Literature search
This study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Diagnostic Test Accuracy criteria 2018 (PRISMA-DTA 2018) [30]. The Web of Science, PubMed, the Cochrane Library and Embase databases were used to search for relevant English language citations published up to February 2019. Our search terms were “tuberculosis,” “pulmonary tuberculosis,” “Chemokine CXCL10,” and “interferon gamma-induced protein 10.” Comprehensive literature search strategies were used based on the following combination of MeSH terms, title/abstracts and all fields for these databases (Supplementary Table). Additionally, the reference lists of the applicable studies, relevant research letters, and reviews were manually searched to find other potentially relevant studies.
Literature selection
Two investigators independently determined literature eligibility. Studies reporting IP-10 levels for the detection of PTB were included according to the following criteria: (1) reporting on individuals with PTB and non-TB (population); (2) provision of IP-10 in whole blood and plasma as index test; (4) Mtb culture as a gold standard, and other reference standard including pathological examination, microscopy and genexpert MTB/RIF test (WHO recommended) [2]; (5) the primary outcomes including diagnostic performance of IP-10 (sensitivity and specificity); (5) randomized controlled trails, prospective and retrospective studies included (study design); (6) more than 10 individuals reported meeting the inclusion criteria. Studies not published in English, other letters (except research letters), conference abstracts, veterinary experiments, reviews and case reports were excluded.
Data extraction
The following data were extracted: the first author, year of publication, country, TB high-burden, study design, age, number of participants (patients with PTB and non-TB subjects), TB site, non-TB status, cut-off for index test (IP-10), diagnostic reference standard, method and condition for the IP-10 assay, HIV-infection status, sensitivity, specificity, true positive (TP), false positive (FP), false negative (FN), and true negative (TN) for IP-10. Two investigators independently extracted data from eligible articles, and disagreements were resolved by discussing and reaching a consensus.
Quality assessment
According to the Cochrane Collaboration, two investigators independently reviewed the methodological quality of eligible articles by Quality Assessment of Diagnostic Accuracy Studies tool-2 (QUADAS-2) [31, 32]. Disagreements were resolved by consensus. Revman (version 5.3) was used to perform the quality assessment.
Data analysis
Excel was used to construct a two-by-two table, including TP, FP, FN, and TN for patients with PTB. Stata (version 14.0) was used to perform the data analysis. The index test had different optimal cut-offs. According to the recommendation of Cochrane Collaboration, the hierarchical summary receiver operating characteristic (HSROC) model by Rutter et al. was utilized when the index test was assessed by applying various thresholds [32, 33]. The HSROC curve was computed with the “metandi” command [34]. Prediction region presented possible point of sensitivity and specificity in the HSROC curve. The summary point showed the pooled sensitivity and specificity under the optimal threshold value. Confidence region reflected the possible summary point.
The main outcomes were the diagnostic performance of IP-10 for detecting PTB by the random effect model, as evaluated by the summary estimates of sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and the area under the curve (AUC). Sensitivity, reflecting the ability of index test to detect patients, calculated by “Sensitivity=TP/(TP+FN)”. Specificity, reflecting the ability of index test to eliminate disease-free, calculated by “Specificity=TN/(FP+TN)”. PLR, a measure of index test for detection potential for disease, could be calculated by the formula “PLR=Sensitivity/(1-Specificity)”. NLR, a measure of index test for detection potential for non-disease, could be calculated by the formula “NLR=(1-Sensitivity)/Specificity”. DOR, a measure for overall accuracy of index test, could be calculated by the formula “DOR=(TP/FN)/(FP/TN)”. AUC, indicated how the index test was accurate, especially exceeded 0.90. 95% confidence interval (CI) was calculated by wilson method and no correction factor applied.
The I2 value was not suitable for the quantification of heterogeneity in accuracy studies [35]. Thus, to explore potential sources of heterogeneity, we used a meta-regression analysis with the “midas” command. The intercept was zero. Seven subgroups were created: TB high-burden country (yes or no), study design type (cohort or not), age (adults or not), IP-10 method (multiplex cytokines assay or ELISA), IP-10 condition (unstimulated or stimulated), and HIV-infection status (yes/some or no).
The Deeks test was used to assess publication bias using the “midas” command [36]. No publication bias existed when studies evenly distributed on the sides of regression line or P value exceeded 0.05 in Deeks’ funnel plot.
The whole process of data analysis was described in Supplementary File.