Hospitalization Fatality Risk of Influenza A(H1N1)pdm09
Hospitalization Fatality Risk of Influenza A(H1N1)pdm09
This systematic review followed the PRISMA [Preferred Reporting Items for Systematic Reviews and Meta-Analyses] guidelines.
We extracted articles with estimates of the HFR of H1N1pdm09 on January 9, 2014, from 2 databases: MEDLINE (PubMed; US National Library of Medicine, Bethesda, Maryland) and EMBASE (Excerpta Medica Database; Elsevier B.V., Amsterdam, the Netherlands). The following free search terms were used to search in "All Fields":
The search was limited to studies published after April 1, 2009, subsequent to the start of the 2009 pandemic. We manually retrieved any additional relevant studies identified.
All titles identified via the search strategy were independently screened by 2 authors (J.Y.W. and B.J.C.). Abstracts of potentially relevant articles and the full texts of manuscripts were reviewed for eligibility. Articles in all languages were selected for assessment if at least 1 statistical estimate of the HFR was presented and described as an estimate for H1N1pdm09. We used Google Translate (Google Inc., Mountain View, California) for articles not written in English. Eligible studies were those that included ≥20 hospitalized cases and in which the authors reported 1 or more population-based estimates of the HFR or sufficient data to calculate an HFR. Studies that reported estimates of the HFR only in population subgroups, such as pregnant women or persons at higher risk of severe illness if infected (e.g., those with chronic health conditions), were excluded. Study quality was not formally assessed, although we did analyze the association of study design factors with heterogeneity in the HFR (see below).
We defined the HFR as the probability of death associated with H1N1pdm09 cases that required hospitalization for medical reasons. The HFR for a cohort of individuals is estimated as the number of H1N1pdm09-associated deaths divided by the number of H1N1pdm09 hospitalizations in the same cohort. It is also possible to estimate the HFR in a population, as the number of H1N1pdm09 deaths divided by the number of H1N1pdm09 hospitalizations in the same population over the same time period.
We defined a cohort study as one which followed the same group of patients throughout the study period, either retrospectively or prospectively. The numbers of hospitalized cases and deaths were obtained from the cohort study, and the deaths were a subset of the hospitalized cases. We defined a discordant-source study as one where the number of hospitalized cases and the number of deaths were collected from surveillance reports or estimated (for instance, by modeling) independently of one another and the deaths might not necessarily have occurred within the group of hospitalized patients, although hospitalizations and deaths were estimated from the same source population over the same time period. The numerator of the HFR could be either counts or estimates of the number of deaths among laboratory-confirmed hospitalized cases, while the denominator could be either counts or estimates of the number of laboratory-confirmed H1N1pdm09 hospitalized cases. Hospitalized cases were those occurring in persons confirmed to have influenza virus infection by reverse-transcriptase polymerase chain reaction, rapid antigen test, or viral culture from an inpatient test. Patients admitted to a hospital after the cessation of viral shedding associated with H1N1pdm09 infection would probably not have been identified in the included studies.
All data were extracted independently and entered onto a standardized form by 2 of the authors (C.M.S. and E.Y.S.). Disagreements were resolved by a third author (J.Y.W.). The primary data extracted were the estimates of the HFR, the estimates or counts of the number of H1N1pdm09-associated hospitalized deaths (numerator), and the estimates or counts of the number of H1N1pdm09 hospitalized cases (denominator). Whenever available, we extracted HFRs stratified by age, hospital type, and hospitalized cases' characteristics, including the proportion of males and the proportion of cases with ≥1 underlying medical disorder. We also contacted individual authors for age breakdowns of hospitalized cases and deaths if this information was not reported in the published paper. If reported, the number of intensive care unit (ICU) admissions and the estimates of HFR for seasonal influenza from the same setting were also extracted. Although age groups differed across studies, children, adults, and the elderly were defined here as persons aged ≤19 years, 20–64 years, and ≥65 years, respectively.
Reported estimates of the HFR were combined using a random-effects model (see the Web Appendix, available at http://aje.oxfordjournals.org/), including a variance-stabilizing transformation before pooling. To achieve variance stabilization, Freeman and Tukey have suggested performing a double arcsine transformation before combining risk estimates. Statistical heterogeneity was assessed by means of the I statistic, with higher values signifying a greater degree of variation. We examined HFR estimates according to study design (cohort studies vs. discordant-source studies), age group, year of virus circulation, a country's per capita gross domestic product (GDP) (http://databank.worldbank.org/), and geographical location, broadly classified as North America, Europe, Central/South America, Asia, and Australia/New Zealand. Meta-regression analyses were conducted using a mixed-effects model (Web Appendix http://aje.oxfordjournals.org/content/182/4/294/suppl/DC1). We defined the first year of the pandemic as the period from 2009 to mid-2010. Relative risks of death comparing pandemic viruses with seasonal viruses were combined using a fixed-effects model (Web Appendix http://aje.oxfordjournals.org/content/182/4/294/suppl/DC1). All analyses were conducted with R, version 3.0.2 (R Foundation for Statistical Computing, Vienna, Austria) and the metafor package.
Methods
This systematic review followed the PRISMA [Preferred Reporting Items for Systematic Reviews and Meta-Analyses] guidelines.
Search Strategy and Selection Criteria
We extracted articles with estimates of the HFR of H1N1pdm09 on January 9, 2014, from 2 databases: MEDLINE (PubMed; US National Library of Medicine, Bethesda, Maryland) and EMBASE (Excerpta Medica Database; Elsevier B.V., Amsterdam, the Netherlands). The following free search terms were used to search in "All Fields":
hospital OR hospitali* OR patient OR inpatient
fatalit* OR mortality OR death OR severity OR seriousness OR lethalit* OR virulence
#1 AND #2
influenza OR flu
pandemic
#4 AND #5
H1N1* OR pH1N1* OR pdmH1N1* OR nH1N1*
#6 OR #7
#3 AND #8
The search was limited to studies published after April 1, 2009, subsequent to the start of the 2009 pandemic. We manually retrieved any additional relevant studies identified.
All titles identified via the search strategy were independently screened by 2 authors (J.Y.W. and B.J.C.). Abstracts of potentially relevant articles and the full texts of manuscripts were reviewed for eligibility. Articles in all languages were selected for assessment if at least 1 statistical estimate of the HFR was presented and described as an estimate for H1N1pdm09. We used Google Translate (Google Inc., Mountain View, California) for articles not written in English. Eligible studies were those that included ≥20 hospitalized cases and in which the authors reported 1 or more population-based estimates of the HFR or sufficient data to calculate an HFR. Studies that reported estimates of the HFR only in population subgroups, such as pregnant women or persons at higher risk of severe illness if infected (e.g., those with chronic health conditions), were excluded. Study quality was not formally assessed, although we did analyze the association of study design factors with heterogeneity in the HFR (see below).
Definition of the HFR
We defined the HFR as the probability of death associated with H1N1pdm09 cases that required hospitalization for medical reasons. The HFR for a cohort of individuals is estimated as the number of H1N1pdm09-associated deaths divided by the number of H1N1pdm09 hospitalizations in the same cohort. It is also possible to estimate the HFR in a population, as the number of H1N1pdm09 deaths divided by the number of H1N1pdm09 hospitalizations in the same population over the same time period.
We defined a cohort study as one which followed the same group of patients throughout the study period, either retrospectively or prospectively. The numbers of hospitalized cases and deaths were obtained from the cohort study, and the deaths were a subset of the hospitalized cases. We defined a discordant-source study as one where the number of hospitalized cases and the number of deaths were collected from surveillance reports or estimated (for instance, by modeling) independently of one another and the deaths might not necessarily have occurred within the group of hospitalized patients, although hospitalizations and deaths were estimated from the same source population over the same time period. The numerator of the HFR could be either counts or estimates of the number of deaths among laboratory-confirmed hospitalized cases, while the denominator could be either counts or estimates of the number of laboratory-confirmed H1N1pdm09 hospitalized cases. Hospitalized cases were those occurring in persons confirmed to have influenza virus infection by reverse-transcriptase polymerase chain reaction, rapid antigen test, or viral culture from an inpatient test. Patients admitted to a hospital after the cessation of viral shedding associated with H1N1pdm09 infection would probably not have been identified in the included studies.
Data Extraction
All data were extracted independently and entered onto a standardized form by 2 of the authors (C.M.S. and E.Y.S.). Disagreements were resolved by a third author (J.Y.W.). The primary data extracted were the estimates of the HFR, the estimates or counts of the number of H1N1pdm09-associated hospitalized deaths (numerator), and the estimates or counts of the number of H1N1pdm09 hospitalized cases (denominator). Whenever available, we extracted HFRs stratified by age, hospital type, and hospitalized cases' characteristics, including the proportion of males and the proportion of cases with ≥1 underlying medical disorder. We also contacted individual authors for age breakdowns of hospitalized cases and deaths if this information was not reported in the published paper. If reported, the number of intensive care unit (ICU) admissions and the estimates of HFR for seasonal influenza from the same setting were also extracted. Although age groups differed across studies, children, adults, and the elderly were defined here as persons aged ≤19 years, 20–64 years, and ≥65 years, respectively.
Statistical Analysis
Reported estimates of the HFR were combined using a random-effects model (see the Web Appendix, available at http://aje.oxfordjournals.org/), including a variance-stabilizing transformation before pooling. To achieve variance stabilization, Freeman and Tukey have suggested performing a double arcsine transformation before combining risk estimates. Statistical heterogeneity was assessed by means of the I statistic, with higher values signifying a greater degree of variation. We examined HFR estimates according to study design (cohort studies vs. discordant-source studies), age group, year of virus circulation, a country's per capita gross domestic product (GDP) (http://databank.worldbank.org/), and geographical location, broadly classified as North America, Europe, Central/South America, Asia, and Australia/New Zealand. Meta-regression analyses were conducted using a mixed-effects model (Web Appendix http://aje.oxfordjournals.org/content/182/4/294/suppl/DC1). We defined the first year of the pandemic as the period from 2009 to mid-2010. Relative risks of death comparing pandemic viruses with seasonal viruses were combined using a fixed-effects model (Web Appendix http://aje.oxfordjournals.org/content/182/4/294/suppl/DC1). All analyses were conducted with R, version 3.0.2 (R Foundation for Statistical Computing, Vienna, Austria) and the metafor package.
