Excess pneumonia and influenza hospitalizations associated with influenza epidemics in Portugal from season 1998/1999 to 2014/2015

Background The aim of this study was to estimate excess pneumonia and influenza (P&I) hospitalizations during influenza epidemics and measure their correlation with influenza vaccine coverage in the 65 and more years old, according to the type/subtype of influenza virus. Methods The study period comprised week 40/1998‐40/2015. Age‐specific weekly P&I hospitalizations (ICD‐9: 480‐487) as main diagnosis were extracted from the National Hospital Discharge database. Age‐specific baseline hospitalization rates were estimated by autoregressive integrated moving average (ARIMA) model without time periods with excess hospitalizations. Excess hospitalizations were calculated by subtracting expected hospitalization rates from the observed during influenza epidemic periods. Correlation between excess P&I hospitalizations and influenza vaccine coverage in the elderly was measured with Pearson correlation coefficient. Results The average excess P&I hospitalizations/season was 19.4/105 (range 0‐46.1/105), and higher excess was observed in young children with <2 years (79.8/105) and ≥65 years (68.3/105). In epidemics with A(H3) dominant, the highest excess hospitalizations were observed among 65 and over. Seasons which influenza B or A(H1)pdm09 dominance the highest excess was observed in children with <2 years. High negative correlation was estimated between excess hospitalizations associated with A(H3) circulation and vaccine coverage in the elderly (r = −.653; 95% CI: −0.950 to −0.137). Conclusion Over 80% of the influenza epidemics were associated with excess hospitalizations. However, excess P&I hospitalizations pattern differed from age group and circulating virus. This ecologic approach also identified a reduction in excess P&I associated with A(H3) circulation with increasing vaccine coverage in the elderly.


| INTRODUCTION
Influenza viruses circulate every year, causing epidemics that are usually benign and mild for the human population but that can complicate into other diseases, like pneumonia. According to Wuerth et al (2016), 1 in the 2002-2011 period, influenza was the fourth causative agent for pneumonia and approximately 9/100 000 of pneumonia hospitalizations had influenza as the etiological respiratory agent. However, as influenza laboratory diagnosis is not usually performed in all suspected cases, using these data underestimate the influenza impact. 2,3 Taking this into consideration, the overall effect of influenza epidemics has been measured through indirect ecologic methods using the Serfling approach 4 and Poisson, negative binomial regression and autoregressive integrated moving average (ARIMA) models to estimate influenzaassociated mortality or hospitalizations rates. [5][6][7][8][9] There are two main approaches in estimating influenza-associated excesses: one based on statistical models that include influenza activity indicators as explanatory covariates; another does not consider covariates and by excluding from the estimating process all parts of the outcome time series where there is evidence of occurrence of some event that might influence the outcome. 9 Both approaches have pros and cons. Using models with covariates allows estimating influenza-associated outcomes by virus type and subtype, but requires robust virological data. 10 In the alternative, this specific data requirement is not needed and can be used provided that consistent mortality or hospitalization time series are available. 10 However, there are limitations also in this approach.
The lack of virological covariates in the model implies the assumption that all excess winter mortality is associated with influenza circulation which may not be appropriate, leading therefore to a mortality overestimation.
The identification of influenza epidemics requires influenza surveillance data, with information on influenza virus type, and influenza epidemic activity period. 11,12 Also, and when available, the identification of other events that contribute to mortality or hospitalizations distribution, like secular trend or seasonality, is desirable so to get a better fit of the model to the time series and improve the quality and the validity of the influenza attributable excess estimate. 13 The influenza impact is particularly evident in specific groups, like the elderly, pregnant and those with chronic disease, with higher risk of complications associated with influenza infection leading to hospitalization or death. 14 For this high-risk group of individuals, yearly vaccination in the autumn is recommended in Portugal and in most EU countries, with the intention of reducing risk of complications, severe disease, and death. 15,16 The influenza vaccine has proven to be moderately effective in reducing medically attended confirmed influenza. 17 Using an ecologic approach, one would expect a reduction in excess hospitalizations/mortality with increasing vaccine coverage and this was already reported in previous influenza-related outcomes studies. 18,19 In Portugal, there are several studies that associate this respiratory infection with excess pneumonia and influenza (P&I) and all-causes mortality. 8,13 However, information on influenza impact on morbidity indicators, such as hospital admissions, is scarce and the knowledge of influenza impact and the role of immunization on hospitalizations is essential for a better resource management and for preparing mitigation measures.
Considering the Portuguese mainland context, this study aims to (i) estimate the excess number of P&I hospitalizations during influenza epidemics from seasons 1998-1999 to 2014-2015 and (ii) to measure their correlation with influenza vaccine coverage in the elderly (65 and more years old), according to the type/subtype of influenza virus predominant in each season.

| METHODS
A time series ecological study was conducted to estimate the baseline of weekly P&I hospitalizations free of influenza epidemics and estimate excess P&I hospitalizations associated with influenza epidemics between 1998/1999 and 2014/2015 seasons.

| Influenza activity
The definition of influenza epidemic periods was based on in-

Predominant virus circulating was provided by Portuguese
Laboratory Network for the Diagnosis of Influenza Infection, namely by National Influenza Reference Laboratory, and was defined according to Influenza Reporting Protocol. 21 The threshold for dominance was set at 60%, and the threshold for codominance is set between 40% and 60%. 21

| Periods potentially associated with excess hospitalizations
These periods include the influenza epidemic (description above on Influenza activity), 2009 pandemic influenza, 22 and heat wave periods (weeks in which two or more consecutive days had an average maximum daily temperature above 32°C with an extra week to account the known delay of impact. 13,23

| Influenza vaccine coverage rates
The influenza vaccine coverage (IVC) rates, for individuals with 65 and more years, were obtained from ECOS (Em Casa Observamos Saúde-At Home We Watch Health), a panel of approximately 1000 households on which a seasonal survey is carried out by computer-assisted telephone interview (CATI). 24 These households were selected randomly from the national telephone directory and recruited considering the representativeness of Portuguese mainland families reachable by telephone.

| Methods to estimate the number of excess hospitalization associated with influenza epidemics
Age-specific baseline hospitalization rates were estimated by ARIMA model, 25 after extracting from the time series the periods potentially associated with excess hospitalizations (defined above and presented on Table S1), using Flubase R package. 26 ARIMA model is composed of 3 terms: first, an autoregressive term (AR), in which the time series is regressed on itself at specific lag times; second, a moving average term (MA), in which the time series is regressed on the regression errors at specific lag times; and finally, an integration term that accounts for the nonstationarity of the time series. If the integrated component is present, the time series is differentiated on itself at specific lag periods; if I = 1, the original time series is transformed in new time series = y(t)−y(t−1).
Using an automatic model identification algorithm included on the package, which recurs to a specific R package named forecast, the final models (Table S2) were selected. 27 Excess hospitalization rates were calculated by subtracting P&I weekly hospitalization rates baseline, obtained through the model fitting, from the observed weekly P&I hospitalization rates during influenza epidemic periods ( Figure 1).
Season excess P&I hospitalization rates 95% confidence level were calculated by approximation to the normal distribution, using F I G U R E 1 Weekly P&I hospitalization rate per 100 000 inhabitants and the respective estimated baseline and baseline upper limit in the absence of the impact of influenza epidemics from 1998 to 2015 as standard error the product of the square root of the number of weeks with excess mortality by the standard deviation of the model residual.
Correlation between season excess rate of hospitalizations and influenza vaccine coverage (last column of Table 1) was measured with Pearson and Spearman correlation coefficient. Confidence interval for coefficient between the vaccine coverage and the season excess P&I hospitalizations rate among the individuals aged 65 years or more was based on bootstrap. 28,29 A significance level of 5% was considered.

| Association between influenza vaccine coverage and seasonal excess P&I influenza-associated hospitalization rates among the elderly
The overall correlation between the influenza vaccine coverage  This fact was reflected in the recommendations issued by health authorities regarding the pandemic monovalent vaccine uptake that did not include the individuals aged 65 or more years of age. 38,39 The excess P&I hospitalizations profile was maintained in seasons with predominance circulation of A (H3) 19 A possible explanation for the difference may be the data source for ILI rates, that is, the use of provisory ILI rates (in the European study) and definitive rates calculated in the end of the season (in the national study).

| DISCUSSION
The results presented in the this study must be interpreted in light of the methods and data limitations. A time series ecological method was used to estimate the seasonal P&I excess hospitalization rates in the study period. As such, the excess P&I hospitalization rate cannot be considered fully attributable to influenza epidemics but only associated with the occurrence of the epidemics. Other covariates as the circulation of other respiratory virus, like the respiratory syncytial virus, were not considered in the model. This fact could have overestimated the impact of the influenza epidemics reported in the present report, mainly for the younger age groups (<2 years). 41,42 Nevertheless, when stratified by influenza virological profile, there were differences between A(H3), A(H1)pmd09, or B predominant seasons, data consistent with previous reports, which suggests that P&I excess hospitalizations were sensible and probably specific in capturing the impact of the influenza epidemics. In this study, an ARIMA model was used and no information about virus type or subtype circulation was considered.
Unlike previous studies, 12,42 the weekly distribution of the influenza virus detection was not included as covariate in the model. This fact did not allow for estimation of the excess P&I influenza-associated hospitalization by (sub)-type influenza virus. However, according to Thompson et al, 10 similar influenza-associated mortality was obtained using ARIMA or Poisson models that included influenza type as covariate. This study also highlighted that, in the absence of robust covariates data, namely on weekly influenza type and subtype data in specific age groups, ARIMA models are a good candidate model to be used in influenza-associated excess studies. 10 Finally, our study did not take into consideration the correlation between circulating influenza virus and vaccine strains. This could have implications on the correlation estimates between vaccine coverage and influenza-associated P&I hospitalizations.
In summary, our type/subtype and age-specific influenzaassociated P&I hospitalizations are in accordance with the literature on influenza excess hospitalization and provide for the first time a measure of the impact of influenza in Portugal for a wide period. Also, this study evidences that there is a negative correlation between influenza vaccination and the influenza-associated P&I hospitalizations, in particular when A(H3) is circulating and in the risk group of the elderly for which the vaccine is recommended. These results are encouraging, specially to validate campaigns for vaccine uptake in seasonal epidemics and in specific age groups in a future pandemic.
Nevertheless, more studies should be performed to evaluate the consistency of these results. Following this line of thought, it is important to continue researches in Portugal by looking into other hospitalizations causes and by incorporating additional information on influenza type/subtype circulation as well as other respiratory virus as a way to fortify these findings.