The study was conducted in accordance with the PRISMA 2020 statement (9) and the protocol was registered on the International Platform for Systematic Reviews (ID: CRD42023400281).

We conducted a comprehensive search in the four major databases: PubMed, Embase, Web of Science, and The Cochrane Library, using the terms “tuberculosis” and “high-throughput nucleotides” and combine them with free words for a comprehensive search. The search was independently performed by two researchers, Yulian Li and Wentao Bian, with the consolidated results being aggregated. The full search strategy can be found in Supplementary Material S1.

Inclusion criteria encompassed randomized trials, cohort or case–control studies, that utilized the metagenomic next-generation sequencing platform (mNGS) on clinical specimens with non-tuberculosis specimens serving as controls. We excluded animal experiments, reviews, conference summaries, case reports, non-English literatures and non-SCI academic journals. There were no restrictions on the year of publication.

Eligible populations included: (1) individuals suspected to be infected with the Mycobacterium tuberculosis complexes by clinicians; (2) Individuals who collect specimens prior to antituberculosis treatment as much as possible; (3) patients with data available through electronic medical records (routine data, such as sex, age, bacteriological or imaging examination, and whether they received anti-tuberculosis drug treatment). The index test was the detection of DNA in clinical specimens from suspected tuberculosis patients via the mNGS platform. Reference criteria included clinical symptoms, imaging reports, bacteriological reports, and anti-TB drug efficacy. According to guidelines jointly developed by the United States Centers for Disease Control and Prevention and the American Society for Infectious Diseases (10), specimens from patients with positive results other than mNGS were considered TB patient specimens. Specimens from patients with negative results on all tests except mNGS were considered non-tuberculosis patient specimens.

Two researchers (YL, WT) independently screened search results based on titles and abstracts. Any disagreements were resolved by a third researcher (JZ).

A 2 × 2 table was used to extract data from the studies: true positive, false positive, true negative, and false negative. Additional extracted data included country, research type, sample size, sequencing platform, sample type, and average age of patients. The process was carried out independently by two researchers (YL, WT), with any disagreements resolved by a third researcher (JZ).

The risk of bias in diagnostic accuracy was assessed using the QUADAS-2 tool (11). One researcher (YL) independently evaluated all included studies. This assessment was reviewed by another reviewer (WT), and any disagreements were resolved by a third researcher (JZ).

The primary outcome was a summary of the combined sensitivity and specificity of all studies. Secondary outcomes included subgroup analysis based on sequencing platform, diagnostic criteria, study type, sample size, and sample types based on pre-collected information. The grouping information was defined as different subgroups and added to the 2 × 2 table. Choose random or fixed effects models depending on the magnitude of heterogeneity. To evaluate secondary outcomes, at least four studies were required to calculate the sensitivity and specificity of previously defined subgroup variables. The Q-test p-value in the forest map was used to test for heterogeneity, with p < 0.05 considered statistically significant. I² was used to describe the size of heterogeneity, and I² > 50% was considered significant heterogeneity.

The subjects of the diagnostic studies were each sample collected from patients suspected of infection (including multiple specimens from the same individual). A receiver operating characteristic (ROC) curve was plotted to assess the average accuracy across all possible thresholds. A diagnostic value was considered high if the area under the curve (AUC) was greater than 0.9. To assess the stability of the results, a sensitivity analysis was performed using a combination of sensitivity and specificity after removing studies at high risk of bias. In addition, a Feigen diagram, plotted according to the disease prevalence rate, was used to evaluate the real-world effect of the diagnostic study. Due to the exclusion of certain studies, discrepancies may exist between the published studies and those not included in the analysis. To account for this, a publication bias test was performed using the funnel chart. p-values greater than 0.05 were considered not to have statistically significant publication bias. STATA14 software was used to analyze the data.

A total of 3,458 articles (comprising 1902 unique records) were retrieved from the database. Following a thorough review of these articles, 17 were included in the meta-analysis (12–28) (Figure 1). The basic characteristics of all included studies are shown in Table 1. All studies were conducted in China, with 13 retrospective studies and 4 prospective studies. The age range of tuberculosis patients in these studies was between 30 and 60 years old. Most studies were sequenced using the BGISEQ platform, with the read length ranging from 200 to 300 bp. Several studies used multiple specimens from the same patient to evaluate sensitivity and specificity. The specimens are divided into pulmonary specimens and extrapulmonary specimens. The study of pulmonary specimens focuses on bronchoalveolar lavage fluid, and extrapulmonary specimens mainly focus on cerebrospinal fluid. Negative specimens for evaluation included those infected with other pathogens or non-infectious specimens, which were identified through diagnostic methods other than mNGS. The final reference criteria was made by clinicians according to clinical guidelines, which required at least one positive indicator except for mNGS.

Figure 2 presents the results of the risk of bias assessment. Eleven studies exhibited a low risk of bias, whereas 6 demonstrated a high risk of bias.

A total of 3,205 specimens were collected from the 17 studies. The combined sensitivity and specificity of the summarized primary results were 0.69 [0.58–0.79] and specificity 1.00 [0.99–1.00] respectively (Figure 3). Because there were 5 studies with a high risk of bias in the patient domain, 3 studies with a high risk of bias in the test domain, 3 studies with an unknown risk of bias in the reference domain, and 7 studies with unclear case flow and timing domains or High risk and large bias, moderately conclusive evidence downgraded (12, 13, 15, 19, 22–26). Supplementary Table S2 shows the subgroup analysis of potential sources for evaluating sensitivity and specificity. The results indicated that mNGS had high sensitivity and specificity but significant heterogeneity (p < 0.05). In the subgroup analysis of specimen types, the sensitivity of pulmonary specimens (0.75) > extrapulmonary specimens (0.61). Interestingly, when the specimen was BALF (bronchoalveolar lavage fluid), the sensitivity was (0.75), which may be due to the fact that BALF was the main clinical specimen in the study of pulmonary specimens. The sensitivity was lowest (0.58) when the specimen was CSF (cerebrospinal fluid). The sensitivity was highest (0.78) when studies with sample sizes <100 were included. However, the sensitivity was lowest when all prospective studies were included for analysis (0.58). Regarding specificity, in the subgroup analysis of specimen types, the specificity was (0.99–1.00) with a slight heterogeneity [48.84–66.52]. When specimens were BALF, specificity heterogeneity was not significant (I² = 48.84, p = 0.08). Heterogeneity in specificity was not significant when all studies with less than 100 specimens were included, although the results require caution (I² = 0.00, p = 0.66). To conclude, the heterogeneity of specificity (0.99) was not significant when all studies using the BGISEQ-50 platform were included (I² = 8.49, p = 0.36).

All subgroup sensitivities showed significant heterogeneity according to previously defined subgroup variables: sequencing platform, diagnostic criteria, study type, sample size, and specimen type, and in terms of specificity, all subgroup specificities showed significant heterogeneity except for the BALF, BGISEQ-50, and sample < 100 subgroups, which may be the source of heterogeneity. In terms of measuring average accuracy at all possible thresholds, the receiver operating characteristic (Figure 4A) area under the curve was 0.98 (>0.9). The publication bias plot was symmetrical as a whole (Figure 4B), and the publication bias was not significant (p = 0.89). Finally, even with the exclusion of six studies with a high risk of bias, heterogeneity in sensitivity and specificity persisted.

According to the World Health Organization’s reported TB incidence of 0.055% in 2022 (29), the accuracy of Fagan’s chart sensitivity was 8%, and the specificity error rate was 0% (Supplementary Figure S2A). When using an environment with a prevalence of 10%, the accuracy of sensitivity was 94% and the specificity error rate was as low as 3% (Supplementary Figure S2B).