Diagnostic accuracy of Mycobacterium tuberculosis cell-free DNA for tuberculosis: A systematic review and meta-analysis

Background Diagnosis of tuberculosis (TB) is still difficult. The purpose of our study was to evaluate the diagnostic accuracy of Mycobacterium tuberculosis cell-free DNA (cfDNA) for diagnosing of TB. Methods We searched relevant databases for studies that used cfDNA to diagnose TB. We evaluated the accuracy of cfDNA compared with the composite reference standard (CRS) and culture. True positive, false positive, false negative, and true negative values for cfDNA were obtained first, then the estimated pooled sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), diagnostic odds ratio (DOR), and the area under the summary receiver operating characteristic (SROC) curve (AUC) of cfDNA for diagnosing TB were calculated with 95% confidence intervals (CIs). Heterogeneity was determined using the I2 statistic. When the heterogeneity was obvious, the source of heterogeneity was further discussed. Results We included 14 independent studies comparing cfDNA with the CRS, and 4 studies compared with culture. The pooled sensitivity, specificity, PPV, NPV, DOR, and AUC of the SROC were 68%, 98%,99%, 62%, 83, and 0.97 as compared with the CRS, respectively. The pooled sensitivity, specificity, PPV, NPV, DOR, and AUC of the SROC were 48%, 91%, 92%, 60%, 5, and 0.88 as compared with culture, respectively. The heterogeneity between studies was significant. Conclusions The accuracy of cfDNA testing for TB diagnosis was good compared with CRS and culture. cfDNA can be used for rapid early diagnosis of TB.


Eligibility criteria
Types of studies. Any type of study design, such as retrospective, prospective, or casecontrol, were eligible for inclusion provided the study assessed the efficacy of cfDNA to diagnose TB. Studies that reported only sensitivity or specificity were excluded.
Participants. Participants of any ethnicity, sex, or age were included provided cfDNA was used to diagnose and confirm TB infection.
Index tests. cfDNA was considered the index test.
Outcomes. The sensitivity and specificity of the index test were considered the primary outcome. Sensitivity is the probability that the index test will be positive in an infected patient. Specificity is the probability that the index test will be negative in a noninfected patient [18].
Comparator test. Single-arm studies were also included if their participants, index test, and outcomes met the inclusion criteria. Comparator tests were not mandatory.
Target conditions. Studies that diagnosed TB using cfDNA with clear reference standards and had their full text accessible were included. True positive (TP), false positive (FP), false negative (FN), and true negative (TN) values for the index test were extracted directly or calculated from the original studies. Studies reported in languages other than English or Chinese, those with <10 participants, conference abstracts without full articles, and case reports were excluded.
Reference standards. MTB culture or a composite reference standard (CRS) was defined as the reference standard. The CRS included MTB AFB smear, culture, clinical manifestations, radiographic changes, other nucleic acid amplification tests, immunological tests, and response to anti-TB treatment.

Literature screening and selection
The candidate studies were imported into ENDNOTE X9.2 literature management software. The same two investigators (GCY and YQS) independently evaluated the imported studies for selecting eligible articles based on inclusion criteria by reviewing their titles and abstracts and then the full text. If there was disagreement between the two investigators, they consulted with a third investigator (YS). First, we gave the controversial literature to a third researcher for independent evaluation, then the three investigators discussed and reported the reasons for inclusion or exclusion, respectively, and then the literature was included or excluded according to the inclusion and exclusion criteria after obtaining agreement.

Data extraction
We extracted the following information from each study: first author name; publication date; country; TP, FP, FN, and TN values for the index test; type of study design; patient selection method; specimen types; target gene; PCR method; specimen condition; TB types; and other parameters. Data was independently extracted by the same two investigators who screened the literature. Any disagreement was resolved by consultation with the third investigator.

Quality evaluation
Two investigators independently evaluated study quality using the revised Quality Assessment of Diagnostic Accuracy Studies tool (QUADAS-2) [19]. Any disagreement was handled as noted above. Publication bias was not assessed, as per the PRISMA-DTA guidelines [16]. The Grading of Recommendations Assessment, Development and Evaluation (GRADE) guideline was used to assess the strength of the body of evidence [20].

Data synthesis and statistical analysis
When different types of TB, target genes or PCR methods were reported in the same article, we considered them as separate studies. TP, FP, FN, and TN values for the index test in each included study were obtained first, then the estimated pooled sensitivity and specificity of cfDNA for diagnosing TB were calculated with 95% confidence intervals (CIs). Calculations were performed using Stata software version 15.0 (Stata Corp., College Station, TX, USA) with the midas command package or Meta-DiSc software version 1.4 (XI Cochrane Colloquium, Barcelona, Spain). Forest plots for sensitivity and specificity were generated using Review Manager (RevMan) software version 5.3 (Nordic Cochrane Centre, Copenhagen, Denmark). We also calculated the pooled positive predictive value (PPV), negative predictive value (NPV), diagnostic odds ratio (DOR), and the area under the summary receiver operating characteristic (SROC) curve (AUC). Heterogeneity was determined using the I 2 statistic (0% indicated no observed heterogeneity; <50% indicated minor heterogeneity; >50% indicated substantial heterogeneity) [21]. Subgroup, meta-regression, and sensitivity analyses were used to explore sources of heterogeneity when significant heterogeneity was observed. Subgroup and meta-regression analyses were performed according to TB type, specimen type, target gene, PCR method, type of study design, specimen condition, and patient selection method. At least 4 eligible studies were required to perform the meta-analysis and meta-regression analyses for predefined variable parameters; these analyses were performed using Stata software version 15.0 with the midas command package [22].

Identification of studies and study characteristics
Using our designed search strategy, 172 candidate articles were obtained from the databases. We identified twelve eligible articles according to the inclusion criteria by screening the title, abstract, and full text (Fig 1) [5,9,15,[23][24][25][26][27][28][29][30][31]. The kappa index of agreement between the two investigators for selection and data extraction was 0.785 (95% CI, 0.581-0.989). Two articles that reported sensitivity without specificity were excluded [32,33]. We also excluded three articles that did not report sensitivity and specificity [34][35][36] and one article that reported data from another included article [37]. The reference criteria used in three articles did not meet the inclusion criteria of this study and were excluded [10,38,39]. One article that simultaneously reported the efficacy of different PCR methods to detect cfDNA for PTB and EPTB diagnosis was considered as four separate studies [5], and another article that simultaneously reported the efficacy of different target gene to detect cfDNA for PTB diagnosis was considered as two separate studies [15]. Another simultaneously reported the efficacy of cfDNA for PTB and EPTB diagnosis [28]. Therefore, 18 studies were included for analysis (14 studies used CRS as the reference standard and 4 studies used MTB culture as the reference standard). Included studies were analyzed separately according to different reference standards. The included studies and their characteristics were presented in Table 1. 15 studies were conducted in TB endemic areas, 13 in China, 1 in India, and 1 in South Africa. Two studies were reported in Chinese [23,28] and the remaining ones were reported in English. All studies focused on adults. Only one study included a subset of HIV-positive patients [31]. Ten were case-control studies, seven were prospective, and one was cohort study. The specimens used in the articles were blood, urine, CSF, pleural fluid, and peritoneal fluid. The number of samples ranged from 19 to 412 and the average sample size was 124.3.

Study quality
The results of the methodological study quality assessment compared with CRS and culture were shown in Fig 2. When using CRS as the reference standard, six studies had nonconsecutive patient selection, one had unclear inappropriate exclusions, and four did not include treatment response in its reference criteria; these were the major sources of bias. The risk of bias from the index test and the flow and timing was judged to be relatively low. When using culture as the reference standard, three studies had nonconsecutive patient selection, one had high bias from the flow and timing; these were the major sources of bias. The risk of bias from the index test and the reference standard was judged to be relatively low. According to the GRADE guidelines, the quality of evidence of this meta-analysis was moderate.

Meta-regression and subgroup analyses
Studies that did not report target genes were excluded from the subgroup and meta-regression analysis. When compared with CRS, the results of the meta-regression analysis are presented in Table 2. TB type (PTB or EPTB), specimen type (plasma or non-plasma), PCR method (real-time PCR or digital PCR), and type of study design (prospective or case-control) had no effect on sensitivity and specificity of cfDNA testing when compared with CRS (P >0.05). Although target gene (IS6110 or non-IS6110 [IS1081, devR]) and specimen condition (fresh or frozen) had no effect on sensitivity (P >0.05), they did significantly affect specificity (P <0.05). Patient selection method (convenience or consecutive) significantly affected sensitivity (P <0.05), but had no effect on specificity (P >0.05). However, when compared with culture, the number of included studies was limited and meta-regression analysis could not be performed.
When compared with CRS, the case-control studies all used plasma specimens for testing, so the results of the subgroup analyses were the same for these two parameters. The sensitivity,
Sensitivity analysis did not identify specific articles as sources of heterogeneity in sensitivity and specificity.

Discussion
This meta-analysis evaluated the effectiveness of using cfDNA testing to diagnose TB. Rapid early diagnosis is essential to improve TB prognosis and control transmission. cfDNA testing had an overall sensitivity of 68%, specificity of 98%, DOR of 83, and AUC of 0.97 for TB diagnosis (including PTB and EPTB) compared with CRS, and sensitivity of 48%, specificity of 91%, DOR of 5, and AUC of 0.88 compared with culture, demonstrating excellent diagnostic performance. There was significant heterogeneity in sensitivity and specificity. The acquisition of the diagnosis in different types of TB is distinctive considering the differences in characteristics between PTB and EPTB, and even between different types of EPTB. Therefore, we analyzed the different types of TB separately.
When compared with CRS, the sensitivity, specificity, DOR, and AUC of cfDNA testing using plasma for diagnosis of PTB were 78%, 97%, 29, and 0.99, respectively; for diagnosis of EPTB, the sensitivity, specificity, DOR, and AUC of cfDNA testing were 65%, 99%, 61, and 0.98, respectively. These results showed that cfDNA testing has good diagnostic performance for both PTB and EPTB. For PTB, the sensitivity of cfDNA testing was higher compared with EPTB, while the specificity was lower; however, the overall diagnostic efficacy was similar. Respiratory specimens, such as sputum and bronchoalveolar lavage fluid (BALF), are typically used to diagnose PTB but obtaining these specimens can be difficult, particularly in children [40]. Moreover, the diagnosis of PTB is difficult when the patient does not cough up sputum or is unable to collect sputum of acceptable quality. Although it may be possible to obtain BALF via fiberoptic bronchoscopy, this invasive procedure may be poorly tolerated by patients with milder disease. The use of non-respiratory specimens such as gastric or stool specimens to diagnose PTB has also been reported but remains controversial [41,42]. Plasma-based cfDNA assay testing offers a new pathway for PTB diagnosis. Although plasma-based testing such as the interferon gamma release assay may provide indirect evidence of PTB without showing evidence of MTB [43], plasma-based cfDNA testing can provide direct evidence of MTB. Plasma specimens are easily accessible and widely applicable. For PTB patients in whom respiratory specimens are difficult to obtain, plasma cfDNA testing is a convenient and effective alternative diagnostic method. However, the diagnostic efficacy of cfDNA testing using plasma for the diagnosis of PTB was reduced when culture was the reference standard, which may be related to the inconsistent use of reference standards. On the other hand, the use of

PLOS ONE
urine testing for cfDNA might be an alternative diagnostic tool for PTB, but our study showed that the diagnostic efficacy of this method was limited when compared to culture. A recent study showed good efficacy of cfDNA testing using urine for the diagnosis of PTB, but the study used Xpert MTB/RIF as the reference standard [39], which might lead to inconsistent results. However, the studies examined in our meta-analysis focused on adults and lacked data on children. Furthermore, the number of studies that examined cfDNA using plasma and urine for the diagnosis of PTB was limited, as was the number of included patients, so our conclusions should be treated with caution. Future large-scale multicenter studies are needed to confirm the diagnostic efficacy of cfDNA testing.
For EPTB, the categories are even richer. The studies included in this meta-analysis reported three common types of EPTB: tuberculous pleurisy, tuberculous meningitis, and abdominal TB. Our results showed that cfDNA testing had the highest diagnostic value for tuberculous pleurisy, followed by tuberculous meningitis and abdominal TB. All three types are paucibacillary and require invasive procedures to obtain specimens. Their diagnosis is difficult because of the low MTB content in the specimens. Diagnostic efficacy of nucleic acid amplification testing, represented by Xpert MTB/RIF, has remained low in most current studies of paucibacillary tuberculosis. cfDNA may be another effective method to diagnose paucibacillary tuberculosis, although current evidence remains limited. The studies included in our meta-analysis used body fluids (pleural fluid, CSF, and ascites fluid) for cfDNA testing, but not plasma specimens. We further analyzed the diagnostic efficacy of cfDNA testing in EPTB using different types of specimens. Diagnostic efficacy was better with testing plasma samples compared with non-plasma samples. However, the number of included studies was small and the results should be treated cautiously. Studies that used plasma specimens for cfDNA testing did not specifically report efficacy according to type of EPTB, so the efficacy of using plasma for cfDNA testing for different types of EPTB remains unclear. Plasma specimens are easily available and more suitable than body fluid specimens that require invasive procedures to obtain. Therefore, cfDNA plasma testing shows promise as a tool for diagnosing EPTB early. However, further studies are needed. For different target genes, the diagnostic efficacy of IS6110 might be superior to that of other non-IS6110 target genes (such as IS1081, devR). However, the number of other non-IS6110 target genes included in this study was limited and this result needed to be treated with caution. Only one study included a subset of HIV-positive patients, but this study did not specifically report the effectiveness of cfDNA in diagnosing TB in HIV-positive patients [31]. The diagnostic efficacy of cfDNA in HIV-positive patients is still needs to be further investigated. No studies reported correlation between radiography image of patients and cfDNA test. There are no clear studies on the correlation between imaging results and cfDNA, and further studies can be conducted in this area in the future. Only a few studies had reported the concentration of cfDNA [15,30], but these studies did not report the effect of different concentrations of cfDNA on the results of the test. Some studies had increased the concentration of cfDNA by increasing the sample volume [30], and the sample volume might be important for cfDNA concentration. However, these are still unclear and need to be confirmed by follow-up studies.
We explored the potential sources of heterogeneity in the studies in our meta-analysis using pre-identified parameters. Meta-regression analysis showed that TB type (PTB or EPTB), specimen type (plasma or non-plasma), PCR method (real-time PCR or digital PCR), and type of study design (prospective or case-control) had no effect on sensitivity and specificity. Target gene (IS6110 or non-IS6110 [IS1081, devR]) and specimen condition (fresh or frozen) affected specificity but not sensitivity. Patient selection method (convenience or consecutive) had an effect on sensitivity but not specificity. Although subgroup analysis showed significant heterogeneity in sensitivity between studies within each subgroup, sensitivity analysis did not identify specific articles as sources of heterogeneity. However, since the heterogeneity in sensitivity was significant, the results must be treated with caution. The included studies also used different CRS references and some did not include radiographic manifestations or treatment response in their CRS criteria. This might be a source of the observed heterogeneity.
There were several limitations in our meta-analysis. First, it was based on individual independent data. Some studies may be missing, despite our best efforts. Second, some studies did not report specific TB and specimen types. Third, the number of studies was limited for some subgroup analyses.

Conclusions
This meta-analysis showed that the performance of cfDNA testing for TB diagnosis (including PTB and EPTB) was good compared with CRS and culture. There was significant heterogeneity in sensitivity and specificity. Diagnostic ability was good for both PTB and EPTB, and the overall diagnostic efficacy for each was similar. For EPTB, subgroup analysis showed that cfDNA had the highest diagnostic value for tuberculous pleurisy, followed by tuberculous meningitis and abdominal TB. Testing efficacy in EPTB was higher with plasma samples compared with non-plasma samples. cfDNA can be used for rapid early diagnosis of TB.