The association between influenza infections in primary care and intensive care admissions for severe acute respiratory infection (SARI): A modelling approach

Abstract Background The burden of severe influenza virus infections is poorly known, for which surveillance of severe acute respiratory infection (SARI) is encouraged. Hospitalized SARI patients are however not always tested for influenza virus infection. Thus, to estimate the impact of influenza circulation we studied how influenza in primary care relates to intensive care unit (ICU) admissions using a modelling approach. Methods We used time‐series regression modelling to estimate a) the number of SARI admissions to ICU associated with medically attended influenza infections in primary care; b) how this varies by season; and c) the time lag between SARI and influenza time series. We analysed weekly adult ICU admissions (registry data) and adult influenza incidence (primary care surveillance data) from July 2007 through June 2016. Results Depending on the year, 0% to 12% of annual SARI admissions were associated with influenza (0‐554 in absolute numbers; population rate: 0/10 000‐0.39/10 000 inhabitants), up to 27% during influenza epidemics. The average optimal fitting lag was +1 week (SARI trend preceding influenza by 1 week), varying between seasons (−1 to +4) with most seasons showing positive lags. Conclusion Up to 12% of yearly SARI admissions to adult ICU are associated with influenza, but with large year‐to‐year variation and higher during influenza epidemics. In most years, SARI increases earlier than medically attended influenza infections in the general population. SARI surveillance could thus complement influenza‐like illness surveillance by providing an indication of the season‐specific burden of severe influenza infections and potential early warning of influenza activity and severity.


| INTRODUC TI ON
Hospital surveillance of severe acute respiratory infection (SARI) 1 is lacking or incomplete in most Western European countries. 2,3 Some countries do monitor laboratory-confirmed influenza hospitalizations or intensive care unit (ICU) admissions and report this to the European Influenza Surveillance Network (EISN). 4 However, this underestimates severe influenza burden as not all hospital patients admitted with respiratory infections undergo laboratory testing for influenza. 5 Additionally, denominator data on the number of patients with symptoms of infectious respiratory illness are generally lacking. 3 The Netherlands, like in most other European countries, has a robust surveillance system for influenza infections in primary care, providing information on timing and duration of the seasonal epidemic. 6 However, the number of serious complications requiring hospitalization is not available through this system. 5 In primary care (and other outpatient settings), influenza epidemics are heterogeneous from season to season. 7,8 This is also reflected by SARI admissions to ICU, 9 with some seasons showing high peak incidence, while other seasons show lower peaks but sometimes higher cumulative incidence over the season, with or without high ICU mortality. 10 However, these extremes in ICU do not always coincide with high burden in primary care. 10 The ratio of SARI in ICU to influenza-like illness (ILI) in primary care is one of the influenza severity parameters proposed by the World Health Organization (WHO). It expresses the number of SARI ICU admissions per observed ILI patient in primary care. 2 But, while ILI is the gold standard for estimating influenza activity in the general population, SARI might be less specific for influenza circulation as it could include a higher background level of respiratory disease by other infections and causes. 11 Thus, to gain better insight into the timing and proportion of SARI ICU admissions that are associated with influenza circulation we used a regression modelling approach. 12,13 Understanding this association will further elucidate the potential of ICU data for strengthening influenza surveillance.

| ME THODS
A long-running robust ILI surveillance system 6

| Intensive care data
Hospital data on weekly admissions to the ICU were retrieved from the National Intensive Care Evaluation (NICE) registry, originally set up for monitoring quality of ICU care. 14 As paediatric ICUs are not included in the registry, the study focuses on the adult population. A SARI admission to ICU was defined as a patient meeting all three of the following criteria: (a) the patient was admitted to the hospital less than two days before ICU admission, (b) the ICU admission was not a readmission to the ICU within the hospitalized period, and (c) the APACHE IV 15

| Influenza-like Illness data
Medically attended ILI incidence data were retrieved from NIVEL Primary Care Database-sentinel general practitioner (GP) practices.
This system covers approximately 0.8% of the Dutch population and is nationally representative for age, sex, regional distribution and population density. 6 Participating GPs report weekly the number and age of ILI patients. The number of patients registered in their practice was used as a denominator for ILI incidence calculation. To confirm influenza circulation, a subset of ILI patients is systematically swabbed for laboratory testing. We calculated influenza circulation as follows: ILI incidence * the proportion of swabs positive for influenza virus. Influenza epidemics are defined within this ILI surveillance as the weeks with ILI incidence exceeding 5.1/10 000 persons for minimally two consecutive weeks.

| Statistical analyses
We used a binomial regression model to associate the number of weekly SARI in adult ICU with the weekly influenza incidence in primary care. As the influenza surveillance data contained pre-defined age groups, we selected the influenza incidence in the 15+ age group as this was the only available age cut-off for child to adult. The number of SARI admissions N SARI w in week w (w = 1, …, 470) was used as outcome We repeated this influenza lag selection for each season and per season selected the influenza lag that showed the best fit (lowest AIC).
All analyses were performed using the statistical package R (version 3.4.0). Model selection was performed in this manner, as R would not run all the possible different model fits at once as this produced too many combinations.
We tested both positive and negative lags between influenza and SARI as the direction of this association is still poorly understood, with ICU admissions possibly being earlier. 16 Influenza circulation may give birth to two distinct populations: vulnerable or fragile persons exposed to influenza in the community may come down with severe illness more quickly than generally healthy persons who may develop ILI symptoms more slowly and/or wait before seeing a GP.
By multiplying the ILI regression coefficients with the observed weekly influenza incidence (lagged according to the season-specific lags), we calculated the influenza-associated proportions of SARI (per week). Further multiplying these weekly proportions by the weekly number of medical ICU admissions produced the estimated absolute numbers of weekly SARI associated with influenza. We then cumulated these weekly SARI numbers by season-year. As the number of ICUs participating in the NICE registry increased over time these absolute numbers were not directly comparable between seasons.
Therefore, we chose 2015 as the index year (there was near-complete national coverage of adult ICUs in NICE) and standardized all estimated numbers to the volume of medical ICU admissions observed in 2015. Year Weekly SARI (%) W eekly influenza incidence Weekly ILI incidence

| The modelled association between influenza circulation and SARI admissions
Observed and modelled SARI numbers are shown in Figure 3.

| Estimated numbers of SARI associated with influenza
On average, 7% of yearly SARI was associated with influenza but with large variations: 0%-12% of SARI was estimated to be influenza-associated depending on the season. The highest pro-  Year SARI incidence rates per season-year: 0/10 000 -0.39/10 000 (Table 2). Overall raw SARI incidence rates were eight-to 78-fold higher (at 2.7 -3.4/10 000) than the estimated influenza-associated SARI rate (Table 2)

| Time lag between SARI and influenza or ILI trends
The overall best fitting influenza lag was on average +1 week for the total study period (ie SARI preceding influenza in the general community by one week showed the best fit). However, the optimal lag varied largely from season to season from −1 to +4 weeks, almost always with positive lags (influenza lagging behind SARI) ( Table 1) (Table 3). 17,22,23 As influenza vaccination uptake by risk groups (with comorbidities and/or the 60 + age group)

| D ISCUSS I ON
has progressively decreased during the study period (from 74%

ACK N OWLED G EM ENTS
We thank Eric van der Zwan for preparing the aggregated NICE data set and for support with data management, and Jeroen Alblas for support with data management. This study was financed from the budget of the RIVM, made available by the Ministry of Health, Welfare and Sport, project number V/150044/19/SS.