Types of studies

We included primary studies that assessed the diagnostic accuracy of Xpert for pulmonary TB and/or rifampicin resistance. Diagnostic accuracy studies are typically cross‐sectional in design. However, we also searched for randomized controlled trials (RCTs) and cohort studies. We only included studies that reported data comparing Xpert to an acceptable reference standard from which we could extract true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. Xpert could be assessed alone or together with other tests.

We excluded studies with a case‐control design because these types of studies are prone to bias, in particular, studies enrolling patients with severe disease and healthy participants without disease. We also excluded studies reported only in abstracts.

Participants

Studies had to include adult or predominantly adult patients suspected of having pulmonary TB or MDR‐TB, with or without HIV infection. We considered adults to be 15 years of age or older. We included studies that assessed the diagnostic accuracy of Xpert using sputum and other respiratory specimens, such as fluid obtained from bronchial alveolar lavage and tracheal aspiration. We included studies from all settings and all countries. Data on age of study participants were available for the majority of included studies. We considered it highly likely that studies that did not report age data involved all or predominantly adults for the following reasons: the vast majority of specimens evaluated with Xpert were sputum specimens and children have difficulty producing sputum; we excluded data on specimens obtained by gastric aspiration, as this specimen collection method is used mostly for investigating TB in children; we excluded studies that specifically evaluated the use of Xpert in children; and we performed a sensitivity analysis by dropping studies that did not report age data to check whether the accuracy results changed (Sensitivity analyses).

Index tests

Xpert was the index test under evaluation.

Target conditions

The target conditions were active pulmonary TB and rifampicin resistance.

Reference standards

For TB, acceptable reference standards utilized solid media: Löwenstein‐Jensen, Middlebrook 7H10 or 7H11, or Ogawa media; or a commercial liquid culture system: such as BACTEC™ 460TB System or BACTEC™ MGIT™ (mycobacterial growth indicator tube) 960 Mycobacterial Detection System, BD, USA; BacT/ALERT® System, bioMérieux, France; or both solid media and a commercial liquid culture system; or VersaTREK® Mycobacteria Detection & Susceptibility, Thermo Fisher Scientific, USA.

For rifampicin resistance, the reference standards were conventional phenotypic drug susceptibility testing methods as recommended by WHO (WHO Policy DST 2008). Acceptable methods used solid media, such as Löwenstein‐Jensen, Middlebrook 7H10 or 7H11, or Ogawa media and/or a commercial liquid culture system, such as BACTEC™ 460TB System or BACTEC™ MGIT™ (mycobacterial growth indicator tube) 960 Mycobacterial Detection System, BD, USA; BacT/ALERT® System, bioMérieux, France; or VersaTREK® Mycobacteria Detection & Susceptibility, Thermo Fisher Scientific, USA.

Electronic searches

Vittoria Lutje, (VL) the Information Specialist from the Cochrane Infectious Diseases Group, searched the following databases on 25 September 2011 using the strategy described in Appendix 1: Cochrane Infectious Diseases Group Specialized Register; MEDLINE; EMBASE; ISI Web of Knowledge; MEDION; LILACS; BIOSIS; and SCOPUS. She also searched the metaRegister of Controlled Trials (mRCT) and the search portal of the WHO International Clinical Trials Registry Platform, to identify ongoing trials. She repeated searches on 15 December 2011. We limited all searches to 2007 onward because the development of Xpert was completed in 2009 and the first paper describing its clinical use was published electronically in 2009 (Helb 2010). VL performed the searches without language restriction.

Searching other resources

To identify additional published, unpublished, and ongoing studies, we performed the following tasks:

• reviewed reference lists of included articles and review articles identified through the above methods;

• contacted Cepheid, the test manufacturer;

• handsearched WHO reports on Xpert;

• contacted researchers at the Foundation for Innovative New Diagnostics (FIND), members of the Stop TB Partnership's New Diagnostics Working Group, and other experts in the field of TB diagnostics.

While preparing this review, the Xpert MTB/RIF Mapping Tool became available (http://xrmt.treattb.org/projects‐listing2). This tool has been developed by the International Union Against Tuberculosis and Lung Disease, through the USAID‐funded TREAT TB initiative, and supported by WHO. The tool allows researchers and policy makers to globally map ongoing research activities related to Xpert. We may use this tool to identify studies for updates of this review.

Selection of studies

Two review authors (KRS and HS) independently scrutinized titles and abstracts identified by electronic literature searching to identify potentially eligible studies (screen 1). We selected any citation identified by either review author during screen 1 for full‐text review. Next, we retrieved the full paper of each potentially eligible article identified in the search. Two review authors (KRS and HS) independently assessed articles for inclusion using predefined inclusion and exclusion criteria (screen 2). In screen 2, we resolved any discrepancies by discussion between the review authors, or if necessary, by decision of a third review author (MP). We maintained a list of excluded studies and their reasons for exclusion.

We named studies according to the surname of the first author and year of publication. For multicentre studies, the study‐naming scheme uniquely identifies multiple study centres from within each study (Boehme 2010a; Boehme 2010b), each of which reported data separately for a distinct population at a given study site. Hence, the number of study centres exceeds the number of studies. (Please note that when cited in the text, "Boehme [year]a" refers to an overall paper, consisting of all centres).

Data extraction and management

We extracted data on the following characteristics: 

• author; publication year; study design; case country of residence; country income status classified by the World Bank List of Economies (World Bank 2011); clinical setting; laboratory setting;

• population, age, gender, HIV status, smear status, and follow‐up;

• reference standard;

• parameter value in the Xpert system associated with determining rifampicin resistance;

• specimen collection (expectorated sputum, induced sputum)

• condition of the specimen (fresh or frozen);

• preparation of the specimen (processed or unprocessed);

• QUADAS‐2 items (Whiting 2011);

• data for two‐by‐two tables for Xpert, including results reported as indeterminate (results reported as invalid, error, or no result);

• time to diagnosis (time from specimen collection until there is an available TB result in laboratory or clinic);

• time to treatment initiation (time from specimen collection until time patient starts treatment).

Whenever possible, we extracted TP, FP, FN, and TN values based on one Xpert result for one specimen provided by one patient. However, in some of the studies, the number of specimens (and Xpert results) exceeded the number of patients, suggesting that a single patient may have provided multiple specimens. We therefore compared pooled sensitivity and specificity for TB detection in all studies with pooled sensitivity and specificity in the subset of studies that provided one Xpert result based on one specimen provided by one patient (see Sensitivity analyses).

When data were available, we extracted FP and TN values for participants without TB disease by their smear and HIV status. Concerning the definition of smear positivity, as the vast majority of included studies performed Xpert in reference laboratories, we assumed these studies adhered to the revised definition of a new sputum smear‐positive pulmonary TB case based on the presence of at least one acid‐fast bacillus in at least one sputum sample in countries with a well functioning external quality assurance system (WHO Policy Smear‐positive TB Case 2007). 

We developed a standardized data extraction form and piloted the form with four studies (22%). Based upon the pilot, we finalized the form. Two review authors (KRS and HS) independently extracted data from each study using the final form. We contacted authors of studies for missing data and clarifications. We entered all data into Excel and SPSS 2006. The final data extraction form is included in Appendix 2.

Assessment of methodological quality

We appraised the quality of included studies with the Quality Assessment of Diagnostic Accuracy Studies (QUADAS‐2) tool (Whiting 2011). QUADAS‐2 consists of four domains: patient selection, index test, reference standard, and flow and timing. We assessed all domains for the potential for risk of bias, and, in addition, we assessed the first three domains for concerns regarding applicability. We used questions, called signalling questions, for each domain to form judgments about the risk of bias. As recommended, we first developed guidance on how to appraise each signalling question and interpret this information. Then, one review author (KRS) piloted the tool with four (22%) of the included studies. Based on experience gained from the pilot, we finalized the tool. Two review authors independently assessed the methodological quality of the included studies with the finalized tool. We presented results in the text, in graphs, and in a table. We did not generate a summary "quality score" because of problems associated with such numeric scores (Juni 1999; Whiting 2005). The domains of the QUADAS‐2 tool and their interpretation are provided in Appendix 3.

Statistical analysis and data synthesis

We performed descriptive analyses for the results of the included studies using SPSS 2006 and present key study characteristics in Characteristics of included studies. We used data reported in the two‐by‐two tables to calculate sensitivity and specificity estimates and 95% confidence intervals (CI) for individual studies and to generate forest plots using Review Manager 5. We chose to use data that were not subject to discrepant analyses (ie unresolved data), since resolved data after discrepant analyses are a potential for risk of bias (Hadgu 2005).

We carried out meta‐analyses to estimate the pooled sensitivity and specificity of Xpert separately for TB detection (I. A. and I. B) and rifampicin resistance detection (II. A.). We determined pooled estimates using an adaptation of the bivariate random‐effects model (Reitsma 2005) to allow for a hierarchical structure for the two multicentre studies (Boehme 2010a; Boehme 2011a). The bivariate random‐effects approach allowed us to calculate the pooled estimates of sensitivity and specificity while dealing with potential sources of variation caused by (1) imprecision of sensitivity and specificity estimates within individual studies; (2) correlation between sensitivity and specificity across studies; and (3) variation in sensitivity and specificity between studies.

We estimated all models using a Bayesian approach with non‐subjective prior distributions and implemented using WinBUGS (Version 1.4.3) (Lunn 2000). Under the Bayesian approach, all unknown parameters must be provided a prior distribution that defines the range of possible values of the parameter and the likelihood of each of those values based on information external to the data. In order to let the observed data determine the final results, we chose to use low‐information prior distributions over the pooled sensitivity and specificity parameters and their between‐study standard deviation parameters. The model we used is summarized in the Statistical Appendix together with the WinBUGS program used to implement it (Appendix 4). Information from the prior distribution is combined with the likelihood of the observed data in accordance with Bayes Theorem to obtain a posterior distribution for each unknown parameter.

Using a sample from the posterior distribution we can obtain various descriptive statistics of interest. We estimated the median pooled sensitivity and specificity and their 95% credible intervals (CrI).The median or the 50% quantile is the value below which 50% of the posterior sample lies. We chose to report the median because the posterior distributions of some parameters may be skewed and the median would be considered a better point estimate of the unknown parameter than the mean in such cases. The 95% CrI is the Bayesian equivalent of the classical (frequentist) 95% CI. (We have indicated 95% CI for individual study estimates and 95% CrI for pooled study estimates as appropriate).The 95% CrI may be interpreted as an interval that has a 95% probability of capturing the true value of the unknown parameter given the observed data and the prior information. We also extracted estimates of the 'predicted' sensitivity and specificity in a future study together with their 95% CrIs. The predicted value gives an idea of heterogeneity at the study level. We can compare the predicted intervals to the pooled intervals to get an idea of the heterogeneity. With a large number of studies, the pooled interval may be narrow. However, if there is considerable variability in sensitivity and specificity estimates between studies, this variability will be reflected in a wide predicted interval despite the large number of studies. We generated the plots using R (version 2.15.1) (R 2008).

Investigations of heterogeneity

Sensitivity analyses

We performed sensitivity analyses by limiting inclusion in the meta‐analysis to: 1) studies that provided data by age that explicitly met the age criterion for participants; 2) studies where consecutive patients were selected; 3) studies where a single specimen yielded a single Xpert result for a given patient; and 4) studies that explicitly represented the use of Xpert for the diagnosis of patients suspected of having TB.

Assessment of reporting bias

We chose not to carry out formal assessment of publication bias using methods such as funnel plots or regression tests because such techniques have not been found to be helpful for diagnostic test accuracy studies (Macaskill 2010). However, Xpert is produced by only one manufacturer and, being a new test for which there has been considerable attention and scrutiny, we believe reporting bias was minimal.