Impact of pediatric influenza vaccination on antibiotic resistance in England and Wales 1 2

Vaccines against viral infections have been proposed to reduce antibiotic prescribing and thereby help control resistant bacterial infections. However, by combining published data sources, we predict that pediatric live attenuated influenza vaccination in England and Wales will not have a major impact upon antibiotic consumption or health burdens of resistance.


Introduction 31
Antibiotic use drives the spread of antibiotic resistance. A substantial proportion of 32 antibiotic prescriptions are unnecessary because they are written to treat conditions that are 33 either self-limiting or non-bacterial in etiology (1). Since influenza is often treated 34 inappropriately with antibiotics, expanding access to influenza vaccines has been proposed as 35 a means of reducing unnecessary prescribing and preventing resistant infections (2). 36 In 2013, England and Wales began rolling out the live attenuated influenza vaccine 37 (LAIV) for 2-16-year-old children (3). Here, we estimate the potential impact on antibiotic 38 prescribing and antibiotic resistance. 39 40

Methods 41
Our age-stratified analysis assumes that some influenza cases lead to general 42 practitioner (GP) consultations and some GP consultations lead to antibiotic prescriptions. 43 We focus on community antibiotic use as the driver of resistance, as hospitalizations for 44 influenza are rare relative to GP consultations (4). 45 Influenza-attributable consultations -For our base-case influenza-attributable 46 consultation rate, we use previous estimates from a time-series statistical attribution analysis 47 6 sinusitis together account for 53% of all inappropriate prescribing, which in turn comprises at 105 least 9-23% of all prescribing in England (1). However, many viral and bacterial pathogens 106 cause these symptoms. By one estimate, influenza causes only 11% of GP consultations for 107 acute respiratory illness in England (4), so it may be optimistic to expect influenza 108 vaccination to substantially reduce antibiotic use in the UK. 109 Our base-case estimate of 726 antibiotic prescriptions per 1000 influenza-attributable 110 consultations is more than double what electronic health records suggest (Methods). One 111 explanation is that this estimate, derived from statistical attribution of antibiotic prescriptions 112 to influenza circulation over 1995-2009 (5), feasibly includes prescribing for infections 113 secondary to influenza infection, such as otitis media, sinusitis and pneumonia. Also, 114 antibiotic use in England has declined since this time-by 22% from 1998 to 2016 (12). 115 Accordingly, our base-case results should be interpreted as the maximum potential reduction 116 by LAIV of antibiotic use. Conversely, in the FluWatch study only 8% of consultations for 117 ILI resulted in influenza or ILI being medically recorded (6), and so electronic health records 118 may not reliably reflect prescribing rates for influenza. 119 In randomized trials, the direct effect of influenza vaccines on vaccinated children has 120 ranged from a 44% reduction in antibiotic prescriptions (Italy) to a 6% increase (United 121 States), both over the 4-month period following vaccination, while estimates of the impact 122 over entire populations (all ages, vaccinated and unvaccinated) range from 11.3 fewer 123 prescriptions per 1000 person-years in Ontario to 3.9 fewer in South Africa and Senegal 124 (Appendix). This variation can be ascribed to differences in: vaccine efficacy and coverage, 125 risk factors among the study population, influenza circulation, existing patterns of antibiotic 126 use, and methodology; estimates of vaccine impact on antibiotic consumption may not be 127 generalizable across settings. 128 7 Our framework estimates the impact of influenza vaccination on resistance using the 129 relationship between influenza circulation and antibiotic use in England and Wales, and can 130 be adapted to other settings for which the strength of this relationship can be quantified. An 131 alternative approach would be to correlate LAIV uptake, rather than influenza circulation, 132 directly with antibiotic use. Challenges with this approach include appropriately controlling 133 for confounding factors in the relationship between vaccine uptake and antibiotic use, and 134 quantifying the effect of herd immunity. However, consistent with our approach, UK-specific 135 empirical estimates have suggested little or no effect of LAIV uptake on prescribing. A self-136 controlled case series study found that 2-4-year-old LAIV recipients in the UK took 13.5% 137 fewer amoxicillin courses in the 6 months following vaccination (13), while an LAIV pilot 138 study detected no difference in prescribing rates for respiratory tract infections between 139 treatment groups (14). No single vaccine is likely to substantially reduce inappropriate 140 antibiotic use in the UK.

Impact of pediatric influenza vaccination on antibiotic resistance in England and Wales
Chungman Chae, Nicholas G. Davies, Mark Jit, and Katherine E. Atkins variation between flu seasons is reported), so we assume that uncertainty in the influenza-attributable GP consultation rates follows a normal distribution. We assume that the standard deviation of any consultation rate derived from this source is always S times the mean rate, where S is estimated from Fig. 2  triangular distribution with B as the peak (mode) and A -C as the 95% highest density interval (using a triangular distribution rather than a normal distribution allows us to account for skew). Then, the probability of a GP visit given PCR-confirmed illness is taken from Table S6 of the same source. We correct for low numbers by assuming a "base proportion" of 12/82 as the measured proportion of PCR confirmable influenza episodes resulting in a GP visit, which comes from the overall number of reported GP visits for 5-64-year-olds with PCR-confirmed influenza. To account for uncertainty in measurement, we draw the "base rate" of GP consultation

Influenza-attributable GP consultations
given PCR-confirmable influenza for 5-64-year-olds from a beta distribution with parameters α = 12 + 1, β = 82 -12 + 1 (i.e. assuming a uniform prior); to account for the observation that this rate is higher in young children and the elderly (Table S6), we add 0.12 to this rate for under-5s and over-65s. The annual influenza-attributable rate of GP consultation for a given age group is then the product of the PCRconfirmable influenza incidence and the rate of GP consultation given PCRconfirmable influenza. and the rest are for acute respiratory infection without fever (2) (hence having a 48% prescription rate), which yields an overall (crude) prescribing rate of 31.3%.
To calculate age-stratified values, we assume prescribing for under-5s is about 20% less, and for over-45s is about 20% more, than prescribing in 5 That is, we draw a value d from a normal distribution with mean 0.2 and standard deviation 0.05, and assume that the relative prescribing rate for under-5s is (1d) times the rate for 5-44-year-olds, while the relative prescribing rate for over-45s is (1 + d) times the rate for 5-44-year-olds.

Impact of LAIV on rates of GP consultation
We use fitted models from Baguelin et al. (7) projecting the impact of LAIV on influenza cases in different age groups, assuming either a 50% uptake (base-case estimate), 30% uptake (low estimate), or 70% uptake (high estimate).

Age-stratified rates for uncertainty analysis
We summarize the base-case and uncertainty-analysis estimates of age-stratified consultation rates, prescription rates, and LAIV impact in the Appendix Table below.
Appendix  To estimate the total number of bloodstream infections caused by a given species in a given country, we begin by taking the maximum of the number of total tested isolates recorded by the ECDC for that country and species. Then we correct that figure according to the estimated population coverage for that country and species to the ECDC (i.e. an estimate of what fraction of the population is covered by the hospitals submitting resistance testing data to national surveillance programs which then report their data to the ECDC). For example, for S. pneumoniae infections in the United Kingdom in 2015, 1095 isolates were tested for penicillin non-susceptibility, 1077 isolates were tested for macrolide non-susceptibility, and 1060 isolates were tested for combined non-susceptibility to both penicillins and macrolides.
Additionally, these isolates were reported as covering an estimated 21% of the entire population of the UK. Accordingly, we estimated the total number of bloodstream

5.3
Alternative scenario 1 -Rather than using the overall antibiotic consumption for each country as the sole predictor in the regression model, we also built a separate series of models where we used as predictors each country's consumption of tetracyclines (J01AA), extended spectrum penicillins (J01CA), beta-lactamase sensitive penicillins (J01CE), and macrolides (J01FA), as these four classes comprise the majority of antibiotics prescribed for sore throat and cough (11). We assume that for a given reduction in the overall prescription rate x, there is a reduction 0.0620x in tetracycline prescribing, 0.4752x in extended spectrum penicillin prescribing, 0.2793x in beta-lactamase sensitive penicillin prescribing, and 0.1835x in macrolide prescribing. This predicted a smaller impact upon resistance than the main scenario (4.1) and comprises the "low-effect" statistical model for the uncertainty analysis ( Figure 2B, main text).

5.4
Alternative scenario 2 -We follow the same procedure as in 4.2, but if any predictor variable is negatively correlated with a resistance related health burden (i.e. the best fitting linear model suggests that decreasing use of that antibiotic would increase resistance), we remove it from the linear regression and rerun the model, continuing this process until all predictors are positively associated with the outcome variable. If more than one variable has a negative association in a given round, all are removed for the next round. This predicted a larger impact upon resistance than the main scenario (4.1) and comprises the "high-effect" statistical model for the uncertainty analysis ( Figure 2B, main text).

Economic calculations
To convert between US and UK health care expenditures, we use hospital-service price level indices for health care purchasing power parity published by the OECD (12) (see their Fig. 1).

Secular trends in antibiotic prescribing rates
Data from NHS Digital show that community antibiotic use in England has decreased by approximately 2.5% per year from 2012 to 2018 (Appendix Figure).