State‐level estimates of excess hospitalizations and deaths associated with influenza

Abstract Background National estimates of influenza burden may not reflect state‐level influenza activity, and local surveillance may not capture the full burden of influenza. Methods To provide state‐level information about influenza burden, we estimated excess pneumonia and influenza (P&I) and respiratory and circulatory (R&C) hospitalizations and deaths in Colorado from local hospital discharge records, death certificates, and influenza virus surveillance using negative binomial models. Results From July 2007 to June 2016, influenza was associated with an excess of 17 911 P&I hospitalizations (95%CI: 15 227, 20 354), 30 811 R&C hospitalizations (95%CI: 24 344, 37 176), 1,064 P&I deaths (95%CI: 757, 1298), and 3828 R&C deaths (95%CI: 2060, 5433). There was a large burden of influenza A(H1N1) among persons aged 0‐64 years, with high median seasonal rates of excess hospitalization among persons aged 0‐4 years. Persons aged ≥65 years experienced the largest numbers and highest median seasonal rates of excess hospitalization and death associated with influenza A (H3N2). The burden of influenza B was generally lower, with elevated median seasonal rates of excess hospitalization among persons aged 0‐4 years and ≥65 years. Conclusions These findings complement existing influenza surveillance. Periodic state‐level estimates of influenza disease burden may be useful for setting state public health priorities and planning prevention and control initiatives.

under-detection of influenza-associated outcomes by surveillance. 4 CDC therefore estimates excess hospitalizations and deaths due to a range of diagnoses during periods of influenza circulation. 5 However, national burden estimates may not specifically inform allocation of resources at the state or local public health level for prevention and response, and national rates may not reflect local populations or influenza circulation. The application of similar estimation methods to state-level data has the potential to provide local burden estimates that complement existing influenza surveillance. We, therefore, adapted CDC's methods and analogous methods from similar studies to estimate influenza excess hospitalizations and deaths in Colorado. [2][3][4][5][6][7][8][9][10][11][12][13][14][15][16] We demonstrate that periodic disease burden estimates are feasible and provide additional information about serious influenza outcomes not routinely available in a state.

| Study design and study population
We estimated the burden of severe influenza in Colorado, a state with a population of 5.5 million people, from July 1, 2007, through June 30, 2016. 17 Outcomes included influenza excess pneumonia and influenza (P&I) and respiratory and circulatory (R&C) hospitalizations and deaths. Estimates of these outcomes were obtained using regression models of weekly hospital discharges and deaths on circulating influenza over 469 weeks. The regression models identified numbers and rates of P&I and R&C hospitalizations and deaths during the influenza season that were in excess of a seasonal baseline. Population denominators were taken from the American Community Survey. 17

| Hospitalization and death data
To estimate influenza excess P&I and R&C hospitalizations and deaths, we first identified P&I and R&C hospitalizations and deaths recorded in statewide non-federal acute care hospital discharge data and death certificate data provided by the Colorado Department of Public Health and Environment. P&I hospitalizations and deaths included records with a respective primary discharge diagnosis or underlying cause of death coded by ICD-9 codes 480-488 or ICD-10 codes J09-J18. R&C hospitalizations and deaths included records with a respective primary discharge diagnosis or underlying cause of death coded by ICD-9 codes 390-519 or ICD-10 codes I00-I99 or J00-J99. These diagnoses groups were aggregated to weekly counts.

| Viral surveillance data
The primary predictor of influenza excess P&I and R&C hospitalizations and deaths was the overall weekly percent of influenza tests in Colorado that were positive for influenza by type and subtype. These data were obtained from a representative sample of sentinel laboratories that provided influenza virus surveillance data to the Colorado Department of Public Health and Environment during the winter respiratory season from October through May of the next year. Tests were ordered at the discretion of the ordering physician and included a combination of molecular assays and antigen detection tests.

| Time-series regression models
Statistical analyses included a basic negative binomial model that was fit separately for P&I and R&C hospitalizations and deaths by F I G U R E 1 Negative binomial model used to estimate excess hospitalizations and deaths associated with influenza A(H1N1), A(H3N2), and B. E(Y i ) is the expected number of hospitalizations or deaths at week i. α is the population offset taken from the American Community Survey 1-year estimates assuming linear weekly growth between years. 17 It is the number of the week in the time series. β 0 is the intercept. β 1 through β 3 are coefficients associated with baseline linear, quadratic, and cubic time trends. β 4 and β 5 are coefficients associated with baseline seasonal changes. β 6 -β 10 are coefficients associated with percentages of specimens testing positive for each influenza virus type and subtype during a given week. Separate coefficients captured the effect of influenza A(H1N1) during pre-pandemic, pandemic, and post-pandemic periods. See the text for model fitting methods E(Y i ) = α*exp(β 0 + β 1 *t i + β 2 *t i 2 + β 3 *t i 3 + β 4 *sin(2t i π/52) + β 5 *cos(2t i π/52) +

Model fitting incorporated the Akaike information criterion (AIC)
and likelihood ratio tests. 20 First, the basic model was constructed with lag periods of 0, 1, 2, or 3 weeks between virus surveillance findings and hospitalization or death, and AIC was used to select the best-fitting lagged model (Table S1). The lag structure (0-3 weeks) was held the same for each influenza type and subtype in an individual model. Next, the significance of β 1 -β 3 was evaluated using the combination of AIC and the likelihood ratio test, and corresponding variables were eliminated if they did not improve model fit. 20        Estimated median seasonal rates of excess R&C hospitalization demonstrated age group-specific epidemiology by influenza type and subtype similar to that for P&I hospitalization but with a larger magnitude (Table 3). Median seasonal rates of influenza excess R&C hospitalization associated with A(H1N1) were highest among persons aged 0-4 years. Those for influenza A(H3N2) were highest among persons aged ≥65 years. Those for influenza B were lower but peaked among persons aged 0-4 years and ≥65 years.

| Evolution of influenza A(H1N1) over time
The model coefficients for percentages of specimens testing positive influenza A(H1N1) from surveillance differed during pre-pandemic, pandemic, and post-pandemic time periods, indicating that the association between virus circulation and influenza excess hospitalizations and deaths differed during these periods (Table   S1). The greatest apparent differences occurred for excess P&I and R&C hospitalizations among persons 5-49 years and 50-64 years old. In these age groups, the coefficients were largest during the pandemic period, suggesting that pandemic influenza A(H1N1) was associated with higher rates of excess hospitalization and death than pre-pandemic virus and that the rates of excess hospitalization and death associated with influenza A(H1N1) were lower in the post-pandemic period. A similar pattern was not apparent in the models for excess P&I and R&C hospitalizations among persons ≥65 years old or in the models for P&I and R&C excess deaths, where the age groups were not as finely stratified.   We also captured the evolving epidemiology of influenza   8,11,12 Our attempts to do so were limited by power and the occurrence of negative estimates. While negative estimates are a known limitation to this approach, they must be interpreted with caution. For example, confidence limits that cross zero should be interpreted as an indicator of lack of statistical significance, rather than as a possible protective effect. 24 To avoid misinterpretation, we allowed for the influenza A(H1N1) coefficient to change at biologically plausible time points coinciding with the emergence and evolution of the pandemic strain. Nevertheless, season-specific influenza burden may be better estimated using other methods. 22 It is also worthwhile noting that the interannual variability in influenza severity by season, type and subtype, and age will impact the overall summary measures.

| D ISCUSS I ON
A challenge to generating model estimates at the state-level is to obtain accurate virus surveillance data. We did not have complete influenza A subtyping data from local surveillance, so we assigned influenza A subtypes based on regional proportions of H1N1 to H3N2. Others have applied regional virus surveillance data to state-level outcomes. 13 We imputed influenza A(H1N1) surveillance findings during the first wave of the 2009 pandemic, because year-round surveillance data were not available. Other challenges may include lack of power to fit age-stratified models for less frequent outcomes.
There are a few potential limitations to our study. One was the inability to account for respiratory syncytial virus (or other pandemic, which may have affected our virus surveillance data. 12 Some authors have used additional indicators of influenza circulation, such as outpatient visits for influenza-like illness, which we did not evaluate. 16,26 Finally, we did not compare our estimates to those extrapolated from surveillance for laboratory-confirmed influenza hospitalization. 22 Current United States methods of influenza disease burden estimation include surveillance for laboratory-confirmed influenza hospitalization captured by the US Influenza Hospitalization Surveillance Network (FluSurv-NET) and adjustment for under-detection through the use of multipliers that account for testing practices and test characteristics. These methods were developed to create more timely national disease burden estimates that better capture season-to-season variation. 22,27 A potential future study would be to develop these newer methods for state-level analyses and compare findings to state-level estimates using the "Serfling-type" models used in this study. 24 In summary, there were nearly 18 000 excess hospitalizations and 1000 excess deaths due to P&I causes, and 31 000 excess hospitalizations and 4000 excess deaths due to R&C causes associated with influenza in Colorado over nine influenza seasons. These local estimates and corresponding rates provide additional information to state public health officials for setting priorities and planning interventions to prevent and control influenza. The methods are feasible and could be used to produce periodic reports of state-level disease burden.

ACK N OWLED G EM ENTS
The authors would like to acknowledge John Hughes, PhD and Deborah Thomas, PhD for their guidance throughout the study.