The neuraminidase inhibitors, oral oseltamivir and inhaled zanamivir, were the predominant medical countermeasure available from emergence of the influenza A H1N1pdm09 virus in early 2009, until the first release of monovalent H1N1 vaccines in October, 2009. Prescribing data from seven countries (Australia, Canada, France, Germany, Japan, UK, USA) suggest at least 18·3 million individuals received oseltamivir between May 1, 2009, and Dec 31, 2009.¹ Country-specific policies for use of neuraminidase inhibitors during the 2009–10 pandemic varied from no use, to targeted use in at-risk patients (most countries), to treatment of all patients with clinical illness (UK). Most use of neuraminidase inhibitors worldwide was in the form of oseltamivir—eg, 97·5% of neuraminidase inhibitors used in the USA.²

There is little prepandemic evidence pertaining to the effectiveness of neuraminidase inhibitors in reducing mortality in patients admitted to hospital with influenza; most evidence comes from observational studies of treatment of seasonal influenza, often in highly specific groups of patients.^3-9 Thus, in 2009–10, neuraminidase inhibitors were used on the basis of rational deduction that they would reduce mortality due to influenza A H1N1pdm09 virus infection rather than on strong pre-existing evidence, although data from treatment of human influenza A H5N1 cases suggested this reduction in mortality might be possible.^10,11 Japanese clinicians used neuraminidase inhibitors widely to treat all people presenting with clinical influenza in 2009–10 and recorded the lowest pandemic mortality rate of any developed country.^12-14 Although a similar treat-all policy existed in the UK in 2009, uptake of neuraminidase inhibitors in patients admitted to hospital with influenza A H1N1pdm09 was low.¹⁵

Two systematic reviews and meta-analyses have examined the effectiveness of neuraminidase inhibitors in reducing mortality due to influenza. Both suggest substantial reductions in mortality by two-thirds to three-quarters compared with no treatment.^16,17 However, limitations are apparent, such as the heterogeneity of studies included and inadequate adjustment for potential confounding. Importantly, neither was able to adjust for the likelihood of a patient receiving antiviral treatment (propensity)—a crucial consideration when antiviral drugs might have been prioritised towards the sickest patients—and neither was able to use a pooled analysis approach with individual participant data.¹⁸

We received replies from 128 (32%) of 401 centres contacted; of these 77 (60%) confirmed willingness to participate and the remainder declined (36 [28%] had no data; three [2%] agreed initially but later withdrew because of lack of capacity for data extraction, institutional review board restrictions preventing sharing of individual participant data, or failure to obtain government approval for data sharing; 12 [9%] had agreed in principle, but were unable to share data within project timescales). No data were requested from nor provided by pharmaceutical companies. After exclusion of duplicate responses (same source dataset), and addition of three further datasets provided through informal contact with domain experts, 80 research groups from 38 countries in six WHO regions contributed anonymised data for 168 117 patients, of whom 24416 had laboratory results indicative of noninfluenza A H1N1 disease. Among the remaining 143 701 laboratory confirmed or clinically diagnosed (without standard study-wide case definition) influenza A H1N1pdm09 cases, 106 138 were outpatients and 2593 had missing information for hospital admission. The remaining 34 970 inpatients were eligible for inclusion (figure 1).

Of the 34 970 inpatients eligible for inclusion, 2095 (6%) had missing information for mortality status, and 3584 (10%) for exposure to neuraminidase inhibitors; 57 (<1%) were unsuitable for inclusion for other reasons (figure 1). Ultimately, we included 29 234 records from 78 studies (two studies provided only outpatient data and were excluded from analysis) of patients admitted to hospital between Jan 2, 2009, and March 14, 2011: 25 001 (86%) laboratory confirmed; 9218 (32%) children; and 1600 (5%) aged 65 years or older. Appendix p 16 show the incidence of cases by month. Full characteristics of the pooled dataset are listed in table 1 with absolute risks of mortality for various exposure categories and subgroups summarised in appendix p 16. Baseline characteristics of each constituent dataset are presented in appendix pp 17–21.

Patients without neuraminidase inhibitor treatment data and therefore excluded from analysis were more likely to be older, to have presented to hospital later, less likely to have a laboratory confirmed diagnosis, and more likely to be treated with antibiotics than were patients included in the analysis (appendix pp 26–27). However, they were less likely to be smokers, obese, or to have an underlying comorbidity. Additionally, their hospital stay was shorter, and they were less likely to have severe outcomes (admission to critical care unit or death), or influenza-related pneumonia (appendix pp 26–27).

After adjustment for propensity score and corticosteroid and antibiotic treatment, the likelihood of mortality in patients treated with a neuraminidase inhibitor was 0·81 (95% CI 0·70–0·93), compared with no treatment (table 2). The OR did not change substantially when only laboratory confirmed cases were included (adjusted OR 0·82 [95% CI 0·70–0·95]). Similarly, we identified significant associations with a reduced mortality risk in adults, pregnant women, and critically ill adult patients (table 2). However, there was no significant association between neuraminidase inhibitor treatment and mortality in children aged 0–15 years (table 2). Post-hoc analyses restricted to children up to 1 year of age and up to 5 years of age did not change this finding (appendix p 27).

Early neuraminidase inhibitor treatment compared with later treatment initiation was associated with an overall significant reduction in mortality risk (adjusted OR 0·48 [95% CI 0·41–0·56]; table 3). The ORs remained essentially unchanged when only laboratory confirmed cases were considered, but risk reduction was higher in pregnant women (table 3). Notably, there was again no significant association between early treatment and mortality in children after adjustment (table 3).

Neuraminidase inhibitor treatment within 2 days of symptom onset compared with none was also associated with a significant reduction in mortality in all patients (adjusted OR 0·50 [95% CI 0·37–0·67]; table 3), with significant risk reductions also noted among laboratory confirmed cases, adults, pregnant women, and adult patients admitted to critical care (table 3). However, there was no significant association with a lower mortality risk in children aged 0–15 years (table 3).

With regard to neuraminidase inhibitor treatment started more than 2 days after symptom onset compared with none, we identified no significant association with mortality in all patients (adjusted OR 1·20 [95% CI 0·93–1·54]), nor in laboratory confirmed cases, adults, pregnant women, or children (table 4). However, we noted an associated mortality risk reduction of about a third (adjusted OR 0·65 [95% CI 0·46–0·93]) in adult patients admitted to critical care.

Information about exact timing of neuraminidase inhibitor treatment from symptom onset was available for 65% (12284 of 18803) of those receiving such treatment. After taking into account clustering by study, propensity score quintiles, and inhospital treatment with antibiotics or systemic corticosteroids, when antiviral use was modelled as a time-dependent covariate to overcome potential immortal time bias (ie, survivor bias), neuraminidase inhibitor treatment was significantly associated with decreased hazard rate of mortality over a 30-day follow-up period (adjusted HR 0·51 [95% CI 0·45–0·58], p<0·0001) as compared with no antiviral treatment. When only treated cases were included, there was an increase in the hazard with each day’s delay in initiation of treatment up to day 5 as compared with treatment initiated within 2 days of symptom onset (adjusted HR 1·23 [95% CI 1·18–1·28], p<0·0001 for the increasing HR with each day’s delay). The unadjusted and adjusted survival curves comparing survival by time to treatment initiation are shown in figure 2 and appendix pp 28–29.

Our results show that neuraminidase inhibitor treatment was associated with reduced mortality in adult patients admitted to hospital with influenza A H1N1pdm09 virus infection. Neuraminidase inhibitor treatment of influenza A H1N1pdm09 at any stage of illness compared with none revealed an associated reduction in the likelihood of mortality (table 2). We identified an associated likelihood of lower mortality when comparing early versus later initiation of treatment and when comparing early treatment with none (table 3, panel). Although we included 4233 patients (14%) without laboratory confirmed influenza A H1N1pdm09, restriction to laboratory-confirmed cases produced near identical estimates, suggesting that the data are not confounded by misclassification bias attributable to other causes (tables 2, 3). Additionally, we noted much the same findings in adults, pregnant women, and adult patients needing admission to critical care. The finding regarding critical care suggests that neuraminidase inhibitors were associated with mortality reduction across the spectrum of severity in adult patients admitted to hospital with influenza A H1N1pdm09. These findings accord closely with previous studies^16,17 but have increased precision and reduced estimates of effectiveness consistent with more complete adjustment for confounders and treatment propensity. They are also consistent with ecological data.^23-25

We were consistently unable to show any association of neuraminidase inhibitor treatment with mortality reduction in children. Possible explanations include lower case fatality proportion in paediatric patients (thus reduced statistical power),^26,27 higher influenza A H1N1pdm09 viral load in children²⁸ than adults leading to reduced drug effectiveness, suboptimum dosing in very young children,²⁹ secondary bacterial infections (eg, meticillin-resistant Staphylococcus aureus), confounding by indication³⁰ (children might have been more likely to have had antivirals prescribed if they had more severe disease or if they failed to respond to other treatments), or a combination of these factors. Since it has been suggested that younger children might be admitted with milder disease compared with older children and adults (precautionary physician behaviour), that the pharmacokinetics and pharmacology of oseltamivir might be different in very young children,²⁹ and that influenza pathogenesis might differ by age,³⁰ we did post-hoc sensitivity analyses separately in children up to 1 year of age and up to 5 years of age, but our findings did not change (appendix p 27). However, we note that these results contrast with those of Louie and colleagues,³² who recently showed a two-thirds reduction in mortality among children treated with neuraminidase inhibitors admitted to hospital with influenza (OR 0·36 [95% CI 0·16–0·83]).

The finding that no treatment was better than late treatment is probably explained by confounding due to illness severity at the point of treatment initiation (ie, confounding by indication). Untreated patients probably had milder disease and patients treated later in the course of their illness might have had delays in hospital admission, delays in diagnosis after admission, or delays in being considered for neuraminidase inhibitor treatment (treatment only started once their condition deteriorated), or combinations of these factors. We advocate early consideration of a diagnosis of influenza in patients admitted to hospital with respiratory infection during periods of known influenza activity, and early instigation of neuraminidase inhibitor treatment based on rapid laboratory confirmation or clinical suspicion.

Our analyses examining the effect of later treatment versus none are especially relevant to the continued clinical debate about the value of delayed therapy. Combining all subgroups of patients, we did not identify any protective association with treatment delayed more than 2 days after symptom onset (table 4). This finding could be explained by confounding by indication. However, we noted that in adult patients admitted to critical care, delayed treatment was associated with reduced likelihood of mortality compared with no treatment (table 4), suggesting that delayed therapy might still be worthwhile in severely ill patients; this finding is plausible since, within this subgroup, treated and untreated patients (who all needed admission to critical care) are likely to have been more balanced in terms of illness severity thereby overcoming confounding by this factor to some extent. Additionally, some patients admitted to critical care might have had prolonged influenza A H1N1pdm09 virus replication in the lower respiratory tract, which might benefit from later initiation of neuraminidase inhibitor treatment. To gain further understanding about overall timing-dependent benefit, we modelled time to start of antiviral treatment using a time-dependent Cox regression model, which showed a significant detrimental survival benefit associated with delay in treatment beyond 2 days after symptom onset (p<0·0001), albeit with overlapping 95% CIs when time to treatment was modelled as a categorical variable; the latter finding suggests that potential differences in treatment benefit between starting on day 3 after symptom onset through to more than 5 days after symptom onset cannot be further clarified through our data. This finding could seem to conflict with the findings in table 4 comparing later neuraminidase inhibitor treatment to no neuraminidase inhibitor treatment but is not surprising, because by comparing only treated patients in figure 2, we possibly eliminated some of the confounding due to indication, which allowed us to identify the potential survival benefits conferred by later treatment, albeit detrimental in proportion to treatment delay.

One of the strengths of this study is the very large number of patients from geographically diverse clinical centres and source populations. We made exhaustive efforts to identify suitable datasets from around the world, but nevertheless cannot comment on the extent to which bias might have been introduced by failing to include centres that did not respond (we cannot say if they had suitable data or not), or that declined to share data; in a worst case scenario, it is possible that less than 20% of potential sites contributed to this analysis. Furthermore, comparatively few cases were from the WHO African (0·1%) and South-East Asia (0·7%) regions, which might limit the extent to which our findings can be generalised.

A clear limitation of our study is that we were unable to adjust specifically for disease severity in our multilevel models because of the heterogeneity of severity measures used across individual datasets. However, we made every effort to include relevant data including severity measures, within each propensity score, but there is still likely to be some residual confounding, particularly due to illness severity at presentation. Likewise, we attempted to control for study-level biases, such as treatment policies, and healthcare seeking behaviour, using multilevel models but there might be residual confounding. A further limitation of our dataset is that 10% of the patients had missing data for exposure to neuraminidase inhibitors and were excluded from the analysis; the characteristics of these patients are compared with those with data for neuraminidase inhibitor exposure in appendix pp 26–27; these patients were more likely to be older, to have presented to hospital later, less likely to have a laboratory confirmed diagnosis, and more likely to be treated with antibiotics.

The decision to adjust for treatment with antibiotics and corticosteroids was taken after consultation with clinical colleagues within the PRIDE study collaboration. This decision results from widespread clinical practice to treat patients admitted to hospital with respiratory illness with corticosteroids and antibiotics. There is particular uncertainty about the possible effect of corticosteroids on the course of severe influenza infection.^33,34 Therefore, it was necessary to separate out the possible effects of antivirals from these other commonly used treatments. We did not do specific analyses to establish the potential effect of antibiotic or corticosteroid use on mortality, but recognise that these factors both warrant further research. Although we were able to adjust for inpatient antibiotics and systemic corticosteroid use, we were unable to adjust for pandemic H1N1 vaccination since 35% (8284 of 23633) of our case series were admitted to hospital before the first availability of vaccine in October, 2009, and 71% (10 967 of 15 349) of data for vaccination status were missing among those admitted after that juncture; however, the available data suggest uptake was no higher than 8% during the study period.

This meta-analysis of individual patient data offers the most rigorous assessment of mortality benefits of neuraminidase inhibitor treatment during the 2009–10 pandemic that is likely to be possible using retrospective observational data. The greatest likelihood of reduced mortality seems to be attributable to treatment started within 2 days of symptom onset. These data offer evidence of the effectiveness of neuraminidase inhibitors during the 2009–10 pandemic and are superior to extrapolations from earlier data on seasonal influenza; they could retrospectively vindicate prepandemic neuraminidase inhibitor antiviral stockpiling decisions made by governments worldwide. Treatment guidance policies should increase emphasis on early empirical neuraminidase inhibitor treatment of adult patients admitted to hospital after presenting with proven or clinically suspected influenza A H1N1pdm09 virus infection. However, most adult patients with suspected or confirmed influenza are not admitted to hospital within 48 h of illness onset. Therefore, the implications of these findings, although based on patients admitted to hospital with influenza A H1N1pdm09, encourage early initiation of neuraminidase inhibitor treatment in outpatients who are appreciably unwell with suspected or confirmed influenza, or at increased risk of complications, including those with influenza A H3N2 or influenza B. Further studies are needed in children to confirm the adequacy of present dose regimens and duration of therapy in terms of clinical efficacy.