2012-2013 Seasonal Influenza Vaccine Effectiveness against Influenza Hospitalizations: Results from the Global Influenza Hospital Surveillance Network

Background The effectiveness of currently licensed vaccines against influenza has not been clearly established, especially among individuals at increased risk for complications from influenza. We used a test-negative approach to estimate influenza vaccine effectiveness (IVE) against hospitalization with laboratory-confirmed influenza based on data collected from the Global Influenza Hospital Surveillance Network (GIHSN). Methods and Findings This was a multi-center, prospective, active surveillance, hospital-based epidemiological study during the 2012–2013 influenza season. Data were collected from hospitals participating in the GIHSN, including five in Spain, five in France, and four in the Russian Federation. Influenza was confirmed by reverse transcription-polymerase chain reaction. IVE against hospitalization for laboratory-confirmed influenza was estimated for adult patients targeted for vaccination and who were swabbed within 7 days of symptom onset. The overall adjusted IVE was 33% (95% confidence interval [CI], 11% to 49%). Point estimates of IVE were 23% (95% CI, −26% to 53%) for influenza A(H1N1)pdm09, 30% (95% CI, −37% to 64%) for influenza A(H3N2), and 43% (95% CI, 17% to 60%) for influenza B/Yamagata. IVE estimates were similar in subjects <65 and ≥65 years of age (35% [95% CI, −15% to 63%] vs.31% [95% CI, 4% to 51%]). Heterogeneity in site-specific IVE estimates was high (I2 = 63.4%) for A(H1N1)pdm09 in patients ≥65 years of age. IVE estimates for influenza B/Yamagata were homogenous (I2 = 0.0%). Conclusions These results, which were based on data collected from the GIHSN during the 2012–2013 influenza season, showed that influenza vaccines provided low to moderate protection against hospital admission with laboratory-confirmed influenza in adults targeted for influenza vaccination. In this population, IVE estimates against A(H1N1)pdm09 were sensitive to age group and study site. Influenza vaccination was moderately effective in preventing admissions with influenza B/Yamagata for all sites and age groups.


Introduction
Influenza vaccination is universally recommended for individuals at increased risk for complications, but the effectiveness of current licensed vaccines has not been clearly established [1,2]. Observational field studies have shown substantial variability in influenza vaccine effectiveness (IVE) by season, strain, and age group [3][4][5][6]. In addition, many of these studies are underpowered for subgroup analyses, complicating estimates of IVE for individual risks. Furthermore, differences in study design and outcome measures limit the ability to compare results across studies, and the external validity of the results is weakened when the study population does not fully represent the different vaccination settings worldwide.
Several networks have been created to provide more representative and robust estimates of IVE [7][8][9][10][11]. These networks use a more standardized approach for data collection, analysis, and reporting of IVE, but most employ passive surveillance and therefore are highly dependent on reporting timeliness and completeness [12][13][14]. Also, few networks include surveillance of severe cases requiring hospitalization.
The Global Influenza Hospital Surveillance Network (GIHSN) was launched in 2012 to address growing awareness that influenza-related hospitalization is a significant burden that remains insufficiently characterized. The GIHSN is a partnership between industry and public health institutions that uses active surveillance and a common core protocol to collect data on the epidemiology of severe influenza, as defined by hospitalization with laboratory-confirmed influenza. The principal aim of the GIHSN is to estimate, when feasible, IVE against hospitalization with influenza. Data collection in the GIHSN is coordinated by regional centers. In the GIHSN's first season (2012-2013), five coordinating centers covering 14 hospitals participated, including the Centro Superior de Investigación en Salud Pública (now FISABIO) (Valencia, Spain); the Reseau National d'Investigation Clinique en Vaccinologie (France), the Research Institute of Influenza (St. Petersburg, Russian Federation), the D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation, and, as a pilot partner, the National Influenza Reference Laboratory (Cappa-Istanbul, Turkey).
Here, we used a test-negative approach [15,16] to estimate IVE against hospitalization with laboratory-confirmed influenza. Validity of the pooled dataset was assessed by quantifying the heterogeneity in the effect estimates across the different study sites.

Study Design
This was a multi-center, prospective, active surveillance, hospital-based epidemiological study carried out during the 2012-2013 Northern Hemisphere influenza season. Data  All patients provided written informed consent. Briefly, data on hospitalized patients with a diagnosis possibly associated with influenza were collected by an active surveillance system composed of healthcare professionals trained to follow a generic study protocol, and influenza was confirmed by reverse transcription-polymerase chain reaction (RT-PCR). At each site, case identification was adapted to the specific local settings of the health care delivery system and type of hospital, although all sites used the same case criteria for definitive inclusion and, in all cases, the study was conducted over a period defined by the weeks with positive specimens for influenza (Table S1)

Study Population
Non-institutionalized adults that were residents of Valencia, Spain or who held a national social security affiliation (France) and were hospitalized for at least 24 h in one of the participating hospitals were considered for inclusion in the GIHSN database. Also, patients admitted at the emergency department (Valencia, France, Russian Federation) and at certain hospital wards (France, Russian Federation) were considered if they had pre-defined chief complaints presumably associated with a previous influenza infection [6]. After informed consent was obtained, patients were screened for the following inclusion criteria: onset of influenza-likeillness (ILI) within 7 days of admission to the hospital; influenza vaccination not contraindicated; not previously positive for influenza virus in the 2012-2013 season; and not hospitalized within 30 days of the current admission. ILI was defined as the presence of at least one systemic symptom (fever or feverishness, malaise, headache or myalgia) and at least one respiratory symptom (cough, sore throat or shortness of breath).

Study Conduct
At enrollment, a nasopharyngeal and a pharyngeal swab were collected and patients were interviewed by a hospital physician, clinical research associate, or both (Russian Federation and France) or a dedicated study nurse (Valencia). Swabs were stored at 220uC. The following data were collected during the interview or by searching clinical records: demographic characteristics; anthropometric measures; information on the ILI episode; dates of symptom onset, hospitalization, and swabbing; antiviral treatment received; intense care unit admission; death during hospitalization; main hospital admission and discharge diagnostics; presence of chronic diseases; pregnancy status; number of hospital admissions in the past 12 months; number of general practitioner consultations in the previous 3 months; smoking habits; and vaccination against influenza in the current (2012-2013) and previous (2011-2012) seasons. Physicians involved in clinical care of patients were also involved in patient recruitment but were not involved in case ascertainment.
Social class was assigned according to occupation as described previously [17]. Functional status before ILI onset was ascertained in patients $65 years of age using the Barthel index [18] and categorized as follows: total dependence, 0-15; severe dependence, 20-35; mild to moderate dependence, 40-90; no dependence, $ 95. Vaccination status during the current season was ascertained from registries, vaccination cards, and interviews with patients, their families, and their physicians. Patients were considered vaccinated if they had received at least one dose of the 2012-2013 seasonal vaccine .14 days before the onset of ILI symptoms. Local vaccination policies and vaccines available at each coordinating site are summarized in Table S2.

Laboratory Confirmation of Influenza
Commercially available (Russian Federation) or in-house (Valencia and France) RT-PCR assays were used to detect Table 1. influenza A (subtypes H3 and H1) and influenza B (Yamagata and Victoria lineages) viruses in swabs (Text S1).

Data Management, Calculations, and Statistical Analysis
Coordinating sites collected anonymized data and checked for missing, inconsistent, or incorrect data. Whenever possible, queries of any inconsistencies or missing data were resolved by the investigators at each of the study sites. Missing data were not replaced for the statistical analyses. Data from each coordinating site were shared with the network coordinating center (FISABIO, Valencia, Spain) through a secured web-based system.
Differences in the distribution of variables were estimated using a chi-square or T-test. A P-value of less than 0.05 was considered to indicate statistical significance.
The primary outcome measure was hospital admission with laboratory-confirmed influenza. Secondary outcome measures were hospital admissions with laboratory-confirmed influenza A(H1N1)pdm09, A(H3N2), or B/Yamagata. IVE was determined in patients $18 years of age who had been swabbed within 7 days of the onset of ILI symptoms and who had been targeted for influenza vaccination because they were obese, pregnant, or $65 years of age, or had recorded comorbidities [19]. In addition, patients were excluded from IVE estimates and analysis if they had received a homeopathic vaccine. IVE was estimated as (12odds ratio [OR]) 6100, where the OR compared the vaccine coverage rate between influenza-positive and influenza-negative patients. Records for which outcome, exposure, or confounding variables were missing were excluded from the multivariate IVE analyses. The adjusted IVE was estimated by logistic regression using a random effects model with study site as a shared parameter for the pooled analysis and including week of symptom onset as a continuous variable, and age group, sex, hospitalization in the previous 12 months, presence of chronic conditions, and smoking habits as potential confounding factors. Parameters not normally distributed were transformed prior to analysis. Polynomial fitting was used for non-linear relationships between week of symptom onset and influenza positivity. The nonlinear relationship between the week of symptom onset (independent variable) and influenza positivity (dependent variable) was modeled as an n th order polynomial, yielding the general polynomial regression model y = b 0 +b 1 x +b 262 +b 363 +…b n x n + Syz i + m i , where the expected value of a dependent variable y (log of the odds of either influenza positivity overall, H1N1, H2N3, B/ Yamagata or B/Victoria) was modeled in terms of the value of the independent variable x (week of onset), b n are the coefficients, Syz i are the effects of the covariates, and m i are the random effects representing between-site variability [20]. Sensitivity analysis was performed by including only samples taken within 4 days of symptom onset. A P-value ,0.05 was considered to indicate statistical significance. Heterogeneity in IVE estimates was assessed using the I 2 statistic [21][22][23]. Potential sources of heterogeneity, including coordinating site, age, and influenza subgroup were examined in ad-hoc analyses. Heterogeneity was defined as low if I 2 statistic ,25%, moderate if 25% to 49%, high if $50% as described previously [22]. Statistical analyses were performed using Stata version 13.1 (College Station, TX).

Patients
A total of 9150 patients were screened by the 14 participating hospitals (Table 1)  broad selection criteria, which were designed to capture the maximum number of patients hospitalized for reasons that have been or could be associated with influenza infection.
Strains isolated at each site. B/Yamagata was the predominant strain isolated from patients in Valencia (63.5% of isolates), while A(H1N1)pdm09 predominated in Moscow (58.8% of isolates). In St. Petersburg, A(H1N1)pdm09 and B/Yamagata predominated and were present at similar frequencies. In France, A(H3N2) and B/Yamagata predominated (Table 2 and Figure 1).
At each site, the distribution of strains in the patients changed as the season progressed ( Figure 1). For example, in Valencia, B/ Yamagata predominated early in the season, with a peak at epidemiological week 2013-7, whereas A(H1N1)pdm09 predominated later in the season, with a peak at epidemiological week 2013-13. In contrast, in St. Petersburg and Moscow, A(H1N1)pdm09 predominated early in the season, while B/ Yamagata predominated later. The pattern in France was different than either of these countries, with several strains coexisting throughout the influenza season.

Patient Characteristics by Influenza Infection Status
Influenza-positive patients were younger than influenza-negative patients admitted to hospital (mean age, 51 vs. 63), less likely to be men, less likely to suffer from comorbidities, and less likely to have been hospitalized in the last year but more likely to have never smoked and more likely to have professional or non-manual skilled jobs (Table 3)  The mean interval between symptom onset and specimen collection was similar for influenza-positive and influenza-negative patients (mean 6 standard deviation = 3.161.6vs. 3.561.7 days), although more influenza-positive than influenza-negative patients were swabbed within 2 days. The risk of being influenza positive decreased by 3% (95% CI, 2% to 4%) (P for trend ,0.0001) for each day elapsed between symptom onset and swabbing.

Patient Characteristics by Vaccination Status
Patients vaccinated during the year of the study (2012-2013) were older than unvaccinated patients (mean, 76 vs. 50 y) ( Table 4). Vaccinated patients were also more likely to be men, suffer from chronic conditions, to have been hospitalized in the last year, to have visited the general practitioner in the last 3 months, to be past smokers, and to have been influenza-vaccinated the previous year (2011-2012).

Patient Characteristics by Study Site
Patients in St. Petersburg (76.9%) and Moscow (94.5%) were mostly ,65 years of age and had either no or one chronic disease ( Table 5), regardless of influenza infection status (Table S3). In Moscow, 72.1% (483/670) of the patients were pregnant women (mean age, 2865 years). The patients in France and Spain were evenly spread across age groups, and at least 70% suffered from one or more chronic condition. The pattern of chronic conditions was similar in Valencia and France (cardiovascular disease, chronic obstructive pulmonary disease, and diabetes), whereas in Moscow and St. Petersburg, the main chronic illness reported was cardiovascular disease. The median (interquartile range) number of chronic illnesses in patients with comorbidities was 1 (1-2) in Valencia, 2 (1-2) in France, 1 (1-1) in St. Petersburg, and 0 (0-1) Moscow. Influenza vaccine uptake was low in Moscow (3.3%) and St. Petersburg (0.8%) but moderate in Valencia (55.4%) and France (53.4%).
All heterogeneity results were similar when assessed using adjusted IVE estimates ( Figures S1, S2, and S3).

Discussion
This study, performed in three different countries during the 2012-2013 influenza season, used a test-negative design to estimate IVE against hospitalization with laboratory-confirmed influenza in adults targeted for vaccination. All patients included in the IVE estimates and analysis had to have been tested for influenza within 7 days of the onset of ILI symptoms. The pooled adjusted IVE was 33% (95% CI, 11% to 49%) against hospitalization. Estimates of IVE for preventing hospital admissions were consistent and moderate across sites and age groups for influenza B/Yamagata (43% [95% CI, 17% to 60%]) but low and non-significant for influenza A(H1N1)pdm09 (30% [95% CI, 2 37% to 64%]) and A(H3N2) (23% [95% CI, 226% to 53%]). IVE estimates for A(H1N1)pdm09 were highly heterogeneous across study sites in patients $65 years of age but not in younger patients. Influenza A(H1N1)pdm09 and B/Yamagata followed by A(H3N2) were the most common strains isolated. These results agree with other interim and preliminary results published for the 2012-2013 influenza season [3,5,[24][25][26][27][28][29].
The low IVE estimates in this study might have been due to genetic drift in influenza at key antigenic sites [5]. Genetic and possible antigenic mismatches have been described in Europe for A(H1N1)pdm09 and A(H3N2) [23][24][25]. Vaccines for the 2012-2013 season containing A/Victoria/361/2011 antigens have been reported to induce antibodies in humans that bind less effectively to most cell-propagated influenza A(H3N2), apparently due to antigenic changes in earlier A/Victoria/361/2011-like vaccine viruses associated with adaptation of the virus to propagation in eggs [30]. Accordingly, vaccines for the 2013-2014 northern hemisphere season are recommended to contain A(H3N2) virus that is antigenically like the cell-propagated prototype virus A/ Victoria/361/2011 [30]. In contrast, in two preliminary analyses of North American data, IVE was moderate and significant against A(H3N2), although this was associated with a good antigenic match between circulating and vaccine A(H3N2) strains [7,31].
The IVE estimates in this study were similar to those reported in sentinel hospital-based studies [32,33] but were lower than reported for general practitioner-attended influenza outcomes [26,29]. This might be because of different effectiveness for different clinical outcomes or because of the generally older age and poorer health of patients requiring hospital admission for influenza infections. Indeed, our study patients were, on average, older and in poorer health than those in the general practitioner sentinel studies. Also, in contrast to some of these general practitioner sentinel studies, our estimates of IVE against influenza A strains were similar across age groups. However, in all reports, including ours, IVE against B/Yamagata influenza was moderate, despite differences in baseline patient characteristics.
The validity of IVE estimates can be influenced by nonspecific case definition, ascertainment, information bias and confounding. To overcome some of these limitations, we used a hospital-based active-surveillance approach to identify eligible patients. Despite each site adapting screening criteria to the particular circumstances of their health care systems and the participating hospitals, all sites consistently applied the network eligibility criteria. In addition, to reduce bias, at all sites, subjects were screened and included in the study without previous knowledge of their exposure or outcome status and belonged to the same population at risk for influenza infection, namely, those targeted for vaccination [34]. All sites used the common GIHSN core protocol and close follow-up and feedback between the coordination center and the different sites to ensure that standard procedures and monitoring were employed throughout the influenza season.
We used a highly specific outcome definition of severe influenza, with influenza infection confirmed by RT-PCR performed in highly qualified central laboratories. To minimize the impact of false negatives on IVE estimates, we excluded patients swabbed more than 7 days after the onset of symptoms [34]. IVE was estimated using the widely used test-negative approach, which has been shown to give consistent results [15], and the analysis of IVE was restricted to periods with similar influenza circulation patterns [16,35]. Furthermore, the IVE was calculated on the basis of ORs determined using a random effects model, which allowed us to take into account potential differences, including type of vaccine, vaccination programs, the levels of immunity across different population and settings, and different use of hospital emergency departments [36,37].
Underlying heterogeneity across study sites may have compromised the accuracy of the overall IVE estimates. We observed high heterogeneity in the estimates of IVE against A(H1N1)pdm09 by site in patients $65 years of age. This was mainly due to opposing directions of IVE estimates in France and Valencia. Identifying the host and pathogen factors that may have contributed to this variability is complicated by limited understanding of the factors that affect annual IVE estimates [38]. One possibility for the heterogeneity is site-specific genetic and antigenic differences between circulating A(H1N1)pdm09 and seasonal vaccine virus- es [25,39,40]. We cannot exclude the possibility that the differences between sites are due, at least to some extent, to different vaccines being used.
The heterogeneity of pooled analyses from existing influenza networks and the relevance of IVE estimates across sites sharing a core standardized protocol remain largely unknown [41,42]. A thorough assessment and exploration of the heterogeneity inherent to multicenter studies is needed to evaluate the robustness of pooled IVE results and the identification of risk factors. One possible framework for understanding the heterogeneity of observational IVE data and how to interpret it is that provided by Beyer et al. who re-analyzed data from a 2010 Cochrane metaanalysis of IVE in the elderly [43]. By rearranging the data according to ''a biological and conceptual framework based on the basic sequence of events throughout the 'patient journey''', they found a mean IVE against complications of 28% (95% CI, 26% to 30%) and against laboratory-confirmed disease of 49% (95% CI, 33% to 62%). They concluded that their findings provide ''substantial evidence for the ability of influenza vaccine to reduce the risk of influenza infection and influenza-related disease and death in the elderly.'' Although both their and our analyses were based on heterogeneous source data, we had similar findings and reached similar conclusions on the effectiveness of influenza vaccines.
The wide confidence intervals observed in our study suggests that small sample sizes may have compromised the precision around risk-specific IVE estimates and the power of statistical tests to detect all sources of heterogeneity. Therefore, random error could have affected our estimates. Accordingly, the IVE estimates should be interpreted with caution. The results of this study support the feasibility of estimating IVE against hospitalization for laboratory-confirmed influenza based on a global active-surveillance hospital-based network. New sites in China and Brazil, and a fully operational site in Turkey will be joining the GIHSN in the 2013-2014 season. This will increase its geographical representativeness and sample size, which will improve the validity and accuracy of data on influenza vaccine effects and their variability. This is especially important for attaining the principal public health objectives of preventing morbidity and premature mortality in people at high risk for complications from influenza.   Text S1 GIHSN laboratory characteristics and procedures.