In total, 1349 records were identified from our literature searches (Figure 1). After removing 623 duplicates, we read titles and abstracts and excluded 682 records. An additional 447 records were non-eligible for various reasons (e.g., studies involving leprosy, Crohn's disease, pneumonia, monocyte chemotactic protein-1, interleukin-12, and interleukin-18), 73 records were animal experiments (mouse, calves, warthogs, etc.), 69 records were reviews, abstracts, and letters, 58 records focused on extra-PTB (pleural TB, TB meningitis, osteoarticular TB, etc.), and 5 records were non-English (Chinese, Russian, Polish, etc.). Then, we reviewed the full texts of 44 articles. Ultimately, 18 articles were included in this study.
Characteristics of included studies
The main characteristics of the 18 articles, comprising 24 trials, are listed in Table 1 and 2  [13-29]. In total, 2836 participants were involved. The year of publication ranged from 2012 to 2018. Nine (50%) studies were from TB high-burden countries (China, South Africa, India, Thailand, and Uganda), and nine (50%) studies were from TB low-burden countries. Study design, TB site, non-TB status, IP-10 cut-off, and reference standards are summarized in Table 1. IP-10 method, condition, HIV-infection status, cut-off values, sensitivity, specificity, TP, FP, FN, and TN of IP-10 for each trial are shown in Table 2.
Quality of included studies
The QUADAS-2 tool reflects the methodological quality of included articles (Supplementary Figure S1). Patient selection bias was unclear for five studies; one study used a case-control design  and four studies did not report the time and consecutiveness of patient enrolment [17, 21, 23, 27]. Additionally, 50% of studies had unclear bias in index tests; in particular, we could not determine whether the results were interpreted in blind conditions [7, 18, 20, 23-25, 27-29]. One study had high risk of bias in the reference standard, which was clinical PTB by clinical presentation and radiological confirmation . Flow and timing bias were unclear in three studies, in which patients were lost in the analysis [19-21]. The applicability concerns were generally low.
A total of 2836 participants, comprising 3633 blood samples were included. The sensitivity for IP-10 was 0.86 (95% CI: 0.80–0.90) and the specificity was 0.88 (95% CI: 0.82–0.92). The pooled PLR was 7.00 (95% CI: 4.76–10.30), and the pooled NLR was 0.16 (95% CI: 0.12–0.23). The pooled DOR was 43.01 (95% CI: 25.80–71.69), indicating that the discriminatory effect of IP-10 was good. Figure 2 shows the HSROC curves for IP-10.
As shown in Table 3, heterogeneity was assessed by a meta-regression analysis. Heterogeneity was not detected with respect to TB high-burden versus TB low-burden countries (P = 0.83), cohort versus other study design types (P = 0.55), adults versus children (with or without adults) (P = 0.59), multiplex cytokine assay versus ELISA to detect IP-10 (P = 0.73), IP-10 stimulation or not (P = 0.72), and HIV infection or not (P = 0.53).
Deeks’ funnel plot showed no statistical significance (P = 0.20), indicating no striking publication bias in this study (Supplementary Figure S2).