Background
Viral respiratory illnesses such as influenza and respiratory syncytial virus (RSV) impose a substantial burden of hospitalization [
1‐
5]. Influenza has a disproportionate impact on the young, older adults, and persons with underlying high risk medical conditions [
1,
2]. Clinical manifestations of influenza virus infection range from asymptomatic or mild upper respiratory illness to severe lower respiratory tract disease or exacerbation of chronic respiratory and cardiac disease. The clinical impact of influenza is not confined to respiratory conditions. In young children in particular, the clinical presentation of influenza can be diverse, and may include gastrointestinal symptoms and febrile convulsions [
6]. In adults, excess cardiovascular outcomes have been attributed to influenza [
7]. The burden of RSV in young children has long been recognized [
3‐
5], and more recent studies have also identified a substantial burden in older adults [
8‐
10]. Clinical manifestations of RSV infection range from sinusitis and otitis media to bronchiolitis and pneumonia.
Because the burden of seasonal influenza hospitalizations is difficult to assess directly, historically it has been estimated by statistically modeling excess wintertime health-related outcomes over the background rates of the same outcomes recorded outside the wintertime period [
11]. The US Centers for Disease Control and Prevention (CDC) has applied a variety of statistical modeling techniques over the past decades to derive these estimates. More recently, virological surveillance-guided multiple regression models that include International Classification of Disease (ICD)-coded outcomes data as additional model variables have been used to derive pathogen-specific estimates of disease burden, including the burden of RSV. The primary advantage of this approach is that the specific and concomitant burden of disease associated with multiple pathogens is estimated while controlling for other disease drivers [
12‐
14].
The burden of influenza varies from season to season and from decade to decade, both overall and in its impact on different age groups, depending in part on the dominant virus type or subtypes in circulation [
15]. Influenza A/H1N1 and A/H3N2 viruses have co-circulated for the past three decades; although A/H3N2 has caused most influenza A illness over this period, A/H1N1 has predominated periodically [
16]. Two distinct lineages of the influenza B virus, Victoria and Yamagata, have co-circulated since at least 1983. In the 1990s, the Victoria lineage was not detected in North America, but re-emerged during 2001 [
17]. Both lineages have subsequently co-circulated or alternated. Suboptimal vaccine protection might occur in seasons where the predominant influenza B virus is from the lineage that is not included in the trivalent seasonal influenza vaccine [
18,
19].
To inform vaccination strategies and to prompt accelerated development of novel vaccine technologies, it is critical to thoroughly characterize evolving epidemiological patterns of virus circulation and the burden of viral illness. In the present study, we estimated the burden of influenza- and RSV-associated hospitalizations in the US between 1997 and 2009. The study had multiple aims. Firstly, it evaluated the relative burden of influenza versus RSV. Secondly, it evaluated the relative burden of influenza A versus influenza B, which is important in assessing the value of the quadrivalent influenza vaccine. Thirdly, it evaluated the burden of illness by age and risk status. Individuals at high risk, including those who are elderly, immunocompromised and chronically ill, generally experience higher rates of severe outcomes from viral infections and are usually prioritized for vaccination [
20]. Finally, to better estimate the burden of influenza and RSV disease, the study used an outcome introduced in a related study of influenza- and RSV-associated mortality [
21], “respiratory disease broadly defined” (“respiratory broad”), which may provide a better balance of sensitivity and specificity compared with estimates derived using classical definitions.
Discussion
In our study, we applied a multiple regression modeling strategy that has long been employed to estimate the seasonal excess in hospitalizations attributable to influenza and RSV. We also tested the effect of using a variety of diagnoses outcomes (with varying sensitivity and specificity) beyond the traditional ICD code ranges of cardiac and respiratory diseases. In addition, we examined pediatric age groups more closely and sought to stratify hospitalization by influenza risk status (e.g. mention of certain chronic underlying illnesses). Using an expanded respiratory disease outcome (respiratory broad) that included respiratory signs/symptoms and other viral diseases, we estimated that an average US winter season has ~300,000 influenza-related hospitalization events. The risk of hospitalization for influenza was most pronounced in young children and older adults, with the highest burden occurring in the oldest age group, age 75+ years. The annual relative impact of influenza A and B varied substantially during the study period. Using the same respiratory broad outcome, the model also attributed 200,000 winter-seasonal hospitalizations to RSV, also mostly in young children and older adults. Our results confirm and expand on the findings of a recent CDC study regarding the burden of influenza- and RSV-related hospitalizations in the US [
2].
In line with what has been reported in other studies, seasons dominated by circulation of A/H3N2 viruses recorded the highest numbers of influenza-related hospitalizations [
1,
2]. The A/H3N2 virus was associated with the majority of influenza-related hospitalizations during our study period. However, we noted that an unusually high 56–74% of influenza hospitalizations were attributable to influenza B during seasons 2000/01, 2002/03, and 2008/09 in which influenza B accounted for 34–47% of all circulating influenza viruses [
24]. This implies that simply considering average burden estimates over time misses the fact that an influenza B-dominated season can be associated with considerable influenza burden. This point is particularly important for pediatric age groups for whom the influenza hospitalization burden was as high in a predominantly influenza B season as in any season dominated by influenza A/H3N2 viruses. These excess hospitalizations mostly occurred among children without known underlying medical conditions listed on their hospital discharge records. This finding is consistent with current US and WHO policies of recommending vaccination for children for influenza regardless of risk status, in addition to the traditional recommendations of high risk and senior populations [
25,
26].
Our statistical modelling approach included time series of influenza and RSV laboratory-confirmed patterns and thus allowed the simultaneous attribution of hospitalizations to each influenza type/subtype and to RSV. This statistical “excess hospitalizations” approach using national time series of hospitalizations has only in the last decade or so been expanded to include RSV. Recently, evidence of RSV burden was reported by two prospective US studies of laboratory-confirmed influenza and RSV which demonstrated a similar (1:1) ratio of hospitalizations attributable to influenza and RSV in seniors [
9,
10]. However, both of those studies were set in local populations (and over fewer seasons) and therefore a national estimate that also allowed computation of the absolute hospital burden was needed. For the entire study period (1997–2009), we found that the average burden of RSV-attributable hospitalizations is indeed substantial, and that its magnitude is nearly half that of influenza in adults aged 65 years and older, and particularly high in those aged 75+ years. Our findings on RSV burden are in excellent agreement with those of Zhou and colleagues [
2], as both studies found a substantial RSV burden in seniors but not as high as that of influenza in an average season. As seniors are rarely tested for RSV, additional data on indirect measures of the RSV burden in seniors are needed to guide RSV vaccine development and set recommendations for the RSV vaccines that are expected in the coming years.
Not surprisingly, in children under 5 years of age, the hospitalization burden of RSV was four times higher than that of influenza (ratio 4:1). We found very low RSV-related hospitalization in children aged 5–17 years, as observed in previous studies [
27,
28]. As for influenza, most pediatric RSV hospitalizations occurred in children without a documented risk condition for influenza; however, we could not investigate known important risk factors such as premature birth and birth month [
29]. Our findings regarding the relative burden of RSV-and influenza-attributable hospitalization across overlapping age groups are also in agreement with other modeling studies [
2,
8], but ours is the only study to investigate the RSV burden both in school children and in older seniors aged 75+ years. A caveat to our data is that most RSV hospitalizations occur in children less than 1 year of age [
30], but we did not stratify the <5 years age group to give more age-specific estimates. Stratification of this age group would also be useful for influenza estimates.
We noted that the relative burden of influenza B versus A was greatest in both pediatric age groups 0–4 and 5–17 years of age; for these age groups, influenza hospitalizations were split approximately equally between influenza B and A. Overall, the number of hospitalizations was low in the 5–17 year age group regardless of influenza type. Zhou and colleagues also found higher rates of hospitalization attributable to influenza B than to influenza A in children 1–4 years of age, but did not study children aged 5–17 years as a separate age group [
2]. In individuals 5–49 years of years, Zhou and colleagues found approximately equal rates of hospitalization attributable to influenza A and B [
2].
An important goal of our study was to generate a national estimate of influenza- and RSV-attributable hospitalizations according to medical risk status across all age segments; this aspect was not evaluated in previous studies of US hospitalization burden [
1,
2]. Influenza vaccine recommendations around the world account for medical risk status based on the presence of co-morbid conditions. We found our findings in children to be surprising. Although an elevated risk of severe health outcomes in high versus low risk children with influenza is well established in the observational literature [
31,
32], we observed the opposite – high risk children had lower hospitalization rates than low risk children. The relative risk of influenza hospitalization in high risk versus low risk children (ratio high risk/low risk) was 0.2 for children 0–4 years of age and 0.6 for those 5–17 years of age. A similar observation has been made by Cromer and colleagues for children 0–4 years of age in an analysis of influenza burden in England, with a relative risk of 0.7 for children <6 months of age and 0.9 for those aged 6 months to 4 years [
33]. However, Cromer and colleagues’ results contrast with ours for older children aged 5–17 years; in their study, high risk children had a higher burden of influenza hospitalization than low risk children (relative risk 5.7) [
33]. In our previous study of influenza- and RSV-related mortality, we also observed higher death rates in low versus high risk children aged 0–4 years, and approximately equal death rates in low and high risk children aged 5–17 years [
21]. There is no obvious explanation for this finding, but it may be related to under-identification and/or under-reporting of risk factors in children, particularly in ICD-coded hospitalization data.
Our study demonstrated the importance of the outcome used to estimate the hospital burden. The traditional P&I and respiratory outcomes are very specific in capturing the influenza-attributable burden, but are not particularly sensitive. In contrast, the classical cardiorespiratory outcome used by Zhou and colleagues [
2] has, by definition, a lower specificity due to the inclusion of many cardiac disease events not related to influenza. Thus, estimates of burden using the P&I and respiratory outcomes will be lower than estimates using the cardiorespiratory outcome. The respiratory broad outcome included any mention of respiratory codes, as well as any mention of respiratory signs/symptoms and unspecified virus illness. The inspiration to add codes outside the typical ICD code range used in classical studies came from a study finding that ICD9 code 079.99 (viral disease, unspecified) was the code most often associated with oseltamivir prescription, despite the fact that suspicion of influenza must have prompted the prescription [
34]. This finding suggests that likely influenza-related admissions would often be missed using the traditional code range of ICD9 460–519 for respiratory diseases. In our study, as expected, influenza-attributable hospitalizations increased as outcomes became more sensitive and less specific (in rank order P&I, classic respiratory, respiratory broad, cardiorespiratory). We suggest that the respiratory broad outcome provides a case definition that is both sensitive and specific, and thus represents a better choice than traditional outcomes.
To further investigate the potential for exclusion of important hospitalization events using classical outcomes, we also applied the model to time series of sepsis and found that approximately 24,000 sepsis hospitalizations in an average season were attributable to influenza or RSV. Even though this is not a high number compared with the joint seasonal burden of 500,000 hospitalizations associated with the two viruses, the economic burden of sepsis is considerable [
35]. This suggests that such outcomes should also be counted and that considerable cost-savings could be made assuming that some sepsis hospitalizations could be prevented by vaccination against influenza or RSV. In principle, such estimations can be done for other disease outcomes with winter-seasonal patterns.
When comparing our hospitalization rates with those of Zhou and colleagues [
2], we see multiple differences in outcome definitions, statistical modeling approach (negative binomial versus linear regression), data source (we used the Agency for Healthcare Research and Quality’s NIS database, while Zhou and colleagues used the State Inpatient Databases and projected rates to non-participating States), and use of secondary as well as primary diagnoses in our study. Using primary mention of the classical cardiorespiratory outcome definition, Zhou and colleagues estimated an annual influenza-attributable hospitalization rate of 64 per 100,000 population [
2]. Using the corresponding outcome definition, our estimates were slightly higher at 80 per 100,000 population. In an earlier CDC study, Thompson and colleagues reported a corresponding estimate of 88 per 100,000 [
1]. For RSV, the hospitalization rate based on primary mention of the cardiorespiratory outcome was estimated at 55 per 100,000 by Zhou and colleagues [
2] and at 49 per 100,000 by our analysis. Thus, there is broad agreement on the overall burden and pattern of influenza and RSV disease when similar comparisons are made.
Our study had several limitations. Firstly, hospitalization data were only available by month for non-governmental researchers, whilst virus surveillance data were available by week. We dealt with this by interpolating monthly data into weekly data. It is possible that this approach may have overestimated the precision of hospitalization rates. In addition, the lack of weekly data may have resulted in the model being unable to discriminate precisely the burden of influenza and RSV in seasons where there was significant overlap of the two epidemics, especially in children where viral coinfections (including with respiratory viruses other than RSV and influenza) are common. Secondly, although the relative distribution of influenza A, influenza B, and RSV is known to vary by age, US age-specific virological data were not available; instead, we used all-age composite virological data to guide the timing of the epidemic in the model. However, in sensitivity analyses (not shown), lagging outcomes by age strata relative to virus circulation did not improve model fit, suggesting that the timing at least was adequate, although the relative sizes of the estimates by age group might be affected. Thirdly, risk status could only be determined by the presence or absence of ICD codes during a single hospitalization episode, and assignment of risk status critically depended upon whether the physician had mentioned among the discharge diagnoses any existing underlying disease. In clinical practice, it is likely that the physician would fail to mention some underlying diseases or factors that are categorized as high risk; indeed some important risk factors such as obesity or pregnancy are not reported reliably unless assessed prospectively. This may have led to mis-assignment of risk status. In contrast, symptoms that were part of the influenza illness may have been listed as underlying disease, leading to overestimation of risk. We could not investigate other known risk factors such as premature birth and obesity, which were unlikely to be mentioned by physicians on discharge. Fourthly, negative burden estimates were not included in the aggregations as they are not biologically plausible. In cases where the coefficient indicates a very small attributable burden from one of the viral terms, the estimate may be negative due to the inherent lack of precision. This problem is particularly acute during periods in which RSV and influenza co-circulate, and for nonspecific outcomes such as cardiorespiratory illness. Setting the negative estimates to zero could have resulted in overestimation of hospitalizations. However, we conducted a sensitivity analysis comparing estimates obtained when we did or did not set negative values to zero. For influenza A/H3N2 and B, there were no negative estimates to set to zero. For A/H1N1, where the burden is considerably smaller, we found that setting the estimates to zero did have a substantial effect. However, the net effect on the total influenza burden was limited, because the A/H1N1 virus causes only a small proportion of the total influenza-attributable hospitalization. Finally, although our study used national data and employed a linear regression model, it was limited by the ICD-coded information on each discharge record. Thus, some true patients may have been missed, while others may have been miscoded and wrongly included. It should be noted that the clinical manifestations of RSV infection are very broad, and that atypical presentation of influenza is common in young children; some of these presentations would have been likely to be assigned ICD codes not included in any of our definitions. Moreover, ICD coding practices and criteria for hospitalization can change over time and thus affect the estimates. For example, diagnosis codes in administrative data are often assigned as part of billing and cost reimbursement procedures, allowing changing financial incentives to bias trends in coding practices. It seems unlikely that such forces will become less important in the future, although improvements in diagnostic testing might make the codes assigned to patients with infectious diseases more accurate. No system will allow complete accuracy in assignment of diagnostic codes. Although the two US laboratory-based studies on RSV burden did not have the problem of assignment of diagnostic codes to contend with, they were limited by the use of local data, the small populations under study and the possibility that a hospitalization was not directly attributable to RSV infection [
9,
10].
There are also some caveats to our statistical methodology. All indirect statistical approaches based on retrospective data have inherent limitations. For example, non-infectious factors and pathogens other than influenza and RSV may influence hospitalization, but specific information on the nature of such influences is unavailable. Although a cyclic term was included in the model to adjust for this, it adds uncertainty to the model estimates in addition to that derived from variability in the data. In addition, when an outcome occurs at a low frequency, the assumption of a normally distributed dependent variable is more likely to be violated. To address the problems associated with low frequency data, some previous analyses have used generalized linear models that relax the assumption of a normal distribution, including discrete probability distributions appropriate for count data such as the Poisson regression with a natural logarithm for the link function [
12,
36]. Use of the logarithm function, however, implies multiplicative effects of respiratory viruses, which is an unrealistic assumption [
37,
38]. It is more likely that co-circulating strains could result in a reduction of the burden due to cross-protection [
39], competitive exclusion or both. This makes the additive model more plausible from a biological and epidemiological perspective [
38‐
40].
Acknowledgements
The authors thank Lone Simonsen and Lewis Kim from Sage Analytica for their contribution to the study and publication, in particular Lone Simonsen for her critical review of the manuscript. Authors are also grateful to all teams of GSK for their contribution to this study and publication, especially Robert Gardner for report writing, Els Tassenoy for her global study management contribution, and Stephanie Wery and Ramandeep Singh for their support during the quality check of the manuscript. The authors would like to thank Business & Decision Life Sciences platform for editorial assistance and manuscript coordination, on behalf of GSK. Bruno Dumont coordinated manuscript development and editorial support. The authors also thank Mary Greenacre (An Sgriobhadair, UK, on behalf of GSK) for providing medical writing support.