• Open Access

Hospitalization of influenza-like illness patients recommended by general practitioners in France between 1997 and 2010

Authors


Dr. Camille Pelat, Department of Infectious Disease Epidemiology, Imperial College, St. Mary’s Campus, Norfolk Place, Paddington, London W2 1PG, UK. E-mail: camille.pelat@gmail.com

Abstract

Please cite this paper as: Pelat et al. (2012) Hospitalization of influenza-like illness patients recommended by general practitioners in France between 1997 and 2010. Influenza and Other Respiratory Viruses DOI: 10.1111/j.1750-2659.2012.00356.x.

Background  The case–hospitalization ratio (CHR) is a key quantity for the management of emerging pathogens such as pandemic influenza. Yet, few running surveillance systems prospectively monitor the CHR during influenza epidemics. Here, we analyze the proportion of recommended hospitalizations (PRH) among influenza-like illness (ILI) patients attended in general practice in France and compare the PRH observed during the 2009–2010 A(H1N1) pandemic with the one of the twelve previous seasons.

Methods  ILI cases were recorded by general practitioners (GPs) involved in surveillance, who indicated for each case whether they recommended hospitalization. We stratify the analysis by age, sex, and viral subtype. We investigate the reasons why GPs recommended hospitalization and the presence of risk factors for pandemic A(H1N1) complications.

Results  The average PRH over the seasons 1997–1998 to 2008–2009 was 3·4‰ (3–3·9). It was three times higher during the 2009–2010 pandemic than during seasonal influenza epidemics (OR = 2·89, 95% CI: 2·28–3·64). The highest increase was among 20–39-year-old women: OR = 11·8 (5·04–29·59). Overall, the principal reasons for recommending hospitalization were “respiratory problems” and “bad general condition.” However, during the pandemic, “age” (mainly associated with infants), “pregnancy,” and “diagnostic” became more frequent than before (P < 0·001). Finally, pregnancy was the reported risk factor for pandemic A(H1N1) complications that had the largest impact on hospitalization recommendation during the pandemic (OR = 38·62, P < 0·001).

Conclusion  Easily implemented in surveillance systems, this protocol has the potential to reveal changes in hospitalization recommendation by GPs. Moreover, if the right data are collected alongside, it could give timely insights into epidemic severity.

Introduction

Accurate and timely assessment of severity through measures such as the case–hospitalization ratio (CHR) is essential for the management of emerging pathogens, such as pandemic influenza.1 The CHR was carefully analyzed to understand the severity of the 2009 A(H1N1) pandemic influenza and to measure its burden on the healthcare systems.2,3 It was also a key quantity entering the calculation of the case–fatality ratio.4 It has been discussed elsewhere how consistent monitoring of influenza CHRs over the long term can help characterize at-risk populations for different viral subtypes, measure the impact of prevention programs, and inform future strategies.5,6

The estimation of the CHR usually relies on outbreak investigation studies of limited size or on the ratio of hospitalized cases provided by one surveillance system over the number of influenza cases estimated by another one.4,7,8 There are few surveillance systems specifically aimed at monitoring this quantity. Since 1997, the proportion of recommended hospitalizations (PRH) among patients attended for influenza-like illness (ILI) in general practice has been monitored in France by one “self-sufficient” surveillance system based on sentinel general practitioners (GPs): the Sentinelles network.9 To achieve this goal, Sentinelles GPs have been recording those ILI cases for which they recommended hospitalization as part of their case reporting.

In the present study, we compare the PRH of the 2009–2010 A(H1N1) pandemic with the twelve preceding seasonal epidemics, stratifying by age, sex, and viral subtype. We also investigate the reasons why GPs recommended hospitalization and the presence of risk factors for pandemic A(H1N1) complications. We finally evaluate the usefulness of this surveillance for public health information by monitoring the precision of the PRH estimate throughout the pandemic.

Materials and methods

ILI surveillance by the Sentinelles network

The Sentinelles network is an epidemiological, real-time, electronic surveillance system, based on voluntary sentinel general practitioners, that has been monitoring the incidence of ILI in general practice since 1984 in France.10Sentinelles GPs are similar to other French GPs with regard to their regional distribution, the proportion of GPs in rural practice, and the type of practice.11 They use the following case definition for ILI: sudden onset of fever (39°C or above) with myalgia and respiratory signs.12 Since January 1997, for each reported ILI case, Sentinelles GPs record if they recommended hospitalization for the patient.

We defined influenza epidemic periods using a periodic regression model fitted on historic non-epidemic data.13,14 This model provides an epidemic threshold that allows the the detection of influenza epidemics. Each influenza season is further characterized by its dominant circulating viral subtype(s), thanks to virological surveillance data:15 A(H1N1), A(H3N2), B, or a combination. One subtype qualifies as dominant if it comprises more than 75% of the season’s influenza isolates. Co-dominance of two subtypes could happen when neither of them accounted for more than 75% of a season’s isolates.

Estimation of the PRH

We estimate the PRH among ILI cases attended in general practice in a given period j (PRHj) as the proportion of reported ILI patients for whom Sentinelles GPs recommended hospitalization.

image

PRHs were computed for each influenza epidemic from 1997–1998 to 2009–2010 by gender and for the following age groups: 0–19, 20–39, and ≥40 years. The GPs could indicate the reason why they recommended hospitalization by selecting from the following list: “respiratory,”“cardiac,” and “other reason.” They could also elaborate in writing any other reason(s). Since October 2009, the answer list was discarded: all answers had to be elaborated in writing. We compare the frequency of each reason between seasonal and pandemic influenza using Fisher’s exact test.

Surveillance during 2009–2010 pandemic

Since October 2009, GPs were asked whether their patients presented one or more risk factors for pandemic influenza complications, to be chosen from the following list: “pregnancy,”“obesity,”“chronic condition,” and “other.” Here, we evaluate with logistic regressions and Wald tests the impact of risk factors for complications on hospitalization odds during the pandemic.

We assess whether the pandemic context made patients consult more by asking, since October 2009, the following question to Sentinelles ILI patients: “Would you consult your GP for the same symptoms if there was not the current pandemic context?” To assess potential changes in the way GPs attended ILI cases, a telephonic survey was conducted by a medical resident on 41 GPs, with the following question: “Would you recommend hospitalization for this patient, for the same symptoms, if there was not the current pandemic context?” On this occasion, GPs were also asked what happened to the patients for whom they recommended hospitalization.

Finally, we use the pandemic to exemplify the real-time estimation of the PRH by the Sentinelles surveillance system. In 2009 in France, the first signal of unusual ILI activity was detected in week 28 (July),16 and the epidemic officially began in week 37 (September). However, pandemic A(H1N1) is believed to have spread in France since week 19 (May).17 To assess how the PRH estimate evolved from May to December, successive weekly estimations were undertaken. For each week from 2009/19 to 52, estimations relied on the cumulative ILI cases reported by then from week 19. We define relative uncertainty (in %) as half the confidence interval divided by the estimate and multiplied by 100. The analysis was carried out with the R software package version 2.13.0 (R Foundation for Statistical Computing, Vienna, Austria). Confidence intervals for means, proportions, and odds ratio were obtained with the profile likelihood method.18 A Bayesian adjustment was made when the number of hospitalizations was zero to move the log-likelihood function away from infinity.19

We finish the analysis by studying two potential factors in the PRH variability: the composition of the Sentinelles GPs and the intensity of influenza seasons, as measured by the cumulative consultation rate for ILI in general practice over each epidemic.

Composition of the Sentinelles network across the study period

We analyze the number, age, sex ratio, and turnover of the GPs who have been reporting to the Sentinelles system between 1997 and 2010. A GP is reporting in a given year Y when he/she reported at least once to the system during year Y. The turnover in year Y is measured as the proportion of reporting GPs that joined the Sentinelles network in year Y among all Sentinelles GPs who have been reporting in year Y. We monitor for each year the number of GPs present from the start of the study period (i.e., who joined in 1997 or earlier) as a measure of “stability” of the network.

We also assess the number of consultations and visits per year per reporting GP and the age distribution of patients, thanks to a subset of GPs who provided their activity report form (RIAP). RIAPs are given to GPs by the French national health insurance, usually annually or quarterly, and are based on reimbursement records. In these forms, the number of consultations and visits is given for the following age groups: 0–15-year-olds, 16–59-year-olds, 60–69-year-olds, and 70 years and older. As we did not have a RIAP per year per reporting GP, the closest in time to each study year (within a limit of 3 years) was chosen.

Relationship between the PRH and ILI consultation rates

We assess the relationship between the PRH and the cumulative rate of ILI consultations by season per 100 000 inhabitants using Pearson’s linear correlation coefficient (ρ) and linear regression models. We repeat this analysis for each age group. To obtain ILI consultation rates overall and by age group, the weekly number of consultations reported by Sentinelles GPs is extrapolated to all French GPs and standardized on the current French population using a procedure in use since 1984.10 It ensures that estimated incidences are independent of the number of reporting GPs. The 2009–2010 season is excluded from the analysis, as it is an outlier both by its magnitude and by its PRH.

Results

Over the thirteen epidemics analyzed herein, GPs reported 76 059 ILI cases and recommended hospitalization for 326 of them. Detailed numbers by season, age group, and gender are provided in Table 1. The largest PRH among ILI patients occurred during the 2009–2010 pandemic period (Figure 1). It was then equal to 9·7‰ (95% confidence interval: 8–11·7), when the average PRH over the seasons 1997–1998 to 2008–2009 was 3·4‰ (3–3·9). This increase was statistically significant, with an odds ratio (OR) equal to 2·89 (95% CI: 2·28–3·64, chi-square test for equal distributions: P < 0·001).

Table 1.   Proportion of recommended hospitalizations among ILI patients (‰) [95% confidence interval], number of recommended hospitalizations (number of ILI patients), as reported by Sentinelles GPs for each influenza season.
SeasonSubtype
OverallBy age group (years)*By gender*
0–1920–39≥40MaleFemale
  1. *Some patients’ age and/or sex might be unknown.

1997–1998A(H3N2)5·3 [3·8–7·3] 36 (6738)2 [0·7–4·3] 5 (2508)1·1 [0·2–3·3] 2 (1853)12·2 [8·3–17·2] 29 (2377)5·2 [3·1–8·1] 17 (3261)5·5 [3·4–8·3] 19 (3477)
1998–1999A(H3N2)3·5 [2·3–5·1] 25 (7136)2·2 [0·8–4·6] 5 (2314)1·9 [0·6–4·4] 4 (2111)5·9 [3·5–9·3] 16 (2711)3·8 [2·1–6·2] 13 (3457)3·3 [1·7–5·5] 12 (3679)
1999–2000A(H3N2)5·3 [3·7–7·2] 35 (6635)1·2 [0·2–3·7] 2 (1677)1·5 [0·4–3·9] 3 (2006)10·2 [7–14·2] 30 (2947)4·8 [2·8–7·5] 16 (3345)5·8 [3·6–8·8] 19 (3290)
2000–2001A(H1N1)1·2 [0·3–3] 3 (2605)0 [0–2·5] 0 (1354)0 [0–4·6] 0 (725)5·7 [1·4–14·8] 3 (525)0·8 [0·1–3·3] 1 (1317)1·6 [0·3–4·8] 2 (1288)
2001–2002A(H3N2)3·7 [2·1–6] 14 (3769)2 [0·5–5·2] 3 (1489)0·9 [0·1–3·9] 1 (1130)8 [3·8–14·3] 9 (1132)3·3 [1·3–6·6] 6 (1835)4·1 [1·9–7·7] 8 (1934)
2002–2003B2·8 [1·2–5·5] 7 (2473)1·5 [0·3–4·6] 2 (1332)1·7 [0·2–7·2] 1 (606)7·6 [2·4–17·6] 4 (525)2·3 [0·6–6·1] 3 (1278)3·4 [1·1–7·9] 4 (1173)
2003–2004A(H3N2)3·6 [2·3–5·3] 22 (6130)2·8 [1·3–5] 9 (3219)1·2 [0·2–3·7] 2 (1648)8·8 [4·6–15·1] 11 (1247)3·9 [2·1–6·5] 12 (3071)3·3 [1·7–5·8] 10 (3004)
2004–2005A(H3N2)3·2 [2–4·8] 21 (6578)0·8 [0·1–2·6] 2 (2396)0 [0–1·8] 0 (1878)8·3 [5·1–12·6] 19 (2293)4 [2·2–6·6] 13 (3262)2·2 [0·9–4·2] 7 (3253)
2005–2006A(H1N1) & B1·5 [0·6–3·1] 6 (3886)1·3 [0·3–3·3] 3 (2326)2·3 [0·4–7·1] 2 (867)1·5 [0·1–6·4] 1 (686)2·1 [0·6–4·8] 4 (1942)1 [0·2–3·2] 2 (1911)
2006–2007A(H3N2)6·6 [4·6–9·1] 33 (4979)4·7 [2·4–8·3] 10 (2122)2·1 [0·5–5·4] 3 (1446)14·4 [9–21·6] 20 (1388)6·7 [3·9–10·5] 16 (2390)6·8 [4–10·5] 17 (2510)
2007–2008A(H1N1) & B1 [0·4–2] 6 (5951)0·8 [0·1–2·5] 2 (2434)0 [0–1·9] 0 (1798)2·4 [0·7–5·5] 4 (1690)1 [0·3–2·7] 3 (2902)1 [0·3–2·6] 3 (2980)
2008–2009A(H3N2)1·5 [0·9–2·6] 13 (8388)0·3 [0–1·3] 1 (3338)1·2 [0·3–3·2] 3 (2453)3·5 [1·7–6·2] 9 (2597)1·7 [0·8–3·4] 7 (4001)1·2 [0·4–2·5] 5 (4282)
2009–2010A(H1N1)9·7 [8–11·7] 105 (10791)10 [7·8–12·7] 63 (6279)7·8 [4·9–11·6] 21 (2685)11·5 [7·3–17·2] 21 (1820)7·6 [5·5–10·2] 40 (5241)11·7 [9–14·8] 63 (5402)
Figure 1.

 Proportion of recommended (PRH) hospitalizations among ILI patients (‰) by season (panels A and B) or by influenza viral subtype (panels C and D). Dashed grey lines: overall PRH; boxes: overall 95% confidence intervals; colored dots: PRH by age or sex (with associated 95% confidence interval).

PRH by viral subtype

The PRH varied substantially according to the dominant circulating influenza subtype(s) (logistic regression likelihood-ratio (LR) test: P < 0·001, for detailed PRH by subtype see Table 2). Compared with the only seasonal A(H1N1) epidemic (2000–2001), the odds of hospitalization were significantly higher when A(H3N2) circulated: OR = 3·44 (1·31–13·92, Wald z-statistic P = 0·034). Co-circulation of A(H1N1) and B or dominance of B viruses did not significantly increase the odds of hospitalization, respectively, OR = 1·06 (0·34–4·65, P = 0·929) and 2·46 (0·68–11·43, P = 0·192). The 2009 A(H1N1) pandemic was associated with hospitalization odds increased by 8·52-fold compared with seasonal A(H1N1) (3·21–34·65, P < 0·001).

Table 2.   Proportion of recommended hospitalizations among ILI patients (‰) [95% confidence interval] reported by Sentinelles GPs during influenza seasons 1997-1998 to 2009-2010, aggregated by dominant influenza viral subtype(s).
Dominant subtype(s)
OverallBy age group (years)*By gender*
0–1920–39≥40MaleFemale
  1. *Some patients’ age and/or sex might be unknown

A(H1N1)1·2 [0·3–3]0 [0–2·5]0 [0–4·6]5·7 [1·4–14·8]0·8 [0·1–3·3]1·6 [0·3–4·8]
A(H1N1) & B1·2 [0·7–2]1·1 [0·4–2·3]0·8 [0·1–2·3]2·1 [0·8–4·5]1·4 [0·6–2·8]1 [0·4–2·2]
A(H1N1) 20099·7 [8–11·7]10 [7·8–12·7]7·8 [4·9–11·6]11·5 [7·3–17·2]7·6 [5·5–10·2]11·7 [9–14·8]
A(H3N2)4 [3·4–4·5]1·9 [1·4–2·6]1·2 [0·8–1·9]8·6 [7·2–10]4·1 [3·3–4·9]3·8 [3·1–4·6]
B2·8 [1·2–5·5]1·5 [0·3–4·6]1·7 [0·1–7·2]7·6 [2·4–17·6]2·3 [0·6–6·1]3·4 [1·1–7·9]
All subtypes except A(H1N1) 20093·4 [3–3·9]1·7 [1·2–2·2]1·1 [0·7–1·7]7·7 [6·6–9]3·5 [2·9–4·1]3·3 [2·7–4]

PRH by age group

The PRH was the highest among patients above 40 years for every viral dominant subtype(s), yet to different extents (Table 2). During the A(H3N2) epidemics, the hospitalization odds of ≥40-year-old patients were 4·44-fold (3·13–6·45) those of 0–19-year-old patients (Wald z-statistic P < 0·001), and 6·96-fold (4·39–11·77) those of 20–39-year-old patients (P < 0·001). During the B and B-A(H1N1) epidemics, the difference between age groups was not statistically significant (respective LR test: P = 0·12 and P = 0·38). During the only seasonal A(H1N1) epidemic, few ILI cases were reported, and GPs recommended hospitalization only for 40 years or older patients, with a PRH of 5·7‰ (1·4–14·8). During the 2009–2010 pandemic, the differences in PRH between age groups flattened: the hospitalization odds of the ≥40-year-olds were 1·15-fold the ones of 0–19-year-olds (0·69–1·86, P = 0·58) and 1·47-fold the ones of the 20–39-year-olds (0·80–2·73, P = 0·21). The PRH of ILI patients over 40 years was not statistically higher during the 2009–2010 pandemic than in seasonal epidemics: OR = 1·5 (0·92–2·32), P = 0·081. Two A(H3N2) epidemics, 2006–2007 and 1997–1998, presented higher PRHs for this age group than the pandemic (Table 1). On the contrary, the PRHs of 0–19- and 20–39-year-old patients were higher during the pandemic than in any previous season, with OR = 6·10 (4·16–9·02, P < 0·001) for the 0–19-year-olds and OR = 6·94 (3·77–12·79, P < 0·001) for the 20–39-year-olds.

Furthermore, the relative ILI attack rate of the 0–19-year-olds was more important during the pandemic: 58% of reported ILI cases belonged to this age group versus 40% during seasonal epidemics, whereas the contrary was observed among ≥40-year-olds: 17% versus 31%. The relative attack rate in intermediate ages remained stable: 25% versus 29%. The hypothesis of equal age distributions in the two periods was significantly rejected (chi-square test P < 0·001). Therefore, this unusually high relative attack rate in the 0–19-year-olds combined with the unusually high PRHs in the younger age groups resulted in a significantly lowered mean age of ILI patients requiring hospitalization during the 2009–2010 A(H1N1) pandemic: 20·5 years (15·1–26·0, median 11 years), when it was 55·6 years (51·9–59·4, median 67·5 years) for seasonal epidemics. This difference was statistically significant (Student’s t-test for equal means: P < 0·001).

PRH by sex

The PRHs of men and women were not statistically different during seasonal influenza epidemics: OR = 0·95 (0·73–1·24, P = 0·713), see values in Table 2. The odds of hospitalization increased by 3·57-fold in women during the pandemic (2·60–4·86, P < 0·001) and by 2·21-fold for men (1·52–3·15, P < 0·001). At this time, the PRH of women was significantly above that of men: OR = 1·53 (1·04–2·30, P = 0·035). The relative ILI attack rate for each gender did not change during the pandemic: women represented 51% of all ILI cases during the pandemic, as much as in seasonal epidemics (Pearson’s chi-squared test P = 0·66).

We further assessed the age × gender categories in which patients’ hospitalization odds increased the most during the pandemic. Women aged 20–39 years had hospitalization odds increased by 11·8-fold during the pandemic (5·04–29·59, P < 0·001). It can be noted that this first category is composed of women of childbearing age. They were followed by 0–19-year-old girls (OR = 8·65, 4·90–15·88, P < 0·001), 0–19-year-old boys (OR = 4·58, 2·68–7·86, P < 0·001), and 20–39-year-old men (OR = 3·92, 1·47–9·58, P = 0·004). While hospitalization odds of the oldest women increased slightly during the pandemic (OR = 1·89, 1·04–3·18, P = 0·025), no change affected the oldest men (OR = 0·86, 0·30–1·94, P = 0·75).

Reasons for hospitalization

There were 278 patients for whom the reason for requiring hospitalization was stated: 174 in seasonal epidemics and 104 during the pandemic (Table 3). The distribution of reasons significantly differed between the two periods (Two-sided Fisher’s exact test: P < 0·001). For seasonal epidemics, most patients (64%) were sent to hospital for respiratory reasons, followed by cardiac reasons (13·2%) and bad general condition (11·5%). During the 2009–2010 pandemic, although the main reasons remained respiratory problems (44·2%) and bad general condition (17·3%), “pregnancy,”“diagnostic,” and “age” became more frequent. The latter concerned seven infants <1 year, two children aged 1 and 11-years old, and an adult of 96 years. One of the two patients sent to hospital for an age reason during seasonal epidemics was under 1 year, and the age of the other one was unknown.

Table 3.   Number of recommended hospitalizations for ILI patients, by category of reason, and percentage over the columns
 Seasonal epidemics 1997–1998 to 2008–2009 (%)2009–2010 A(H1N1) pandemic (%)Overall (%)
Age2 (1·1)10 (9·6)12 (4·3)
Bad general condition20 (11·5)18 (17·3)38 (13·7)
Cardiac23 (13·2)5 (4·8)28 (10·1)
Confusion6 (3·4)1 (1·0)7 (2·5)
Dehydration2 (1·1)2 (1·9)4 (1·4)
Diagnostic0 (0·0)9 (8·7)9 (3·2)
Meningitis5 (2·9)3 (2·9)8 (2·9)
Pregnancy1 (0·6)6 (5·8)7 (2·5)
Respiratory111 (64·0)46 (44·2)157 (56·6)
Risk Factor2 (1·1)2 (1·9)4 (1·4)
Social2 (1·1)2 (1·9)4 (1·4)
Total174104278

Surveillance during 2009–2010 pandemic

The presence/absence of risk factors for pandemic influenza complications was described for 8878 ILI patients during the 2009–2010 season. The PRH was 5·2‰ (3·7–6·9) in the 8152 patients without risk factors and 56·5‰ (41·2–74·8) in the 725 ones presenting at least one risk factor (OR = 11·56, 7·45–17·92, P < 0·001). Specifically, pregnancy multiplied the odds of hospitalization by 38·62 (15·05–87·26, P < 0·001), coexistence of chronic condition and obesity by 20·69 (4·82–61·37, P < 0·001), chronic condition alone by 7·12 (3·79–12·65, P < 0·001), and obesity alone by 5·22 (0·29–25·00, P = 0·107) (Table 4).

Table 4.   Proportion of recommended hospitalizations (PRH) for ILI patients by risk factor during the 2009-2010 A(H1N1) pandemic.
Risk factorPRH (‰) [95% confidence interval]# recommended hospitalizations (# ILI patients)
None5·2 [3·7–6·9]42 (8152)
Pregnancy166·7 [75·3–297·5]7 (42)
Chronic condition & obesity96·8 [25·2–232·1]3 (31)
Chronic condition35·5 [20·6–56·1]15 (422)
Obesity26·3 [3·3–110·8]1 (38)
Other77·7 [45·5–121·1]15 (193)
At least one risk factor56·5 [41·2–74·8]41 (726)
Overall9·3 [7·5–11·5]83 (8878)

When asked by their GPs, 6383 of 7758 patients (82%) thought that they would have consulted for the same symptoms in a regular seasonal epidemic. This fraction was significantly larger in the subgroup of patients requiring hospitalization: 70 of 72 patients, that is, 97%, (OR = 7·44, 2·34–45·33, P = 0·005). The 41 GPs interviewed over the phone accounted for 60 of the 105 recommended hospitalizations during the pandemic. When asked their personal views, they thought that they would have recommended hospitalization for only 32 of the 60 patients (53%) during a regular epidemic. When asked about what happened to these 60 patients, they answered that 27 (45%) were admitted to hospital, 20 (33·33%) only had a consultation there and were rapidly discharged, two did not go to the hospital (3·33%), and one did not go and died at home (1·67%). What happened of the remaining 10 patients was unknown (16·67%). The weekly prospective monitoring of the PRH during the 2009–2010 season is shown in Figure 2. Estimations culminated at 25·6‰ (10·3–51·2) in week 29, then decreased to reach 11·3‰ (8–15·3) in week 44 (last week of October) and remained quite stable until week 52 (10·2‰, 8·5–12·2). Note that this final estimate differs slightly from the one in Table 1, which is based only on data reported from week 37 to 52. The relative uncertainty of the PRH estimate was above 300% in week 19. It steeply decreased to reach 32% in week 44, that is, 2 weeks before the epidemic peak, which occurred in week 46 and 18% in week 52.

Figure 2.

 Real-time weekly estimation of the proportion of ILI patients for whom GPs recommended hospitalizations (plain line) and its 95% confidence interval (dashed lines), for each week between week 2009/19 and 2009/52, based on cumulative cases from week 2009/19.

Influence of the Sentinelles network’s composition and of epidemic intensity

After decreasing from 339 in 1997 to 220 in 2002, the number of GPs reporting to the Sentinelles network augmented to reach 372 in 2010 (Table 5 and Figure 3). The mean turnover (the proportion of new reporting GPs) was 23% per year (range: 8%-36%). The proportion of GPs present from the start of the study period (1997) decreased from 100% in 1997 to 23% in 2009 and increased to 25·5% in 2010 as the total number of Sentinelles GPs decreased. The proportion of women fluctuated between 11% in 1999 and 18% in 2001, 2008, 2009, and 2010, with an overall upward trend. The median age also augmented from 45-year-olds in 1997 to 53·5-year-olds in 2010. Across the study period, the median number of consultations and visits per reporting GP per year was quite stable (from 4690 in 1997 to 4665 in 2010) (Figure 3). The median proportion of patients also remained roughly stable for all age groups, with a slight decrease (from 24% to 22%) in the 0–15-year-olds and a slight increase (from 54% to 55%) for the 16–59-year-olds.

Table 5.   Characteristics of the GPs reporting to the Sentinelles network surveillance system from 1997 to 2010
YearNumber of GPsAge (in years)SexYear joined the Sentinelles network
Min1st quartileMedianMean3rd quartileMaxUnknownWomenMenUnknown% of women1997 or earlierCurrent yearAfter 1997 and before current year
199733931404545·325068240299012339
199827532424646·23506923424011225223
199925232434747·425170428223111221247
200023631434847·9852693322031141833518
200126830424747·1752691482191181588030
200222031434847·6153762321871151096051
200323729444948·4953776342021141048350
2004228304550495478635193015965478
200524231465150·0855791731211013907577
200625530465250·74568016362190148854113
2007283284752·551·36578115362470139268123
200837929465351·3958821868311018106138135
20094522946·255351·7758831881371018104136212
2010372304753·552·3598414673050189536241
Figure 3.

 Characteristics of the GPs reporting to the Sentinelles network surveillance system from 1997 to 2010 (A: number and turnover, B: sex, C: age, D: activity). E–H: age distribution of Sentinelles GPs’ patients in four age groups. Panel D to H are based on a subset of GPs that provided an activity report form within 3 years of each study year.

Finally, no significant correlation was found between the PRH and the cumulative seasonal consultation rate for ILI in general practice: ρ = 0·32 (P-value: P = 0·30). Furthermore, this relationship did not have the same direction in all age groups and, for each of them, fell far from statistical significance (see Figure 4).

Figure 4.

 Proportion of ILI patients for whom Sentinelles GP recommended hospitalization versus the cumulative rate of ILI consultations by influenza season: (A) among the 0–19-year-olds, (B) the 20–39-year-olds, (C) the 40 years and older, (D) overall.

Discussion

Interpretation of the results

Sentinelles GPs recommended hospitalization of ILI patients three times more frequently during the 2009–2010 A(H1N1) pandemic than during seasonal influenza epidemics (OR = 2·89, 95% CI: 2·28–3·64). The highest increase was among 20–39-year-old women: OR = 11·8 (5·04–29·59). Alone, this indicator does not allow understanding what magnitude of increase was owing to severity and what was precautionary, but this increase is consistent both with 2009 A(H1N1) presenting more severely in young adults, especially if pregnant, as was reported early in the course of the pandemic20,21 and with the pandemic guidelines recommending hospital-based diagnosis for pregnant ILI patients.22

Some pitfalls arise in the interpretation of the PRH presented in this work. First, the positive predictive value of the Sentinelles definition of ILI was reported around 40% for seasonal influenza, implying that the PRH based on these clinical data could be an underestimation of the true influenza PRH.12 Second, this ILI definition has a limited sensitivity as it misses the asymptomatic cases and the symptomatic cases that do not feel bad enough to consult a GP. This leads to overestimating the influenza PRH by biasing the case ascertainment to the most severe ones. Third, it might be worthy to note once again that, herein, we monitor the proportion of recommended hospitalizations. Some of the patients might not go to the hospital, and some might only have a consultation there. In that respect, the present measure overestimates the true CHR. On the other hand, some patients for whom GPs did not recommend hospitalization could have been hospitalized afterward, for example, if their condition persisted or got worse after their consultation with the GP, making the present measurement an underestimation of the CHR.

Finally, the specific hospitalization criteria set up for pandemic A(H1N1) patients may have encouraged French GPs to send ILI patients to hospital,23 biasing our numerator, while the heavy media coverage and fear of the pandemic might have boosted the public concern and health-seeking behaviors, biasing our denominator. In particular, guidelines incited French GPs to send pregnant ILI patients to hospital consultations dedicated to viral identification of pandemic A(H1N1).22 It is thus possible that some Sentinelles GPs reported ambiguously directing a patient to hospital for consultation or hospitalization. The fact that GPs thought that they would have required hospitalization during a regular epidemic for only 53% of the patients seen during the pandemic is in accordance with this hypothesis (even if the magnitude of this bias does not explain the tripling of the PRH observed during the pandemic). Thus, the PRH observed during the 2009–2010 pandemic does not readily reflect severity but, instead, the increased burden on the healthcare system.

Enhancing surveillance for a better interpretation of the PRH

To disentangle severity from patients’ behavior and GP’s practice inside the PRH and allow comparison between seasons, it is necessary to collect data both on the GP’s tendency to recommend hospitalization (numerator) and on the patients’ tendency to seek heath care (denominator). Questions in that respect could be asked to GPs and to their patients in each electronic case report. During the pandemic, we used a telephonic survey to assess the former and included the latter in the online questionnaire. The formulation and the suggested answer list should be allowed to change dynamically with the context (seasonal or pandemic in particular) while ensuring that the right data for inter-annual comparability are collected.

To infer from the PRH the true hospitalization ratio of ILI patients seen by general practitioners, and thus the true burden of the disease on hospitals, it would be also necessary to know (i) how many patients for whom hospitalization was recommended went to hospital and how many of them were admitted and (ii) how many ILI patients were hospitalized without GP recommendation. For question (i), it could be thought of GPs logging back into the system to complete a previous report. However, because reported cases are currently anonym in the system, this would imply rethinking the collection design. An annual cross-sectional survey among GPs’ ILI patients could instead be contemplated to collect information on both points (i) and (ii). To this respect, a pilot survey was launched during the 2009 pandemic to shed light on the future of 60 patients for whom GPs recommended hospitalization. GPs were contacted by phone in the days following their declaration. They knew whether or not their patients went to the hospital for 50 patients (83·33%). Among those, 27 were admitted, 20 had a consultation, and only three did not go to the hospital. This pilot protocol has an encouraging response rate and should be considered for yearly implementation.

Finally, to assess the case–hospitalization ratio of confirmed influenza patients seen by general practitioners, both the PRH and the necessary side data described earlier should be collected on virologically confirmed patients. The Sentinelles network only implements virological surveillance in Corsica on a small number of patients, so currently these developments are not contemplated.

Influence of the Sentinelles network’s composition and of epidemic intensity

We assessed whether the composition of the Sentinelles network changed over the whole study period. We found that the average age of reporting GPs slightly and smoothly increased over the whole study period. The proportion of women also slightly increased, although with more fluctuations. Both increases are also observed in the French medical population as a whole.24,25 No noticeable trend or seasonal fluctuations in the number of consultations and visits or in the age distribution of patients was evidenced. Thus, it seems unlikely to us that the composition of the Sentinelles GPs explains the observed variations in the PRH, in particular the 2009–2010 increase. We also found no significant linear correlation between the PRH and the cumulative consultation rate for ILI in general practice, for all age groups and overall. Hence, the intensity of influenza seasons does not seem to influence the probability that GPs recommend hospitalization for ILI patients.

Conclusion

To sum up, the surveillance protocol presented herein does not provide immediately interpretable CHRs unlike ad hoc surveys based on virologically or serologically confirmed cases. Nevertheless, with the right data collected alongside to help with interpretation and assess inter-annual comparability, this system can help measuring the severity of epidemics on a scale defined by the GPs’ concern about their patients’ health. Furthermore, nested in a routine surveillance system, this protocol allows the continuous collection of baseline data on influenza-like illness severity. Besides, by linking hospitalizations to ILI cases at the individual level, the data presented here allows direct estimation of the probability of hospitalization among cases, and there is no delay between GPs’ and hospitals notifications to be corrected for. Finally, the Sentinelles network was able to provide a PRH estimate with a relative uncertainty of ± 32% 2 weeks before the peak of the 2009–2010 pandemic, and with a relative uncertainty of ± 18% at the end of the pandemic. In conclusion, if data collection can be enhanced to help disentangling severity from behaviors and practices, this protocol is a promising tool for generating timely insights into the severity of influenza epidemics and their burden on healthcare systems.

Acknowledgements

We would like to thank all the general practitioners of the Sentinelles network who participated in the surveillance. This work was supported by the French medical research institute (Institut National de la Santé et de la Recherche Médicale) and by the University Pierre et Marie Curie – Paris 6. We also thank Cécile Souty for her analyses of the Sentinelles network composition.

Conflict of interest

The authors declare that they have no conflict of interest.

Ancillary