Types of studies

We will include cross‐sectional studies, cohort studies, and randomised controlled trials that assessed the accuracy of at least one of the defined index tests for pulmonary tuberculosis. We will only include studies that used a microbiologic reference standard (defined below). We will include studies from all settings and time periods. Randomized controlled trials will be included because these studies may report sensitivity and specificity in addition to patient health outcomes. For randomised studies that compare different screening strategies, we will evaluate each arm as a separate cohort. Data on the index test(s) must be available to be extracted as true positive, false positive, true negative, and false negative against the reference standard(s) so that we can construct two‐by‐two contingency tables.

Studies applying index tests multiple times to an individual within a short timeframe (e.g. within a single hospital admission), will be considered diagnostic rather than using a screening approach, and these studies will be excluded. Studies in which children with negative screening test results were not subjected to the reference standard will be included. As shown in Figure 2, this often occurs in tuberculosis prevalence studies when it is assumed that those with a negative screen (e.g. no CXR abnormalities) do not have active tuberculosis. However, this leads to poor enumeration of true negative and false negative test results. Therefore, we will only include such studies that partially verified tuberculosis status if they were conducted prospectively and enrolled a consecutive or random sample of eligible children. Our rationale for specifying these strict design criteria is to enable us to calculate positive predictive values from such studies and include them in a separate set of analyses. Due to the direct relationship between prevalence and predictive values, to reduce potential variation in prevalence between studies, we will only include studies in the analyses of positive predictive value if the studies were done in the same setting (e.g. community or accessing healthcare), target population, and for the same purpose (e.g. contact tracing).

We will include cohort studies with children with active tuberculosis identified after the time point that the screening test was applied. Especially with studies performed in settings of intended use, the collection of specimens and conduct of the reference standard may occur some time after the screening test was done. In low resource‐settings, this process may take weeks. However, a longer time between the index test and the reference standard would make us less sure that the target condition did not change between the two tests. We will address this issue in the QUADAS‐2 domain 4: Flow and timing and with a sensitivity analysis (see Sensitivity analyses).

We will exclude case reports and case‐control studies, the latter because of the high risk of bias in diagnostic accuracy studies (Rutjes 2006).

Participants

We will include studies enrolling HIV‐positive and HIV‐negative children younger than 15 years old not known to have active tuberculosis prior to screening. We will exclude studies if they include older children and we are unable to extract data for children younger than 15 years old from the publication. We will include children in the general population in high burden settings, children living in areas of high tuberculosis burden, and high‐risk groups, including children younger than five years old; children living with HIV; children with recent exposure to a person with active tuberculosis; and household, school, or other contacts of a person with active tuberculosis. We will include studies in which children are screened only once and studies that report longitudinal screening with repeated screening tests at predetermined intervals.

Index tests

For symptom screening, we will include studies that assess symptoms of tuberculosis or combinations of symptoms as described by the primary study authors. Symptoms of active tuberculosis in children may include cough, fever, decreased appetite, weight loss or failure to thrive, and fatigue or reduced playfulness. Older children may experience symptoms similar to those in adults and include persistent cough, haemoptysis, and weight loss, fever, night sweats and fatigue. The threshold was presence or absence of symptoms.

For CXR screening, we will include studies that utilize conventional radiography, digital radiography, and computed radiography. We will include all classification systems for identification of CXR abnormalities, including automated interpretation of radiography using deep learning or artificial intelligence technology. We will categorize all CXR screening results as follows. We will use an author defined threshold for CXR results. Essentially this is an implicit threshold utilized by the CXR reader.

• Normal.

• Any CXR abnormality, i.e. abnormalities suggestive of tuberculosis and other abnormalities.

• Abnormalities suggestive of tuberculosis.

For Xpert MTB/RIF and Xpert Ultra, we will include studies in which the index tests are evaluated in expectorated or induced sputum, gastric aspirate specimens, nasopharyngeal aspirate specimens, and bronchoalveolar lavage specimens. Tuberculosis bacilli in sputum can be swallowed and detected in stool so we will also include studies assessing stool specimens. We will include studies assessing more than one type of respiratory specimen collected at the same time and extract 2 x 2 data separately for each specimen type.

Xpert MTB/RIF provides the following printed test results:

• MTB (M tuberculosis) DETECTED; Rif (rifampicin)resistance DETECTED;

• MTB DETECTED; Rif resistance NOT DETECTED;

• MTB DETECTED; Rif resistance INDETERMINATE;

• MTB NOT DETECTED;

• INVALID (the presence or absence of MTB cannot be determined);

• ERROR (the presence or absence of MTB cannot be determined);

• NO RESULT (the presence or absence of MTB cannot be determined).

Xpert Ultra also gives the following semi‐quantitative classifications of M tuberculosis bacterial burden from the sample: trace, very low, low, moderate, and high. For this review, Xpert MTB/RIF and Xpert Ultra results will be categorized as:

• Positive: 'MTB DETECTED,' including 'trace' results from Xpert Ultra

• Negative: 'MTB NOT DETECTED'

• Inconclusive: 'INVALID,' 'ERROR,' or 'NO RESULT'

We will not evaluate rifampicin resistance in this review.

As shown in Figure 2, with two parallel screening tests, the parallel strategy will entail any of the individual components of the strategy being positive resulting in a positive parallel strategy screen and all individual components being negative resulting in a negative parallel strategy screen. For studies assessing parallel screening tests, if data for the individual components of the parallel strategy against the reference standard is also available, these data will also be extracted for analysis.

Target conditions

The target condition is active pulmonary tuberculosis.

We anticipate that some studies may evaluate the index tests for active tuberculosis and not explicitly state 'pulmonary tuberculosis', the target condition in this review. We will include these studies because the most common type of active tuberculosis in children is lung disease; hence, most screening studies in children evaluate tests for pulmonary tuberculosis and diagnose tuberculosis using respiratory specimens. If data are sufficient, we will perform a sensitivity analysis limiting inclusion to those studies that explicitly evaluated the index tests for pulmonary tuberculosis.

Reference standards

We will utilize two reference standards, a microbiological and a composite reference standard.

Electronic searches

We will search the following databases without language restriction, using the search terms and strategy described in Appendix 1.

• MEDLINE and MEDLINE in Process (OVID), from 1946.

• Embase (OVID), from 1947.

• Cochrane Central Register of Controlled Trials (CENTRAL), published in the Cochrane Library.

• Scopus (Elsevier, from 1970).

We will also search ClinicalTrials.gov, the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/trialsearch), and the International Standard Randomized Controlled Trials Number (ISRCTN) registry (www.isrctn.com/) for trials in progress.

Searching other resources

To identify any relevant published data not identified with our electronic search, we will contact experts in the field, and check the references of relevant reviews from the past ten years. With the studies selected for inclusion in this review, we will perform forward and backward reference checking to identify any additional eligible studies.

Selection of studies

We will use Covidence to manage the selection of studies (Covidence 2017). Two review authors (BV and TN) will independently screen all titles and abstracts from the electronic searches to identify potentially eligible studies. We will obtain full‐text articles of potentially eligible studies, and the two review authors (BV and TN) will independently assess the full‐text articles for study eligibility using the predefined inclusion and exclusion criteria. We will resolve any disagreements by discussion or with a third review author (AMM). As needed, we will contact study authors to clarify the study methods and other information. Studies excluded during the full‐text review will be listed in the 'Characteristics of excluded studies' table with a summary of reasons for exclusion. We will illustrate the study selection process in a PRISMA flow diagram (Moher 2009).

Data extraction and management

We will design a data extraction form and pilot it on at least two included studies. After reviewing the piloted forms with the other review authors, we will finalize the form. Two review authors will use the data extraction form to independently extract data from the included studies. We will discuss any inconsistencies with a third review author. We will enter the extracted data into an Excel database (Excel 2013) on password‐protected computers and secured in the cloud storage Dropbox for future review updates.

We will extract the following information for each included study.

Assessment of methodological quality

Two review authors will independently assess the methodological quality of the included studies using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS‐2) instrument, which we will adapt for this review (Whiting 2011). The preliminary tool with signalling questions tailored to this review is in Appendix 2. As recommended, we will assess each of the four domains (patient selection, index tests, reference standards, and flow and timing) for risk of bias and the first three domains for concerns regarding applicability.

We will judge each item as 'yes' (adequately addressed), 'no' (inadequately addressed), or 'unclear' when there is insufficient information reported to make an assessment. One review author will pilot the tool on two included studies. We will then make any necessary revisions to finalize the tool. We will resolve disagreements between the two review authors' independent assessments through discussion or additional input from a third review author. We will present results of the quality assessment in text, tables, and graphs.

Statistical analysis and data synthesis

We will perform descriptive analyses of the included studies and present their key characteristics in the 'Characteristics of included studies' table and a summary table. We will present individual study estimates of sensitivity and specificity graphically on forest plots and in receiver operating characteristics (ROC) space using Review Manager 5 (RevMan 2020).

We will consider one index test result per child per time point. However, for studies assessing serial screening over time for individuals, separate screens may be assessed if they are also compared against serial confirmatory tests over time (i.e. multiple screens for one individual). Within each group listed in Objectives, we will perform analyses by index tests and reference standards. For symptom screening as the index test, we plan to perform analyses by single symptoms and multiple symptoms (such as the WHO four‐symptom screen), where data are available.

We will consider combining categories depending on the number of studies and screening definitions found in each category. We will also stratify the analyses by the type of reference standard used, microbiological or composite. Separate analyses will be performed for studies that verify participants regardless of their index test results (i.e. complete verification) and those that only verify test positives (i.e. partial verification).

When there are sufficient data, we will perform meta‐analyses to estimate summary values of sensitivity and specificity using a bivariate model (Chu 2006; Reitsma 2005). We chose the bivariate model because we anticipate dealing with binary test results or studies that used the same threshold because they applied the threshold recommended by the test manufacturer. Also, we note that the bivariate model is appropriate to use for index tests such as Xpert MTB/RIF and Xpert Ultra, which apply a common positivity criterion (Macaskill 2010). When we are unable to fit the models due to sparse data or few studies, we will simplify the models to univariate random‐effects logistic regression models to pool sensitivity and specificity separately (Takwoingi 2015). For studies that verify only test positives, we will pool positive predictive values using a univariate random‐effects logistic regression model. We will perform meta‐analyses using the meqrlogit command in Stata version 16 (Stata 16).

For test comparisons, we will exclude studies where one index test was used as the reference standard for another index test in the comparison e.g. if Xpert MTB/RIF was used as the reference standard for CXR. We will perform comparative meta‐analyses by first including all studies with relevant data in indirect comparisons to make use of all available data. We will then perform additional analyses by restricting the analyses to only comparative studies that made direct comparisons between the index tests within the same study population. Comparative meta‐analyses will be performed using bivariate meta‐regression by adding test‐type as a covariate to bivariate models. We will assess model fit using likelihood ratio tests to compare models with and without the covariate terms. We will calculate absolute differences in sensitivity and specificity using the model parameters. We will obtain 95% confidence intervals and P values for the absolute differences using the delta method and Wald tests, respectively.

For subgroups or screening definitions that do not have sufficient data for a meta‐analysis, we will summarize findings using descriptive methods.

Investigations of heterogeneity

We will visually inspect forest plots and summary ROC (SROC) plots for heterogeneity. When data allow, we will evaluate potential sources of heterogeneity using subgroup analyses and bivariate meta‐regression.

For subgroup analyses, we will assess the following subgroups: children aged 0 to 4 years, children aged 5 to 14 years, HIV positive children, and HIV negative children.

For meta‐regression analyses, we will include high tuberculosis burden country (yes or no) and single or initial screening versus more than one screening and consider each source of heterogeneity as a single covariate in a bivariate model.

Sensitivity analyses

When data allow, we will perform sensitivity analyses and explore the effect of risk of bias and study characteristics on the accuracy of index test results by limiting inclusion in the meta‐analyses to the following.

• Studies that only used consecutive or random selection of participants.

• Studies with an appropriate interval between the index test and the reference standard.

• Studies that avoid incorporation bias (inclusion of index test as part of the reference standard).

• Studies that explicitly evaluated the index tests for pulmonary tuberculosis.

Assessment of reporting bias

We will not formally assess reporting bias using funnel plots or regression tests as these have not been reported as helpful for diagnostic test accuracy studies (Macaskill 2010).

Assessment of certainty of the evidence

We will assess the certainty of evidence using the GRADE approach for diagnostic studies (Balshem 2011; Schünemann 2008). As recommended, we will rate the certainty of evidence as either high (not downgraded), moderate (downgraded by one level), low (downgraded by two levels), or very low (downgraded by more than two levels) based on five domains: risk of bias, indirectness, inconsistency, imprecision, and publication bias. For each outcome, the certainty of evidence starts as high when there are high‐quality observational studies (cross‐sectional or cohort studies) that enrolled participants with diagnostic uncertainty. If we find a reason for downgrading, we will use our judgement to classify the reason as either serious (downgraded by one level) or very serious (downgraded by two levels).

Three review authors will discuss judgements and apply GRADE in the following way (Schünemann 2020a; Schünemann 2020b).