Introduction

The World Health Organization (WHO) has identified the need for a non-sputum-based triage test to rule out active tuberculosis (ATB) disease [1]. The WHO consensus meeting report describes that such a triage test should have 90% sensitivity and 70% specificity at minimum to end tuberculosis (TB) by 2035 [1]. In clinical practice, a triage test to rule out ATB requires high negative predictive value (NPV). WHO has also described the need for a test to predict progression from latent TB infection (LTBI) to ATB with >75% specificity and >75% sensitivity [2]. Further, the Foundation for Innovative New Diagnostics (FIND) and the New Diagnostics Working Group of the Stop TB Partnership have proposed a need for a prognostic test for TB risk that requires a positive predictive value (PPV) > 5.8% at a 2-year cumulative incidence of ATB of 2% (http://www.finddx.org/wp-content/uploads/2016/05/TPP-LTBIprogression.pdf).

Sputum culture is considered the gold standard for ATB diagnosis but takes 6–7 days for a positive diagnosis and up to 42 days for a confirmed negative diagnosis. Current sputum-based tests in clinical practice (e.g., smear microscopy, culture, and PCR-based assays) do not meet the desired target product profiles (TPPs), lack the sensitivity to reliably distinguish ATB from LTBI, and are prone to producing false negative results because sufficient bacilli-containing sputum samples can be difficult to obtain, especially from children and from individuals co-infected with HIV [3–12]. Sputum-bacilli tests cannot be used to identify patients with a high risk of progression because diagnosis with ATB is defined by the presence of bacilli in sputum [12]. A stool-based diagnostic test for ATB in children was shown to have 31.9% sensitivity at 99.7% specificity. These performance statistics are well suited for a diagnostic test for ruling in a patient with ATB, but not for a triage test to rule out ATB or for a progression test [13].

Recently, diagnostic gene signatures based on host immune response have been repeatedly demonstrated to accurately distinguish infection from other non-pathogenic inflammatory conditions [14], and to distinguish bacterial and viral infections [15,16]. Particularly for ATB, several host-response-based gene signatures have been proposed over the last decade for distinguishing patients with ATB from healthy controls and patients with LTBI and other diseases (ODs), and to predict progression from LTBI to ATB [17–29]. Collectively, these studies profiled whole blood or peripheral blood mononuclear cells (PBMCs) from samples that span a broad range of clinical conditions, including different age groups (children, adolescents, and adults), infection types (LTBI and ATB), and control (noninfectious) conditions. The number of genes in these signatures varies dramatically [17,20,23]. Notably, a variety of computational techniques—including support vector machine, random forest, linear discriminant analysis, logistic regression, and difference of means—have been applied to identify these signatures.

Despite extensive efforts, none of these TB gene signatures has been translated into a point of care (POC) diagnostic for several reasons. First, none of these signatures except 1 has been validated in prospective independent cohorts. As more gene signatures have been described, there has been a dearth of studies comparing these signatures with each other to verify if host-response-based signatures are appropriate for translation to clinical practice. Second, a majority of these gene signatures are composed of a large number of genes. Although a higher number of genes in a signature tends to increase accuracy [24], implementation of such a signature as a simple and cost-effective POC test is very difficult using current technology for measuring gene expression. Virtually all commercially available platforms for measuring transcriptional host response are limited by the number of genes they can measure. For instance, the Cepheid GeneXpert system, arguably the most widely used platform in TB diagnostics, can measure only up to 10 genes. Also, 1 or 2 of these genes need to be control genes, which further reduces the number of genes that can be used in a diagnostic signature to 8 or 9. Third, and most importantly, the generalizability of these transcriptional signatures to real-world patient populations in various clinical contexts is questionable. For instance, a 16-gene signature for predicting progression from LTBI to ATB developed using a single cohort from a single country was shown to lack generalizability to cohorts from other countries on the same continent [29]. In contrast, a 3-gene signature developed using heterogeneous cohorts from multiple countries has been shown to be more generalizable to other countries in retrospective [17] and prospective validation [11,30]. Another factor with substantial impact on the generalizability of these signatures is that the different statistical models used for creating these signatures are difficult to generalize across different populations and different measurement technologies. For instance, models based on K-nearest-neighbors clustering are difficult to generalize due to high sensitivity to batch effects and scaling within data [31].

The proliferation of host-response-based gene signatures despite the challenges described above raises several questions. First, do these signatures perform similarly to each other in different clinical contexts in different patient populations? If yes, the second question is, do 1 or more of these signatures have the potential to move towards translation into clinical practice cost-effectively? Third, an overarching question is, does host response to Mycobacterium tuberculosis (Mtb) have the potential to achieve the generalizability required to be used as a non-sputum-based triage test that meets the TPPs described by the WHO and other groups for ending TB by 2035?

We sought to answer these questions through a systematic comprehensive analysis of gene signatures for diagnosis of TB. To estimate the diagnostic accuracy of each of the published signatures, we reconstructed the classification model associated with each gene signature using the same discovery cohort as the original publication to the best of our ability. We then evaluated the accuracy of each signature in distinguishing patients with ATB from those with LTBI or ODs and healthy controls using publicly available gene expression datasets. We also evaluated whether these signatures could predict progression from LTBI to ATB prior to sputum conversion. We used specificity, sensitivity, comparison with various TPPs for ending TB by 2035, weighted mean area under the receiver operating characteristic curve (AUROC), PPV, NPV, false positive rate (FPR), heterogeneity across datasets, and diagnostic odds ratio (DOR) to evaluate the accuracy of various gene signatures.

Methods

Gene signatures for comparison

In June 2018 we performed an extensive search of published gene signatures for distinguishing patients with ATB from healthy controls or patients with LTBI or ODs. We searched the National Center for Biotechnology Information (NCBI) repository of publications (PubMed) for all publications describing a gene signature for the diagnosis of ATB. We included all blood-based gene signatures that were specifically designed to diagnose ATB. We did not exclude any studies because of study criteria or date. Search terms included the following: TB (tuberculosis) gene signature, TB (tuberculosis) transcriptional signature, and TB (tuberculosis) diagnostic. We identified 11 publications describing 16 gene signatures for the diagnosis of ATB (Table 1) [17–27]. We note that we only considered gene signatures that were described for diagnosis of ATB compared to healthy controls or patients with LTBI or ODs. For instance, we did not include the 4-gene RISK4 signature by Suliman et al. [29] or the 16-gene correlates of risk signature by Zak et al. [28] as both signatures are designed to predict progression from LTBI to ATB, not to diagnose ATB as a triage test.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Gene expression signatures compared within this study.

https://doi.org/10.1371/journal.pmed.1002786.t001

Transcriptome datasets used for comparing signatures

We searched NCBI GEO and European Bioinformatics Institute (EBI) ArrayExpress in June 2018 using the following search terms: TB (tuberculosis) gene expression, TB (tuberculosis) microarray, TB (tuberculosis) blood microarray, TB (tuberculosis) RNAseq, TB (tuberculosis) blood RNAseq, TB (tuberculosis) peripheral blood mononuclear cells gene expression, TB (tuberculosis) peripheral blood mononuclear cells microarray, and TB (tuberculosis) peripheral blood mononuclear cells RNAseq. We included all datasets that measured transcriptomes from the blood of patients with ATB and at least 1 other group of individuals. We did not exclude datasets based on collection date or sample number. We excluded datasets profiled using quantitative PCR because they did not have enough coverage to capture all genes across the 16 signatures included in our analysis. We identified 24 datasets containing 3,083 transcriptome profiles from whole blood or PBMCs of patients with ATB and healthy controls or patients with LTBI or ODs from 14 countries through an extensive search of 2 public data repositories (NCBI GEO and EBI ArrayExpress) (Table 2). These datasets also include the 8 datasets that were used to derive the 16 gene signatures. We used these 24 datasets to compare each of the 16 gene signatures for their ability to distinguish ATB from all other groups (healthy controls, LTBI, and ODs). In most instances the method of diagnosis of patients with ATB could be confirmed (as indicated under “diagnosis method” in Table 2); in some cases the method of diagnosis could not be identified from the information available publicly. The datasets included in this analysis were collected in 14 different countries and measures on 18 different platforms. In total these datasets include 3,083 individuals, of whom 944 were patients with ATB at the time of sample collection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Transcriptome datasets used for comparison of 16 gene signatures for diagnosis of ATB.

https://doi.org/10.1371/journal.pmed.1002786.t002

Results

Our systematic search of the literature for published transcriptional signatures diagnosing ATB against other clinical conditions identified 16 transcriptional signatures (Table 1) that distinguished patients with ATB from 1 or more of the following: healthy controls, patients with LTBI, or patients with ODs. Next, we searched 2 public data repositories (NCBI GEO and EBI ArrayExpress) for gene expression datasets that profiled whole blood or PBMC samples comparing patients with ATB to healthy controls or patients with LTBI or ODs. We identified 24 independent datasets consisting of 3,083 transcriptome profiles from 14 countries (Table 2). Note that 8 of these 24 datasets were used to derive 1 or more of the 16 gene signatures. Therefore, in order to ensure that the discovery cohort(s) of each signature did not bias its overall performance, we removed the corresponding discovery cohort(s) for each signature when computing the overall performance of each signature across all datasets. For example, for the Verhagen10 signature, we removed GSE41055 when estimating the overall AUROC and PPV, whereas for Sweeney3 we removed GSE19491, GSE37250, and GSE42834.

Overall, 630 genes were described across the 16 signatures, with the number of genes in a signature ranging from 3 [17,22,23] to 393 [20] (Fig 1). A majority of the genes (81%) were in only 1 signature. Every gene signature partially overlapped with at least 1 other signature. One gene signature (Maertzdorf4) did not include a unique gene as it was derived from another gene signature (Maertzdorf15).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Distribution of genes across the signatures included in this study.

Each row represents a gene signature for active tuberculosis diagnosis. Each column represents 1 gene. The number at the end of a signature name represents the number of genes in the given signature. Genes present in only 1 signature are red; those in 2 or more signatures are blue.

https://doi.org/10.1371/journal.pmed.1002786.g001

Comparison of accuracy across only datasets with culture-confirmed diagnosis of ATB

For each of the 16 gene signatures, we built a classification model by following the methods described in the corresponding paper to the best of our ability. We found that the classification model we created for each signature closely reproduced the AUROC reported in the corresponding paper, though not exactly (S1 Table). We applied these “locked” models “as is” to other datasets to assess the generalizability of each model for accurately distinguishing patients with culture-confirmed ATB across independent datasets.

Across the 9 independent datasets with culture-confirmed ATB, 11 gene signatures had weighted mean AUROC > 0.8, which suggests that many of the host-response-based signatures tended to have generalizability (Table 3). However, 2 gene signatures (Verhagen10 and Anderson51) had weighted mean AUROC ≤ 0.6. All but 2 gene signatures (Berry86 and Berry393) had high NPV (>98% at 2% prevalence) in culture-confirmed datasets. Signatures with low AUROC or NPV (Verhagen10, Leong24, Anderson51, Berry86, and Berry393) suggest that several signatures did not generalize to independent cohorts with culture-confirmed diagnosis of ATB. Arguably, the lack of generalizability in these signatures may be expected for several reasons. First, some signatures were derived using samples from children (Verhagen10 and Anderson51). Second, a signature may not have been derived to be generalizable. For instance, Anderson et al. [19] described 2 gene signatures (Anderson42 and Anderson51) from the same dataset of children under 5 years of age for distinguishing patients with ATB from patients with LTBI (Anderson42) or with ODs (Anderson51), which may not generalize to adults. In contrast, the Sweeney3 signature, which was derived using 3 independent cohorts of adults, generalized to children (age ≤ 5 years) and adolescents (age 12–18 years). Across all culture-confirmed datasets, the Sweeney3 and Sambarey10 signatures had the highest accuracy in distinguishing patients with ATB from healthy controls and patients with LTBI or ODs (specificity 74% at 90% sensitivity). Both signatures were the only signatures to meet the minimal WHO TPP for a triage test in cohorts of patients with culture-confirmed diagnosis of ATB.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Weighted mean AUROC, specificity at 90% sensitivity, and NPV at 2% prevalence for ATB versus all other conditions across all datasets and across only culture-confirmed datasets for each of the 16 gene signatures.

https://doi.org/10.1371/journal.pmed.1002786.t003

Comparison of accuracy for diagnosis of ATB using bivariate meta-analysis

The comparison of the 16 signatures can also be performed as a bivariate meta-analysis by combining sensitivity and specificity from diagnostic tests across different datasets. We used the R package mada to compare the 16 signatures by computing their DOR, heterogeneity in DOR, and overall FPR. A clinically useful generalizable triage test should have low FPR and high DOR with no heterogeneity across characteristics of patient populations such as genetic background of host, Mtb strain, age, HIV co-infection, and bacillus Calmette–Guérin vaccination.

Four signatures (Sweeney3, Kaforou52, Kaforou44, and Anderson51) had no heterogeneity irrespective of what datasets were used for analysis (only culture-confirmed or all datasets); the remaining 12 signatures showed heterogeneity (Table 4). Among the 4 signatures with no heterogeneity, the Sweeney3 signature had the highest DOR with the lowest FPR (Table 4) irrespective of what datasets were used, and was the only signature using fewer than 10 genes. Interestingly, when only culture-confirmed datasets were used, the daCosta3 signature had the highest DOR (32.44), with no heterogeneity and a 26% FPR. However, when datasets with clinical diagnoses of ATB were included, the daCosta3 signature had increased heterogeneity (13.63%) and a very high FPR (45%). On the other hand, the Maertzdorf15 signature had substantial heterogeneity (19.07%) when using only culture-confirmed datasets, which decreased to no heterogeneity when datasets with clinical diagnoses were included, without substantial changes in DOR or FPR. These changes in heterogeneity and FPR depending on which datasets are used in the analysis further suggest that certain signatures may not be generalizable to broad patient populations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Comparison of 16 gene signatures for diagnosis of ATB using bivariate meta-analysis.

https://doi.org/10.1371/journal.pmed.1002786.t004

Signature performance predicting progression 6 months prior to ATB diagnosis

Predicting progression from LTBI to ATB prior to sputum conversion is an important step in reducing overall incidence of ATB. Previously, a 16-gene (CoR [28]) and a 4-gene (RISK4 [29]) signature have been described to identify individuals with a high likelihood of progression. CoR was derived from and validated in the Adolescent Cohort Study (ACS; GSE79362) with validation AUROC = 69% with 66% sensitivity and 81% specificity. However, this signature was shown to have poor generalizability in the GC6-74 cohort from other African countries in a follow-up study. Therefore, we excluded CoR from our comparison. The second signature, RISK4, was derived from the GC6-74 cohort and was shown to have AUROC = 69% in the ACS [29]. Using the ACS, we compared the 16 gene signatures for their ability to predict progression from LTBI to ATB 6 months prior to sputum conversion. Because CoR was derived from the ACS cohort and shown to be not generalizable to cohorts from other African countries, we excluded the CoR signature from the comparison. We also excluded the RISK4 signature from further comparison as its performance characteristics are previously described in the ACS cohort [29].

Seven out of 16 signatures had AUROC > 0.8 and PPV > 5.8% at 2% prevalence in the ACS cohort (Table 5). Higher numbers of genes in a signature again did not correspond to a substantial increase in the AUROC. Only 2 of these signatures (Sweeney3 and Jacobsen3) had fewer than 10 genes. Interestingly, although the daCosta3 signature had the highest PPV (14.60%) for predicting progression from LTBI to ATB, it had AUROC = 0.56. Sweeney3 had the lowest number of genes with the highest AUROC (0.86, 95% CI 0.78–0.93) and PPV (13.6%), which exceeded the FIND TPP for the progression test. Overall, these results demonstrated that, in patients with LTBI, host response is able to identify those at high risk of progression to ATB prior to sputum conversion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. AUROC, PPV, and NPV for progression from LTBI to ATB in the ACS cohort up to 180 days prior to diagnosis.

https://doi.org/10.1371/journal.pmed.1002786.t005

Discussion

In this study, we compared 16 gene signatures for distinguishing patients with ATB from healthy controls or patients with LTBI or ODs using 24 independent datasets of >3,000 whole blood or PBMC transcriptome profiles from 14 countries. Collectively, these datasets represented real-world heterogeneity observed in patients with TB. For instance, the samples collected across 14 countries represented diversity in both host and pathogen genetics. Similarly, some datasets profiled samples from children whereas others profiled samples from adults, which represented heterogeneity in host response due to age. These data also represented heterogeneity in clinical practice as patients were diagnosed using different criteria (e.g., sputum culture versus sputum microscopy).

Across these biologically and technologically heterogeneous data, our comparison found that several gene signatures distinguished patients with ATB with moderate to high accuracy, although almost all signatures included a large number of genes, which severely restricts their ability for cost-effective translation to clinical practice at the POC. Importantly, our analysis found that a higher number of genes in a signature did not translate into higher accuracy across biologically heterogeneous data. Only 2 gene signatures (Sweeney3 and Sambarey10) satisfied the WHO TPP for a non-sputum-based triage test to identify which patients need further testing for confirming ATB, when comparing the signatures using only datasets from patients with culture-confirmed diagnosis of ATB.

When we included additional datasets that diagnosed ATB using other means (e.g., sputum microscopy or clinical diagnosis), the accuracy of 13 signatures decreased such that no signature satisfied the WHO TPP for a non-sputum-based triage test. For the 2 signatures that satisfied the WHO TPP for a non-sputum-based triage test when using only culture-confirmed datasets, the reduction in the AUROC and specificity of Sweeney3 was minimal (AUROC = 0.85, specificity = 66%, and sensitivity = 90%), whereas Sambarey10 had a substantial reduction of 15% in specificity. It is possible that inclusion of patients with ATB that was not diagnosed using positive culture caused the reduction in accuracy and may underestimate the accuracy of these signatures.

The inclusion of children in our analyses could have decreased the overall performance regarding misclassification of cases because of the challenges in diagnosis of ATB in children. GSE39939 and GSE39940, both of which were part of the same study [19], contained 157 and 334 samples (491 samples total), respectively, suggesting that the number of children in these datasets was sufficient. Out of the 491 samples, 190 samples were from children with ATB, of which 44 patients with ATB were sputum-negative; the remaining 146 children with ATB were sputum-positive. When using children with sputum-positive ATB, we did not see a substantial decrease in overall performance compared to adults. However, when children with clinically diagnosed or culture-negative ATB were included in our analysis, every gene signature had lower accuracy, similar to when including adults with clinically diagnosed or culture-negative ATB. These results suggest that gene signatures for diagnosis of ATB based on host response are not substantially affected by age, but by the possible inaccuracy of culture-negative clinical diagnosis of ATB.

We found that 7 signatures (Sweeney3, Kaforou27, Kaforou52, Jacobsen3, Anderson42, Sambarey10, and Kaforou44) identified adolescents with LTBI who progressed to ATB up to 6 months prior to sputum conversion. Among these signatures, only Sweeney3 has been prospectively validated in an active screening cohort to further demonstrate that host response to Mtb is detectable in blood samples earlier [13]. These results further suggest that host-response-based gene signatures could have substantial impact on the diagnosis of incipient TB, which is defined as an asymptomatic phase with early disease [35]. Incipient TB may last for up to 1 year approximately, during which a patient may be intermittently infectious by shedding bacilli in the sputum. Although the exact definitions of LTBI and incipient TB may be different, patients with LTBI progressing towards ATB in principle are similar to patients with incipient TB. Our results suggest that host-response-based gene signatures should be further explored for diagnosis of incipient TB in larger cohorts.

The ability of gene signatures for diagnosis of ATB to predict progression from LTBI to ATB prior to sputum conversion and to diagnose ATB in active screening is additional indirect evidence that suggests our estimates of accuracy for each gene signature may be underestimates. Many of the samples labeled as LTBI but classified as ATB may be progressors with subclinical ATB. Collectively, these results highlight the need for assessing the host-response-based TB diagnostics in larger prospective cohorts.

An important contributing factor to the lower generalizability of several signatures is likely the choice of underlying classification model. For instance, signatures using K-nearest-neighbors clustering as a model had the overall worst performance because of the lack of co-normalization of data across datasets and platforms. Models based on K-nearest-neighbors clustering benefit from co-normalized data; however, it is impractical, and may be very difficult, if not nearly impossible, to co-normalize data from different clinics using different technologies.

Overall, when considering the feasibility of translating a gene signature as a POC test (e.g., number of genes, required specificity at 90% sensitivity, and robustness across datasets from different geographic regions and clinical contexts), the Sweeney3 signature consistently ranked among the best signatures. The signature has also been prospectively validated in at least 2 independent cohorts using reverse transcription PCR [11,30]. Collectively, the Sweeney3 signature has been now shown to (1) predict progression to ATB 6 months prior to sputum conversion, (2) distinguish ATB in active screen, (3) track treatment response, and (4) stratify patients with ATB at the time of diagnosis with high likelihood of subclinical ATB after treatment. This robustness of Sweeney3 may be due to the fact that it was derived using 3 independent cohorts that represented broad biological and technical heterogeneity. This is in line with the observation by Suliman et al. [29] that a gene signature derived using a homogeneous cohort from 1 country was not broadly applicable to patients from other countries with similar genetic background. It is important to note that Sweeney3 is the only prospectively validated signature among the 16 signatures compared here. Importantly, the ability of the Sweeney3 gene signature to identify patients with ATB in different clinical contexts and achieve the WHO TPP using only 3 genes provides strong evidence for the potential of host-response-based diagnostics to impact clinical practice.

Our analysis has a few limitations. First, we did not have access to the exact published models, or any hyper-parameters used to build the models, for some of the gene signatures compared here. Therefore, we re-implemented these models to the best of our ability by following the details in the corresponding papers. Hence, in the process of trying to build models that reproduced as closely as possible the results reported in the corresponding paper, the choices we made and hyper-parameters we inferred may have been different from the those in the original models. This could have resulted in overfitting, which in turn may have resulted in reduced generalizability of models in independent cohorts and underestimation of their accuracy. We recommend that when diagnostic signatures are published, the corresponding models should be made available, along with a list of hyper-parameters and coefficients to enable reproducibility and comparison between models. Second, for some studies we were not able to use the original training data as they were not available. We chose another dataset that was similar to the discovery cohort described in the corresponding paper. This choice again may have resulted in underestimation of accuracy. Therefore, if a model was extremely sensitive to training data, overfitting may have happened. This limitation points to the need for sharing underlying data used for building a classification model. Third, none of the datasets used in our analysis included patients with nontuberculous mycobacteria infections. Therefore, it is not possible to evaluate whether the signatures compared here can differentiate patients with nontuberculous mycobacteria infections or ATB. Fourth, LTBI was defined using either a tuberculin skin test (TST) or IGRA, which could have different implications for progression to ATB. It is possible that patients identified as having LTBI using different diagnostic criteria could have different transcriptome profiles. Our results showed that a few gene signatures demonstrated consistently high accuracy across datasets irrespective of how LTBI was defined, suggesting that host response to ATB is sufficiently different and robust to overcome the heterogeneity in clinical practice of how LTBI is defined. Importantly, our work described here points to future studies of how existing data could be used to identify differences in the transcriptomes of patients with LTBI diagnosed with TST or IGRA.

Despite the limitations in re-implementation of current ATB diagnostic models, our comparison of 16 blood-based gene signatures strongly suggests the potential of using host-response-based gene signatures for diagnosis of ATB. The fact that a subset of signatures perform with clinically useful accuracy across multiple datasets with no heterogeneity further suggests that they should be explored in larger prospective cohorts for estimating their impact on clinical practice, instead of creating more gene signatures. Further studies should compare these signatures to investigate whether they are correlated with each other (identify the same patients or miss the same patients). If different signatures correctly diagnose different patient populations, it may be advisable to integrate these signatures in a single diagnostic model. However, before validating these signatures in prospective cohorts, they must be published and “locked” for other researchers to investigate such as we have done here. It is highly unlikely that any of these signatures will be measured using RNA sequencing or microarrays in resource-poor areas where TB is prevalent. Therefore, the prospective studies for these signatures should be performed using technologies that are cost-effective when used on a large scale. When host-response-based ATB diagnostics are validated in prospective studies, they should be designed to facilitate the identification of a threshold that can be used in clinical practice. Future prospective trials will also have to understand whether 1 threshold is sufficient across different clinical contexts (e.g., progression from LTBI to ATB, active screening for ATB, and treatment response) or multiple, better-tuned thresholds for each clinical context are needed.

Supporting information