2012/13 influenza vaccine effectiveness against hospitalised influenza A(H1N1)pdm09, A(H3N2) and B: estimates from a European network of hospitals

While influenza vaccines aim to decrease the incidence of severe influenza among high-risk groups, evidence of influenza vaccine effectiveness (IVE) among the influenza vaccine target population is sparse. We conducted a multicentre test-negative case-control study to estimate IVE against hospitalised laboratory-confirmed influenza in the target population in 18 hospitals in France, Italy, Lithuania and the Navarre and Valencia regions in Spain. All hospitalised patients aged ≥18 years, belonging to the target population presenting with influenza-like illness symptom onset within seven days were swabbed. Patients positive by reverse transcription polymerase chain reaction for influenza virus were cases and those negative were controls. Using logistic regression, we calculated IVE for each influenza virus subtype and adjusted it for month of symptom onset, study site, age and chronic conditions. Of the 1,972 patients included, 116 were positive for influenza A(H1N1)pdm09, 58 for A(H3N2) and 232 for influenza B. Adjusted IVE was 21.3% (95% confidence interval (CI): -25.2 to 50.6; n=1,628), 61.8% (95% CI: 26.8 to 80.0; n=557) and 43.1% (95% CI: 21.2 to 58.9; n=1,526) against influenza A(H1N1) pdm09, A(H3N2) and B respectively. Our results suggest that the 2012/13 IVE was moderate against influenza A(H3N2) and B and low against influenza A(H1N1) pdm09.

While influenza vaccines aim to decrease the incidence of severe influenza among high-risk groups, evidence of influenza vaccine effectiveness (IVE) among the influenza vaccine target population is sparse. We conducted a multicentre test-negative case-control study to estimate IVE against hospitalised laboratoryconfirmed influenza in the target population in 18 hospitals in France, Italy, Lithuania and the Navarre and Valencia regions in Spain. All hospitalised patients aged ≥18 years, belonging to the target population presenting with influenza-like illness symptom onset within seven days were swabbed. Patients positive by reverse transcription polymerase chain reaction for influenza virus were cases and those negative were controls. Using logistic regression, we calculated IVE for each influenza virus subtype and adjusted it for month of symptom onset, study site, age and chronic conditions. Of the 1,972 patients included, 116 were positive for influenza A(H1N1)pdm09, 58 for A(H3N2) and 232 for influenza B. Adjusted IVE was 21.3% (95% confidence interval (CI): -25.2 to 50.6; n=1,628), 61.8% (95% CI: 26.8 to 80.0; n=557) and 43.1% (95% CI: 21.2 to 58.9; n=1,526) against influenza A(H1N1) pdm09, A(H3N2) and B respectively. Our results suggest that the 2012/13 IVE was moderate against influenza A(H3N2) and B and low against influenza A(H1N1) pdm09.

Background
Antigenic drifts of influenza viruses expose the population to new but related influenza variants on a regular basis [1]. On the basis of a yearly revised composition of seasonal influenza vaccines, the World Health Organization (WHO) considers annual Influenza vaccination as the most efficient measure against influenza [2]. Every year, the seasonal influenza vaccine licensure is obtained based on immunogenicity data [3]. While these immunogenicity data are thought to be valid for healthy adults [4], the development of correlates of protection suited to vulnerable populations is still to be achieved [5].
The population targeted for influenza vaccination in Europe includes those at increased risk of exposure to influenza virus as well as of developing severe disease, especially disease resulting in hospitalisation or death [6]. Target groups for vaccination usually include adults over 59 or 64 years of age and people of any age with certain underlying medical conditions [7,8]. Measuring influenza vaccine effectiveness (IVE) in each influenza season is important for the following reasons: to identify vaccines types and brands with low IVE; to decide on alternative preventive strategies if early estimates of IVE are low (e.g. preventive use of antivirals among vulnerable individuals); and to help decide on the next season's vaccine content. Repeated evidence of suboptimal IVE among the population targeted for annual influenza vaccination would also further advocate the need for vaccines that are more effective in this population. Moreover, there are ongoing scientific debates about the effect of repeated vaccination on the immunological response induced by the seasonal influenza vaccine [9][10][11] and further evidence is needed.
In 2011, we launched a pilot study to estimate the IVE against laboratory-confirmed influenza hospitalisation using a network of hospitals in the European Union (EU) [12]. During the 2012/13 influenza season, co-circulation of influenza A(H1N1)pdm09, A(H3N2) and B/Victoria-and B/Yamagata-lineage viruses was reported in Europe [13]. The objective of the study presented here was to measure the 2012/13 seasonal IVE against hospitalisation with subtype-specific laboratory-confirmed influenza in a hospital network in four EU countries: France, Italy, Lithuania and Spain.

Methods
We conducted a case-control study using the test-negative design [14] in 18 hospitals located in five study sites: France (five hospitals), Italy (two), Lithuania (two), and the Navarre (four) and Valencia (five) regions in Spain. Each study site adapted a generic protocol [15] to the local context (Table 1).

Study population
The study population was all community-dwelling adults (18 years of age or older), belonging to the target groups for vaccination as defined locally [16][17][18][19][20], admitted to one of the participating hospitals with no contraindication for influenza vaccination. Patients were excluded if they had previously tested positive for influenza virus in the 2012/13 season or resided outside the hospital catchment area (for the 11 hospitals with known catchment area).
Study teams actively screened all patients admitted for potentially influenza-related conditions. These conditions included the following: acute myocardial infarction or acute coronary syndrome; heart failure; pneumonia and influenza; chronic pulmonary obstructive disease; myalgia; altered consciousness, convulsions, febrile-convulsions; respiratory abnormality; shortness of breath; respiratory or chest symptoms; acute cerebrovascular disease; sepsis; and systemic inflammatory response syndrome. Among them, study teams invited patients with an onset of influenza-like illness (ILI) symptoms (one systemic and one respiratory symptom) within the past seven days to participate. Those accepting to participate were swabbed and tested for influenza. Reverse transcription polymerase chain reaction (RT-PCR) was used to detect influenza viruses and to classify them as influenza A(H3N2), influenza A(H1N1)pdm2009 or influenza B. Patients positive for influenza were classified as cases of a given influenza type/subtype and those testing negative were controls.
We defined the study period as at least 15 days after the beginning of each site-specific seasonal influenza vaccination campaign until the end of the influenza season as declared by local influenza surveillance systems. For each of the influenza type/subtype analyses, we excluded the controls with onset of symptoms before the week of the first laboratory-confirmed case or after the week of the last laboratory-confirmed case. We used the International Organization for Standardization's week numbers [21] to ensure consistency across study sites.
We considered patients as vaccinated against seasonal influenza if they had received at least one dose of the 2012/13 influenza vaccine more than 14 days before onset of ILI symptoms. Patients not vaccinated or vaccinated less than 15 days before ILI onset were considered as unvaccinated.

Data collection
We collected data on the ILI episode, demographics, chronic diseases ( Table 2), number of hospitalisations in the previous 12 months, number of consultations at the general practitioner (GP) in the previous three months, smoking status, vaccination against influenza in 2012/13 and 2011/12 and, for those aged 65 years and over, functional status before ILI onset using the Barthel score [22]. The data were gathered from hospital medical records, face-to-face interviews with the patient and/or patient's family and laboratory databases. The vaccination status was obtained from vaccination registers in two study sites, interview with the patients and/or patient's family in two sites and contact with the patient's physician in one site.

Data analysis
Study sites transmitted anonymised datasets to the pooled analysis coordinator, through a passwordsecured web-based platform. We ran a complete case analysis, excluding records for which laboratory results, vaccination status or potential confounding variables were missing.
To test for heterogeneity between study sites, we used Cochran's Q-test and the I 2 index [23]. The Q-test provides a p value that indicates the presence or not of heterogeneity. The I 2 index quantifies the proportion of the variance attributable to differences between study sites. It is common to consider that I 2 around 25%, 50% and 75% indicate low, medium and high heterogeneity, respectively.
We conducted separate analyses for each type/subtype of influenza. We estimated the pooled IVE as 1 minus the odds ratio (OR) (expressed as a percentage) of being vaccinated in cases versus controls, using a one-stage method with study site as fixed effect in the model [24].
We assessed the presence of effect modification by comparing the time-and study site-adjusted OR (assuming that the test-negative design case-control study is a density case-control study implying adjustment for the time of symptom onset) across strata of characteristics using the homogeneity test. We considered a variable as a confounder when the percentage change between the unadjusted and adjusted OR was greater than 15%.
We conducted a multivariable logistic regression analysis. In addition to study site and month of symptom onset, we adjusted the models for the covariates identified as potential confounders in the stratified analysis as well as the presence of at least one underlying condition and the age that we modelled as a restricted cubic spline with four knots [25]. The likelihood ratio test was used to decide on the final models. We conducted stratified analyses by age group (less than 65 years, 65-79 years and 80 years and above).
To study the effect of previous influenza vaccination on laboratory-confirmed influenza, we conducted a stratified analysis using four vaccination status categories: vaccination in none of the seasons (2011/12 and 2012/13), 2012/13 vaccination only, 2011/12 vaccination only and vaccination in both seasons and computed and compared IVE for each of these categories using vaccination in none of the seasons as a reference.
We carried out sensitivity analyses excluding the weeks when less than 10% of the patients included were positive for influenza, excluding patients who received antivirals between the onset of symptoms and swabbing and by restricting the analysis to patients swabbed within four days of symptoms onset. To avoid the inclusion of patients with acute manifestation of chronic respiratory illnesses rather than respiratory infection, we restricted our analysis to patients with no underlying respiratory conditions.
We ran all analyses with Stata v12 (Stata Corp LP, College Station, TX, United States).
The p values associated with the Q-test and the I 2 index using models adjusted for age, month of symptom onset and chronic condition, testing for heterogeneity between study sites, were respectively 0. 19 (Table 6).

Discussion
Our results suggest that in the population targeted for the influenza vaccination, the 2012/13 IVE for laboratory-confirmed hospitalised influenza was 21.3% against A(H1N1)pdm09, 61.8% against A(H3N2) and 43.1% against B.
The adaptation of a generic protocol by 18 European hospitals enabled us to pool data and obtain a sample of 1,972 hospitalised ILI patients targeted for influenza vaccination. In a season with co-circulation of the three viruses, this large sample size allowed us to compute type-/subtype-specific estimates of IVE against hospitalised influenza and to further attempt to stratify by age group. However, stratified analyses led to estimates with broad confidence intervals. Consequently, some results of the stratified analyses can only be used to generate hypotheses.
The test-negative design has been mainly discussed and validated for GP-based studies [26,27]. It is assumed that by restricting the study population to patients consulting for ILI, the health-seeking behaviour confounding effect (associated with propensity to get vaccinated and to go to the GP in case of influenza) is controlled for. Since in our study sites all people needing hospitalisation are likely to be hospitalised, we believe that confounding due to health-seeking behaviour is minimised.
In hospital-based studies, several outcomes could be used. If we were to measure IVE against influenza confirmed-severe acute respiratory infection (SARI), we would need to make sure that for both cases and controls a respiratory infection was the cause of admission. We have chosen a broader case definition and a more sensitive inclusion criteria to cover a larger part of the influenza disease burden. As a consequence, some of the ILI in the seven days before admission may correspond to an exacerbation of underlying respiratory conditions. This could lead to an overestimation of the IVE. Restricting our analysis to patients with no underlying respiratory conditions provides similar results and does not support this hypothesis. Furthermore, we adjusted for the presence and number of previous hospitalisations for underlying conditions.
The inclusion of patients swabbed more than four days after symptoms onset or after antiviral treatment had started could have led to misclassification biases if viral clearance occurred before swabbing. However, analyses confined to patients swabbed within four days of symptom onset and to patients who did not receive antiviral treatment did not change the results.
Studies using the test-negative design may underestimate the IVE when the ratio of controls to cases is high, especially if the laboratory tests have low specificity [28]. In our study, all cases were confirmed by RT-PCR, which has high specificity [29]. In the analyses restricted to weeks when the control to case ratio was lower than 9:1 resulted in very similar IVE estimates.
The data quality was high with only 49/2,021 records with missing outcomes or exposures in the database. We believe that ascertainment of vaccination status through patient interviews in two of the five study sites has not introduced differential information bias as data were collected before laboratory testing.
Due to the small sample size in some study sites, the test of heterogeneity may have had no power to detect heterogeneity even if differences exist between study sites. Different IVE across study sites could be due to variations in circulating strains, different vaccines by study site or different measured and unmeasured confounding factors. Further typing of circulating strains would be valuable to discuss site-specific IVE with regard to the level of matching between vaccine and locally circulating strains. Different access to vaccination according to age and underlying condition and to hospitalisation [30] could partly explain variations in IVE across study sites. Finally, the presence of random errors cannot be ruled out due to low sample size by study site. A larger sample size would be needed to carry out a two-stage pooled analysis [24].
Our results suggest that, in people belonging to target groups for vaccination, the 2012/13 IVE varied by subtype and age group. However, we cannot exclude the possibility that the variability of IVE results by age group mainly reflects sample size limitations. Small stratum-specific sample sizes (and very small number of cases) lead to unstable results and do not allow for biological interpretation of age-specific results.
Our results would suggest that IVE against A(H3N2) was higher among older age groups. This observation would be in contradiction to the principles of immune senescence. In addition to the sample-size limitations, and as discussed above, we cannot exclude a selection bias for our controls, which we adjusted for. However we used the same control group for the three subtypes and age-specific results vary by subtype. We consider that it is unlikely that confounding factors would differ by subtype.
When looking at the effect of repeated vaccination (over two consecutive seasons), similar patterns were observed for influenza A (H3N2) [32] suggested that a natural infection in season 1 produces antibodies that have a larger potential to form high post-vaccination titres in season 2 than vaccineinduced antibodies. Smith et al. [33] hypothesised that large antigenic distances between vaccines in seasons 1 and 2, and between vaccine in season 1 and epidemic strain in season 2, significantly increase the risk of infection among repeated vaccinees compared with those receiving the vaccine in season 2 only. Considering the antigenic differences between the 2011/12 vaccine and the 2012/13 circulating strains, this hypothesis could explain our results, suggesting a higher IVE against influenza A(H3N2) and B among patients vaccinated in 2012/13 only compared with those vaccinated in 2011/12 and 2012/13. Further studies, including a longer history of vaccine uptake and natural infections would be of great value to better understand the effect of repeated vaccination on the immunological response to a new influenza seasonal vaccine and the level of clinical protection conferred to individuals.
Our results suggest a low IVE against A(H1N1)pdm09, especially among the elderly [34]. A total of 14 cases of influenza A(H1N1)pdm09 occurred among patients older than 80 years. While the majority of these cases (n=12) were vaccinated patients, small numbers make the IVE estimates hard to interpret in that age group. The IVE was similar for those vaccinated in 2011/12 only or in both seasons. There was no effect for those vaccinated in 2012/13 only. The recommended A/California/7/2009(H1N1)pdm09-like virus strain was the same for the 2011/12 and 2012/13 vaccines and matched the 2012/13 circulating strains (some A/California/06/2009 also reported). Long-lasting immune response induced by trivalent inactivated vaccines was previously described [35] and some recent results suggest that frequent previous vaccinations may be effective for the current influenza season [11]. The absence of protection among patients vaccinated in 2012/13 only is difficult to understand and interpret; it may reflect the presence of associated (and unmeasured) negative confounders for which repeated vaccination may be a surrogate. In addition, other studies [36][37][38] suggest a decreasing effect in the season difficult to reconcile with a long-term effect between seasons. Considering the small sample size in some of the vaccination groups in our study, we cannot exclude the possibility that this observation is due to chance.
Increasing the number of study sites in this network would allow a sufficient sample size to be reached early enough in the season to prompt the use of alternative prevention measures if a low IVE against hospitalised cases is observed among the target group. Early estimates of IVE against hospitalised influenza are also a useful complement to guide the decisionmaking of WHO experts regarding the composition of the next season's vaccines. A larger sample size and good documentation of vaccine brands used would allow the computing of brand-specific IVE. To further study the effect of previous seasonal vaccination will require documenting past vaccination over several seasons. In addition, ways to measure past natural immunity may also be needed to better understand the complex immunity of influenza natural infection and vaccination.