduction number was less than one, and therefore transmission was not sustainable.

Fit of model to Kikwit data
Although there was low per capita transmission from the community to HCW (median Rch=0. 16), this represents a considerable risk to HCW: on average, each eight community cases infected one HCW. Overall, the net reproduction number at the start of the study period was 2.98 (2.11-4.36), with a major contribution from HCW, despite their low number. The timing and shape of the decrease in transmission depended on the occupation of cases ( Figure   4 and Table 1). We inferred an early and rapid decrease in HCW-related transmission, however we found that within-community transmission decreased several weeks later. The

Comparison to type 2 scenario
Although it was not possible to fit the model directly to a type 2 outbreak, simulations of the type 2 scenario resulted in similar characteristics to those observed ( Figure 2), matching patterns of cumulative proportion of HCW infected either through time or in total, as well as characteristics of number infected at each stage of the outbreak (Supplement S5).

Ahead-of-time HCW-targeted vaccination strategies
These strategies had greater effect on total outbreak size in type 1 scenarios, whereas reactive and combined strategies had greater effect in type 2 scenarios ( Figure 5). Impact of ahead-oftime HCW-targeted strategies depended on the coverage achieved: 50% coverage of HCW decreased the total number of cases in type 1 outbreaks (type 1: 121 (50-243), type 2: 813

Sensitivity analysis on HCW-related transmission in type 2 scenario
The overall size of outbreaks was smaller because of the lower total reproduction number (Supplementary Section S7). The general pattern of effect of different vaccination strategies was the same as the type 2 scenario parameterised to [21], which gives confidence in the generalisability of our findings.

Discussion
We used as much information on HCW-related transmission as possible to classify EVD outbreaks into two broad types: the first, where the infection is catalysed by HCW and community transmission is low, was observed in Kikwit (1995)  Data from Kikwit (1995) provide uniquely detailed information on likely source of infection and timing of symptoms. However, some data were missing, and the suggested routes of infection may not be correct. We did not consider transmission after death, or that some individuals (such as carers) may be more likely to be infected, which could affect estimates of transmission rate.
When generalising our findings to the current context, diagnosis and testing may now occur sooner than during the 1995 Kikwit outbreak. In addition, the rapid change in HCW-related transmission may partially have resulted from an increase in use of PPE. In the current context, HCW may have more rapid access to PPE, or have improved awareness of EVD, and therefore the change in transmission rate could be different. This would decrease the impact of HCW-targeted vaccination in type 1 scenarios, and therefore our findings may be on the upper end for ahead-of-time HCW vaccination.
We used data for all the outbreaks and sub-outbreaks within a larger epidemic that had information on the occupation of cases. It is possible that there is some misclassification of occupation, or that the definition of HCW changed from one outbreak to another, especially during the long West African epidemic. Incorrect classification could affect the reproduction numbers attributed to each group, although we do not have evidence of systematic misclassification. We conducted sensitivity analyses on the number of HCW and found that it did not affect the findings of vaccine impact.
Our model framework could not test other potential vaccination strategies, such as ring vaccination, because we did not track specific contacts that individuals make. We assumed that individuals mix randomly within occupation groups, with no heterogeneity within         (1-φ )γ (1-v )   Simulations without vaccination are shown in grey, and each colour represents a vaccination strategy (Table 3): reactive mass vaccination with (a) prime-boost vaccine or (b) single dose vaccine; ahead-of-time HCW vaccination only, with coverage in HCW of (c) 10%, (d) 30% or (e) 50%; ahead-of-time HCW vaccination plus reactive mass vaccination, with coverage in HCW of (f) 10%, (g) 30% or (h) 50%. We give the 75% CI due to high variation in the simulation sets, and 95% CI are given in the Supplement (S6.5) Note different y-axes.