Methods
Findings are reported according to the Quality of Reporting of Meta-Analysis (QUOROM) statement and the Standards for Reporting of Diagnostic Accuracy (STARD) statement.
Search Strategy
The literature was systematically searched using predetermined inclusion criteria. Studies were included that reported the sensitivity and/or specificity of an influenza rapid test to detect the presence of 2009 pandemic influenza (H1N1) infection or contained sufficient information to calculate the sensitivity and specificity based on diagnosis of clinical specimens using the rRT-PCR as a gold standard reference test. No language restrictions were applied. Studies were identified eligible for inclusion by searching the databases MEDLINE (NLOM, Bethesda, MD, USA) and EMBASE (Elsevier, Amsterdam, the Netherlands) using PUBMED and OVID interfaces, respectively. Publication dates were restricted to between 1/1/2009 and 1/15/2010, inclusive. Search terms for each database included the following: "influenza diagnostic," "influenza rapid test," "rapid test H1N1," and "influenza rapid". Subsequently, the title and abstract of each potential study were screened to determine potential eligibility, which was then confirmed by a review of the full text. References from eligible studies were also examined for additional potential studies, and papers referencing eligible studies were identified using Google Scholar and considered for inclusion.
Data Synthesis and Meta-analysis
Data synthesis was performed according to guidelines on systematic reviews of diagnostic accuracy studies. The bivariate logit-normal random effects meta-analyses were conducted to summarize the overall sensitivity and specificity of rapid tests. Compared to fixed effects models, the random effects models typically provide conservative estimates with wider confidence intervals because it assumes that the meta-analysis includes only a sample of all possible studies. In addition, the random effects models appropriately account for the difference in study sample sizes, both within-study variability (random error) and between-study variability (heterogeneity). In general, the bivariate approach offers some advantages over separate univariate random effects meta-analysis by accounting for the correlation between sensitivity and specificity. This correlation will exist if the different studies use different test-thresholds and thus are operating at different points along the underlying receiver operating characteristic (ROC) curve for the test. However, one study reported that the differences between univariate and bivariate random effects models for summarizing pooled sensitivity and specificity are trivial based on extensive simulations. Thus, we utilized the univariate logit-normal random effects meta-analyses to generate forest plots (i.e., graphical display designed to illustrate the relative strength in meta-analysis of multiple quantitative scientific studies addressing the same question) with overall and rapid test-specific pooled estimates for both sensitivity and specificity. Parameters used to summarize diagnostic accuracy include the following: sensitivity and specificity directly estimated from the univariate and/or bivariate random effects models; positive and negative likelihood ratio, positive and negative predictive values, and the diagnostic odds ratio (DOR) derived from parameter estimates from the bivariate random effects models accounting for potential correlation between sensitivity and specificity estimates. In addition to reporting pooled sensitivity and specificity, which are often regarded as intrinsic properties of a diagnostic test, we also report other metrics because they are clinically more meaningful in some settings. Sensitivity is estimated by the proportion of positive tests among those with the disease of interest, whereas specificity is estimated by the proportion of negative tests among those without the disease. The positive (or negative) likelihood ratio is estimated by the ratio of the proportion of positive (or negative) tests in the diseased versus non-diseased subjects. The positive (or negative) predictive value is estimated by the proportion of subjects with a positive (or negative) test who have (or do not have) the disease. The DOR, commonly considered a global measure of test performance, is estimated by the ratio of the odds of a positive test result in diseased subjects to the odds of a positive test result in non-diseased subjects.
The Begg- and Mazumdar-adjusted rank correlation test and the Egger et al. regression asymmetry test were used to assess publication bias for sensitivity and specificity, respectively. The Cochran's Q-test was used to detect heterogeneity. Location (US versus non-US) and rapid test manufacturer were included as covariates to examine their possibility as factors causing heterogeneity. Tests for small-study effects were employed only when at least four studies were available. The univariate logit-normal random effects meta-analyses were implemented in R version 2·12·1 (http://cran.r-project.org/) meta package, and the bivariate random effects models were fitted using the NLMIXED procedure in sas version 9·2 (SAS Institute, Cary, NC, USA). The summary ROC curve was plotted based on the regression line of sensitivity on the false-positive rate (1–Sp) in logit scale using the estimates from the bivariate random effects models rather than the line proposed by Rutter and Gatsonis.