This review is reported in line with the Preferred Reporting Items for Systematic reviews and Meta-Analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA) ²⁴ recommendations ( Table 1). The systematic review protocol was registered with the International Prospective Register of Systematic Reviews (PROSPERO) on 02 January 2021 with registration number CRD42021224155 and published in the BMJ Open ²⁵ .

We considered cross-sectional and case-control studies, prospective and retrospective cohort studies, and randomised control trials evaluating diagnostic and/or prognostic performance of human host-blood transcriptomic signatures of TB (index tests). Eligible studies included PLHIV in the signature test and/or validation cohorts. Studies that only reported signature discovery cohort performance, or treatment response and failure monitoring cohorts, were not considered. PLHIV of all ages, ethnicities, and in all settings were considered. Studies which did not report any measures of signature performance (sensitivity and specificity, or reported data which enable the reconstruction of a two-by-two table for test accuracy calculation for PLHIV), did not clearly state the case definition of TB disease, did not report primary data, or did not independently report signature performance in PLHIV, were excluded.

The primary TB disease endpoint (target condition) was defined by a positive microbiological test, such as mycobacterial culture or the Xpert MTB/RIF assay (reference standards), in sputum or other bodily fluid sample. Microbiologically-confirmed extra-pulmonary TB disease was also considered. The secondary TB disease endpoint was defined by non-microbiologically-confirmed, presumptive TB diagnosed via composite clinical features. TB disease diagnosed within one month of the index test was presumed to be prevalent disease (diagnostic studies). Prognostic studies were defined as prospective studies in which participants were followed up for progression to incident TB disease with measurement of a transcriptomic signature from blood samples collected at enrolment. Eligible studies included healthy individuals, latent Mtb-infected individuals, or individuals with other respiratory or systemic diseases as a control group. Latent Mtb infection was defined by a positive TST or IGRA.

We systematically searched PubMed ( MEDLINE), WOS Core Collection, Biological Abstracts, and SciELO Citation Index (via Web of Science), Africa-Wide Information and General Science Abstracts (via EBSCOhost), Scopus, and Cochrane Central Register of Controlled Trials databases for full-text articles published in English between 1 January 1990 and 31 December 2020 using Medical Subject Headings (MeSH) and keyword search terms for "Diagnosis", "Messenger RNA", "Biomarkers/blood”, “Tuberculosis”, and “HIV”. The search strategy, including publication date range, were prespecified and published in a systematic review protocol ²⁵ . We reviewed reference lists of eligible articles and performed forward citation tracking using Science Citation Index (via Web of Science) to identify further articles and reports missed by the electronic database search ²⁶ .

Two reviewers (SCM and SV) independently conducted the literature search and screened the search outputs for potential inclusion using EndNote bibliographic software to manage references, as previously described ²⁷ . After removal of duplicates, the selection process included an initial screening of titles and abstracts for relevance, followed by full text review for eligibility. The two reviewers resolved any disagreements or uncertainties by discussion. Data elements of included studies were then independently extracted by the two reviewers. Corresponding authors of potentially eligible studies were contacted to provide deidentified participant-level data to reconstruct two-by-two tables or summary performance data for the PLHIV subgroup. Studies without summary or participant-level data available for the PLHIV subgroup were excluded.

We performed a narrative synthesis of the eligible study cohorts and signatures, including study design, cohort and signature characteristics, and diagnostic and prognostic performance of signatures stratified by study control groups (healthy, latent-Mtb infected, or other disease), and diagnostic reference standards (microbiological or composite clinical). Studies and cohorts were designated by the first author name and year of publication (e.g. Author2019a) and signatures by first author and number of transcripts (e.g. Author11). Signature area under the curve (AUC), sensitivity, and specificity were summarised in forest plots (R forestplot package ²⁸ ). For studies with available participant-level data for the PLHIV subgroup, we were able to recalculate AUC (R pROC package ²⁹ ), and benchmark sensitivity and specificity against the WHO TPP minimum performance criteria for a triage (70% specificity and 90% sensitivity) ¹⁶ or prognostic (75% specificity and 75% sensitivity) ¹⁸ test. 95% Confidence intervals for AUCs and sensitivity and specificity, were calculated using the DeLong ³⁰ and Wilson binomial ³¹ methods, respectively. For studies in which participant-level data were not available, we report summary AUC, sensitivity, and specificity estimates, and 95% confidence intervals, for the PLHIV subgroup as published in the original papers. Most of these estimates were not specifically benchmarked against the WHO TPP minimum performance criteria for a triage test.

The methodological quality and applicability concerns of included studies was assessed by the two reviewers using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool ³² and graphically represented using traffic-light plots (R robvis package ³³ ). Risk of bias and applicability concerns for individual study was evaluated in four domains relating to (1) patient selection, (2) measurement of the index test, (3) measurement of the reference standard, and (4) study flow and timing of investigational and diagnostic procedures. Risks of bias are reported within each domain of each study as low risk (no risks of bias identified), some concerns (one risk or unclear risk of bias identified), or high risk (more than one risk or unclear risk of bias identified). Overall risk of bias for each study was reported as low risk (no risks of bias identified in any domains), some concerns (some concerns identified in one domain), or high risk (some concerns identified in two or more domains or high risk of bias in any domain). We assessed the cumulative quality of evidence synthesised by the systematic review using the “Grading of Recommendations Assessment, Development and Evaluation” (GRADE) approach ³⁴ with classification based on study design and limitations, indirectness, inconsistency, imprecision, and publication bias ^{35, 36} .

We performed the literature search in January 2021, identifying 1,580 unique records published between 1 January 1990 and 31 December 2020, of which 98 full-text articles were assessed for eligibility, and 12 studies ^{13, 22, 37– 47} met all criteria for inclusion ( Figure 1). The main reasons for study exclusion were absence of PLHIV in study cohorts (n=29), absence of an independent test or external validation cohort which included PLHIV (n=20), and inappropriate index test (n=16) or study design (n=10). Nine of 10 studies excluded for inappropriate design were commentaries or reviews. In addition, deidentified participant-level data to reconstruct two-by-two tables, or summary performance data for PLHIV, were not available and no data were received from corresponding authors for 9 records ^{19, 48– 55} .

The 12 eligible studies included 10 independent test or validation cohorts featuring PLHIV ( Table 2), cumulatively evaluating diagnostic performance of 31 transcriptomic signatures ( Table 3) incorporating over 700 unique transcripts. All independent test and validation cohorts enrolled PLHIV from outpatient clinics or hospital inpatients, most with suspicion (symptom-positive) or high risk (initiating antiretroviral therapy) of prevalent TB ( Table 2). No studies prospectively enrolled participants without significant risk factors from community settings with a high burden of subclinical TB. Cohort sizes were generally small, with a maximum of 97 TB cases and 176 controls with HIV in the Kaforou2013 ³⁷ study, and 67 TB cases and 134 controls with HIV in the Södersten2020 ⁴⁶ study. All other cohorts included less than 50 TB cases with HIV. Only one of the 12 studies (Darboe2019) ⁴² reported transcriptomic signature prognostic performance for incident TB disease (disease recurrence after completion of TB treatment). The remainder reported diagnostic performance for differentiating patients with pulmonary TB disease from healthy controls, from controls with latent TB infection, or from other diseases. Most studies were conducted in Africa, with 8 out of 10 cohorts enrolling participants in South Africa. Only one cohort (Södersten2020) ⁴⁶ included participants from another continent (Peru, South America). Many studies used the same microarray datasets, GSE37250 ³⁷ and GSE39941 ³⁸ , from the Kaforou2013 (Malawi and South Africa) and Anderson2014 (Kenya, Malawi and South Africa) studies respectively, for external validation of signatures. Three cohorts measured signature scores using near-point-of-care RT-qPCR platforms (Darboe2019 and Penn-Nicholson2019a,b) ^{42, 45} , while only one study (Södersten2020) ⁴⁶ used a point-of-care device, the Cepheid Xpert-MTB Host-Response (HR) Prototype, to measure the Sweeney3 signature ^{19, 49} in biobanked whole blood RNA samples. No included studies validated signature performance for diagnosing extrapulmonary TB. However, one cohort (Anderson2014b) ³⁸ included 17 HIV-infected children with non-microbiologically-confirmed, presumptive clinical TB with clinical and radiologic features that prompted empirical treatment. All other study cohorts, with the exception of Anderson2014, enrolled adults and used a microbiological reference standard (culture or Xpert MTB/RIF).

All 31 signatures evaluated were discovered in case-control cohorts, eight of which were nested within prospective cohorts ( Table 3). Only one of the 31 signatures (Rajan5) ⁴⁴ was discovered in a cohort exclusively consisting of PLHIV. Seven signatures (Duffy10 ⁴³ , Gliddon3 ⁵⁵ , Gliddon4 ⁵⁵ , Kaforou27 ³⁷ , Kaforou44 ³⁷ , Kaforou53 ³⁷ , and Sweeney3 ^{19, 49} ) were discovered using the Kaforou2013 cohort ³⁷ and two signatures (Anderson42 and Anderson51) ³⁸ using the Anderson2014 cohort ³⁸ , which included participants both with and without HIV. The 31 signatures were predominantly discovered in adult cohorts, with only 6 out 31 derived in paediatric (Gjøen7 ⁵¹ , Anderson42, and Anderson51) or adolescent (Darboe11 ⁵⁶ , Penn-Nicholson6 ⁴⁵ , and Zak16 ⁴⁰ ) cohorts. Signature discovery cohorts were geographically diverse, with recruitment in Asia (3 signatures), Europe (10), North America (4), and South America (1); however there was a predominance in Africa (16), particularly South Africa (15).

The 40 most frequent transcripts, all included in 3 or more signatures, are listed in Table 4. We used the INTERFEROME database ⁵⁷ to classify interferon-stimulated genes (ISGs): We defined ISGs as genes significantly up or down regulated in expression (>1.5 fold change) in any human samples treated with Type-I IFN, relative to control samples. Almost all (38/39) of the most common transcripts with available gene annotations were classified as ISGs. The six transcripts most frequently included in signatures were Guanylate Binding Protein (GBP) 5 (11 signatures), GBP6 (10 signatures), Complement C1q B Chain (C1QB) and Fc fragment of IgG receptor Ia (FCGR1A) (7 signatures each), and Basic Leucine Zipper ATF-Like Transcription Factor 2 (BATF2) and GBP2 (6 signatures each).

In the patient selection domain, 5 of the 10 independent test and external validation cohorts consecutively or randomly enrolled participants, with the remaining cohorts not reporting sampling method ( Figure 2). Eight of the 10 cohorts utilised a case-control design ( Table 2), with exclusion of participants with uncertain diagnosis introducing a high risk of spectrum bias and potentially inflating diagnostic accuracy. The Penn-Nicholson2019b cohort ⁴⁵ used a prospective design in recruiting symptomatic clinic attendees, but excluded probable and uncertain TB cases from analysis. Only one prospective diagnostic accuracy study, Turner2020 ¹³ , measured signature scores and tested performance in all enrolled participants with clinically suspected tuberculosis, including those with uncertain diagnosis, representative of the target population of symptomatic clinic attendees.

In the index test measurement domain, transcriptomic signature scores were interpreted without knowledge of the reference standard (i.e. blinded) in the Turner2020 ¹³ and Penn-Nicholson2019b ⁴⁵ cohorts, with unclear reporting for the other studies. Due to the early stage of biomarker development and diverse signature measurement and score calculation methodologies, no studies used pre-specified signature score thresholds. The risk of bias in the reference standard domain was deemed to be low in all cohorts with use of appropriate and standardised microbiological confirmatory TB testing (Mtb culture and Xpert MTB/RIF) likely to correctly classify the target condition. Reference standard results were interpreted without knowledge of the results of the transcriptomic signature scores (i.e. blinded) for all included studies. In the study flow and timing domain, all studies used an appropriate interval between index test and reference standard sample collection, and all participants received the same reference standard tests. However, only the Turner2020 ¹³ study included all participants in analysis. In terms of applicability concerns, selection of participants, and measurement and interpretation of index tests and reference standard matched the review question for all included studies. Overall 9 of the 10 cohorts had a high risk of bias, and one study (Turner2020 ¹³ ) had some concerns due to lack of a pre-specified test threshold prospectively applied to each signature ( Figure 2).

Nine independent test or external validation diagnostic cohorts with PLHIV subgroups were included in the systematic review ( Table 2). Three cohorts evaluated diagnostic performance of 12 signatures for discriminating HIV-infected adults with prevalent TB disease from latent-Mtb infected individuals or healthy controls with HIV ( Figure 3), and 6 cohorts evaluated diagnostic performance of 29 signatures for discriminating HIV-infected adults or children with prevalent TB disease from those with HIV and other respiratory or systemic diseases ( Figure 4).

Only the Kaforou27 (Kaforou2013a and Rajan2018 cohorts), Penn-Nicholson6 (Penn-Nicholson2020a cohort), and Rajan5 (Rajan2018 cohort) signatures met the WHO TPP minimum performance criteria for a triage test for differentiating adults with prevalent TB disease from latent-Mtb infected individuals or healthy controls with HIV in these case-control studies ( Figure 3). However, the upper bounds of the 95% confidence intervals for 4 other signatures met the TPP benchmarks: Darboe11 (Penn-Nicholson2020a cohort), Duffy10 (Kaforou2013a cohort), Kaforou53 (Rajan2018 cohort), and Sweeney3 (Kaforou2013a and Rajan2018 cohorts).

Signatures which are able to distinguish PLHIV with prevalent TB disease from those with other respiratory or systemic diseases are more clinically useful in the inpatient or outpatient setting, as compared to the community. Seven out of 29 signatures met the WHO TPP minimum performance criteria for a triage test in this context ( Figure 4): Anderson51 in the Anderson2014b paediatric cohort; Kaforou44 in the Kaforou2013b adult cohort; and Maertzdorf4, Qian17, Roe1, Roe3, and Sweeney3 in the Turner2020 adult cohort. A further 8 signatures (Anderson39.OD, de Araujo1, Kaforou25, Kaforou39, Penn-Nicholson6, Rajan5, Suliman2, and Zak16) had upper 95% confidence interval bounds for sensitivity which met the 90% TPP benchmark (with specificity set at 70%) in the Turner2020 prospective diagnostic accuracy cohort. However, the PLHIV subgroup of this cohort only included 17 TB cases and 27 controls with other respiratory diseases. In other cohorts, the Anderson51 (Anderson2014a cohort), Penn-Nicholson6 (Penn-Nicholson2020b cohort), and Sweeney3 (Kaforou2013b cohort) signatures also approached the WHO minimum TPP triage test criteria. Only one signature (Sweeney3) was tested using a point-of-care device, the Cepheid Xpert-MTB-HR Prototype in the Södersten2020 cohort on biobanked whole blood RNA samples from symptomatic inpatients with HIV. Sweeney3 achieved sensitivity of 91.0% (95%CI 81.8–95.8) and specificity of 55.8% (95%CI 47.2–64.1), falling short of the WHO TPP benchmark. No signatures met the WHO TPP benchmark criteria for a non-sputum confirmatory diagnostic test in PLHIV (98% specificity and 80% sensitivity) ¹⁶ in any cohorts.

Participant-level data were available for only 3 signatures in the Anderson2014b cohort, which also included 17 children with non-microbiologically confirmed, presumptive clinically-diagnosed TB disease ( Table 2). All signatures performed poorly in differentiating TB from other diseases in this subset ( Figure 5).

Only the Södersten2020 study reported signature performance stratified by CD4 cell count, with lower Sweeney3 specificity in the 56 inpatients with CD4 cell count less than 200 (66.7%, 95%CI 46.7–82.0) as compared to 129 inpatients with CD4 greater than or equal to 200 (94.7%, 95%CI 88.9–97.5) ⁴⁶ . Conversely, sensitivity was higher in the inpatients with CD4 cell count less than 200 (93.8%, 95%CI 79.9–98.3) as compared to the inpatients with CD4 greater than or equal to 200 (75.0%, 95%CI 50.5–89.9). No participant-level data were available to perform further subgroup analyses by CD4 cell count, HIV plasma viral load, TPT or ART status, or other variables in any cohorts.

There was considerable clinical and methodological heterogeneity between cohorts, with participants recruited in diverse settings, with dissimilar eligibility criteria, and distinct composition of the control groups. Studies also used different signature measurement methods (microarray, RNA sequencing, and RT-qPCR), different methods of signature score calculation, and there were no standardised signature score thresholds. Due to the significant heterogeneity in study design and index test measurement, and limited participant data, a meta-analysis was not deemed appropriate.

Only one study (Darboe2019) ⁴² evaluated transcriptomic signature prognostic performance for recurrent TB disease in adults with HIV who had recently completed TB therapy. There were no eligible studies evaluating performance for prediction of progression to incident TB disease in PLHIV without prior TB disease. We stratified the prognostic performance of the Darboe11 signature by time from measurement until TB disease recurrence ( Figure 6). Signature prognostic performance was best within 90 days of signature measurement, and waned thereafter. Sensitivity and specificity of the Darboe11 signature did not meet the minimum WHO TPP performance criteria for a prognostic test (75% sensitivity and 75% specificity) ¹⁸ in any time window. Sample size was not sufficient to perform subgroup prognostic performance analyses, however Darboe11 signature scores were higher in individuals with detectable plasma HIV viral load (>400 copies/mL) as compared to those with an undetectable plasma viral load (<400 copies/mL; p<0.0001) ⁴² .

A total of 10 cohorts, 2 prospective cohorts (29 TB cases; 51 other respiratory diseases) and 8 case-control studies (353 TB cases; 606 controls), were included in this systematic review of the diagnostic and prognostic accuracy of host blood transcriptomic signatures in PLHIV. All studies used reliable reference standards for definitive TB diagnosis. However, we adjudged that there was a very serious risk of bias due to exclusion of participants with indeterminate (non-microbiologically confirmed) TB in numerous studies, removing diagnostic uncertainty, and resulting in reduced diversity of clinical TB disease. Several case-control studies also included healthy asymptomatic controls, healthy Mtb-sensitised (latent Mtb infected) individuals (IGRA or TST positive), or individuals with other uncommon diseases, not reflective of the target population or setting, further exacerbating the spectrum bias. The inclusion of severe TB cases and healthy controls (or controls with inappropriate other diseases) may have resulted in misleadingly high diagnostic accuracy in some of these studies. Other limitations included uncertainty regarding consecutive recruitment, blinding status not clearly stated, and lack of a priori score thresholds.

Indirectness is synonymous with applicability, generalisability, translatability, and external validity of the evidence ³⁶ . Included studies evaluated diagnostic performance among adults or children, within clinical outpatient and hospital inpatient settings, prospectively among symptomatic clinic attendees or within matched case-control cohorts. Some of these settings and populations are not appropriate or relevant to clinical practice, and results are unlikely to be generalisable. The lack of diagnostic uncertainty, spectrum bias, and inappropriate control selection is a concern for external validity of these results. Point-of-care device translatability has only been tested for one signature (Sweeney3) among PLHIV, with unsatisfactory diagnostic accuracy ⁴⁶ . Technical variability and operator reliability have not been tested on point-of-care platforms for any tests.

With regards to downstream effects, false negative signature results (patients incorrectly classified as not having TB) may have serious consequences, with delayed TB diagnosis resulting in Mtb transmission to close contacts, and increased risk of morbidity and mortality. While consequences are less serious, false positive results (individuals incorrectly classified as having TB) may result in costly further investigations, or 6 months of curative therapy with potential adverse effects and without apparent benefit. Incorrect diagnosis of TB may result in missed or delayed alternate diagnosis and treatment, with potential downstream consequences. Misdiagnosis of TB may also result in stigmatisation from family and community, and psychological distress. There is no uncertainty regarding true positive and true negative results.

We found very serious risks of inconsistency, with significant unexplained heterogeneity of diagnostic sensitivity and specificity estimates for signatures in different validation cohorts and settings. No signatures consistently met the WHO TPP criteria in all or most cohorts, suggesting publication bias toward more optimistic signature performance, particularly in discovery cohorts. There was also a very serious risk of imprecision, with small sample sizes and wide confidence intervals for estimates of test accuracy among PLHIV, with no pooling of data. Data was not available or accessible for the PLHIV subgroup in numerous cohorts. In summary, the data included in this review provides very low quality evidence and we would not recommend any changes to clinical practice based on these results.