Burden of influenza, respiratory syncytial virus, and other respiratory viruses and the completeness of respiratory viral identification among respiratory inpatients, Canada, 2003‐2014

Background A regression‐based study design has commonly been used to estimate the influenza burden; however, these estimates are not timely and many countries lack sufficient virological data. Alternative approaches that would permit a timelier assessment of the burden, including a sentinel surveillance approach recommended by the World Health Organization (WHO), have been proposed. We aimed to estimate the hospitalization burden attributable to influenza, respiratory syncytial virus (RSV), and other respiratory viruses (ORV) and to assess both the completeness of viral identification among respiratory inpatients in Canada and the implications of adopting other approaches. Methods Respiratory inpatient records were extracted from the Canadian Discharge Abstract Database from 2003 to 2014. A regression model was used to estimate excess respiratory hospitalizations attributable to influenza, RSV, and ORV by age group and diagnostic category and compare these estimates with the number with a respiratory viral identification. Results An estimated 33 (95% CI: 29, 38), 27 (95% CI: 22, 33), and 27 (95% CI: 18, 36) hospitalizations per 100 000 population per year were attributed to influenza, RSV, and ORV, respectively. An influenza virus was identified in an estimated 78% (95% CI: 75, 81) and 17% (95% CI: 15, 21) of respiratory hospitalizations attributed to influenza for children and adults, respectively, and 75% of influenza‐attributed hospitalizations had an ARI diagnosis. Conclusions Hospitalization rates with respiratory viral identification still underestimate the burden. Approaches based on acute respiratory case definitions will likely underestimate the burden as well, although each proposed method should be compared with regression‐based estimates of influenza‐attributed burden as a way of assessing their validity.


| INTRODUCTION
Despite significant efforts to assess the burden and severity of the 2009 pandemic strain (A(H1N1)pdm09) in real time, several gaps in global influenza surveillance capacity became apparent, notably the completeness of case ascertainment, which the World Health Organization (WHO) believed to be quite low globally. 1 In response, the WHO issued a briefing note in 2009 advising that statistical models similar to the methodology introduced by Thompson et al in 2003 2 are required to more completely estimate the burden associated with influenza, 1 and later funded an international study to use a statistical model to estimate the global burden associated with A(H1N1)pdm09. 3 Subsequently, to facilitate real-time assessment of the burden, enhance influenza surveillance globally, and promote standardization, the WHO developed influenza surveillance guidelines in consultation with member countries. 4 Since 2003, a regression approach has been widely used in many countries, 3,5 including Canada to estimate the excess burden associated with influenza and other respiratory viruses. [6][7][8][9] Because regression estimates are not available in real time, researchers have proposed other methods 10,11 to use surveillance data to predict the full burden in near real time, only some of which have been evaluated at year-end against a regression approach. 6,[12][13][14] Despite the focus on developing methods for more timely estimates of the burden associated with influenza, other respiratory viruses cause considerable disease burden in Canada. 15,16 In this study, our objectives were (i) to update our estimates of the average annual respiratory hospitalization rate attributable to influenza, RSV, and other respiratory viruses in Canada by age group and disease severity; (ii) to quantify the degree to which the rates of hospitalizations coded as due to those virus rates underestimate this burden; and (iii) to discuss issues related to adopting emerging approaches based on near real-time data that would permit more timely assessment of the burden of influenza and other respiratory viruses. The proportion of hospitalizations attributed to influenza and other respiratory viruses for which the respiratory virus was identified provides an approximate measure of completeness of viral identification, estimates that help hospital administrators interpret their data, as the corresponding multiplier (reciprocal of completeness) has been used in this way with near real-time surveillance data. 10 As estimation of the hospitalization burden attributable to influenza for the pediatric population has been a challenge, we explored this issue by comparing the number of admissions with the influenza virus identified to the estimated excess attributed to influenza for various diagnostic groups. We also assessed the potential impact of restricting testing to various clinical case definitions on the burden estimates.

| Data source
Hospital discharge records for all patients admitted to an acute care hospital for a respiratory condition between September 2003 and August 2014 were extracted from the Canadian Institute of Health

| Measures of viral activity
Hospitalizations with an identified respiratory virus were classi-

| Creation of the dependent variables for the regression model
A weekly time series was created for each subgroup for which we sought to estimate the excess burden associated with influenza, RSV, and ORV activity. Records were aggregated accordingly by age group and diagnostic group, discharge status, and severity as required for estimation of the excess burden. Diagnostic categories included all respiratory conditions (J00-J99) as a primary diagnosis and the following subcategories: primary pneumonia and influenza (PNI) (J10-J18), primary influenza-like illness or unspecified acute respiratory infection (J06.9, J11), primary respiratory with any mention of an acute respiratory condition (ARI) (J00-J22, J44.0) (as either a primary or secondary ARI diagnosis), and primary respiratory without any mention of an ARI dx as either a primary or secondary diagnosis. We interpreted the term "SARI" from the WHO surveillance guidelines to correspond to the ICD-10 diagnostic codes for either ILI (J11) or ARI (J00-J22, J44.0), although the WHO SARI case definition has not been formally evaluated against ICD-10. 4 All diagnosis codes were determined after discharge.

| Statistical analysis
The weekly number of hospitalizations (by date of admission) for each subgroup was modeled as a function of viral activity, seasonality, and trend using a quasi-Poisson regression model similar to previously published estimates of the influenza burden in Canada (see Appendix S2 for details) 6,7 and internationally. 3  to the uncertainty in the denominator. When the lower 95% CI of the estimated excess was less than the number with viral identification, the upper 95% CI of the completeness estimate was replaced with 100%. Negative estimates of excess hospitalizations can arise due to chance alone if the true (though unknown) number of hospitalizations due to the virus is less than the usual variation in the total number of hospitalizations. In this case, the negative estimate is usually not statistically significant. A negative excess in hospitalizations associated with influenza activity could also arise due to the fixed number of hospital beds with limited surge capacity and the resulting need to prioritize admissions. In a previous study, we found non-respiratory hospitalizations were lower than the estimated baseline during periods of influenza activity. 15 Various methodological issues related to the regression approach have been discussed elsewhere. 6,15 Analyses were performed with sas (version 9.1; Cary, NC, USA).

| RESULTS
Over the study period, there was an average of 180 000 hospitalizations for respiratory conditions per year (700 per 100 000 population) of which influenza, RSV, and ORV were identified in 1.1%, 2.6%, and 0.4%, respectively, and attributed to an estimated 4.7%, 5.3%, and 3.8% of respiratory hospitalizations, respectively. For inpatients under the age of 2 years, 25% of primary respiratory hospitalizations had a viral identification (Table 1)

| Influenza-attributed hospitalizations by diagnostic subgroup
Influenza-attributed hospitalization rates were highest in the oldest and youngest age groups although the impact of the diagnostic subgroup varied with age ( Figure 2). In the pediatric population, the only diagnostic group other than influenza (J10) for which we found a notable excess associated with influenza activity was the ILI and other unspecified acute respiratory infection (J11, J06.9) categories estimated at 3.3 (95% CI: 2.7, 3.8) per 100 000 population. The average annual rate of influenza hospitalizations (J10, virus identified) was 11.7/100 000, while the estimated rate of excess influenza-attributed respiratory hospitalizations was considerably lower at 3.8 (95% CI: −1.6, 9.2) ( Table 2).
Restricting the category to any ARI resulted in an estimated excess rate of 8.0 (95% CI: 4.3, 11.8). These otherwise apparently contradictory estimates can be explained by the statistically significant offset (or negative excess) for non-ARI respiratory admissions with the rate estimated at −4.6 (95% CI: −7.0, −2.2)/100 000. The burden attributable to influenza for the pediatric age groups is therefore assumed to be best represented by the sum of influenza hospitalizations (J10) plus the influenza-attributed ILI and unspecified acute respiratory infection     for viral identification and viral attribution by status are provided in Table 4.

| INTERPRETATION
This study confirms the very high rates of RSV hospitalizations in young children and that viral identification likely occurs in a majority of the hospitalizations due to influenza or RSV in the pediatric population in Canada. Among adults, influenza likely accounts for more than half of all hospitalizations due to a respiratory virus; however, completeness of viral identification for influenza is much lower among adults, and declines with increasing age. Completeness of influenza viral identification was higher for patients admitted to a special care unit than for inpatient deaths. One-quarter of respiratory hospitalizations attributed to influenza did not have an ARI diagnosis recorded.

| Other Canadian studies and validation
This study updates previous estimates of rates of respiratory hospitalizations attributable to influenza, RSV, and other respiratory viruses 15,16 for a period when ICD-9 was used, and more recent studies In our previous studies, we have had similar difficulties estimating the excess influenza burden for pediatric inpatients from all respiratory admissions. The offsetting (or negative) excess associated with influenza activity for non-ARI respiratory hospitalizations, which we have seen before among non-respiratory hospitalizations for adults, 15 suggests that the concept of excess influenza hospitalizations may not be appropriate for the pediatric age groups. However, as we found a statistically significant excess for RSV among non-ARI respiratory hospitalizations (results not shown), we have likely underestimated the full burden for influenza in the pediatric age group. We detected an excess associated with influenza among the ILI and unspecified acute respiratory infection category (J11, J06.9), so we conclude that the viral identification for influenza is not complete.

| Comparison with other studies
American studies suggest a more substantial underdiagnosis of influenza for the US population. A similar regression-based study by Ortiz

| Limitations
This is an ecological study design, and the wide confidence intervals for estimates of the rate of excess respiratory hospitalizations attributable to influenza in the pediatric age groups, as well as for other categories, illustrate the limitations of this approach. The relatively small number of excess hospitalizations attributable to influenza compared to RSV among pediatric inpatients contributes to the imprecision. 16 We have concluded that the ecological estimate of excess respiratory hospitalizations for the pediatric age groups (ie, the influenza-attributed respiratory estimate) is an underestimate of the true burden due to evidence of an offset (negative excess) in non-ARI respiratory hospitalizations. As we have seen evidence of a similar offset effect 22 in non-respiratory admissions for adults, 15 it is possible that there was a reduction in other respiratory hospitalizations to accommodate the influenza surge. Our estimates of completeness are only approximate as the ecological estimates of burden are a measure the net excess rather than the number of hospitalizations truly due to the virus. As a result, there will be some discrepancy between this approach and the one used by Reed et al. 10 Low viral testing rates among diagnostic groups associated with a specific respiratory virus, such as croup and parainfluenza virus, 16,23 could result in large confidence intervals, or biased results. Noting that a viral identification was recorded in only 2% of croup hospitalizations, we are unsure of whether we have captured the full ORV-attributed burden. The actual national burden may differ slightly from our estimates as data from the province of Quebec were not available from the DAD. Variation from season to season would likely overshadow regional differences, although completeness of viral identification could vary considerably across provinces, regional health units, or even by hospital, and over time. Hence, the estimates of completeness may not apply to Quebec, or necessarily to any individual province, health unit, or hospital.
As we do not currently have information on which inpatients were tested, we could not evaluate whether the excess influenza-attributed ILI and unspecified acute respiratory infection hospitalizations correspond to patients for whom testing was not done, or for patients for whom the specimen was test negative. As time since symptom onset and sample procurement and handling can influence the sensitivity of viral identification, it is possible that hospitalizations for which the specimen tested negative were still due to one of the tested viruses. 24 In which case, the operational sensitivity is likely lower than the manufacturer's estimate of sensitivity as used in the study by Reed et al. 10 As the ILI diagnostic category is not typically used for patients once the virus is identified, we cannot assess the impact of using only the clinical ILI/SARI diagnostic category for testing purposes as suggested by the WHO sentinel surveillance program recommendations. 4 The WHO ILI and SARI case definitions are not identical to the ICD-10 codes for ILI and ARI. The ICD-10 definition of ARI is much broader than the ICD-10 ILI definition, while the WHO SARI definition is similar to their ILI case definition (acute respiratory infections with recent onset of cough and fever ≥38°C) but restricted to hospitalized cases. Hence, we were not able to fully assess the potential impact of using a SARI approach on our burden estimates. In a study of emergency department visits where only the clinical or pre-test diagnoses are recorded, we found that 58% of ILI visits were likely due to influenza, but that an ILI diagnosis was present in only 7% of excess respiratory emergency department visits attributed to seasonal influenza. 8 The ARI diagnostic category captured 84% of influenza-attributed ED visits. However, only 10% of all ARI ED visits were likely due to seasonal influenza. Similar results were found in a laboratory-based study during a period of peak influenza activity. 25

| CONCLUSIONS
Though overall, the morbidity burden for influenza, RSV, and ORV is similar; the age distribution is very different with the rate of RSV- When dealing with near real-time data, delays in reporting make the interpretation of weekly data considerably more complex. [26][27][28] Use of an ARI case definition to identify cases for viral identification will result in the identification of a much larger proportion of hospitalizations due to influenza than use of an ILI definition, although at a cost of considerably higher testing rates. 25 An advantage of a sentinel surveillance program is that in capturing the test negative inpatient results, rates can be extrapolated to the full population. Even with a SARI case definition as proposed by the WHO, ensuring representation will be a challenge, with considerable underrepresentation of elderly inpatients with chronic respiratory conditions (such as COPD) expected as symptoms of an acute respiratory infection were either not present or overlooked in our data. For countries with limited or irregular viral identification data, the routine testing of inpatients from sentinel hospitals meeting either the ILI, ARI, or SARI case definitions should provide the quality weekly virological data required for regression-based estimates of excess mortality and morbidity associated with influenza.