The impact of quality-adjusted life years on evaluating COVID-19 mitigation strategies: lessons from age-specific … – BMC Public Health

The purpose of the analyses presented here was to evaluate distinct vaccine uptake strategies in the context of COVID-19. We aim to achieve qualitative results by exploring counterfactual scenarios driven by vaccine uptake. Therefore, we exploited the unfolding of the Belgian COVID-19 crisis with the induction of SARS-COV-2 in February 2020 and the emerging Alpha, Delta, and Omicron (BA.1 and BA.2) VOCs. Each simulation spans the first two years of the COVID-19 pandemic, running from March 2020 to February 2022. It includes age-specific uptake of first, second, and booster doses of adenovirus and mRNA-based vaccines. The uptake scenarios being examined vary between August 2021 and February 2022, which is the period our results primarily focus on.

We extended a previously published stochastic transmission model for SARS-CoV-2 in Belgium by Abrams et al.[27], by including COVID-19 vaccination, emergence of different VOCs and waning immunity. Our transmission model is a discrete-time age-structured compartmental model with a chain-binomial transition process between various disease compartments that can be categorised into susceptible, exposed, infected, recovered, and death states. Overall, after exposure to the pathogen and acquiring infection, an individual becomes infectious after a latent period and moves to a pre-symptomatic state. Subsequently, individuals develop symptoms or remain asymptomatic, before recovering. Symptomatic infections start mild and have an age-specific probability of progressing to serious illness, implying hospitalisation with or without admission to the ICU. We also account for disease-related mortality of hospitalised cases. The original model formulation is duplicated into a two-strain compartmental structure (see Fig.1), and transitions between multiple copies of the two-strain model (see Fig.2) allowed for waning immunity against infection and severe disease. Further elaboration on the construction of the model, specifically based on multiple substructures, is presented in the subsequent paragraphs.

The model structure proposed by Abrams et al.[27] including ten 10-year age groups has been adapted to a two-strain version with a common susceptible class and a duplication of all infection-related health states. Our model structure (see Fig.1) enabled co-circulation of two variants at the same time with distinct properties with respect to susceptibility, latent period, disease severity, hospital length of stay, mortality, and vaccine-related protection. To cover the newly emerging Delta VOC, we re-used the health states of the dominated original strain in the simulation after book keeping all states. A similar transition was made with the Omicron VOC when the Alpha VOC was fully dominated by the Delta VOC. More information about model dynamics and parameters is provided in the Supplementary Information. Our model operates starting from March 1st, 2020, and accounts for the emergence of new pathogen strains and the administration of various vaccine doses. In the early stages, these factors are represented in small amounts, with heterogeneity and randomness playing critical roles. Even slight variations can become amplified over time. Later on, after COVID-19 vaccination is introduced and attains high coverage, the size of the remaining susceptible population becomes small. This makes the stochastic nature of infection and subsequent processes like hospitalisation and death increasingly significant, especially since the model is calibrated using age-specific incidence data for each of these stages. These elements highlight the importance of the stochastic nature of our compartmental model in accurately reflecting and predicting evolving dynamics.

Health states and transitions in the two-strain transmission model. The model structure is described in the main text and model parameters are listed in the Supplementary Information

We used the reported social contact rates of 42 Belgian CoMix survey waves between April2020 and March2022[28, 29] as proxy for effective contacts that allow disease transmission according to the social contact hypothesis[30]. CoMix has been designed as a collection of surveys in which a panel of participants retrospectively reports all social contacts made from 5:00 AM on the day preceding the survey up to 5:00 AM on the day of the survey. A contact was defined as an in-person conversation of at least a few words or a skin-to-skin contact[28]. Changes in transmission that are not directly attributable to changes in contact behaviour are captured in age-specific proportionality factors. They represent, for example, changes in compliance to (social distancing) measures, seasonality effects, and shifts in the location-specific contact intensity (e.g., contacts inside are more risky than contacts outside). For each wave, we estimated age-specific proportionality factors to translate social contact rates into transmission rates that capture age-specific susceptibility and infection-related risk behaviour associated with social contacts[31].

The introduction and presence of VOCs in the model population are taken into account in the parameter estimation process based on the baseline genomic surveillance of SARS-CoV-2 in Belgium by the National Reference Laboratory[32]. To simulate the replacement of the original strain in 2021, we aggregated all Alpha, Beta, and Gamma VOC samples that were identified, which we refer to hereafter as Alpha VOC infections, when estimating the penetration of the VOCs into the Belgian population. We attributed the growth advantage of the Alpha VOC completely to transmissibility and ignored the potential effect of immune escape. We assumed that there was no change in the probability of hospital admission for the (aggregated) Alpha VOC. Conflicting post hoc observations have been reported on the severity of this VOC[33]. Therefore, we have chosen to highlight the significant role that increased transmissibility potential plays in hospitalisations and mortality, regardless of any direct effect of the variant on severity.

For the Delta VOC, we account for increased transmissibility and adopted an adjusted hazard ratio for hospitalisation of 2.26 relative to the Alpha VOC based on a cohort study conducted in the UK[34]. Due to the lack of age-specific information to align the reported 95% confidence interval of [1.32;3.89] with our age-specific model design, we opted to use the estimated mean value without considering parameter uncertainty. This adjusted hazard ratio was essential to match the reported incidence of hospitalisations with genomic surveillance data on the Delta VOC[32].

With the emergence of the Omicron VOC, studies[35, 36] indicated a change in the incubation period and the serial interval, which contributes to its transmission advantage. This had a large impact on the estimated reproduction number and the effect of restrictive measures. As such, we included a VOC-specific latent period in our transmission model, which was inferred specifically for the Omicron VOC during the calibration process. Furthermore, Omicron-specific hazard ratios for hospitalisation were pivotal to capture the trends observed in 2022. We adopted age-specific hazard ratios for hospital attendance with the Omicron VOC compared to the Delta VOC, from a cohort study in the UK[37]. More specifically, we used for our 10-year age bands: 1, 0.89, 0.67, 0.57, 0.54, 0.42, 0.32, 0.42, 0.49 and 0.49. The simulation period covers both Omicron sub-lineages BA.1 and BA.2, the latter became dominant in Belgium on February 28th, 2022. The differences in transmission for BA.2 are absorbed in the wave-specific proportionality parameters for February/March 2022.

All the levels of protection adopted are summarised in Table1. We used a leaky vaccine implementation approach in which vaccination with 74% effectiveness implies that for a vaccinated individual the probability of acquiring infection is 74% lower compared to a non-vaccinated individual of the same age. Vaccine-induced immunity against infection is implemented as a step function in terms of protection against infection 21days after the first dose of vaccine. Protection induced by second and booster vaccine doses is assumed to be fully achieved 7days after vaccine administration (i.e., depending on the maximal effectiveness of the vaccine as reported in Table1). We consider the differences between mRNA- and adenovirus-based vaccines in how they induce immunity and in terms of protection. We assumed that vaccinated individuals who acquire infection are at a lower risk of hospital admission with COVID-19 and all booster doses in Belgium are mRNA-based vaccines. Given our model structure, reported protection levels against hospital admission were applied as protection against severe disease, which ultimately leads to hospital admission. Vaccinated individuals (with or without a booster) who acquire infection do not have a lower risk of transmitting the disease. This assumption is challenged in the sensitivity analysis.

Vaccine-induced protection and waning immunity have been included through duplication of the two-strain compartmental structure with uptake-based and time-specific transitions (see Fig.2). This model structure allowed us to explicitly keep track of vaccine type and dose-specific vaccine uptake and to differentiate protection against infection and severe disease between vaccine type and number of doses. The duplicated two-strain compartmental structure also allowed differential waning immunity against infection and severe disease.

We integrated waning immunity into our model by establishing a series of steps transitioning from complete protection to a state of diminishing immunity over an average period of 90 days. In the framework of the compartmental model, the waning rate is defined as the fraction of individuals transitioning from full protection per time unit, which inversely correlates with the average protection duration. Consequently, we incorporated submodels for diminishing vaccine-induced immunity, featuring levels of reduced protection as detailed in Table1. Initially, infection-induced immunity offers 100% protection, assuming individuals in the Recovered state are not susceptible to reinfection. Therefore, our model accounts for the decrease in infection-induced immunity by moving individuals from the Recovered to the Susceptible compartment within a submodel, which still affords a degree of protection against future infections. We assumed the effect of a booster dose independent of the immunity state upon vaccination, i.e., with or without prior infection or a specific vaccine scheme. We accounted for waning immunity after the booster dose with a dedicated submodel and an average transition time of 90 days. Note that even with waning immunity, vaccinated individuals maintain partial protection against subsequent infection and severe disease upon infection. VOC-specific protection levels for the booster dose have been derived from the literature (see Table1).

Vaccine uptake in the model is based on age-specific data at the national level reported by the Belgian Scientific Institute for Public Health, Sciensano[38]. By August 2021, on average 90% of the population aged over 20 years had completed their two-dose regimen with mRNA or adenovirus-based vaccines. On the contrary, about 10% of the 0-19-year-olds received two doses of an mRNA vaccine at that time. It is important to note that in August 2021, mRNA vaccines were only authorized for use in children aged 12 years and older by the European Medicine Agencys Committee for Medicinal Products for Human Use. Subsequently, in 2022, the authorisation was extended to younger children, initially to those aged 6 years and older, and later to infants as young as 6 months of age. The decision to administer booster doses at the end of 2021 was based on the evaluation by the European Medicines Agency that indicated an increase in antibody levels following a booster dose administered about 6 months after the second dose in individuals aged 18 to 55 years. Based on this evidence, first booster doses were recommended in Belgium for people 18 years and older at least 6 months after the second dose.

Full details on the type- and dose-specific vaccine uptake by age we included in the model is presented in Fig.S2. We did not explicitly account for risk-group vaccination, since our model structure did not facilitate more subpopulations with differential risk and potentially a more severe episode of COVID-19 disease once infected (i.e., a higher probability of hospitalisation and/or a higher probability of death, if hospitalised). In our analysis, we primarily considered age as the main determinant of risk and severity. The reported uptake of Pfizer-BioNtech (Comirnaty) and Moderna (Spikevax) vaccines are aggregated into one mRNA vaccine type. The relatively low number of reported Johnson & Johnson (Ad26.COV2.S) and Curevac (CV07050101) vaccines were aggregated in the model with the adeno-based AstraZeneca vaccine (ChAdOx1 or Vaxzevria) based on similarities in protection and waning immunity. Third doses (i.e. first booster dose) are included in the transmission model as a separate submodel with all health-related compartments. A comprehensive summary of vaccine uptake we included in our model is depicted in Fig.3, which presents also the scenarios discussed in the subsequent sections of the Methods.

Overview of the duplicated two-strain model structure to account for vaccine type- and dose-specific immunity against infection and severe disease in combination with differential waning immunity over time. The grey boxes embody the transmission structure included in Fig.1 while only the Susceptible and Recovered are shown here (with (R_i) representing (R_a) and (R_b)). More information on the waning states is included in Table1

We used Bayesian methods to fit our transmission model to multiple data sources, including daily hospital admissions and bed occupancy, early seroprevalence, genomic surveillance, and mortality data. In order to capture the full extent of the intrinsic variability of the model, we relied on Markov Chain Monte Carlo (MCMC) sampling with 60 chains in the calibration procedure. An adaptive Metropolis-within-Gibbs algorithm was used as MCMC sampler, and parameter priors were based on permutations of previously converged calibration results. The model parameters related to hospital incidence and VOC prevalence were estimated by gradually extending the time horizon over consecutive calibration runs for the stochastic model. The absence of age-specific data on daily hospital discharges and transitions between general wards and ICU hampered a likelihood approach to accommodate hospital occupancy in general and in the ICU. Therefore, the fitting of the model was performed using a multi-step procedure. First, all transmission-related model parameters were estimated while calibrating the model to the observed incidence data on hospitalisation, early seroprevalence and genomic surveillance as described above. Next, all parameters related to hospital and ICU occupancy (including discharge rates) were estimated based on minimising a least squares criterion for the distance between the observed and generated loads. Finally, the estimated mortality-related parameters are inferred again using a likelihood-based approach, distinguishing whether a hospital discharge was due to mortality or recovery. This multi-step procedure has been performed multiple times, of which the final iteration is described in TableS2. Finally, we selected the 40 best performing MCMC chains of the last step to derive parameter estimates for our simulation study.

We used hospital admissions with COVID-19 as a primary source of information to capture the burden of disease. During the development of the model, we observed that around 10%-20% of the admissions with the Alpha and Delta VOCs were primarily due to other pathologies, but patients who tested positive when admitted were transferred to the COVID-19 wards and counted in the COVID-19 hospital load. With the Omicron VOC, the difference between admissions with COVID-19 and for COVID-19 increased even more. Given our focus on hospital capacity, hence occupancy, hospital admissions with COVID-19 were most informative in combination with reported estimates for hospital stay.

We estimated a transmission advantage of the Alpha VOC compared to the original strain of 32% (95% CrI: 24-39%). For the Delta VOC, the transmission advantage compared to the Alpha strain was estimated to be 87% (95% CrI: 71-106%). For Omicron, we estimated an almost instant transition from the exposed to the pre-symptomatic infectious health state (which is in line with the shorter serial interval we referred to previously) and a transmission advantage compared to the Delta VOC of 35% (95% CrI: 9-70%). A comprehensive overview of the model parameters is presented in TableS3 of the Supplementary Information.

The baseline scenario consisted of all estimated parameters during the calibration of the compartmental model and fitted the national trends of SARS-CoV-2 pandemic in Belgium. This includes, for example, the emergence and dominance of the Omicron variant from December 2021 and the observed decrease in hospital admissions and deaths at that time. The full model output from March 2020 is presented in Fig.S4. The transmission model was based on bi-weekly social contact survey data, which allows for including adjusted behaviour over time in the model. The survey data represented changing contact rates, while the estimated proportionality factors captured differences in, for example, contact intensity, susceptibility and infectivity. These factors were age-specific and part of the parameter estimation process (see Fig.S5).

To estimate the burden of disease, we included the loss of QALYs from a published study on the model-based cost-effectiveness of SARS-CoV-2 vaccination along with physical distancing in the United Kingdom[22]. Disease morbidity estimates were obtained by multiplying the model-based incidence of mild infections, and hospitalised and ICU admitted patients with the QALY loss values in Table2. Disease-related mortality based on the quality-adjusted life expectancy[39] is obtained by combining the age-specific model estimations on mortality with the Belgian life expectancy for 2019 reported by Statbel[40] and the latest age-specific Belgian population norms based on EQ-5D-5L[25].

We explored retrospective counterfactual scenarios based on vaccine uptake in the presence or absence of the Omicron VOC. None of these scenarios explicitly included the importation of infected cases as a result of international travel except for the introduction of VOCs. We started from the final calibration of the model and the reported vaccine uptake scheme and explored proportionally increased uptake of two doses in 511-year-old children and first booster doses in adults over 18-years. Vaccine uptake levels and timing could be explored more in detail with additional objectives and trade-offs, although this analysis aims to provide a basis for predominantly qualitative interpretations. We allow for stochastic variation in the transmission process by running each of the 40 estimated model parameter sets 10 times, hence incorporating 400 model realisations in the final comparison. The number of realisations was determined through a process of model exploration and consideration of the trade-off between model realisations and computational feasibility due to model complexity.

To explore the impact of the uptake of the COVID-19 vaccine, we evaluated an adjustment of the uptake of the first booster dose in adults and an increase in the level of childhood vaccination. First, we changed the uptake of the first booster dose so that it matches the age-specific two-dose uptake levels by 1 March 2022 (see Fig.3). That is, we assumed that all those eligible for a first booster dose effectively received an mRNA booster dose. The reported first booster dose uptake in the Belgian adult population was 76%, so the additional uptake in the scenario analysis was rather limited. Secondly, we defined a scenario in which we arbitrarily included only 60% of the reported uptake of the first booster dose. The 40% reduction is applied uniformly across all age groups. A third adoption scenario focused on children aged 5-11 years in July-August of 2021. Vaccination in this age class was was not licensed at the time in Belgium, although we explore possible outcomes if 5- to 17-year-old children had been vaccinated simultaneously. More specifically, we aligned the uptake of the mRNA vaccine for children aged 5-9 and 10-11 years with the reported uptake of children aged 12-15 and 16-17 years, respectively. This approach required scaling the reported uptake to match the number of age bins in each group. For example, for each reported first dose in 12-15-year-old children (i.e. 4 age bins), we included (frac{5}{4}) dose within the 5-9-year-olds (i.e. 5 age bins) at the given point in time. The time between two consecutive doses is assumed to be three weeks, and the resulting vaccine uptake is presented in Fig.3. The final uptake of two doses of vaccine is approximately 80% in the group of 10-19 years and 40% in the group of 0-9 years (which corresponds to 80% in the group of 5-9 years).

Reported and scenario-based uptake of COVID-19 vaccines over time in adults above the age of 20y (top) and 0-19-year-old children (bottom). The uptake is presented in terms of the absolute number of doses (left axis) and as a percentage of the target group (right axis)

To assess the impact of the Omicron VOC on the epidemic trajectory, we performed simulations of our three adjusted vaccine uptake scenarios without the presence of the Omicron VOC, while keeping all other model parameters constant.

In our main analysis, we adopted a conservative approach that assumed no infectiousness-related protection from the COVID-19 vaccine, which potentially underestimates the effect of the intervention. As such, we opt to minimise the risk of overestimating the intervention-related benefits for this exploratory analysis. However, household studies conducted in Denmark[41] and the UK[42] have reported a 31%-45% decrease in the risk of SARS-CoV-2 transmission among vaccinated individuals. Therefore, as a sensitivity analysis taking into account these findings, we performed a comprehensive model calibration assuming a 30% reduction in infectiousness for vaccinated individuals and exploring the effect on the vaccine scenarios.

As robustness analyses, we performed the full model calibration with invariant proportionality factors across different consecutive CoMix waves. A single set of age-specific parameters was not possible as the link between observed contact rates and disease transmission was not constant throughout 20202022 due to differences in contact intensity, duration, and location, among other things. Therefore, we aggregated the CoMix waves into five groups based on the distancing measures that were in place, the (school) holiday periods, and the model fit. For the time period without CoMix data (i.e. for March and SeptemberNovember 2020), we still required time-specific q-factors. Details are provided in Supplementary TableS1.

Part of this epidemiological mathematical modelling study was carried out to inform the Belgian government and the general public about COVID-19 trends and possible interventions. This study is based on data sources from the Belgian Institute of Public Health (Sciensano) in combination with published estimates and data sets (e.g., CoMix). Funding agencies did not have a role in study design, data collection, data analysis, data interpretation, reporting, or in the writing of this manuscript. Data preparation and statistical analyses were performed using R (version 4.2.2; R Foundation for Statistical Computing, Vienna, Austria) on MacOS 12.5 and using R (version 4.0.2) on Rocky Linux 8.8 on the VSC cluster.

Read the original here:

The impact of quality-adjusted life years on evaluating COVID-19 mitigation strategies: lessons from age-specific ... - BMC Public Health