Parallel use of low‐complexity automated nucleic acid amplification tests and lateral flow urine lipoarabinomannan assays to detect tuberculosis disease in adults and adolescents living with HIV

Abstract

Background

Low‐complexity automated nucleic acid amplification tests (LC‐aNAATs) are molecular World Health Organization (WHO)‐recommended rapid diagnostic tests (also known as mWRDs) widely used to diagnose tuberculosis disease. The lateral flow urine lipoarabinomannan assay (LF‐LAM) is recommended by the WHO to assist in diagnosing tuberculosis disease amongst people with HIV. Previous systematic reviews have assessed the diagnostic accuracy of LC‐aNAATs and LF‐LAM used in isolation for the detection of tuberculosis, but in clinical practice the tests may be used in parallel (i.e. LC‐aNAAT in combination with LF‐LAM).

Objectives

To compare the diagnostic accuracy of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine versus LC‐aNAATs on respiratory samples alone for detection of tuberculosis disease in adults and adolescents with HIV who present with presumptive tuberculosis.

Search methods

We searched Cochrane CENTRAL, MEDLINE, Embase, Science Citation Index‐Expanded, Biosis Previews, Conference Proceedings Citation Index – Science, Scopus, WHO Global Index Medicus, ProQuest Dissertations & Theses, ClinicalTrial.gov, and the WHO International Clinical Trials Registry up to 3 November 2023.

Selection criteria

We included studies that allowed assessment of the diagnostic accuracy of parallel testing and LC‐aNAAT on respiratory samples in the same study group. Participants were adults and adolescents (defined as 10 years of age and older) with HIV who presented with presumptive tuberculosis. The reference standards we used for the detection of tuberculosis disease were microbiological or composite.

As well as published studies, we included unpublished data if the data provided by study authors on request were the final data and could be used to compare diagnostic accuracy of parallel testing to one of the component tests.

Data collection and analysis

Two review authors independently extracted data using a standardised form and assessed methodological quality using QUADAS‐2 and QUADAS‐C tools. We performed bivariate random‐effects meta‐analysis using a Bayesian approach to estimate sensitivity, specificity, and absolute differences between index tests. We performed subgroup analyses based on the presence of signs and symptoms, CD4 cell count, and clinical setting, as well as separate analyses for those with a positive screen for tuberculosis, advanced HIV, or serious illness.

Main results

In 27 studies involving 12,651 participants, of whom 2368 (19%) had tuberculosis based on a microbiological reference standard, the parallel use of respiratory LC‐aNAAT and urine LF‐LAM had a pooled sensitivity of 77.5% (95% credible interval (CrI) 73.4 to 81.3) and specificity of 89.4% (95% CrI 85.8 to 92.3). Compared to respiratory LC‐aNAAT alone, parallel testing had 6.7 (95% CrI 3.8 to 10.7) percentage points higher sensitivity (low certainty) and −6.8 (95% CrI −9.5 to −4.7) percentage points difference in specificity (low‐certainty evidence), using a microbiological reference standard.

In 23 studies, 11,109 participants, of whom 3723 (34%) had tuberculosis based on a composite reference standard, parallel testing had a pooled sensitivity of 67.6% (95% CrI 59.9 to 74.6) and a pooled specificity of 96.2% (95% CrI 92.8 to 98.1). Compared to respiratory LC‐aNAAT alone, parallel testing had 16.0 (10.7 to 22.9) percentage points higher sensitivity (low‐certainty evidence) and −3.5 (95% CrI −6.6 to −1.7) percentage points difference in specificity (very low certainty evidence), using a composite reference standard.

Authors' conclusions

In the diagnosis of tuberculosis disease in people with HIV who present with presumptive tuberculosis, parallel testing (LC‐aNAAT on respiratory samples and LF‐LAM on urine) improves sensitivity at the cost of reduced specificity compared to LC‐aNAAT on respiratory samples alone. The gain in sensitivity should be weighed against the loss of specificity, taking into consideration the varying tuberculosis prevalence in different settings. For low‐prevalence settings, using the tests in parallel may lead to a large increase in false‐positive results. In settings with high tuberculosis prevalence, the benefit of identifying additional patients with tuberculosis at the point‐of‐care likely outweighs the relatively lower risk of overtreatment of those without tuberculosis.

Funding

Internal sources: Liverpool School of Tropical Medicine, UK

External sources: Foreign, Commonwealth and Development Office (FCDO), UK. Project number 300342‐104; WHO, TB Prevention, Diagnosis, Treatment, Care & Innovation (PCI), Global TB Programme

Registration

Protocol available via https://doi.org/10.1002/14651858.CD016070, version published 13 May 2024

Plain language summary

In adults and adolescents with HIV, is using two different rapid tests together ('parallel testing') to diagnose tuberculosis (TB) disease more accurate than using only one?

This review looked at adults and adolescents with HIV who have signs and symptoms of tuberculosis (TB) disease. Diagnosis can be done using respiratory samples (fluids or mucus collected by coughing or using tubes to remove fluid from the throat, lungs, or stomach) or urine. We wanted to know whether using two rapid tests together ‐ an automated test on a respiratory sample (LC‐aNAAT) and a urine strip test (LF‐LAM) ‐ is better than using LC‐aNAAT alone. Using these two tests together is called parallel testing.

Key messages

‐ Parallel testing finds more cases of TB than using just the LC‐aNAAT alone.

‐ However, parallel testing may also lead to more false TB diagnoses in people who do not actually have TB, and this undesirable effect is likely more pronounced in settings where TB is less common.

‐ In places where TB is common, the benefit of parallel testing (i.e. detecting more people who actually have TB) may outweigh the harms (i.e. misdiagnoses, unnecessary treatment, and delayed diagnosis of other illnesses).

Why is it important to improve the diagnosis of TB amongst people with HIV?

People with HIV are at a higher risk of developing TB than those without HIV. TB in patients with HIV often presents with non‐specific symptoms, and detecting the bacteria (Mycobacterium tuberculosis) with standard sputum (spit) tests can be difficult. This is partly due to challenges in coughing up a sputum sample, which may contain very few bacteria, and because TB may occur in parts of the body other than the lungs. As a result of delayed or missed diagnoses, TB is the leading cause of hospital admission and death amongst people living with HIV worldwide.

What is parallel testing?

Parallel testing refers to using more than one test at the same time. If any one of the tests returns a positive test result, we consider the parallel testing to have a positive test result. We studied parallel testing using LC‐aNAAT on respiratory samples in combination with LF‐LAM on urine.

What did we want to find out?

‐ How accurate is parallel testing for all forms of TB in adults and adolescents (aged 10 years and older) with HIV who have signs or symptoms of TB?

‐ How accurate is parallel testing for identifying TB compared to using respiratory LC‐aNAAT alone?

What did we do?

We searched for studies that had investigated the accuracy of both respiratory LC‐aNAAT and urine LF‐LAM for TB in adults and adolescents living with HIV, and we combined the results of these studies.

What did we find?

The review included 27 studies that involved 12,651 participants, of whom 2368 (19%) had confirmed TB.

The included participants were from 19 low‐ and middle‐income countries where HIV and TB are common. Many studies were from countries in sub‐Saharan Africa. One study was conducted with adolescents, but the other studies were of adults.

Accuracy of parallel testing

The results of these studies indicate that in a group of 1000 people, of whom 200 (20%) actually have TB, then:

‐ parallel testing would correctly identify 155 as having TB, but would miss 45; for the remaining 800 people who do not have TB, parallel testing would correctly identify 715 as not having TB, but would misdiagnose 85 as having TB.

Parallel testing compared to respiratory LC‐aNAAT alone

The results of these studies indicate that if a parallel test strategy was used instead of a respiratory LC‐aNAAT test alone, in a group of 1000 people, of whom 200 (20%) actually have TB, then:

‐ parallel testing would accurately diagnose 13 more patients who actually have TB who would have been missed by LC‐aNAAT on respiratory samples used alone (i.e. 13 fewer 'false negatives' with parallel testing). On the other hand, using parallel testing would misdiagnose 54 more people who do not have TB (54 more 'false positives' with parallel testing), compared to LC‐aNAAT alone.

What are the limitations of the evidence?

The evidence for parallel testing in adolescents is limited. Although we included a large number of adults from different settings, only a small number of adolescents were part of this review.

Some studies in this review excluded patients who could not produce a sputum or urine sample. This may have resulted in an underestimation or overestimation of the accuracy of parallel testing.

Because it is difficult to detect TB in patients with HIV, the method for verifying whether a person truly has TB or not (what we call the 'reference standard') may not be perfect. This means that we might have underestimated or overestimated the accuracy of parallel testing. Some studies using a different reference standard (called a 'composite reference standard') showed different results, though these results were very uncertain.

How up to date is this evidence?

The evidence is based on searches conducted up to 3 November 2023.

Summary of findings

Summary of findings 1. Parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone to diagnose tuberculosis disease in adults and adolescents with HIV using a microbiological reference standard.

Review question: How does the diagnostic accuracy of the parallel use of LC‐aNAAT and LF‐LAM compare to LC‐aNAAT alone for detecting tuberculosis disease in adults and adolescents with HIV who have been subject to a microbiological reference standard?
Study design	Fully paired diagnostic accuracy studies
Population	Adults and adolescents (defined as 10 years of age and older) with HIV who present with presumptive tuberculosis
Index test A, parallel	Parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine
Index test B, LC‐aNAAT	LC‐aNAAT on respiratory samples
Target condition	Tuberculosis disease (all forms)
Reference standard	Microbiological reference standard
Consequence of index test results	For any who test positive on either LF‐LAM or a respiratory LC‐aNAAT, treatment is recommended.
Accuracy of parallel testing (95% CrI)	Sensitivity 77.5% (73.4% to 81.3%); specificity 89.4% (85.8% to 92.3%)
LC‐aNAAT (95% CrI)	Sensitivity 68.0% (60.8% to 74.9%); specificity 96.7% (95.7% to 97.6%)
LF‐LAM (95% CrI)	Sensitivity 39.1% (32.6% to 45.9%); specificity 91.9% (88.7% to 94.4%)
	Point estimate	95% CrI	Number of participants (studies)		Certainty of the evidence (GRADE)
Summary difference in sensitivity	6.7%	3.8% to 10.7%	2368 (27)		⨁⨁◯◯a,b LOW
Summary difference in specificity	−6.8%	−9.5% to −4.7%	10,283 (27)		⨁⨁◯◯a,b LOW
Difference in test resultsc	Results per 1000 participants tested (95% CI)
	Prevalence of tuberculosis disease: 20%*
	Parallel testing			LC‐aNAAT alone
True positives	155 (147 to 163)			136 (122 to 150)
	13 more with parallel (8 to 21 more)
False negatives	45 (37 to 53)			64 (50 to 78)
	13 fewer with parallel (21 to 8 fewer)
True negatives	715 (686 to 738)			774 (765 to 780)
	54 fewer with parallel (76 to 37 fewer)
False positives	85 (62 to 114)			26 (20 to 35)
	54 more with parallel (37 to 76 more)

*A summary of findings table for settings with lower and higher tuberculosis prevalence (2.5% and 40%) is available in Supplementary material 19.

Abbreviations: CI: confidence interval; CrI: credible interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification tests; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aRisk of bias: as assessed by review authors using QUADAS‐2 and QUADAS‐C, several studies had high risk of bias in the patient selection domain because the studies had inappropriate exclusion criteria. Several studies had high risk of bias in the reference standard domain because we thought the reference standard used was unlikely to correctly classify the target condition. Differences in sensitivity and specificity did not significantly change in sensitivity analyses that we conducted by excluding studies with high risk of bias and high applicability concerns. Therefore, we downgraded only 1 level for sensitivity and 1 level for specificity.

^bInconsistency: the individual study estimates for difference in sensitivity and specificity fall outside our 95% CrI and the 95% prediction interval is wider than the CrI. Therefore, we downgraded sensitivity and specificity by 1 level.

Indirectness: generally, we judged applicability as 'low concern', therefore we did not downgrade for indirectness.

Imprecision: we did not rate imprecision because the target of certainty rating was defined by the bounds of the 95% CrI.

Publication bias: we did not downgrade for publication bias as, in addition to our systematic literature search, we searched in clinical trial registries and had extensive communication with authors to retrieve additional studies/data. Moreover, there was a WHO public call for data. Often, data had to be reanalysed for the purpose of this review, so study authors would not have known the result prior to expressing willingness to share their data. For those study authors who did not report data despite our efforts to contact them, we did not have the impression that they did not share data or reply because they were biased towards a certain result.

^cDifference in test results: the summary difference pooled the study‐specific differences and not the difference between the pooled sensitivities and specificities. Consequently, there may be some discordance between the reported differences and the subtraction of parallel and component test results.

Summary of findings 2. Parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone to diagnose tuberculosis disease in adults and adolescents with HIV using a composite reference standard.

Review question: How does the diagnostic accuracy of the parallel use of LC‐aNAAT and LF‐LAM compare to LC‐aNAAT alone for detecting tuberculosis disease in adults and adolescents with HIV who have been subject to a composite reference standard?
Study design	Fully paired diagnostic accuracy studies
Population	Adults and adolescents (defined as 10 years of age and older) with HIV who present with presumptive tuberculosis
Index test A, parallel	Parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine
Index test B, LC‐aNAAT	LC‐aNAAT on respiratory samples
Target condition	Tuberculosis disease (all forms)
Reference standard	Composite reference standard
Consequence of index test results	For any who test positive on either LF‐LAM or a respiratory LC‐aNAAT, treatment is recommended.
Accuracy of parallel testing (95% CrI)	Sensitivity 67.6% (59.9% to 74.6%); specificity 96.2% (92.8% to 98.1%)
LC‐aNAAT (95% CrI)	Sensitivity 46.8% (38.6% to 55.2%); specificity 99.9% (99.8% to 100%)
LF‐LAM (95% CrI)	Sensitivity 38.6% (30.7% to 47.0%); specificity 96.3% (93.0% to 98.1%)
	Point estimate	95% CrI	Number of participants (studies)		Certainty of the evidence (GRADE)
Summary difference in sensitivity	16.0%	10.7% to 22.9%	3723 (23)		⨁⨁◯◯a,b LOW
Summary difference in specificity	−3.5%	−6.6% to −1.7%	7386 (23)		⨁◯◯◯a,b VERY LOW
Difference in test resultsc	Results per 1000 participants tested (95% CI)
	Prevalence of tuberculosis disease: 20%*
	Parallel testing			LC‐aNAAT alone
True positives	135 (120 to 149)			94 (77 to 110)
	32 more with parallel (21 to 46 more)
False negatives	65 (51 to 80)			107 (90 to 123)
	32 fewer with parallel (46 to 21 fewer)
True negatives	770 (742 to 785)			799 (799 to 800)
	28 fewer with parallel (53 to 13 fewer)
False positives	30 (15 to 58)			1 (0 to 1)
	28 more with parallel (13 to 53 more)

*A summary of findings table for settings with lower and higher tuberculosis prevalence (2.5% and 40%) is available in Supplementary material 20.

Abbreviations: CI: confidence interval; CrI: credible interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification tests; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aRisk of bias: as assessed by QUADAS‐2 and QUADAS‐C, several studies had high risk of bias in the patient selection domain because the studies had inappropriate exclusion criteria. All studies had high risk of bias in the reference standard domain because of incorporation bias. Therefore, we downgraded by 1 level for sensitivity and specificity. Given that the observed specificity of each of the tests appears to be biased towards 100%, we downgraded specificity by an additional level.

^bInconsistency: the individual study estimates for difference in sensitivity and specificity fall outside our 95% CrI and the 95% prediction interval is wider than the CrI. Therefore, we downgraded sensitivity and specificity by 1 level.

Indirectness: generally, we judged applicability as 'low concern', therefore we did not downgrade for indirectness.

Imprecision: we did not rate imprecision because the target of certainty rating was defined by the bounds of the 95% CrI.

Publication bias: we did not rate down for publication bias as we, in addition to our systematic literature search, included searching in clinical trial registries and have had extensive communication with authors to retrieve additional studies/data. Also, there was a WHO public call for data. Often, data had to be reanalysed for the purpose of this review, so study authors would not have known the result prior to expressing willingness to share their data. For those authors that did not report data despite our efforts to contact them, we did not have the impression that they did not share data or reply because they were biased towards a certain result.

^cDifference in test results: the summary difference pooled the study‐specific differences and not the difference between the pooled sensitivities and specificities. Consequently, there may be some discordance between the reported differences and the subtraction of parallel and component test results.

Background

Tuberculosis is caused by Mycobacterium tuberculosis and is transmitted through respiratory aerosols released from people with tuberculosis disease. Tuberculosis disease is challenging to diagnose, particularly in people with HIV. Diagnosis is often missed or delayed, resulting in worse prognosis and outcomes for patients. Globally, 1.25 million deaths in 2023 resulted from tuberculosis, including 161,000 deaths amongst people with HIV, making it a leading cause of death in this population [1]. A common reason for the high mortality amongst people with HIV is late treatment initiation due to diagnostic delay.

It is the unique pathophysiology and clinical manifestations of tuberculosis in people with HIV that make diagnosis particularly difficult in this population [2]. First, signs and symptoms are often non‐specific, such as fever, weight loss, and fatigue, which may mimic other opportunistic infections like pneumocystis pneumonia, histoplasmosis, and non‐tuberculous mycobacterial infection, or conditions like HIV‐associated cancers. Cough may or may not be present. Second, people with HIV and a compromised immune system may present with paucibacillary (i.e. low bacterial burden) pulmonary tuberculosis or disseminated tuberculosis disease, which reduces diagnostic yield in biological samples like sputum [3]. Third, sputum can be particularly difficult to obtain in people with HIV [4] and cannot be used to diagnose extrapulmonary tuberculosis.

The World Health Organization (WHO) recommends that people with HIV who report any one of cough, fever, weight loss, and night sweats (i.e. the WHO four‐symptom screen) should be evaluated further for tuberculosis disease [5]. In people with HIV, additional tuberculosis screening tools that are recommended include chest X‐ray, C‐reactive protein (CRP), and molecular WHO‐recommended rapid diagnostics (also known as mWRDs) [5].

Tuberculosis diagnosis historically has relied heavily on evaluation of sputum to directly detect the presence of mycobacteria or mycobacterial components in the specimen. Low‐complexity automated nucleic acid amplification tests (LC‐aNAATs) on respiratory samples are part of the molecular WHO‐recommended rapid diagnostics widely used to diagnose tuberculosis [6]. However, this approach may lack sensitivity in people unable to produce sputum, with paucibacillary pulmonary tuberculosis, or extrapulmonary disease, which are more common in people with HIV than in those without HIV. The antigen‐based lateral flow urine lipoarabinomannan assay (LF‐LAM) is recommended by the WHO to assist in diagnosing tuberculosis amongst people with HIV [7]. The LF‐LAM is conducted on urine and categorises patients as having tuberculosis disease or not having tuberculosis disease. Thus, it does not identify the site of tuberculosis disease. LF‐LAM has the advantage of being performed on urine (a relatively easy specimen for collection) and of being conducted at the point‐of‐care without the need for electricity or special equipment. Moreover, in contrast to a sputum‐based testing strategy, LF‐LAM test sensitivity increases inversely with the degree of immunosuppression [8, 9, 10, 11, 12, 13].

Previous systematic reviews have assessed the diagnostic accuracy of LC‐aNAATs and LF‐LAM used in isolation for the detection of tuberculosis, including Xpert MTB/RIF and Xpert Ultra assays (Cepheid, Sunnyvale, USA) for pulmonary tuberculosis [14, 15] and extrapulmonary tuberculosis [16], and urine LF‐LAM for people with HIV [17]. While LF‐LAM was found to have limited sensitivity when used as a single test [17], some studies have found that urine LF‐LAM and respiratory LC‐aNAAT may be complementary for the diagnosis of tuberculosis disease in people with HIV [4, 8, 10, 18]. In the 2016 Cochrane review on the diagnostic accuracy of LF‐LAM, sensitivity increased by 13% when both LF‐LAM and sputum Xpert MTB/RIF were used rather than the use of sputum Xpert alone, while the specificity decreased by 4% [19]. However, the results were based on only a few studies, and analysis was restricted to participants able to produce sputum.

In this Cochrane review, we evaluated the incremental diagnostic accuracy of using LC‐aNAAT on respiratory samples in parallel with urine LF‐LAM (i.e. LC‐aNAAT in combination with LF‐LAM, with positivity defined as either test being positive) for the diagnosis of any tuberculosis disease (without distinguishing the specific site of disease) amongst people with HIV. We define incremental accuracy as the change in diagnostic accuracy (sensitivity and specificity) when a single test is combined with another test, compared to the single test alone. Our focus is on the diagnosis of tuberculosis disease in adults and adolescents with HIV.

Target condition being diagnosed

Tuberculosis predominantly affects the lungs (pulmonary tuberculosis), but it can affect any part of the body, such as the lymph nodes, pleura, urogenital system, brain, or spine (i.e. extrapulmonary tuberculosis). Extrapulmonary tuberculosis can occur alone or together with pulmonary tuberculosis. People with HIV have higher rates of extrapulmonary tuberculosis including mycobacteraemia (i.e. mycobacteria cultured from blood) [2]. Moreover, people with HIV have a higher risk than people without HIV of disseminated tuberculosis disease (i.e. affecting multiple organ systems) [20]. In this review, our target condition was any form of tuberculosis disease in people with HIV.

Index test(s)

We assessed three index tests: (1) LC‐aNAAT conducted on respiratory samples, (2) LF‐LAM performed on urine, and (3) parallel testing of LC‐aNAAT on respiratory samples and LF‐LAM on urine, in which we considered either test being positive as a positive result for the combination.

We compared parallel testing versus LC‐aNAAT on respiratory samples.

LC‐aNAAT

LC‐aNAATs are a subgroup of molecular WHO‐recommended rapid diagnostic tests, and the technologies include the Xpert MTB/RIF and Xpert MTB/RIF Ultra (Cepheid, Sunnyvale, USA); Truenat MTB and MTB Plus (Molbio Diagnostics, Goa, India); STANDARD M10 (SD Biosensor, Republic of Korea); and IRON‐qPCR (BIONEER, Republic of Korea).

The Xpert MTB/RIF [21] and Xpert MTB/RIF Ultra (Xpert Ultra) [22] assays simultaneously detect M tuberculosis complex and rifampicin resistance in a single step, where most of the sample processing and the amplification process are combined in a closed system. Xpert MTB/RIF and Xpert Ultra cartridges are used with the GeneXpert system. Xpert MTB/RIF and Xpert Ultra provide a result within two hours of starting the test, with minimal hands‐on technical time [21, 22]. A Cochrane review on the diagnostic accuracy of Xpert MTB/RIF for pulmonary tuberculosis found a pooled sensitivity of 85% (95% credible interval (CrI) 82 to 88) and specificity of 98% (95% CrI 97 to 98) [23]. As the Xpert MTB/RIF sensitivity decreased in people with HIV to 81% (95% CrI 75 to 86) [23], the Xpert Ultra was designed to detect the M tuberculosis complex with greater sensitivity. Amongst people with HIV, a comparative Cochrane review found a pooled sensitivity of 87.6% (95% CrI 75.4 to 94.1) for Xpert Ultra versus 74.9% (95% CrI 58.7 to 86.2) for Xpert MTB/RIF; pooled specificity was 92.8% (95% CrI 82.3 to 97.0) for Xpert Ultra versus 99.7% (95% CrI 98.6 to 100.0) for Xpert MTB/RIF [15]. Xpert assays can also be used to diagnose extrapulmonary tuberculosis; sensitivity varies across different extrapulmonary sampling sites, being highest in cerebrospinal fluid [16].

Truenat MTB and Truenat MTB Plus have been developed for the automated detection of the M tuberculosis complex with both sample processing and amplification in one step, and results obtained within one hour [24, 25]. Truenat MTB is a quantitative test that provides a colony‐forming unit (CFU) value [25], while Truenat MTB Plus is semi‐quantitative and gives four grades (high, medium, low, and very low) that are based on CFUs but do not specify the actual count [24]. Instead of using a cartridge, Truenat employs a chip onto which the processed sample is applied. A reflex test to detect rifampicin resistance is available with the platform. One multicentre trial evaluating the diagnostic accuracy of these assays for pulmonary tuberculosis reported a sensitivity of 73% for Truenat MTB and 80% for Truenat MTB Plus [26].

STANDARD M10 and IRON‐qPCR are both validated as fully automated platforms with an all‐in‐one cartridge for nucleic acid extraction and amplification. Results are available within one hour. IRON‐qPCR technology includes targets for M tuberculosis complex (and resistance mutations for isoniazid, rifampicin, and fluoroquinolones) to be run on processed sputum or on isolates (culture growth) [27]. STANDARD M10 requires a simple sputum pre‐treatment process but can include cartridges for simultaneous detection of M tuberculosis and non‐tuberculous mycobacteria and drug‐resistance against rifampicin and isoniazid [28].

We will refer to the Truenat MTB and MTB Plus assays as “Truenat” and Xpert MTB/RIF and Xpert Ultra as “Xpert” unless it is necessary to distinguish between specific test versions.

LF‐LAM

LF‐LAM detects lipoarabinomannan (LAM), a lipopolysaccharide in mycobacterial cell walls, which is released from metabolically active or degenerating bacterial cells during tuberculosis disease [29]. LAM is found in the bloodstream of people with tuberculosis disease and subsequently filtered across the glomerular basement membrane of the kidneys into urine [30]. To note, LF‐LAM detection in urine is not site‐specific, meaning a positive LF‐LAM result alone does not tell us whether the patient has pulmonary or extrapulmonary tuberculosis, or both. Evidence suggests that concentrations of LAM in urine increase progressively with the degree of immunosuppression, probably due to higher rates of mycobacterial dissemination or overall mycobacterial burden [31]. Renal M tuberculosis disease may contribute to the presence of LAM in urine, as evidenced by autopsy studies [32], but urinary tract disease is not necessary for having LAM antigen in urine.

LF‐LAM is performed manually by applying 60 μL of unprocessed urine to the sample pad of the Determine TB LAM Ag test and leaving the strip test to incubate at room temperature for 25 minutes [33]. The strip is then inspected visually. The intensity of any visible band is graded by comparing it with the intensities of the bands on a manufacturer‐supplied reference scale card. Of note, the reference scale was revised in January 2014. Prior to January 2014, the reference scale card included five bands (grade 1 representing a very low intensity band to grade 5 representing a high/dark intensity band). After January 2014, the manufacturer revised the reference scale card to have four reference bands, such that the band intensity for the new grade 1 corresponded to the band intensity for the previous grade 2. Under the current manufacturer recommendations (using the revised four‐band reference card), only tests with patient bands that have the same intensity as grade 1 or higher are considered positive [33].

A 2019 Cochrane review assessed the diagnostic accuracy of a single LF‐LAM for the diagnosis of tuberculosis disease in people with HIV [17]. Amongst participants with symptoms, the LF‐LAM pooled sensitivity was 42% (95% CrI 31 to 55) and pooled specificity was 91% (CrI 85 to 95) [17]. Amongst unselected participants not assessed for signs and symptoms of tuberculosis, LF‐LAM pooled sensitivity was 35% (95% CrI 22 to 50) and pooled specificity was 95% (95% CrI 89 to 96) [17].

Clinical pathway

LC‐aNAATs and LF‐LAM are used for the diagnosis of tuberculosis disease. Figure 1 shows the clinical pathway and presents the context in which the index tests might be used. Individuals to be evaluated for tuberculosis are adults and adolescents living with HIV with presumptive tuberculosis. 'Presumptive tuberculosis' refers to a person with signs or symptoms suggestive of tuberculosis or who has a positive or abnormal screening test for tuberculosis. For people with HIV, the WHO recommends using tools such as the WHO four‐symptom screen, chest X‐ray, CRP, and molecular WHO‐recommended rapid diagnostics for tuberculosis screening [5]. When molecular WHO‐recommended rapid diagnostics are used as an option for screening, the WHO highlights the importance of following up with a diagnostic assessment including LF‐LAM and a repeat molecular WHO‐recommended rapid diagnostic test to prevent the potential harm of overdiagnosing and overtreatment [5]. People with HIV are also recommended to undergo tuberculosis diagnostic testing if people are seriously ill (i.e. any of the following danger signs: respiratory rate 30 breaths per minute or greater, heart rate 120 beats per minute or greater, unable to walk unaided, or temperature 39 °C or greater), if admitted with advanced HIV (i.e. defined as PLHIV with CD4 cell count less than 200 cells/mm³ or a WHO stage 3 or 4 condition), or with low CD4 cell count regardless of symptoms (i.e. CD4 less than 200 cells/mm³ for inpatients and CD4 less than 100 cells/mm³ for outpatients) [34].

1.

Clinical pathway for LC‐aNAAT and LF‐LAM in parallel for diagnosis of tuberculosis amongst adults and adolescents living with HIV

This figure was created by the author team based on current WHO guidelines [5, 7, 34, 98].

Abbreviations: CRP: C‐reactive protein; DR‐TB: drug resistant TB; LAM: lipoarabinomannan; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay; mWRD: molecular World Health Organization‐recommended diagnostic test; TB: tuberculosis

Footnotes

^aPeople aged 10 years and above, HIV‐positive, and attending a health facility
^bPresumptive TB refers to a person with signs or symptoms suggestive of tuberculosis or who has a positive or abnormal screening test for tuberculosis. This includes WHO‐recommended symptom screen (i.e. WHO four‐symptom screen) or signs and symptoms of extrapulmonary TB (or both). Other screening tools include: CRP with a cut‐off of greater than 5 mg/L, chest X‐ray, molecular WHO‐recommended rapid tests. The following populations further qualify for TB testing regardless of TB screening result: inpatients in medical wards where the TB prevalence is greater than 10% (molecular WHO‐recommended rapid tests including LC‐aNAAT should be used); inpatients with advanced HIV disease (CD4 less than 200 cells/mm³ or WHO clinical stage 3 or 4 events) or being seriously ill (respiratory rate greater than 30 breaths/minutes or temperature greater than 39 °C or heart rate greater than 120 beats/minute or unable to walk unaided) (LF‐LAM should be used as an add‐on to clinical judgement in combination with other tests); outpatients who are seriously ill or have a CD4 less than 100 cells/mm³ or a WHO clinical stage 3 or 4 event (LF‐LAM should be used as an add‐on to clinical judgement in combination with other tests).
^cAmongst inpatients, LF‐LAM may assist in the diagnosis of active TB in HIV‐positive adults and adolescents if they have signs and symptoms of TB, or have advanced HIV disease, or are seriously ill. Amongst outpatients, LF‐LAM may assist in the diagnosis of active TB in HIV‐positive adults and adolescents if: signs and symptoms of TB or seriously ill; or CD4 less than 100 cells/mm³; or a WHO clinical stage 3 or 4 event.
^dThe LF‐LAM strip‐test that can detect the LAM antigen in urine at the point of care
^eLC‐aNAAT for TB diagnosis and detection of resistance to isoniazid or rifampicin (or both). LC‐aNAATs evaluated in this review include Xpert MTB/RIF and Xpert MTB/RIF Ultra (Cepheid, Sunnyvale, USA); Truenat MTB and MTB Plus (Molbio Diagnostics, Goa, India); STANDARD M10 (SD Biosensor, Republic of Korea); and IRON‐qPCR (BIONEER, Republic of Korea).
^fFor people testing positive on either LF‐LAM or LC‐aNAAT, tuberculosis treatment is recommended.

In resource‐limited settings in particular, molecular WHO‐recommended rapid diagnostics such as the LC‐aNAATs are recommended as the initial diagnostics for tuberculosis and detection of rifampicin resistance rather than smear microscopy, culture, and phenotypic DST [34]. LF‐LAM is recommended to assist in the diagnosis of tuberculosis in certain subpopulations of people with HIV, with current WHO guidelines describing the use of LF‐LAM by patient setting (i.e. inpatient and outpatient settings) [7].

Tuberculosis treatment is recommended for people testing positive on either LF‐LAM or an LC‐aNAAT. Of note, it is likely that urine for LF‐LAM results will be available before a respiratory sample for LC‐aNAAT results or that only one of the tests result will be performed; therefore, clinicians are advised to base treatment initiation decisions on the result of the first positive available test, often LF‐LAM, while waiting for results of other diagnostic tests. Additional investigations should be performed to assess drug resistance using rapid molecular tests (e.g. with Xpert or Truenat). Phenotypic (culture with drug susceptibility testing) and other molecular methods (e.g. line probe assays (LPA)) may also be used to assess drug resistance. If people test negative on LF‐LAM, additional investigations should be prioritised as the sensitivity of LF‐LAM is insufficient to rule out tuberculosis. If people test negative on both LF‐LAM and LC‐aNAAT, tuberculosis is still not ruled out, and further investigations for tuberculosis are recommended. These investigations may include chest imaging, additional clinical assessments, additional molecular WHO‐recommended rapid diagnostic tests, and culture.

The downstream consequences of tuberculosis parallel testing for the person tested include the following.

• True positive: people would benefit from diagnosis and appropriate treatment.

• True negative: people would be spared unnecessary treatment and would benefit from reassurance and pursuit of an alternative diagnosis.

• False positive: people would probably experience anxiety and morbidity caused by additional testing, unnecessary treatment, and possible harms; possible stigma associated with having tuberculosis; and the risk that a false‐positive result may halt further diagnostic evaluation for the illness causing their symptoms.

• False negative: increased risk of morbidity and mortality and delayed treatment initiation; risk of ongoing tuberculosis transmission.

Role of index test(s)

Regarding the role of the tests, LC‐aNAAT on respiratory samples is suggested as an initial test for tuberculosis disease and detection of rifampicin resistance rather than smear microscopy, culture, and phenotypic drug‐susceptibility testing [6]. For LF‐LAM, the WHO states that: "LF‐LAM should be used as an add‐on to clinical judgement in combination with other tests. It should not be used as a replacement or triage test" [6]. The LC‐aNAAT on respiratory samples and LF‐LAM may be performed in parallel (i.e. LC‐aNAAT in combination with LF‐LAM) where a positive result on either test implies a positive result for the combination [6].

Alternative test(s)

Phenotypic tests

Smear microscopy

Smear microscopy is the examination of smears for acid‐fast bacilli. It is widely used as a diagnostic tool for tuberculosis, with the advantages of being a simple, rapid, and low‐cost technique. The main limitation is the reported low sensitivity (e.g. 32% amongst people with HIV [4]). Around 5000 to 10,000 bacilli per millilitre must be present in the specimen for tuberculosis bacteria to be visible by microscopy [35]. Sensitivity can be increased using fluorescence microscopy instead of light microscopy of Ziehl‐Neelsen‐stained smears (e.g. for people with HIV sensitivity increased from 36% to 73%), with similar specificity [36]. Due to the low sensitivity, the WHO recommends replacing smear microscopy with molecular WHO‐recommended rapid diagnostics, such as the Xpert or Truenat assays, as the initial test for all people being evaluated for tuberculosis [6].

Culture

Mycobacterial culture remains the most sensitive modality to diagnose tuberculosis, with only around 100 bacilli per millilitre required for a positive culture [35]. In addition, mycobacterial culture additionally makes it possible to conduct phenotypic drug susceptibility testing. The disadvantage of culture is that it often takes more than two weeks and up to eight weeks before cultures become positive if specimens contain few bacilli. Higher sensitivity (i.e. a lower limit required for detection) is demonstrated by liquid media (89.4%) compared to solid LJ media (74.0%), but contamination rates for liquid media systems are higher than for solid media [37]. Moreover, all culture methods require a highly equipped laboratory and highly skilled laboratory staff.

Other molecular tests and genotyping

The loop‐mediated isothermal amplification test (TB‐LAMP) (Eiken Chemical, Tokyo, Japan) has been recommended by the WHO for diagnosing pulmonary tuberculosis since 2016 [38]. It is a manual and simple test that can be performed directly on sputum samples, with a result provided in less than one hour. The amplification process utilises at least four different sets of primers and is carried out in a single step, comprising a strand displacement reaction at 65 ºC for 15 to 60 minutes. The sensitivity of this test in adults from different settings varies from 76% to 80%, and specificity from 97% to 98% [38]. Several non‐respiratory sample types, such as cerebrospinal fluid, pleural fluid, and lymph node aspirate, can be tested for focal extrapulmonary disease using LC‐aNAAT, with diagnostic accuracy varying depending on sample type [16]. Additionally, moderate‐complexity aNAAT are available for the detection of M. tuberculosis complex and the identification of mutations within the rpoB, inhA, and katG genes [39].

Rationale

Respiratory samples may be difficult to obtain, particularly in people with HIV, and this population may also have paucibacillary or extrapulmonary disease. A systematic review and meta‐analysis of the diagnostic yield of urine LAM using individual participant data found that, while nearly all participants were able to produce a urine sample for testing, one in five participants was not able to produce sputum [4]. From clinical studies, we expected that there could be an incremental value in combining LC‐aNAAT on a respiratory sample and LF‐LAM on a urine sample to increase diagnostic yield for tuberculosis in people with HIV [8, 9, 40, 41, 42]. In a meta‐analysis by Broger and colleagues [4], in hospitalised people with HIV, the yield for microbiological diagnosis of tuberculosis using LF‐LAM (Determine TB LAM) and Xpert assays were similar (51% versus 47%), while the tests in combination had a yield of 71%, which supports the idea of implementing combined testing strategies.

This Cochrane review estimated the change in diagnostic accuracy (i.e. sensitivity and specificity) when using respiratory LC‐aNAAT and urine LF‐LAM in parallel to diagnose tuberculosis disease amongst people with HIV, compared to using respiratory LC‐aNAAT alone. The findings informed WHO guideline development [43].

Objectives

To compare the diagnostic accuracy of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine versus LC‐aNAAT on respiratory samples alone for detection of tuberculosis disease in adults and adolescents with HIV who present with presumptive tuberculosis.

Secondary objectives

To investigate the following sources of heterogeneity: signs and symptoms of tuberculosis; clinical setting; CD4 cell count; screening test positivity for tuberculosis disease (e.g. by chest X‐ray, molecular WHO‐recommended rapid diagnostics, or CRP); serious illness; and advanced HIV.

Methods

Criteria for considering studies for this review

Types of studies

We included comparative diagnostic accuracy studies that evaluated the accuracy of a parallel combination of respiratory LC‐aNAAT and urine LF‐LAM (henceforth referred to as ‘parallel testing’) and the diagnostic accuracy of respiratory LC‐aNAAT alone, in the same study group (Table 3), against a microbiological reference standard (that did not include the index tests) or a composite reference standard. Fully paired studies (i.e. both index tests evaluated on the same participants) and studies designed with randomisation of participants to two or more index test groups were eligible for inclusion. Other study designs were excluded because of a high risk of confounding.

1. Eligible comparisons and minimal study design requirements.

Population	Index test A	Index test B (comparator)	If no IPD available: accuracy data on the following tests should be reported	If IPD available: data on the following tests should be available
Adults and adolescents with HIV	Respiratory LC‐aNAAT + LF‐LAM	Respiratory LC‐aNAAT	Respiratory LC‐aNAAT + LF‐LAM Respiratory LC‐aNAAT	Respiratory LC‐aNAAT LF‐LAM Reference standard test

Abbreviations: IPD: individual participant data; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay
^aOnly applicable to fully paired designs

We included studies that had not originally been designed as comparative diagnostic accuracy studies (e.g. randomised controlled trials), but from which we could extract data to determine the diagnostic accuracy of the index tests (i.e. two‐by‐two table data). Additionally, we included studies that were originally designed to assess the accuracy of LF‐LAM, with LC‐aNAAT and mycobacterial culture as reference standards. If the authors of such studies made data available for estimation of comparative diagnostic accuracy, we reanalysed them for the purpose of this review (Table 3).

We included studies where there was an attempt to collect urine for LF‐LAM and respiratory samples for LC‐aNAAT testing for all participants. Respiratory samples included sputum, induced sputum, tracheal aspiration, and fluid from bronchial alveolar and gastric lavage. We also included studies that had made an attempt to obtain fully paired data for parallel testing, but where data on some participants were unavailable (e.g. either a urine sample for LF‐LAM or a respiratory sample for LC‐aNAAT was not available, or there was an available sample but no final result). We included these studies because respiratory samples can be difficult to obtain in people with HIV, especially in those with severe disease [4, 8]; excluding those unable to produce samples (or in whom one test did not lead to a valid result) may underestimate the net sensitivity of parallel testing [17].

We included published studies and preprints. We included unpublished studies if the data provided by study authors on request were the final data; otherwise, we classified the study as ongoing. We included data provided by manufacturers and investigators in response to the public call for data by the WHO Global TB Programme, which was posted online on 17 January 2024.

We excluded studies enroling tuberculosis cases and non‐cases from separate populations (also called two‐gate or case‐control studies). We also excluded reviews, commentaries, and editorials. We did not have language restrictions.

Participants

We included studies of adults and adolescents (defined as 10 years of age and older) with HIV who presented with presumptive tuberculosis. Presumptive tuberculosis is based on the presence of signs and symptoms compatible with tuberculosis (e.g. a positive WHO four‐symptom screen). We also included participants with a positive screening test for tuberculosis (e.g. abnormal chest X‐ray findings, elevated C‐reactive protein (CRP), or positive molecular WHO‐recommended rapid diagnostic test). We included participants from studies that evaluated the index tests in people with HIV irrespective of tuberculosis symptoms; for example, studies of people with HIV being assessed for initiation of antiretroviral treatment or people with HIV admitted to medical wards. Although participants in these studies are not assessed for signs and symptoms as a condition of study inclusion, they often have characteristics suggestive of tuberculosis where further tuberculosis testing is recommended; studies may have high rates (more than 80%) of participants with a positive WHO four‐symptom score [17] or a positive tuberculosis screening test [44, 45], or be conducted in settings with high prevalence rates of tuberculosis (more than 10%) and a high likelihood of participants having tuberculosis prior to being tested for it.

Data for children below 10 years of age will be reported in a separate Cochrane review [46].

Index tests

The principal focus of this review is to estimate the incremental accuracy of the parallel use of respiratory LC‐aNAAT and urinary LF‐LAM. We define parallel use as both tests being conducted concurrently on the same person, and a positive result on either test being interpreted as a positive result for the combination of tests. In clinical practice, a urinary LF‐LAM test may be performed first, with the LF‐LAM result then known to those interpreting the respiratory LC‐aNAAT (although an LC‐aNAAT result is often an objective, automatically generated result). The other index tests are the respiratory LC‐aNAAT (comparator test) and the urinary LF‐LAM (part of parallel testing).

LC‐aNAATs were eligible for inclusion when performed on respiratory samples.

For LC‐aNAATs in general, results are automatically generated, indicating “DETECTED” for a positive result or “NOT DETECTED” for a negative result, or it will be indicated if the test result is invalid. For example, with Xpert MTB/RIF and Xpert Ultra, the user is provided with a printable test result as follows.

• MTB (M. tuberculosis) DETECTED; Rif (rifampicin) resistance DETECTED

• MTB DETECTED; Rif resistance NOT DETECTED

• MTB DETECTED; Rif resistance INDETERMINATE

• MTB NOT DETECTED

• INVALID/ERROR/NO RESULT (the presence or absence of M. tuberculosis cannot be determined.)

Xpert Ultra incorporates a semi‐quantitative classification for results: trace, very low, low, moderate, and high. We considered a trace result to mean MTB (M. tuberculosis) DETECTED [47].

For urinary LF‐LAM, at the time of writing (2024), the only commercially available test is Determine TB LAM Ag (Abbott, Chicago, IL, USA). We considered Determine TB LAM Ag test results at grade 1 and above as positive (on the current manufacturer scale of 1 to 4). For studies that used the prior reference scale card with band intensities graded on a scale of 1 to 5, we considered the test to be positive at grade 2 and above, corresponding to the current recommended positivity threshold. We excluded studies that did not use a positivity threshold that corresponded to the manufacturer's recommendations. The SILVAMP TB LAM (Fujifilm, Tokyo, Japan) has been evaluated in several studies [48, 49, 50], but demonstrated significant lot‐to‐lot variability in a large, multisite clinical study, resulting in the test not being commercialised or made available for clinical use [51]. We did not include data on SILVAMP TB LAM.

Target conditions

The target condition was any form of tuberculosis disease: pulmonary tuberculosis, extrapulmonary tuberculosis, or both. In this review, since the aim of parallel testing was to identify people with any type of tuberculosis disease, we did not regard pulmonary tuberculosis and extrapulmonary tuberculosis as separate target conditions.

Reference standards

We required studies to ascertain the presence or absence of tuberculosis disease using at least one of the following two reference standards.

Microbiological reference standard

For a microbiological reference standard, we required the following.

• Presence of tuberculosis disease was defined as a positive M. tuberculosis culture on any sample type or a positive nucleic acid amplification test (NAAT) result for the detection of tuberculosis on any sample other than LC‐aNAAT on respiratory samples.

• Absence of tuberculosis disease was defined as negative M. tuberculosis cultures and negative NAAT performed on any sample type other than LC‐aNAAT on respiratory samples.

We included studies that used either culture, NAAT (other than respiratory LC‐aNAAT), or a combination of both as the microbiological reference standard, with positivity defined as either test being positive. We used automated liquid culture methods (such as Mycobacterial Growth Indicator (MGIT)) and solid culture methods using Löwenstein Jensen's (LJ) medium as reference standards. The molecular method to detect tuberculosis could also be used as a reference standard, including LC‐aNAAT on non‐respiratory samples such as urine.

For microbiological reference standards, we considered as higher quality, a reference standard in which two or more specimen types were evaluated for tuberculosis diagnosis, and as lower quality, a reference standard when only one specimen type was evaluated for TB diagnosis. Of note, we excluded respiratory LC‐aNAAT and urine LF‐LAM results from the microbiological reference standard to avoid incorporation bias.

We expected all studies to obtain sputum culture(s) and some studies to obtain additional specimens from other body sites for culture or NAAT. Based on current WHO recommendations and clinical practice, NAAT on extrapulmonary samples (e.g. LC‐aNAAT on cerebrospinal fluid (CSF), urine, stool) is also sometimes undertaken for the purpose of categorising tuberculosis. We included non‐respiratory NAAT to improve the microbiological reference standard as there are risks of misclassification bias if we ignore these extrapulmonary microbiological tests, given that the target condition is both pulmonary and extrapulmonary tuberculosis.

Composite reference standard

Recognising that a microbiological reference standard, especially if relying on respiratory samples alone, may fail to detect tuberculosis in people with HIV, we also included studies using a composite reference standard.

For a composite reference standard, we required the following.

• The presence of tuberculosis disease was defined as a positive microbiological reference standard (as defined above) or positive smear microscopy, or author‐defined clinical criteria for tuberculosis, or a clinical decision to initiate treatment for tuberculosis.

• Absence of tuberculosis disease was defined as a negative microbiological reference standard and a negative composite reference standard (i.e. not meeting criteria for any of the microbiological or clinical components of a positive composite reference standard).

The composite reference standard is based on the results of microbiological tests, imaging studies, histology, and clinical characteristics, and includes at least one component test that was positive, according to the definition of the primary study authors. Criteria may include persisting symptoms (e.g. cough longer than two weeks, fever, night sweats, or weight loss), radiographic findings consistent with tuberculosis, positive sputum smear microscopy, histology suggestive of tuberculosis, or clinical decision to start tuberculosis treatment. We expected the definition of the composite reference standard to be heterogeneous across studies. Considerable uncertainty surrounds a clinical diagnosis of tuberculosis. Therefore, a reference standard that uses clinical tuberculosis is considered a lower quality reference standard and may incorrectly classify patients without tuberculosis disease as having tuberculosis disease. Moreover, the composite reference standard is at high risk of bias because the LC‐aNAAT on respiratory samples and LF‐LAM could be part of the composite reference standard (e.g. when the decision to start tuberculosis treatment directly depends on the results of LC‐aNAAT on respiratory samples) leading to incorporation bias.

Search methods for identification of studies

We attempted to identify all relevant studies regardless of language or publication status (published, unpublished, in press, and ongoing).

Electronic searches

The Cochrane Infectious Diseases Group Information Specialist (Vittoria Lutje) developed and performed the literature search on 3 November 2023, using the search strategy described in Supplementary material 1. We searched the following electronic databases.

• Cochrane Central Register of Controlled Trials (CENTRAL), up to issue 10 in 2023

• MEDLINE (Ovid, from 1946)

• Embase (Ovid, from 1947)

• Science Citation Index Expanded (SCI‐EXPANDED, from 1900), Conference Proceedings Citation Index – Science (CPCI‐S, from 1900), and BIOSIS Previews (from 1926), all three using the Web of Science platform

• Scopus (Elsevier, from 1970)

• WHO Global Index Medicus

• ClinicalTrials.gov (clinicaltrials.gov) and the search portal of the WHO International Clinical Trials Registry Platform (WHO ICTRP, trialsearch.who.int) for ongoing trials

• ProQuest Dissertations & Theses A&l (from 1861) for relevant dissertations

We did not use search filters for diagnostic accuracy studies [52]. We limited the search to studies published after 2010; Xpert MTB/RIF was first endorsed in 2010 for clinical use, and other LC‐aNAATs and LF‐LAM first became available several years later.

Searching other resources

We examined reference lists of included studies and any relevant reviews and studies that we identified. We contacted experts in the field and leading research groups working on new diagnostics such as SMART4TB, STOP TB, and FIND (Foundation for Innovative New Diagnostics). In addition, we checked studies submitted to WHO and shared with the review team after a WHO call for data on 17 January 2024.

We checked the included studies for retractions using Retraction watch and for errata using PubMed.

Data collection and analysis

Selection of studies

We uploaded titles and abstracts retrieved from the electronic literature search to EndNote to remove duplicates. Subsequently, we uploaded them to Covidence for study selection [53]. Two review authors independently screened all titles and abstracts to determine potentially eligible studies. The same review authors independently assessed full‐text articles. We resolved disagreements through discussion and, if necessary, involved a third review author as arbiter. We listed reasons for exclusion. Review authors who co‐authored any of the potentially eligible studies did not participate in assessing the eligibility of their own study [54]. We illustrated the study selection process in a PRISMA 2020 diagram [55]. We assessed studies included in the review for retractions using Retraction watch and for errata using PubMed (Supplementary material 21).

Data extraction and management

Using a data extraction form developed in advance (Supplementary material 6), two of the review authors independently extracted data on the following characteristics.

• General study characteristics (e.g. publication year, study design, setting)

• Participants (e.g. age, sex/gender, median CD4 cell count)

• Target condition (e.g. pulmonary tuberculosis, extrapulmonary tuberculosis, or both)

• Index tests (e.g. type of LC‐aNAAT and LF‐LAM, specimen types, positivity threshold)

• Reference standard (e.g. microbiological or composite reference standard, types of tests)

• Study results: two‐by‐two table data (true positives, false negatives, false positives, and true negatives). We recorded the number of participants in whom diagnostic testing was attempted (i.e. attempt for sputum and urine collection, transportation to lab, performance of index tests) but who did not have a result for any reason along the diagnostic testing pathway.

For studies where we could not extract these data directly from the publications, we reached out to study authors for data. They reported back in the format of three‐by‐three‐by‐two tables designed by review authors for the purpose of our review (Supplementary material 7). Some study authors shared individual‐participant data (IPD).

Two review authors independently extracted data from each study. We resolved disagreements through discussion, and, if necessary, by involving a third review author as arbiter. Review authors who co‐authored any of the included studies did not participate in data extraction of their own study. We contacted study authors for missing data and clarifications. We extracted data using Microsoft Excel.

Assessment of methodological quality

We assessed the methodological quality of included studies using QUADAS‐2 [56] and QUADAS‐C [57] tools (Supplementary material 8; Supplementary material 9). QUADAS‐2 is a tool for assessing the risk of bias in studies and the applicability of single‐test accuracy estimates; QUADAS‐C is an extension to QUADAS‐2 for assessing risk of bias in comparative accuracy estimates. Review authors independently assessed each study. We resolved disagreements through discussion and, if necessary, by involving a third review author as arbiter. Review authors who co‐authored any of the included studies did not participate in the assessment of their own study.

For the reference standard, some studies relied on sputum culture alone as the microbiological reference standard, which had the risk of tuberculosis disease being missed for the following reasons: i) the population of people with HIV may be characterised by low quality sputum specimens; ii) paucibacillary pulmonary disease; iii) sputum scarcity; or iv) extrapulmonary tuberculosis without pulmonary tuberculosis. Consequently, true‐positive index test results may be misclassified as false positives using sputum culture alone as a reference standard, and therefore underestimate specificity for LF‐LAM and underestimate the incremental value of the parallel strategy. The sensitivity of the microbiological reference standard can be increased by evaluating multiple specimens, including evaluating specimens from extrapulmonary sites of disease. If only one specimen was used for culture, we considered the risk of bias to be high in the reference standard domain. The quality of the reference standard is addressed in the reference standard domain of the QUADAS‐2 tool (Supplementary material 8).

Because of the uncertainties that surround a clinical diagnosis of tuberculosis, a composite reference standard based on additional tests and clinical characteristics (and no tuberculosis by the microbiological reference standard) may incorrectly classify people without tuberculosis as having tuberculosis disease. If patients are correctly classified as negative by the index test but incorrectly classified as positive by the composite reference standard, this will lead to an underestimation of the index tests’ sensitivity. In the absence of any other risk of bias, we answered ’unclear’ in the reference standard domain of the QUADAS‐2 tool if the composite reference standard involved clinical features to confirm tuberculosis diagnosis (Supplementary material 8). This is consistent with the 2016 Cochrane review [19], which also included composite reference standards. Variability in definitions is an inherent characteristic of the composite reference standard.

Some composite reference standards may introduce incorporation bias by including the results from respiratory LC‐aNAAT, LF‐LAM, or both as part of the criteria for confirming tuberculosis; for example, if the decision to start tuberculosis treatment was based directly on a positive respiratory LC‐aNAAT result. This would likely lead to an overestimation of diagnostic specificity. If the reference standard incorporated any of the LC‐aNAAT on respiratory samples, LF‐LAM tests, or both, we considered the risk of bias to be high in the reference standard domain of the QUADAS C tool (Supplementary material 9).

Statistical analysis and data synthesis

See Supplementary material 10 for the data analysis plan. For each study, we estimated sensitivity, specificity, and their absolute differences (parallel testing minus component test) between index tests under comparison, together with 95% confidence intervals (CIs). We presented these visually in forest plots and summary receiver operating characteristic (SROC) plots.

If at least three studies were available for a given analysis, we performed bivariate random‐effects meta‐analysis [58, 59] using a Bayesian approach to estimate summary sensitivity, summary specificity, the 95% CrI and the 95% prediction interval (PI) of each test.

To estimate the absolute summary difference in sensitivity and the absolute summary difference in specificity between tests, we also used the Bayesian bivariate random‐effects model, but with slight changes to the original approach. Instead of modelling at the lower level (within‐study) sensitivity and specificity and computing the difference between summary estimates across tests, we directly modelled the difference in sensitivity and difference in specificity within each study. This was accomplished by assuming that the difference in true positive (TP) (i.e. TP of parallel testing minus TP of component test) and the difference in true negative (TN) (i.e. TN of parallel testing minus TN of component test) both follow a binomial distribution. This approach naturally accounts for the dependence between the two tests being compared (i.e. the parallel testing strategy and the component test) and ensures that a difference in sensitivity cannot be lower than zero and a difference in specificity cannot be greater than zero, as would be expected when using two tests in parallel. As the binomial distribution does not allow negative integers (TN of parallel testing minus TN of component test is likely to be negative), we used positive values to fit the model and edited the specificity calculation to report a negative difference estimate.

We conducted these analyses in R statistical software version 4.3.2 using the package rjags [60]. We assessed model convergence using the Gelman‐Rubin statistic and trace plots; no convergence problems were detected.

We specified prior distributions as follows.

• Mean logit‐transformed sensitivity and mean logit‐transformed specificity: normal distribution with mean 0 and variance of 4, corresponding to a roughly uniform distribution on the probability scale

• Corresponding between‐study standard deviation: gamma distribution with a shape parameter of 2 and rate parameter of 0.5

• Correlation between the logit‐transformed sensitivities and specificities across studies: uniform distribution from −1 to 1

Overall, we analysed index test results as either test ‘positive’ or ‘negative’. For the parallel testing, we considered any positive component test result as a ‘positive’ result for the combination.

Approach to unavailable and nondeterminate test results

In some participants, the LC‐aNAAT or LF‐LAM test result was unavailable because the participant was unable to produce a sputum or urine sample or produced an inadequate sample for testing (i.e. unavailable test results). To avoid selection bias, we analysed the total number of participants with positive and negative reference standard results, irrespective of participants’ ability to provide a specific sample for the index tests.

For some specimens, testing may not have resulted in a valid result (i.e. a nondeterminate test result). For LC‐aNAATs, nondeterminate results for the detection of tuberculosis included tests classified as “INVALID”, “ERROR”, or “NO RESULT”. For LF‐LAM, the nondeterminate results for the detection of tuberculosis were the number of tests where the control band did not appear or where the test band intensity could not be interpreted.

For analytical purposes of evaluating a parallel testing strategy (in which either sputum LC‐aNAAT or urine LF‐LAM may contribute diagnostic data), we included unavailable or nondeterminate test results on one component test as ‘not positive’. For the primary analysis, we analysed participants with a ‘not positive’ as test‐negative. We performed a sensitivity analysis in which we excluded those participants with at least one unavailable or nondeterminate index test result (see below). When possible, we clarified the reasons for unavailable index test results (e.g. no sample obtained; sample inadequate for testing) and recorded the number of nondeterminate test results.

We did not include data on indeterminate results for the detection of rifampicin resistance.

Investigations of heterogeneity

We conducted all investigations of heterogeneity using diagnostic accuracy data based on the microbiological reference standard. We investigated heterogeneity in comparative accuracy estimates by performing subgroup analysis with the covariates.

• Signs and symptoms of tuberculosis (i.e. pulmonary, extrapulmonary or both) versus asymptomatic

• CD4 cell count (CD4 ≤ 200 versus CD4 > 200)

• Clinical setting (inpatient versus outpatient)

• Participants screening positive for tuberculosis ((e.g. chest X‐ray abnormal, CRP elevated, screening test positive with molecular WHO‐recommended rapid diagnostics)

• Advanced HIV (only data available for this stratum)

• Seriously ill (only data available for this stratum)

• Type of LC‐aNAAT (analysis restricted to within‐study comparisons)

For between‐study subgroup analyses, we computed the probability that the subgroup difference was nonzero and inferred that there was a substantial difference if this probability exceeded 95%.

Sensitivity analyses

We conducted all investigations of heterogeneity using diagnostic accuracy data. We performed the following sensitivity analyses by limiting inclusion in the meta‐analysis based on the following criteria.

• Exclusion of studies at high risk of bias (QUADAS‐C) and high concerns regarding applicability (QUADAS‐2) in at least one domain

• Exclusion of participants with at least one unavailable or non‐determinate index test result

• Alternative prior distributions that are less informative: increasing the variance of the normal distributions over the summary log odds of the sensitivity or specificity to 100 and using a uniform distribution ranging from zero to three over the between‐study standard deviation on the log odds scale

Assessment of reporting bias

We did not carry out a formal assessment of publication bias using methods such as funnel plots or regression tests because such techniques have not been helpful for diagnostic test accuracy studies [61]. However, we had extensive outreach to researchers to obtain data from unpublished studies, as well as a WHO public call for data, to limit the risk of bias.

Assessment of the certainty of the evidence

We assessed the certainty of the evidence using the GRADE approach for comparative test accuracy questions [62], which involves rating the risk of bias, indirectness, inconsistency, imprecision, and risk of publication bias across the body of evidence. We assessed the certainty that the true difference in sensitivity and true difference in specificity lie within the bounds of the 95% CrI (a minimally contextualised rating). We used the QUADAS‐2 and QUADAS‐C judgments to rate risk of bias and indirectness. We examined variability in study‐specific estimates (of difference in sensitivity or specificity) and overlap in CIs, and downgraded for inconsistency when point estimates of individual studies were outside our target of certainty rating. We did not downgrade for imprecision as our target of certainty rating was defined by the 95% CrIs of summary estimates. We did not rate down for publication bias as we regarded the risk of publication bias to be low.

Differences between protocol and review

See Table 4.

2. Differences between protocol and review.

Methods section	Difference from protocol
Objective	In our protocol, the objective was to compare the diagnostic accuracy of parallel testing versus any of the tests alone. The revised objective is to compare the accuracy of parallel testing versus LC‐aNAAT on respiratory samples. As current guidelines do not recommend LF‐LAM as a stand‐alone test [7] (as illustrated in the clinical pathway (Figure 1)), we did not include GRADE judgments for the comparison against LF‐LAM alone. However, results for comparing parallel testing versus LF‐LAM alone are included in supplementary files for (1) transparency, as we had prespecified this comparison in the protocol and (2) because it may be useful to some clinicians operating in circumstances in which LF‐LAM is being used as the sole diagnostic test (see Supplementary material 14; Supplementary material 15; Supplementary material 17; Supplementary material 18).
Data analysis	In our protocol, we wrote that meta‐analytic estimates of the difference in sensitivity and specificity would be obtained by adding the index test type as a covariate to the bivariate model. This produced 95% CrIs of those differences that contained impossible values, namely difference in sensitivity < 0 and difference in specificity > 0. To address this problem, we decided to directly model the difference in sensitivity and the difference in specificity within each study, as described in Statistical analysis and data synthesis. Sensitivity analyses that could not be performed because studies were too few for meta‐analysis are indicated in Supplementary material 11.
Assessment of the certainty of evidence	We originally planned to use the null value (Δsens/Δspec = 0) as our judgment threshold for rating certainty. However, the null value is an uninformative threshold. Because of the way our parallel testing strategy is defined (any positive component test result implies a positive parallel testing strategy), the difference in sensitivity cannot be below zero and the difference in specificity cannot be above zero. Therefore, we instead rated the certainty that the true difference in sensitivity or specificity lies within the bounds of the 95% CrIs.

CrI: credible interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

Results

Results of the search

We identified and screened a total of 456 records for inclusion, of which 130 were assessed as full‐text reports. We excluded 95 studies, with reasons provided in Supplementary material 3; four studies are ongoing with no final data (Supplementary material 4); and we included 31 unique studies (Supplementary material 2). All included studies provided data comparing the parallel use of respiratory LC‐aNAAT and urine LF‐LAM to LC‐aNAAT on respiratory samples alone for the detection of tuberculosis. Figure 2 shows the flow of studies in the review.

2.

Study flow

Of the 31 included studies, we could extract data directly from the study publication for two studies (Peter 2015 [12]; Shah 2014 [63]). On request, study authors provided additional summary data reanalysed for this review for 15 studies (Åhsberg 2023 [64]; Bjerrum 2015 [11]; Boloko 2022 [65]; Bonnet 2023 [66]; De Vasconcellos 2021 [67]; Huerga 2023 [68]; Kebede 2021 [69]; Liu 2022 [70]; Olbrich 2023 [71]; Reeve 2023 [72]; Shapiro 2022 [73]; Spooner 2022 [74]; Stead 2023 [75]; Szekely 2022 [76]; Van Hoving 2019 [77]) and shared individual participant data for three studies (Benjamin 2019 [78]; Broger (Cohort4) 2020 [79]; Gils 2023 [80]). In 11 studies (Calligaro 2017 [81]; García 2020 [82]; Gupta‐Wright 2018a [83]; Huerga 2017 [84]; Huerga 2019 [85]; Huerga 2020a [86]; Huerga 2021 [87]; Kasaro 2020 [88]; Lawn 2017 [89]; Peter 2016 [90]; Yoon 2019 [91]), study authors agreed that we could re‐analyse summary data based on the individual participant data shared by them for previous meta‐analyses by Broger and colleagues [4]. All studies were classified as having a fully paired design for this comparative diagnostic accuracy review. The original study designs for included studes are listed in Table 5. Data analyses that we had planned but did not undertake are detailed in Supplementary material 11.

3. Overview of included studies and data syntheses.

Study	Clinical setting and country	Study design and participant sampling	Index tests	Study target condition	TB prevalence per total sample size, % (n/N)	TB prevalence per total sample size, % (n/N)	Clinical presentation (% symptoms)	CD4 count (median cells/mm3)
					MRS	CRS
Åhsberg 2023	Inpatients (Ghana)	Consecutive and prospective enrolment as part of a multicentre cluster‐randomised trial to assess urine Determine LF‐LAM coupled with healthcare worker training	Xpert MTB/RIF or Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, EPTB	5.1 % (4/79)	30.2% (42/139)	Symptomatic (100% W4SS positive)	91
Benjamin 2019	Outpatients (Brazil)	Consecutive and prospective enrolment into a multicentre diagnostic accuracy study evaluating urine Determine LF‐LAM in conjunction with sputum smear microscopy or sputum Xpert MTB/Rif	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	25.8% (39/151)	27.2% (41/151)	Severe immunosuppression and presumptive TB (93% with cough)	42
Bjerrum 2015	Inpatients (15%) and outpatients (85%) (Ghana)	Consecutive and prospective enrolment into a single‐centre urine Determine LF‐LAM diagnostic accuracy study	Xpert MTB/RIF, Determine LF‐LAM	PTB	16.1% (31/192)	32.8% (63/192)	Unselected (91% W4SS positive)	152
Boloko 2022	Other (admission to the emergency room and the medical ward) (South Africa)	Prospective and consecutive enrolment into a cohort study evaluating yield of blood Xpert MTB/RIF Ultra based on available biobanked samples	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	70.7% (466/659)	88.6% (584/659)	High suspicion of TB (69% with cough)	62
Bonnet 2023	Outpatients (Côte d’Ivoire, Uganda, Cambodia, and Vietnam)	A diagnostic performance and yield study assessing different TB screening strategies (W4SS, CXR, Determine LAM, sputum Xpert MTB/RIF), based on the STATIS trial. Participants were prospectively and randomly assigned to the extensive TB screening arm.	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	NA	9.1% (48/525)	Unselected (62% W4SS positive)	28
Broger (Cohort4) 2020	Outpatients (Vietnam)	A urine Determine LF‐LAM diagnostic accuracy study based on biobanked urine samples from consecutively and prospectively enroled participants from Vietnam (Cohort4)	Xpert MTB/RIF, Determine LF‐LAM	PTB	66.7% (38/57)	71.9% (41/57)	High suspicion of TB (100% pulmonary symptomatic disease)	54
Calligaro 2017	Other (community‐based screening in a vehicle) (South Africa, Zimbabwe)	Consecutive and prospective enrolment as part of a multicentre randomised controlled trial to assess a novel package of TB diagnostics (sputum Xpert MTB/RIF, urine Determine LF‐LAM, and sputum culture)	Xpert MTB/RIF, Determine LF‐LAM	PTB	11.7% (29/248)	40.5% (102/252)	Unselected (95% with cough)	234
De Vasconcellos 2021	Other (intensive care unit) (South Africa)	Consecutive and prospective enrolment into a single centre cohort to assess urine Determine LF‐LAM diagnostic accuracy and clinical utility	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, EPTB	NA	46.7% (7/15)	Critically ill with high suspicion of TB (100% in mechanical ventilation)	156
García 2020	Inpatients (30%) and outpatients (70%) (Guatemala)	Consecutive and prospective enrolment into a multicentre prospective cohort assessing the efficacy of untreated and α‐mannosidase‐treated urine Determine LF‐LAM to detect TB	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	17.9% (51/285)	NA	Symptomatic (100% W4SS positive)	129
Gils 2023	Other (container‐based clinic in a Lesotho hospital setting and a mobile clinic in a South African primary healthcare setting) (Lesotho, South Africa)	Consecutive and prospective enrolment into a multicentre advanced HIV package of care implementation study	Xpert MTB/RIF, Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB	9.7% (59/609)	NA	Symptomatic (100% W4SS positive)	Not reported. 29.9% had a CD4 count below 200 cells/mm3
Gupta‐Wright 2018a	Inpatients (Malawi, South Africa)	Consecutive and prospective enrolment as part of a multicentre randomised controlled trial to assess urine Determine LF‐LAM and urine Xpert MTB/RIF as an intervention	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	5.8% (74/1266)	16.3% (210/1287)	Unselected (90% W4SS positive)	231
Huerga 2017	Inpatients (76%) and outpatients (24%) (Kenya)	Consecutive and prospective enrolment into a single‐centre urine Determine LF‐LAM diagnostic yield study	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	33.0% (113/342)	70.7% (335/474)	Symptomatic (100% had either cough lasting > 2 weeks or cough of any duration combined with any other W4SS symptom)	109
Huerga 2019	Outpatients (Mozambique)	Consecutive and prospective enrolment into a multicentre urine Determine LF‐LAM diagnostic yield study	Xpert MTB/RIF, Determine LF‐LAM	PTB	28.4% (52/183)	28.2% (67/238)	Symptomatic (100% W4SS positive)	40
Huerga 2020a	Outpatients (Malawi)	Consecutive and prospective enrolment into a multicentre urine Determine LF‐LAM diagnostic yield study	Xpert MTB/RIF, Determine LF‐LAM	PTB	NA	13.4% (65/485)	Symptomatic (100% W4SS positive)	341
Huerga 2021	Inpatients (Malawi)	Consecutive and prospective enrolment into a single‐centre cohort to assess urine Determine LF‐LAM diagnostic yield	Xpert MTB/RIF, Determine LF‐LAM	PTB	NA	23.5% (91/387)	Unselected (90% W4SS positive)	173
Huerga 2023	Outpatients (Uganda, Kenya, Mozambique, and South Africa)	Consecutive and prospective enrolment into a multicentre urine FujiLAM and Determine LAM diagnostic accuracy study	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, EPTB	6.3% (92/1471)	20.5% (306/1494)	Symptomatic or advanced HIV (65% with TB symptoms; 35% with advanced HIV)	528 (symptomatic participants) and 128 (participants with advanced HIV)
Kasaro 2020	Outpatients (Zambia)	Consecutive and prospective enrolment into a multicentre before‐and‐after study using a sputum Xpert MTB/RIF algorithm intervention	Xpert MTB/RIF, Determine LF‐LAM	PTB	14.2% (70/493)	14.6% (82/563)	Symptomatic or newly diagnosed with HIV (NR)	299
Kebede 2021	Inpatients (Ethiopia)	A diagnostic yield analysis assessing sputum Xpert MTB/RIF Ultra and urine Determine LF‐LAM, based on a single‐centre prospective cohort assessing the role of empirical TB treatment	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB	19.2% (10/52)	NA	Symptomatic (100% W4SS positive)	46.2% had a CD4 count below 200 cells/mm3
Lawn 2017	Inpatients (South Africa)	A urine Determine LF‐LAM diagnostic accuracy and yield analysis, based on a single‐centre prospective observational study determining the prevalence and predictors of TB amongst admitted patients with HIV	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	30.7% (129/420)	NA	Unselected (92% W4SS positive)	149
Liu 2022	Inpatients (China)	Consecutive and prospective enrolment into a single‐centre urine Determine LF‐LAM diagnostic accuracy study	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	10.6% (9/85)	NA	Symptomatic or seriously ill or advanced HIV (58% W4SS positive)	87
Olbrich 2023	Inpatients (37%) and outpatients (63%) (India, Malawi, Mozambique, South Africa, Tanzania)	Consecutive and prospective enrolment into a multicentre diagnostic accuracy study evaluating multiple tests for childhood TB	Xpert MTB/RIF Ultra, Determine LAM	PTB, EPTB	13.0% (6/46)	58.7% (27/46)	Suspicion of PTB (100% had at least one of: chest radiograph suggestive of TB, weight loss, cough, fever, signs and symptoms of EPTB)	317
Peter 2015	Outpatients (South Africa, Tanzania, Zambia)	A urine Determine LF‐LAM diagnostic accuracy study, nested within a multicentre prospective study assessing feasibility, accuracy, and clinical effect of sputum Xpert MTB/RIF	Xpert MTB/RIF, Determine LF‐LAM	PTB	31.9% (92/288)	54.5% (157/288)	Symptomatic (100% W4SS positive)	210
Peter 2016	Inpatients (Tanzania, Zambia, Zimbabwe, South Africa)	Consecutive and prospective enrolment as part of a multicentre randomised controlled trial to assess urine Determine LF‐LAM as an intervention	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	30.8% (144/467)	59.3% (293/494)	Symptomatic (100% W4SS positive)	81
Reeve 2023	Unclear (South Africa)	Prospective and consecutive enrolment into a diagnostic accuracy study of sputum Xpert MTB/RIF and Ultra and the diagnostic yield of urine Determine LF‐LAM	Xpert MTB/RIF, Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB	12.1% (122/1011)	NA	Unselected ART‐naive participants (55% W4SS positive)	299
Shah 2014	Inpatients (59%) and outpatients (41%) (Uganda)	Prospective enrolment into a multicentre comparative diagnostic accuracy study	Xpert MTB/RIF, Determine LF‐LAM	PTB, mycobacteraemia	49.5% (103/208)	NA	Symptomatic (100% W4SS positive)	63 (participants with TB) and 280 (participants without TB)
Shapiro 2022	Unclear (South Africa)	Consecutive and prospective enrolment into a multicentre study comparing parallel testing of two different non‐sputum samples as an alternative to sputum testing as part of the PROVE‐TB cohort	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, (EPTB)	128% (79/616)	41.8% (242/579)	Unselected (97% had any of cough, weight loss, fevers/chills, loss of appetite, fatigue)	205
Spooner 2022	Outpatients (South Africa)	Consecutive prospective sampling into a multicentre diagnostic accuracy study of sputum TB‐LAMP and urine Determine LF‐LAM	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB	10.2% (73/713)	12.7% (91/719)	Unselected (82% W4SS positive)	262
Stead 2023	Inpatients (South Africa)	Consecutive and prospective enrolment into a multicentre urine Xpert MTB/RIF Ultra and urine Determine LF‐LAM diagnostic accuracy and yield study	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	28.6% (68/238)	49.2% (117/238)	Symptomatic (100% W4SS positive)	76
Szekely 2022	Inpatients (42%) and outpatients (58%) (Malawi, South Africa, Tanzania, Thailand, Uganda, Vietnam, and Zambia)	Consecutive and prospective enrolment into a multicentre diagnostic accuracy study of the urine FujiLAM	Xpert MTB/RIF Ultra, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	13.4% (218/1630)	34.3% (486/1416)	Unselected inpatients and symptomatic outpatients (93% W4SS positive)	372
Van Hoving 2019	Other (admitted to emergency centre) (South Africa)	Consecutive and prospective enrolment into a single‐centre urine Determine LF‐LAM diagnostic accuracy study	Xpert MTB/RIF, Determine LF‐LAM	PTB, EPTB, mycobacteraemia	36.0% (148/411)	55.0% (226/411)	Symptomatic (100% W4SS positive)	87
Yoon 2019	Outpatients (Uganda)	Consecutive and prospective enrolment into a multicentre diagnostic performance and yield study of TB diagnostic tests (sputum Xpert MTB/RIF and urine Determine LF‐LAM) alone and combined with screening strategies	Xpert MTB/RIF, Determine LF‐LAM	PTB	11.4% (49/431)	NA	Unselected (90% W4SS positive)	149

Abbreviations and definitions: CRS: composite reference standard; Determine LF‐LAM: Determine TB LAM Ag strip test (Abbott, Chicago, USA); EPTB: extrapulmonary tuberculosis; MRS: microbiological reference standard; n: number of participants diagnosed with TB; N: total number of participants; NA: not applicable; PTB: pulmonary tuberculosis; TB: tuberculosis disease; W4SS: WHO‐recommended four‐symptom screen; Xpert MTB/RIF: Xpert MTB/RIF low‐complexity automated NAAT assay used with the GeneXpert system (Cepheid, Sunnyvale, USA); Xpert MTB/RIF Ultra: Xpert MTB/RIF ultra low‐complexity automated NAAT assay used with the GeneXpert system (Cepheid, Sunnyvale, USA)

Description of included studies

Figure 2 shows the study identification and selection process for the review. Table 5 provides an overview of the included studies and syntheses. Study‐specific characteristics are detailed in Supplementary material 2. A summary of key characteristics for all included studies is provided in Supplementary material 12. None of the included studies had corrections or retractions (see Supplementary material 21).

The 31 studies were conducted in 19 different low‐ and middle‐income countries. Fourteen of the 19 countries (74%) were high‐burden countries for tuberculosis and HIV [92], and 27 of 31 (87%) studies enroled participants from 12 countries located in sub‐Saharan Africa. One study contributed data on adolescents only, aged 10 to 15 years with a median age of 12 (Olbrich 2023). The other 30 studies included people with HIV with a median age between 31 and 46 years.

Types of LC‐aNAAT evaluated were Xpert MTB/RIF (21 studies), Xpert Ultra (7 studies), a pragmatic mix of the two assays (1 study), and both Xpert MTB/RIF and Xpert Ultra (2 studies). All studies of LF‐LAM involved the Determine TB LAM assay.

Twenty‐seven studies compared the accuracy of the parallel strategy of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone and involved 12,651 participants, 2368 (19%) with tuberculosis according to the microbiological reference standard. Tuberculosis prevalence, according to the microbiological reference standard, ranged from 5% in Åhsberg 2023 to 71% in Boloko 2022. The number and type of samples obtained for the reference standard varied. In Åhsberg 2023, non‐respiratory Xpert or culture was not standardised (relied on routine care), and in Gupta‐Wright 2018a, sputum culture was not standardised (relied on routine care), such that the reference standard could have been exclusively based on either respiratory or non‐respiratory samples alone.

Twenty‐three studies compared the accuracy of the parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone and involved 11,109 participants, 3723 (34%) with tuberculosis according to a composite reference standard. Tuberculosis prevalence according to the composite reference standard ranged from 9% in Bonnet 2023 to 89% in Boloko 2022. The definition of the composite reference standard differed between studies and was based on different combinations of microbiological tests, histology, radiographic findings, clinical assessment, or decision to initiate tuberculosis treatment. For 20 of 23 studies, the composite reference standard included initiation of tuberculosis treatment. The LC‐aNAAT or LF‐LAM was incorporated in the composite reference standard in 21 of 23 studies: LC‐aNAAT in 21 studies and LF‐LAM in 11 studies.

Methodological quality of included studies

QUADAS‐2

Figure 3 shows our assessment of the risk of bias (as high, low, or unclear) and concerns about the applicability of the results to the objective of the review (as high, low, or unclear) for each of the 31 included studies.

3.

Risk of bias and applicability concerns in single‐test accuracy estimates (QUADAS‐2)

Review authors' judgments about each domain for each included study. For the index test domains, though assessed separately for each index test, we reported the highest (i.e. worst) judgment here.

In the patient selection domain, we considered 11 studies (35%) to be at high risk of bias for the following reasons: nine studies excluded participants who could not produce sputum samples (and/or urine) (Benjamin 2019; Bjerrum 2015; Broger (Cohort4) 2020; Calligaro 2017; Kebede 2021; Peter 2015; Peter 2016; Shah 2014; Stead 2023); one study enroled participants who had started tuberculosis treatment before index testing (De Vasconcellos 2021); and one study enroled a subset of participants with a known positive sputum Xpert ultra result for convenience that could have inflated sensitivity (Shapiro 2022). Regarding applicability, we had high concern about three studies (10%): one study because participants were evaluated exclusively as part of a community screening strategy and did not resemble people with presumed HIV/TB co‐infection (Calligaro 2017); one study exclusively enroled participants that had screened negative with Xpert MTB/RIF (Kebede 2021); and one study enroled additional Xpert Ultra positive participants by convenience and also enroled HIV negative participants (> 90% HIV positive) (Shapiro 2022). We were unclear about the applicability of two studies (6%) (Boloko 2022; Olbrich 2023).

In the index test domain, we judged that all studies had a low risk of bias for the index tests. For LC‐aNAATs, the results are automatically generated. For LF‐LAM, all studies prespecified the grade used for positivity and interpreted the test at the recommended positivity threshold without knowledge of the results of the reference standard. Regarding applicability, with respect to LC‐aNAAT, we had low concern about all studies. For LF‐LAM, we judged four studies as unclear in terms of applicability: in three of them, stored urine samples were tested (Boloko 2022; Broger (Cohort4) 2020; García 2020), and in one study, there was missing information (Åhsberg 2023).

In the reference standard domain, 27 studies evaluated diagnostic accuracy against a microbiological reference standard, 23 studies against a composite reference standard, and 20 studies included both.

For the microbiological reference standard, we considered 12 studies (44%) to have a low risk of bias because the studies attempted to collect sputum for culture and specimens from different sites for extrapulmonary tuberculosis (e.g. urine for Xpert), and the reference standard was interpreted without knowledge of the results of the index test. We considered 12 studies (44%) to be at high risk of bias: 10 studies did not include testing of any extrapulmonary specimens (Bjerrum 2015; Broger (Cohort4) 2020; Calligaro 2017; Gils 2023; Huerga 2019; Kasaro 2020; Kebede 2021; Peter 2015; Reeve 2023; Yoon 2019), one study did not have a standardised approach for sputum culture (Gupta‐Wright 2018a), and one study did not have a standardised approach for testing extrapulmonary samples (Spooner 2022). We considered three studies (11%) to have an unclear risk of bias: one study included a positive blood Xpert Ultra as part of tuberculosis cases (Boloko 2022), and in two studies, the approach for testing extrapulmonary samples was unclear (Liu 2022; Shapiro 2022).

For the composite reference standard, 21 (91%) studies had a high risk of bias: 19 studies involved clinical judgement to confirm tuberculosis where it was unclear if this correctly classified participants and the studies interpreted the reference standard with knowledge of the index test results (Åhsberg 2023; Bjerrum 2015; Boloko 2022; Bonnet 2023; Calligaro 2017; De Vasconcellos 2021; Gupta‐Wright 2018a; Huerga 2017; Huerga 2020a; Huerga 2021; Huerga 2023; Kasaro 2020; Olbrich 2023; Peter 2015; Peter 2016; Shapiro 2022; Stead 2023; Szekely 2022; Van Hoving 2019); two studies interpreted the reference standard with knowledge of the index test (Benjamin 2019; Spooner 2022). One study had an unclear risk of bias as the reference standard included clinical judgement, and it was unclear if blinding to index tests was performed (Broger (Cohort4) 2020).

Regarding the applicability of the reference standard, we had low concern in terms of 28 of the 31 studies (90%) because these studies performed a test to identify M. tuberculosis species (speciation) for the microbiological reference standard, made an attempt by design to ensure a minimum of standard testing, and had appropriate criteria for composite reference standard. We had high concern about two studies (6%) as they lacked protocol‐directed testing of either sputum culture (Gupta‐Wright 2018a) or of both sputum and extrapulmonary samples (Liu 2022). One study (3%) had unclear applicability because we could not tell whether the study performed speciation (Kasaro 2020).

In the flow and timing domain, we considered nine studies (29%) to be at high risk of bias because not all participants received the same reference standard (García 2020; Gupta‐Wright 2018a; Huerga 2017; Huerga 2021; Liu 2022; Spooner 2022; Stead 2023) or were included in the review analysis (De Vasconcellos 2021; Huerga 2019; Liu 2022). Three studies (10%) had unclear risk of bias: in two studies, it was unclear if all participants received the same reference standard (Huerga 2020a; Huerga 2023), and in one study, the time interval between the index test and reference standard was more than seven days (Van Hoving 2019).

QUADAS‐C

Figure 4 shows the risk of bias for 31 studies comparing parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine versus LC‐aNAAT on respiratory samples alone. All studies used a paired diagnostic accuracy design for the purpose of this review.

4.

Risk of bias in comparative accuracy estimates (QUADAS C)

Review authors' judgements about each domain for each included study

In the patient selection domain, we considered 10 studies (32%) to be at high risk of bias as the risk of patient selection bias for both index tests was considered high, mainly because studies excluded participants that could not produce samples for testing (Benjamin 2019; Bjerrum 2015; Broger (Cohort4) 2020; De Vasconcellos 2021; Kebede 2021; Peter 2015; Peter 2016; Shah 2014; Shapiro 2022; Stead 2023).

In the index test domain, we considered all but one study (97%) to have a low risk of bias as we judged the risk of bias for the index tests as low in the QUADAS‐2 assessment. The studies provided sufficient description that the results of LF‐LAM were blinded to the LC‐aNAAT results. For one study (Shapiro 2022), it was unclear if LF‐LAM was interpreted by someone blinded to the LC‐aNAAT result. We found it unlikely that conduction and interpretation of one index test would have any unwanted effect or advantage for subsequent index tests.

In the microbiological reference standard domain, we considered 12 studies (44%) to be at high risk of bias as we judged that the reference standard used was likely to misclassify participants for LF‐LAM and LC‐aNAAT in QUADAS 2 (Bjerrum 2015; Broger (Cohort4) 2020; Calligaro 2017; Gils 2023; Gupta‐Wright 2018a; Huerga 2019; Kasaro 2020; Kebede 2021; Peter 2015; Reeve 2023; Spooner 2022; Yoon 2019). Two studies (7%) had an unclear risk of bias for the comparison as we did not know if the reference standard for the index test was likely to classify the participants correctly (Liu 2022; Shapiro 2022). All studies avoided incorporating LC‐aNAAT and LF‐LAM as part of the reference standard.

In the composite reference standard domain, we considered 22 of 23 studies (96%) to be at high risk of bias: 21 studies as we considered the risk of bias for the index test to be high in QUADAS‐2, and 21 studies incorporated either LF‐LAM, LC‐aNAAT, or both in the reference standard to classify participants as tuberculosis cases. We considered one study (4%) to have an unclear risk of bias for the comparison as the risk of bias for this domain was unclear in QUADAS 2 (Broger (Cohort4) 2020).

In the flow and timing domain, we considered nine studies (29%) to be at high risk of bias because they had high risk of bias for the individual test in QUADAS 2 (Gupta‐Wright 2018a; Huerga 2017; Huerga 2021; Liu 2022; Spooner 2022; Stead 2023), the same reference standard was not used across index test groups comparison (Gupta‐Wright 2018a; Huerga 2017; Huerga 2021; Huerga 2023; Liu 2022; Spooner 2022), or the proportions of unavailable test results differed considerably between LC‐aNAAT and LF‐LAM (Gupta‐Wright 2018a; Huerga 2017; Huerga 2020a; Huerga 2021; Huerga 2023; Lawn 2017; Liu 2022; Stead 2023). Test results were unavailable mainly because of the inability of participants to produce samples, in particular sputum. Test results were sometimes unavailable due to nondeterminate test results for LF‐LAM or invalid LC‐aNAAT, although these were often reported to be low. We considered one study to have an unclear risk of bias (De Vasconcellos 2021).

Findings

Comparative diagnostic accuracy based on the microbiological reference standard

Results for the comparative diagnostic accuracy of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine versus LC‐aNAAT on respiratory samples alone for the detection of tuberculosis using a microbiological reference standard are presented in Table 6.

4. Parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone for detection of tuberculosis using a microbiological reference standard.

Participants	Studies (participants)	% Pooled sensitivity (95% CrI)	% Pooled specificity (95% CrI)
ALL
LF‐LAM	27 (12,651)	39.1 (32.6 to 45.9)	91.9 (88.7 to 94.4)
LC‐aNAAT	27 (12,651)	68.0 (60.8 to 74.9)	96.7 (95.7 to 97.6)
Parallel	27 (12,651)	77.5 (73.4 to 81.3)	89.4 (85.8 to 92.3)
Summary difference between parallel and LC‐aNAAT		6.7 (3.8 to 10.7)	−6.8 (−9.5 to −4.7)
SENSITIVITY ANAYSES
Excluding studies with high risk of bias
LF‐LAM	6 (3110)	43.3 (30.1 to 57.9)	86.2 (78.4 to 91.4)
LC‐aNAAT	6 (3110)	59.9 (45.9 to 72.6)	95.0 (91.8 to 97.2)
Parallela	6 (3110)	73.4 (61.7 to 82.9)	82.5 (74.3 to 88.5)
Summary difference between parallel and LC‐aNAAT		13.8% (7.6 to 22.9)	−12.2 (−19.5 to −7.3)
Excluding participants with at least one unavailable or nondeterminate index test result
LF‐LAM	27 (10,685)	38.7 (32.5 to 45.2)	91.9 (88.7 to 94.4)
LC‐aNAAT	27 (10,685)	75.1 (70.7 to 79.3)	96.2 (95.1 to 97.2)
Parallela	27 (10,685)	80.4 (76.8 to 83.6)	89.3 (85.7 to 92.2)
Summary difference between parallel and LC‐aNAAT		4.4 (2.8 to 6.5)	−6.9 (−9.1 to −4.5)
Alternative prior distributions
LF‐LAM	27 (12,651)	38.8 (31.6 to 46.3)	92.2 (88.5 to 94.8)
LC‐aNAAT	27 (12,651)	68.2 (60.2 to 75.8)	96.8 (95.7 to 97.7)
Parallela	27 (12,651)	77.5 (73.3 to 81.6)	89.6 (85.7 to 92.7)
Summary difference between parallel and LC‐aNAAT		6.1 (3.1 to 10.6)	−6.6 (−9.6 to −4.3)

Abbreviations: CrI: credible Interval; IPD: individual participant data; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aParallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination

Diagnostic accuracy of the parallel use of respiratory LC‐aNAAT and urine LF‐LAM

See Figure 5. The sensitivity of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine ranged from 33% to 100% and the specificity from 57% to 100%. The pooled sensitivity of parallel testing was 77.5% (95% CrI 73.4 to 81.3) and the specificity 89.4% (95% CrI 85.8 to 92.3) (27 studies; 12,651 participants, 2368 (19%) with tuberculosis).

5.

Forest plots of the diagnostic accuracy for the parallel use of respiratory LC‐aNAAT and urine LF‐LAM, respiratory LC‐aNAAT alone, and urine LF‐LAM alone, using a microbiological reference standard

The squares represent the sensitivity and specificity of one study; the black line represents its CI.

Abbreviations: CI: confidence interval; FP: false positive; FN: false negative; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay; TP: true positive; TN: true negative

Diagnostic accuracy of the component tests

See Figure 5. The sensitivity of respiratory LC‐aNAAT alone ranged from 24% to 100%, and specificity from 91% to 100%. The pooled sensitivity of LC‐aNAAT alone was 68.0% (95% CrI 60.8 to 74.9) and specificity 96.7% (95% CrI 95.7 to 97.6) (27 studies; 12,651 participants, 2368 (19%) with tuberculosis).

Across studies, the sensitivity of urine LF‐LAM alone ranged from 0% to 75%, and specificity from 59% to 100%. The pooled sensitivity of LF‐LAM alone was 39.1% (95% CrI 32.6 to 45.9) and specificity 91.9% (95% CrI 88.7 to 94.4) (27 studies; 12,651 participants, 2368 (19%) with tuberculosis).

Incremental accuracy of parallel testing versus LC‐aNAAT alone

See Figure 6. For studies included, the absolute difference in sensitivity for the parallel use of respiratory LC‐aNAAT and urine LF‐LAM versus respiratory LC‐aNAAT alone ranged from 0% in some studies (Bjerrum 2015; Kebede 2021; Liu 2022; Olbrich 2023; Spooner 2022) to an increase of 38% in one study (Gupta‐Wright 2018a). The absolute difference in specificity for parallel testing versus LC‐aNAAT alone ranged from 0% (Broger (Cohort4) 2020; Kebede 2021) to a decrease of 34% (Huerga 2019).

6.

Forest plot of study‐specific estimates of difference in sensitivity and specificity between the parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone for tuberculosis in adults and adolescents with HIV, based on the microbiological reference standard

The black squares represent the point estimate of one study; the black horizontal line, its 95% CI. The QUADAS‐C risk‐of‐bias judgements by study are included on the right.

Parallel test: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination.

Abbreviations: CI: confidence interval; CrI: credible interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

The pooled difference in sensitivity between parallel testing versus LC‐aNAAT alone was 6.7% (95% CrI 3.8 to 10.7; 95% prediction interval (PI) 0.6% to 45.9), and the pooled difference in specificity between parallel testing versus LC‐aNAAT alone was −6.8% (95% CrI −9.5 to −4.7; 95% PI −32.8 to −1.1). See Figure 7 and Supplementary material 13.

7.

Forest plot of pooled difference in sensitivity and specificity (all studies combined) by index test: LF‐LAM, LC‐aNAAT, and their parallel use

The squares represent the pooled sensitivity and specificity; the black horizontal line, its 95% CrI. The pooled difference in sensitivity and specificity between parallel testing and LC‐aNAAT alone is indicated by a line connecting two squares. This pooled difference may not correspond to the difference between the pooled single‐test accuracy estimates (see Statistical analysis and data synthesis).

Parallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted concurrently on the same person, and a positive result on either test implies a positive result for the combination

Abbreviations: CRS: composite reference standard; CrI: credible interval; LCa‐NAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay; MRS: microbiological reference standard; N: total number of participants

Comparative diagnostic accuracy based on the composite reference standard

Results for the comparative diagnostic accuracy of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM on urine versus LC‐aNAAT on respiratory samples alone for the detection of tuberculosis using a composite reference standard are presented in Table 7.

5. Parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone for detection of tuberculosis using a composite reference standard.

Participants	Studies (participants)	% Pooled sensitivity (95% CrI)	% Pooled specificity (95% CrI)
ALL
LF‐LAM	23 (11,109)	38.7 (30.8 to 47.0)	96.3 (93.0 to 96.3)
LC‐aNAAT	23 (11,109)	46.7 (38.5 to 55.2)	99.9 (99.8 to 100)
Parallela	23 (11,109)	67.7 (59.9 to 74.6)	96.2 (92.8 to 98.1)
Summary difference between parallel and LC‐aNAAT		16.0 (10.7 to 22.9)	−3.5 (−6.6 to −1.7)
SENSITIVITY ANALYSES
Excluding participants with at least one unavailable or nondeterminate index test result
LF‐LAM	23 (8922)	38.2 (30.9 to 46.0)	96.1 (92.8 to 98.0)
LC‐aNAAT	23 (8922)	53.8 (45.5 to 62.0)	99.9 (99.8 to 100)
Parallela	23 (8922)	69.7 (61.5 to 77.1)	96.1 (92.7 to 98.0)
Summary difference between parallel and LC‐aNAAT		12.5 (8.4 to 17.6)	−3.6 (−6.7 to −1.8)
Alternative prior distributions
LF‐LAM	23 (11,109)	38.3 (29.5 to 47.6)	96.8 (93.2 to 98.7)
LC‐aNAAT	23 (11,109)	45.6 (36.5 to 54.9)	100 (99.9 to 100)
Parallela	23 (11,109)	67.8 (59.0 to 75.5)	96.8 (93.1 to 98.7)
Summary difference between parallel and LC‐aNAAT		15.7 (9.8 to 23.5)	−3.2 (−6.7 to −1.3)

Abbreviations: CrI: credible interval; IPD: individual participant data; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aParallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination.

See Figure 8 for forest plots of the diagnostic accuracy of parallel testing, LC‐aNAAT alone, and LF‐LAM alone based on a composite reference standard. The pooled sensitivity of parallel testing was 67.6% (95% CrI 59.9 to 74.6) and pooled specificity 96.2% (95% CrI 92.8 to 98.1) (23 studies; 11,109 participants, 3723 (34%) with tuberculosis).

8.

Forest plots of the diagnostic accuracy for the parallel use of respiratory LC‐aNAAT and urine LF‐LAM, respiratory LC‐aNAAT alone, and urine LF‐LAM alone, using a composite reference standard

The squares represent the sensitivity and specificity of one study; the black line, its 95% CI.

Abbreviations: CI: confidence interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay; TP: true positive; FP: false positive; FN: false negative; TN: true negative

Incremental accuracy of parallel testing versus LC‐aNAAT alone

See Figure 9. For studies included, the absolute difference in sensitivity for the parallel use of respiratory LC‐aNAAT and urine LF‐LAM versus respiratory LC‐aNAAT alone ranged from 2% in some studies (Benjamin 2019; Broger (Cohort4) 2020) to an increase of 55% in others (Åhsberg 2023; Gupta‐Wright 2018a). The absolute difference in specificity for parallel testing versus LC‐aNAAT alone ranged from 0% (Broger (Cohort4) 2020; De Vasconcellos 2021; Gupta‐Wright 2018a; Spooner 2022) to a decrease of 36% (Huerga 2019).

9.

Forest plot of study‐specific estimates of difference in sensitivity and specificity between the parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone for tuberculosis in adults and adolescents with HIV, based on the composite reference standard

The black squares represent the point estimate of one study; the black horizontal line, its 95% CI. The QUADAS‐C risk‐of‐bias judgements by study are included on the right.

Parallel test: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination

Abbreviations: CI: confidence interval; CrL: credible interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

The pooled difference in sensitivity and pooled difference in specificity between parallel testing versus LC‐aNAAT alone when using a composited reference standard were 16.0% (95% CrI 10.7 to 22.9; 95% PI 2.3 to 60.3) and −3.5% (95% CrI −6.6 to −1.7; 95% PI −47.2 to −0.1), respectively. See Figure 7 and Supplementary material 16.

Subgroup analysis

See Table 8 for the comparative diagnostic accuracy for parallel testing (LC‐aNAAT and LF‐LAM) versus LC‐aNAAT alone by subgroup, based on the microbiological reference standard.

6. Parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone for detection of tuberculosis using a microbiological reference standard, by subgroups.

Participants	Studies (participants)	% Pooled sensitivity (95% CrI)	% Pooled specificity (95% CrI)
Symptomatic
LF‐LAM	23 (8644)	40.0 (32.5 to 48.1)	91.9 (87.9 to 94.7)
LC‐aNAAT	23 (8644)	68.1 (59.7 to 75.9)	96.7 (95.6 to 97.7)
Parallela	23 (8644)	78.5 (73.8 to 82.8)	89.3 (85.1 to 92.6)
Summary difference between parallel and LC‐aNAAT		7.0 (3.8 to 11.7)	−6.8 (−10.1 to −4.4)
Asymptomatic
LF‐LAM	8 (1370)	11.5 (4.1 to 25.2)	97.3 (94.1 to 99.0)
LC‐aNAAT	8 (1370)	47.4 (27.8 to 67.8)	98.9 (97.9 to 99.5)
Parallela	8 (1370)	53.5 (34.8 to 71.7)	96.1 (92.7 to 98.2)
Summary difference between parallel and LC‐aNAAT		4.9 (1.0 to 14.9)	−2.5 (−5.4 to −0.9)
CD4 ≤ 200
LF‐LAM	13 (2798)	49.7 (39.5 to 59.4)	88.1 (81.5 to 92.7)
LC‐aNAAT	13 (2798)	65.8 (54.7 to 76.0)	95.5 (93.2 to 97.2)
Parallela	13 (2798)	78.7 (72.9 to 83.8)	85.6 (78.5 to 90.8)
Summary difference between parallel and LC‐aNAAT		9.2 (4.0 to 17.4)	−9.4 (−14.7 to −5.6)
Inpatients
LF‐LAM	10 (3951)	49.7 (40.2 to 59.9)	89.0 (83.4 to 92.9)
LC‐aNAAT	10 (3951)	63.5 (47.0 to 79.1)	96.9 (95.3 to 98.2)
Parallela	10 (3951)	77.7 (68.5 to 85.9)	87.2 (81.3 to 91.6)
Summary difference between parallel and LC‐aNAAT		11.7 (4.6 to 23.0)	−9.5 (−14.5 to −5.9)
Outpatients
LF‐LAM	11 (5892)	28.8 (21.1 to 38.3)	94.3 (89.5 to 96.9)
LC‐aNAAT	11 (5892)	71.7 (65.0 to 77.9)	97.1 (95.4 to 98.1)
Parallela	11 (5892)	76.0 (69.8 to 81.7)	92.0 (86.2 to 95.4)
Summary difference between parallel and LC‐aNAAT		3.6 (1.9 to 5.8)	−4.4 (−8.2 to −2.3)
Screening positive only
LF‐LAM	6 (1901)	34.4 (22.9 to 48.7)	89.1 (76.0 to 95.5)
LC‐aNAAT	6 (1901)	71.2 (54.1 to 83.9)	95.2 (91.3 to 97.3)
Parallela	6 (1901)	80.8 (67.2 to 90.2)	84.5 (72.0 to 91.9)
Summary difference between parallel and LC‐aNAAT		7.3 (2.1 to 18.0)	−9.1 (−19.8 to −3.8)
Advanced HIV only
LF‐LAM	12 (2945)	44.9 (35.1 to 55.2)	89.8 (82.2 to 94.5)
LC‐aNAAT	12 (2945)	73.6 (64.2 to 82.4)	96.5 (94.4 to 98.0)
Parallela	12 (2945)	81.9 (75.5 to 87.6)	87.1 (78.4 to 92.8)
Summary difference between parallel and LC‐aNAAT		8.1 (3.8 to 14.0)	−8.4 (−15.2 to −4.2)
Seriously ill only
LF‐LAM	6 (577)	47.8 (30.6 to 67.0)	82.4 (71.4 to 90.6)
LC‐aNAAT	6 (577)	67.9 (47.8 to 85.0)	94.0 (89.4 to 97.1)
Parallela	6 (577)	78.7 (62.8 to 90.7)	79.7 (68.7 to 88.5)
Summary difference between parallel and LC‐aNAAT		13.3 (4.4 to 25.4)	−14.4 (−24.2 to −7.4)

Abbreviations: CrI: credible Interval; IPD: individual participant data; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aParallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination

Comparative accuracy stratified by symptomatic versus asymptomatic participants

We identified 23 studies that reported data for participants with symptoms suggestive of tuberculosis; the studies involved 8644 participants, 1557 (18%) with tuberculosis. Most studies defined symptomatic participants based on a positive WHO four‐symptom screen, but some included prolonged cough or being seriously ill. Amongst symptomatic participants, the pooled sensitivity of parallel testing was 78.5% (95% CrI 73.8 to 82.8) and specificity 89.3% (95% CrI 85.1 to 92.6) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 7.0% (95% CrI 3.8 to 11.7; 95% PI 0.5 to 48.2) and −6.8% (95% CrI −10.1 to −4.4; 95% PI −36.6 to −0.9), respectively.

Eight studies reported data for asymptomatic participants, involving 1370 participants, 55 (4%) with tuberculosis. Asymptomatic participants were enroled based on low CD4 cell count, admission to hospital, or initiation of antiretroviral therapy. Amongst asymptomatic participants, the pooled sensitivity of parallel testing was 53.5% (95% CrI 34.8 to 71.7) and the pooled specificity was 96.1% (95% CrI 92.7 to 98.2%) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 4.9% (95% CrI 1.0 to 14.9; 95% PI 0.6 to 24.1) and −2.5% (95% CrI −5.4 to −0.9; 95% PI −16.6 to −0.3), respectively.

There was no evidence that the difference in sensitivity amongst symptomatic participants exceeds that of asymptomatic participants in whom there was an alternative indication for tuberculosis testing (the probability that the between‐subgroup difference was more than zero was 69%). However, regarding the difference in specificity, there was evidence of between‐subgroup difference (the probability that the between‐subgroup difference was less than zero was 99%).

Comparative accuracy stratified by CD4 cell count

We identified 13 studies that reported data for participants with CD4 ≤ 200 cells per µL involving 2798 participants, 484 (17%) with tuberculosis. Amongst participants with CD4 ≤ 200 cells, the pooled sensitivity of parallel testing was 78.7% (95% CrI 72.9 to 83.8) and the specificity was 85.6% (95%CrI 78.5 to 90.8) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 9.2% (95% CrI 4.0 to 17.4; 95% PI 0.6 to 57.4) and −9.4% (95% CrI −14.7 to −5.6; 95% PI −40.0 to −1.6), respectively.

Two studies reported data for participants with CD4 greater than 200; it involved 148 participants, 11 (7%) with tuberculosis. This low number of studies precluded the meta‐analysis and investigation of heterogeneity we had planned in our protocol.

Comparative accuracy stratified by inpatient versus outpatient

We identified 10 studies that reported data for inpatients; the studies involved 3951 participants, 641 (16%) with tuberculosis. Amongst inpatients, the pooled sensitivity of parallel testing was 77.7% (95% CrI 68.5 to 85.9) and specificity was 87.2% (95% CrI 81.3 to 91.6) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 11.7% (95% CrI 4.6 to 23.0; 95% PI 0.9 to 61.5) and −9.5% (95% CrI −14.5 to −5.9; 95% PI −32.3 to −2.2), respectively.

Eleven studies reported data for outpatients; the studies involved 5892 participants, 752 (13%) with tuberculosis. Amongst outpatients, the pooled sensitivity of parallel testing was 76.0% (95% CrI 69.8 to 81.7) and specificity 92.0% (95% CrI 86.2 to 95.4%) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 3.6% (95% CrI 1.9 to 5.8; 95% PI 0.8 to 13.0) and −4.4% (95% CrI −8.2 to −2.3; 95% PI −29.3 to −0.5), respectively.

There was evidence that the difference in sensitivity amongst inpatients exceeds that of outpatients (the probability that the between‐subgroup difference was larger than zero was 98%). Regarding difference in specificity, there was also evidence of between‐subgroup difference (the probability that the between‐subgroup difference was less than zero was 97%).

Comparative accuracy in participants screening positive for tuberculosis

Six studies reported data for participants screening positive for tuberculosis; the studies involved 1901 participants, 346 with tuberculosis (18%). Participants were reported screening positive by chest X‐ray in four studies (Åhsberg 2023; Huerga 2023; Olbrich 2023; Van Hoving 2019), by CRP in one study (Reeve 2023), and by chest X‐ray or CRP in one study (Shapiro 2022). Amongst participants screening positive, the pooled sensitivity of parallel testing was 80.8% (95% CrI 67.2 to 90.2) and pooled specificity 84.5% (95% CrI 72.0 to 91.9) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 7.3% (95% CrI 2.1 to 18.0; 95% PI 0.5 to 46.9) and −9.1% (95% CrI −19.8 to −3.8; 95% PI −50.6 to −0.9), respectively. We did not have data for participants screening negative for tuberculosis.

Comparative accuracy in participants with advanced HIV

Twelve studies reported data for participants with advanced HIV (Åhsberg 2023; Huerga 2019; Huerga 2023; Kebede 2021; Liu 2022; Olbrich 2023; Reeve 2023; Shapiro 2022; Spooner 2022; Szekely 2022; Van Hoving 2019; Yoon 2019); the studies involved 2945 participants, 525 with tuberculosis (18%). Amongst participants with advanced HIV, the pooled sensitivity of parallel testing was 81.9% (95% CrI 75.5 to 87.6) and pooled specificity was 87.1% (95% CrI 78.4 to 92.8) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT was 8.1% (95% CrI 3.8 to 14.0; 95% PI 1.1 to 36.3) and −8.4% (95% CrI −15.2 to −4.2; 95% PI −49.4 to −0.8), respectively. We did not have data for participants without advanced HIV.

Comparative accuracy for seriously ill participants

Six studies reported data for seriously ill participants (Åhsberg 2023; Huerga 2023; Kebede 2021; Olbrich 2023; Shapiro 2022; Van Hoving 2019); the studies involved 577 participants, 166 with tuberculosis (29%). Amongst seriously ill participants, the pooled sensitivity of parallel testing was 78.7% (95% CrI 62.8 to 90.7) and pooled specificity was 79.7% (95% CrI (68.7 to 88.5) (Table 8). The pooled difference in sensitivity and pooled difference in specificity between parallel testing and LC‐aNAAT were 13.3% (95% CrI 4.4 to 25.4; 95% PI 2.0 to 42.3) and −14.4% (95% CrI −24.2 to −7.4; 95% PI −44.2 to −3.0), respectively. We did not have data for participants who were not seriously ill.

Comparative accuracy by type of LC‐aNAAT

There were data only for Xpert MTB/RIF and Xpert MTB/RIF Ultra as LC‐aNAAT. Investigation of heterogeneity comparing accuracy for these two types of LC‐aNAAT was restricted to within‐study comparisons for two studies (Gils 2023; Reeve 2023), see Table 9.

7. Within‐study comparison of Xpert MTB/RIF versus Xpert Ultra.

	Absolute difference in sensitivity			Absolute difference in specificity
	Point	Lower 95% CI	Upper 95% CI	Point	Lower 95% CI	Upper 95% CI
Gils 2023
Parallela (Xpert Ultra and LF‐LAM) vs Xpert Ultra alone	6.8%	2.7%	16.2%	−9.1%	−11.8%	−7.0%
Parallela (Xpert MTB/RIF and LF‐LAM) vs Xpert MTB/RIF alone	6.8%	2.7%	16.2%	−9.1%	−11.8%	−7.0%
P for interaction	0.94	No evidence of a difference between Xpert Ultra and Xpert MTB/RIF		0.72	No evidence of a difference between Xpert Ultra and Xpert MTB/RIF
Reeve 2023
Parallela (Xpert Ultra and LF‐LAM) vs Xpert Ultra alone	0.8%	0.1%	4.5%	−1.1%	−2.1%	−0.6%
Parallela (Xpert MTB/RIF and LF‐LAM) vs Xpert MTB/RIF alone	0%	0%	3.5%	−1.1%	−2.1%	−0.5%
P for interaction	0.92	No evidence of a difference between Xpert Ultra and Xpert MTB/RIF		0.41	No evidence of a difference between Xpert Ultra and Xpert MTB/RIF

Abbreviations: CI: confidence interval; LC‐aNAAT: low‐complexity automated nucleic acid amplification test; LF‐LAM: lateral flow urine lipoarabinomannan assay; P: probability; vs: versus

^aParallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination

Nine studies reported data for Xpert Ultra as LC‐aNAAT; the study involved 5752 participants, 658 (11%) with tuberculosis. The pooled sensitivity of parallel testing was 77.1% (95% CrI (66.4 to 85.8) and pooled specificity was 88.5% (95% CrI 82.0 to 93.1) (Table 10). The pooled difference in sensitivity and pooled difference in specificity between parallel use of respiratory Xpert Ultra and urine LF‐LAM versus respiratory Xpert Ultra alone were 7.5% (95% CrI 3.0 to 15.2; 95% PI 0.7 to 44.2) and −7.4% (95% CrI −13.1 to −3.8; 95% PI −36.6 to −1.1), respectively.

8. Parallel use of Xpert Ultra and LF‐LAM versus Xpert Ultra alone for detection of tuberculosis using a microbiological reference standard.

	Studies (participants)	% Pooled sensitivity (95% CrI)	% Pooled specificity (95% CrI)
Xpert Ultra only
LF‐LAM	9 (5752)	34.4 (22.9 to 48.7)	89.1 (76.0 to 95.5)
Xpert Ultra	9 (5752)	67.8 (51.3 to 81.1)	96.8 (95.2 to 98.0)
Parallela	9 (5752)	77.1 (66.4 to 85.8)	88.5 (82.0 to 93.1)
Summary difference between parallel and LC‐aNAAT		7.5 (3.0 to 15.2)	‐7.4 (‐13.1 to ‐3.8)

Abbreviations: CrI: credible interval; LF‐LAM: lateral flow urine lipoarabinomannan assay

^aParallel: LC‐aNAAT on respiratory samples and LF‐LAM on urine conducted simultaneously on the same person, and a positive result on either test implies a positive result for the combination

Sensitivity analysis

See Table 6 for sensitivity analyses based on the microbiological reference standard.

When excluding studies with high risk of bias (QUADAS‐C) and high concern regarding applicability (QUADAS‐2), the pooled difference in sensitivity increased to 13.8% (95% CrI 7.6 to 22.9) and the pooled difference in specificity was −12.2 (95% CrI −19.5 to −7.3) (6 studies, 3110 participants).

When we excluded 1966 participants (from 27 studies) with at least one unavailable or nondeterminate index test result, the pooled difference in sensitivity decreased to 4.4% (95% CrI 2.8 to 6.5), while the pooled difference in specificity was −6.9% (95% CrI −9.1 to −4.5) (27 studies, 10,685 participants).

In sensitivity analysis with alternative prior distributions, the 95% CrI did not change the results.

See Table 7 for our sensitivity analyses based on the composite reference standard.

Unavailable and nondeterminate test results

Across 27 studies, 1966 of 12,651 (15.5%) participants had at least one unavailable or nondeterminate index test result.

The proportion of participants with unavailable respiratory LC‐aNAAT results ranged from 2% in Shapiro 2022 to 63% in Stead 2023, while the proportion of participants with unavailable urine LF‐LAM results varied from 0 % in Shapiro 2022 to 13% in Spooner 2022. See Supplementary material 2 for a description of unavailable index test results by study in the flow and timing domain.

For Xpert MTB/RIF, the median proportion of nondeterminate index test results was 27/1210 (2.2%), reported by four studies (Åhsberg 2023; Boloko 2022; Calligaro 2017; Lawn 2017). For Xpert Ultra, the median proportion of nondeterminate index test results was 9/1452 (0.6%), reported by three studies (Åhsberg 2023; Huerga 2023; Olbrich 2023). For LF‐LAM, the median proportion of nondeterminate index test results was 19/5805 (0.3%) reported by eight studies (Bjerrum 2015; Boloko 2022; Broger (Cohort4) 2020; Huerga 2023; Lawn 2017; Olbrich 2023; Peter 2015; Van Hoving 2019).

Comparative diagnostic accuracy of parallel testing versus LF‐LAM alone

Results for the comparative accuracy of parallel testing versus LF‐LAM alone are shown in Supplementary material 14; in Supplementary material 15 for the microbiological reference standard; and in Supplementary material 17; Supplementary material 18 for the composite reference standard.

Discussion

This Cochrane review summarises the evidence on the diagnostic accuracy of the parallel use of LC‐aNAAT on respiratory samples and LF‐LAM test on urine compared to LC‐aNAAT on respiratory samples, for the detection of tuberculosis disease in adults and adolescents with HIV who present with presumptive tuberculosis. The review is based on published studies and preprints available up to 3 November 2023, unpublished studies identified through a WHO call for data on 17 January 2024, and unpublished data provided by investigators.

Summary of main results

We included 31 fully paired comparative accuracy studies: 27 studies based on a microbiological reference standard (involving 12,651 participants, 2368 (19%) with tuberculosis) and 23 studies based on a composite reference standard (involving 11,109 participants, 3723 (34%) with tuberculosis).

The main findings, together with the hypothetical consequences for a cohort of 1000 people where 20% have tuberculosis, can be viewed in Table 1 and Table 2. We also prepared summary of finding tables for low and high tuberculosis prevalence levels (2.5% and 40%, respectively), which can be viewed in Supplementary material 19 and Supplementary material 20.

Against a microbiological reference standard, the use of parallel testing resulted in a pooled sensitivity of 77.5% and pooled specificity of 89.4%. Compared to respiratory LC‐aNAAT alone, using a parallel testing strategy may increase sensitivity by an absolute value of 6.7% (95% credible interval (CrI) 3.8 to 10.7; low‐certainty evidence) and may decrease specificity by 6.8% (95% CrI 9.5 to 4.7; low‐certainty evidence).

Applied to a hypothetical cohort of 1000 adults and adolescents with HIV, where 200 have microbiologically confirmed tuberculosis disease, parallel testing will identify 13 more (95% CrI 8 to 21 more) people with tuberculosis (true positives) who would have been missed with respiratory LC‐aNAAT alone. On the other hand, parallel testing will wrongly diagnose 54 more (95% CrI 37 to 76 more) people without tuberculosis (false positives) compared to respiratory LC‐aNAAT alone.

Against a composite reference standard, parallel testing resulted in a pooled sensitivity of 67.6% and pooled specificity of 96.2%. Compared to respiratory LC‐aNAAT alone, a parallel‐test strategy may increase sensitivity by 16.0% (95% CrI 10.7 to 22.9; low‐certainty evidence) and may decrease specificity by 3.5% (95% CrI 6.6 to 1.7; very low certainty evidence), compared to respiratory LC‐aNAAT alone.

Applied to a hypothetical cohort of 1000 adults and adolescents with HIV where 200 have confirmed tuberculosis by a composite reference standard, parallel testing will identify 32 more (95% CrI 21 to 46 more) people with tuberculosis (true positives) who would have been missed with respiratory LC‐aNAAT alone. Conversely, parallel testing will wrongly diagnose 28 more (95% CrI 13 to 53 more) people without tuberculosis (false positives) than respiratory LC‐aNAAT alone.

Interpretation of findings

The gain in sensitivity with parallel testing reflects the added value of LF‐LAM testing for the diagnosis of tuberculosis disease in adults and adolescents with HIV. The loss in specificity may reflect additional false‐positive results, or additional cases not identified by the reference standard, for example, respiratory culture‐negative tuberculosis disease.

Accuracy of the reference standard

The overall proportion of tuberculosis cases across studies varied from 19% based on the microbiological reference standard to 34% when based on a composite reference standard, highlighting major differences between the two types of reference standards.

The microbiological reference standard could underdiagnose people with tuberculosis, as microbiological confirmation of disease is challenging, particularly in people with HIV. Immunocompromised patients are more likely to have extrapulmonary tuberculosis disease or paucibacillary pulmonary disease, which may exist concurrently with disseminated tuberculosis, where a sputum‐based microbiological reference standard will be less accurate [2]. Across individual studies, the specificity of the parallel testing strategy was lowest when the microbiological reference standard was based on sputum culture alone or with only a few extrapulmonary samples (Åhsberg 2023; Huerga 2019), suggesting limitations in the microbiological reference standard.

The improved specificity of parallel testing when evaluating against a composite reference standard suggests that many of the participants who were negative for tuberculosis by the microbiological reference standard may have been treated for tuberculosis in clinical practice. On the other hand, a composite reference standard including clinical judgement and based on tuberculosis treatment may have wrongly classified participants as having tuberculosis when they did not. Notably, in most studies, clinicians were not blinded to the respiratory LC‐aNAAT or LF‐LAM test results. Therefore, clinical decisions (and categorisation of participants as having tuberculosis) may have been informed by the index test results, leading to incorporation bias and overestimation of index test specificity. These factors may explain the lower estimated reduction in specificity with the parallel testing strategy when using a composite reference standard compared to a microbiological reference standard (−3.5% versus −6.8%, respectively).

There was a higher estimated gain in sensitivity of parallel testing (compared to LC‐aNAAT alone) when using a composite compared to a microbiological reference standard (16.0% versus 6.7%), despite the fact that the sensitivity of parallel testing was estimated to be lower (67.6%) than with a microbiological reference standard (77.5%). This was driven by the relative decline in individual test sensitivity of LC‐aNAAT when measured against a composite reference standard (68.0% versus 46.8%), while the sensitivity of LF‐LAM remained similar for microbiological and composite reference standards. Consequently, the incremental value of parallel testing increases with a composite reference standard.

Quality of the included studies

Comparing parallel testing with respiratory LC‐aNAAT alone, the increase in sensitivity was highest in studies with low risk of bias in the patient selection domain (QUADAS‐C), namely those that included participants regardless of their ability to produce sputum samples (e.g. Gupta‐Wright 2018a; Lawn 2017; Stead 2023), see Figure 6. These studies had lower estimates of sensitivity for respiratory LC‐aNAAT, with associated higher incremental increases in sensitivity with the addition of LF‐LAM in the parallel testing strategy. Conversely, studies with high risk of selection bias (excluding participants unable to produce sputum or urine) reported limited increase in sensitivity with the addition of LF‐LAM as part of a parallel testing strategy, compared to respiratory LC‐aNAAT alone (e.g. Bjerrum 2015; Peter 2015; Shapiro 2022), see Figure 6. Participants with sputum scarcity may represent a highly selected population, potentially more likely to have disseminated or extrapulmonary diseases and to be more severely ill, with limited ability to produce good‐quality respiratory samples. To the extent that the inability to produce samples is correlated with the severity of tuberculosis disease, this approach to participant selection could have reduced the gain in sensitivity achieved with parallel testing.

In order to minimise the effect of selection bias, our main analyses included all participants enroled in the studies, regardless of available or valid LC‐aNAAT results (the results we did not know were imputed as negative). In sensitivity analysis, we excluded participants with at least one unavailable or nondeterminate component test, and, as expected, the sensitivity of respiratory LC‐aNAAT increased as the denominator decreased, while the sensitivity of LF‐LAM did not change. Consequently, the estimated difference in sensitivity between parallel and respiratory LC‐aNAAT was reduced, while the difference in specificity remained similar.

We judged a reference standard to have a low risk of bias if it included multiple specimens from different sites (and testing with mycobacterial culture, or LC‐aNAAT for non‐respiratory samples). We considered the potential for misclassification of individuals with tuberculosis when sputum culture was not performed, or when extrapulmonary samples were not evaluated. In these instances, misclassification bias could lead to lower estimates of index test specificity if the index test can correctly detect people with tuberculosis who were missed by the reference standard.

In sensitivity analysis, we excluded the studies at high risk of bias (QUADAS‐C) and for which we had high concerns regarding applicability (QUADAS‐2). We expected and found an increase in the sensitivity gain (from 6.7% to 13.8%). However, the difference in specificity also increased (from ‐6.8% to ‐12.2%); this was mainly driven by one study that included severely sick participants and used a limited number of extrapulmonary samples for the reference standard (Åhsberg 2023). For the composite reference standard, almost all studies had a high risk of bias due to incorporation bias.

Explanation of heterogeneity

We found considerable heterogeneity across studies, with prediction intervals for both sensitivity and specificity being wider than the CrI around the pooled estimates. We hypothesised this heterogeneity might be attributable to the variability in accuracy of the individual index tests, risk of bias, and study design (i.e. methodological diversity), as well as differences in setting and participant characteristics (i.e. clinical diversity).

When investigating heterogeneity, we observed a higher increase in sensitivity with a parallel testing strategy for inpatients compared to outpatients (11.7% versus 3.6%). The decrease in specificity for a parallel testing strategy, compared to respiratory LC‐aNAAT alone, was largest for the following subgroups: inpatients; seriously ill patients; patients with advanced HIV, including patients with CD4 count less than 200 cells, where microbiological confirmation is known to be challenging [2]. The loss in specificity for parallel testing compared to LC‐aNAAT also differed between analysis of asymptomatic participants (−2.5%) compared to symptomatic participants (−6.8%). We observed that the individual test sensitivity of LF‐LAM generally increased when the sensitivity of LC‐aNAAT decreased, which is an indication of how the two tests may complement each other, while the sensitivity of parallel testing remained relatively consistent across subgroups. Nonetheless, the pooled estimates of the incremental change with parallel testing compared to LC‐aNAAT alone did not differ substantially across most subgroups.

The manufacturers (Cepheid Inc, Sunnyvale, USA), of the Xpert assays have replaced Xpert MTB/RIF in most settings with Xpert Ultra, a newer generation LC‐aNAAT that is anticipated to have increased sensitivity. In total, nine studies in this review evaluated Xpert Ultra, and 21 evaluated Xpert MTB/RIF. In two within‐study subgroup analyses, the type of LC‐aNAAT (Xpert Ultra or Xpert MTB/RIF) made no difference to the results (Table 9). We may not have seen the expected increase in Xpert Ultra sensitivity compared to Xpert MTB due to the target condition being all forms of tuberculosis and to the inclusion of participants unable to produce sputum (i.e. imputed as a negative result for LC‐aNAAT).

Implementation considerations

The additional patients with tuberculosis identified by parallel testing using LC‐aNAAT and LF‐LAM in combination will likely receive treatment they may otherwise not have received. The benefit conferred by the increased sensitivity of a parallel‐test strategy including LF‐LAM as a point of care test will likely also be earlier administration of treatment. Reasons for this observed increase in diagnostic sensitivity with the addition of urine LF‐LAM testing could be the test's ease of use, application on non‐sputum samples (for people unable to produce sputum), and ability to identify extrapulmonary or disseminated forms of tuberculosis as a non‐site‐specific test, despite relatively modest individual test sensitivity. A pertinent question is whether this will lead to an improvement in patient‐important outcomes. In randomised trials, the use of LF‐LAM testing as part of a tuberculosis diagnostic strategy has already been demonstrated to decrease mortality in hospitalised adults with HIV, compared to diagnostic strategies that do not include LF‐LAM testing [93]. However, parallel testing will also lead to overtreatment of people who do not have tuberculosis, with the possible harm of missing other conditions falsely diagnosed as tuberculosis and the risk of adverse clinical events due to unnecessary tuberculosis treatment [94]. In a high‐prevalence setting, the risk of overtreatment is likely to be acceptable, given the detection of an important number of people with tuberculosis who may otherwise be missed. However, in low‐prevalence settings, the number of additional false positives may be relatively high (Table 1; Table 2; Supplementary material 19; Supplementary material 20). The need and use of the reference scale card to interpret LF‐LAM results may further have implications for adoption and implementation of the tests in clinical practice [64, 95].

As the focus of this review is diagnostic accuracy, we did not assess other REASSURED (real‐time connectivity, ease of specimen collection, affordable, user‐friendly, rapid/robust, equipment‐free, deliverable to end‐users) criteria [96]. Nor did we assess the time to sample collection and time to test results. Nonetheless, these factors play an important role and should be considered in policy development when selecting diagnostic tests and testing strategies that aim to achieve high diagnostic yield and broad population coverage, ultimately improving patient‐important outcomes [97, 98]. Other patient‐important outcomes including time to diagnosis and treatment initiation, proportion diagnosed and treated, and mortality [93, 95] are also important considerations for policy decisions.

Differences from previous reviews

Cochrane reviews on diagnostic tests for tuberculosis have typically assessed the tests individually or compared two or more tests when a new version of a test emerged [15, 26]. Only one Cochrane review, by Shah and colleagues, has assessed the accuracy of a parallel combination of tests [19]. Our findings are consistent with this review by Shah and colleagues [19], which found the parallel combination of LF‐LAM and sputum Xpert MTB/RIF (diagnosis affirmed by either test being positive) to be 75% sensitive (77.5% in our review), with a difference of 13% (6.7% in our review) compared to Xpert MTB/RIF alone (three studies; 909 participants, 36% with tuberculosis). Pooled specificity of parallel testing, according to Shah and colleagues [19], was 93% (89.4% in our review), a difference of 4% (6.8%) compared to Xpert MTB/RIF alone (three studies; 909 participants, 36% with tuberculosis). Furthermore, our estimates of LF‐LAM accuracy were consistent with those of a previous Cochrane review in which the sensitivity and specificity of LF‐LAM were estimated to be 42% and 91%, (8 studies; 3449 participants, 37% with tuberculosis) [17].

Our individual test accuracy results for LC‐aNAAT, LF‐LAM and parallel testing mirror the findings of the individual participant data review by Broger and colleagues [4], which found a sensitivity of 41% (95% Crl 15 to 66) for LF‐LAM, 61% (95% Crl 25 to 88) for sputum Xpert, and 75% for LF‐LAM and Xpert in combination. We took a different approach to further compare the diagnostic accuracy of a parallel‐test strategy versus one test alone, and we provided individual test accuracy estimates including both sensitivity and specificity.

Strengths and weaknesses of the review

The findings in this review are based on a comprehensive search and strict inclusion criteria, as well as standardised duplicate study inclusion, data extraction, and assessment of methodological quality using the QUADAS‐2 and QUADAS‐C tools. An important strength of our review is the inclusion of a large set of paired comparative accuracy studies on parallel testing versus respiratory LC‐aNAAT. For the microbiological reference standard, we carefully harmonised the case definition across studies to ensure the index tests were not included. We also used a composite reference standard that included clinically diagnosed tuberculosis to explore its influence on sensitivity, specificity, and comparative accuracy. We communicated extensively with authors of primary studies to obtain unpublished data, which were reanalysed in our review, and to clarify information that we then used in the methodological quality assessment.

Our review also has some limitations. We could not assess the risk of publication bias using funnel plots. However, with our extensive outreach to researchers and a WHO public call for data, we believe the risk of publication bias is low. Our literature search was performed on 3 November 2023, and although newer studies might have emerged since the search, the inclusion of abstracts, unpublished data, and responses from the WHO call likely mitigate this concern. We reanalysed data provided by the study authors; only two studies provided published data that we extracted directly. Our re‐analysis of data could be considered a strength of the review, as we were able to include more data, but it is possible that these studies represent lower quality data and these quality issues might not have been fully captured by QUADAS‐2.

Our meta‐analytical model addresses the dependence between the two index tests being compared, but it does not take into account the paired nature of data from paired accuracy studies. We are also uncertain whether our strategy of treating the unavailable results as 'not positive' and imputing them as negative was the most appropriate strategy.

For some investigations of heterogeneity, we only had data for one of the subgroups (e.g. participants with CD4 ≤ 200 cells per µL) and did not have sufficient data for meta‐analysis for the opposing subgroup (e.g. participants with CD4 > 200 cells). Therefore, some sources of heterogeneity remain unexplored.

Applicability of findings to the review question

Overall, we had low concern about the applicability of the included studies to our review question, as assessed by QUADAS‐2. However, only one study included young adolescents aged 10 to 15 years (Olbrich 2023). Several studies included participants aged 15 or 16 years and above (Huerga 2017; Huerga 2019; Huerga 2023; Kasaro 2020; Kebede 2021; Shapiro 2022), although the median age suggested few adolescents were included (Supplementary material 2). This may reduce generalisability to adolescents with HIV, who are at high risk of tuberculosis but are often neglected [99, 100]. While we included all commercially available LC‐aNAATs in our search, we identified only studies evaluating Xpert MTB/RIF and Ultra in combination with LF‐LAM. This may limit the applicability of our results to settings where Xpert MTB/RIF and Ultra are available. The applicability of studies using frozen samples for LF‐LAM testing is unclear as no study has made a direct comparison of the performance of fresh versus frozen urine samples.

Authors' conclusions

Implications for practice

Parallel testing of respiratory LC‐aNAAT (low‐complexity automated nucleic acid amplification test) and urinary LF‐LAM (lateral flow urine lipoarabinomannan assay) to diagnose tuberculosis improves sensitivity compared to LC‐aNAAT on respiratory samples alone, but at the cost of reduced specificity. The magnitude of gain in sensitivity should be weighed against the loss of specificity, taking into account different clinical scenarios with varying prevalence of tuberculosis disease. In low‐prevalence settings, using the tests in parallel may lead to a large increase in false‐positive results and overtreatment. In high‐prevalence settings that have limited patient follow‐up and limited extrapulmonary tuberculosis investigation, the benefit of identifying additional patients with tuberculosis at the point‐of‐care likely outweighs the relatively lower risk of overtreatment of people without tuberculosis. Other consequences of false‐positive results include delayed alternative diagnosis, likelihood of anxiety and morbidity caused by additional testing, unnecessary treatment, and adverse events related to unnecessary assessment and treatment; and stigma associated with a diagnosis of tuberculosis.

The use of LF‐LAM in addition to LC‐aNAAT can facilitate rapid diagnosis of tuberculosis, which is crucial for initiating timely treatment, especially for people with HIV, who are at a higher risk of tuberculosis‐related morbidity and mortality. Moreover, our findings suggest that the parallel‐test strategy may be particularly useful to increase sensitivity in inpatient settings compared to outpatient settings. Clinicians must still consider the risk of false‐negative test results and the consequences of increased morbidity and mortality, delayed treatment initiation, and continued risk of tuberculosis transmission.

Implications for research

In this review, only two studies reported the accuracy of a parallel strategy as fully paired comparative diagnostic accuracy studies. Future studies should report the accuracy of the combination of tests as well as that of the individual tests. Studies in settings outside sub‐Saharan Africa are lacking, as are studies that focus on adolescents aged 10 to 18 years. For a non‐site‐specific test like LF‐LAM, it is important to use a reference standard that includes at least two different specimens (e.g. sputum, blood, or urine). In addition, for presumed extrapulmonary tuberculosis, it is important to use appropriate specimens from the suspected sites of involvement. When evaluating a parallel‐test strategy with an assay that is not site‐specific, future studies may consider systematic sampling of multiple sites to improve the quality of the reference standard. Moreover, future studies should include patients unable to produce sputum or urine in the analysis of parallel testing and report the proportion of unavailable samples. Further research on the effective implementation of parallel testing within routine clinical practice is needed. This includes accounting for turnaround time, test robustness, failures, user‐friendliness, impact on drug susceptibility testing (e.g. in cases of positive LAM, unavailable sputum), and adoption.

Supporting Information

Supplementary materials are available with the online version of this article: 10.1002/14651858.CD016070.

Supplementary materials are published alongside the article and contain additional data and information that support or enhance the article. Supplementary materials may not be subject to the same editorial scrutiny as the content of the article and Cochrane has not copyedited, typeset or proofread these materials. The material in these sections has been supplied by the author(s) for publication under a Licence for Publication and the author(s) are solely responsible for the material. Cochrane accordingly gives no representations or warranties of any kind in relation to, and accepts no liability for any reliance on or use of, such material.

Supplementary material 1 Search strategies

Supplementary material 2 Characteristics of included studies

Supplementary material 3 Characteristics of excluded studies

Supplementary material 4 Characteristics of ongoing studies

Supplementary material 5 Data package

Supplementary material 6 Data extraction form

Supplementary material 7 Data reporting form for primary study authors

Supplementary material 8 QUADAS‐2 review specific tailoring and guidance

Supplementary material 9 QUADAS‐C review specific tailoring and guidance

Supplementary material 10 Data analysis plan

Supplementary material 11 Analyses reported in the Data Analysis Plan that were not performed

Supplementary material 12 Summary characteristics of included studies

Supplementary material 13 SROC plot for parallel use and LC‐aNAAT alone using a microbiological reference standard

Supplementary material 14 Study‐specific estimates of difference in sensitivity and specificity between parallel testing and LF‐LAM alone using a microbiological reference standard

Supplementary material 15 SROC plot for parallel use and LF‐LAM alone using a microbiological reference standard

Supplementary material 16 SROC plot for parallel use and LC‐aNAAT alone using a composite reference standard

Supplementary material 17 Study‐specific estimates of difference in sensitivity and specificity between parallel testing versus LF‐LAM alone using a composite reference standard

Supplementary material 18 SROC plot for parallel use and LF‐LAM alone using a composite reference standard

Supplementary material 19 Summary of finding tables for low and high tuberculosis prevalence levels: parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone using a microbiological reference standard

Supplementary material 20 Summary of finding tables for low and high tuberculosis prevalence levels: parallel use of LC‐aNAAT and LF‐LAM versus LC‐aNAAT alone using a composite reference standard

Supplementary material 21 Errata and retraction

New

Additional information

Contributions of authors

All authors contributed significantly to this review.

S Bjerrum was the corresponding author.

SB co‐ordinated contributions from the co‐authors and wrote the final draft of the review.

SB, JÅ, BY, and MS screened papers against eligibility criteria.

BY conducted the literature searches with Vittoria Lutje, Cochrane Infectious Diseases Group Information Specialist.

SB, JÅ, BY, and MS appraised the quality of papers.

SB, JÅ, BY, MS, BS, HP, AR, AK, and LO extracted data for the review and sought additional information about papers.

TB, IO, and CMD reanalysed previously shared individual participant data for this review, with permission from study authors.

BY, ND, IS, and MWD performed analyses.

MWD, SB, and JÅ provided figures and tables for the review.

SB, BY, JÅ, MWD, LO, RN, TB, IO, AK, CD, IS, DJ, AL, and MS worked on the methods sections and commented on the draft review.

BY, AL, and MS provided expert content for the review.

All authors checked and approved the final version of the review.

Declarations of interest

Development of the Cochrane review was made possible, in part, with financial support from the United States Agency for International Development (USAID) administered by the World Health Organization (WHO) Global TB Programme, Switzerland. SB, BY, JÅ, RN, LO, DJ, and MS received funding from WHO as consultants to the review. AL served as consultant to the WHO without funding..

SB, TB, JÅ, LO, DJ, and MS have been involved in primary studies that could have been included in this review. These authors were not involved in the assessment of study eligibility, data extraction, assessment of methodological quality, or GRADE assessments for their own studies, in accordance with Cochrane policy [54].

RN is the Chair of the TB Proof board and an independent consultant to the WHO. The Division of Infectious Diseases and Tropical Medicine of the Ludwig‐Maximilians University Munich, where LO is based, received Xpert MTB/RIF Ultra cartridges for free for research use from Cepheid.

TB reports inventorship on patent applications in the field of tuberculosis detection but no personal ownership and is a shareholder of Avelo Ltd.

All review authors have no additional financial involvement with any organisation or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the review, apart from those disclosed, and declare no additional relevant interests.

Sources of support

Internal sources

• Liverpool School of Tropical Medicine, UK

  Cochrane Infections Disease Group

External sources

• Foreign, Commonwealth and Development Office (FCDO), UK

  Project number 300342‐104

• World Health Organization, TB Prevention, Diagnosis, Treatment, Care & Innovation (PCI), Global TB Programme, WHO APW reg. #. 2023 ‐ 1425004, Switzerland

  Consultancy funding to lead this review. SB, BY, JÅ, RN, LO, DJ, and MS received funding from the WHO as consultants to the review.

Registration and protocol

Protocol (2024) doi.org/10.1002/14651858.CD016070 [101]. This protocol is aligned with a protocol in children registered as 'Parallel use of low‐complexity automated nucleic acid amplification tests and lateral flow urine lipoarabinomannan assays to detect tuberculosis disease in children' [46]. Both reviews informed the WHO guideline development group meeting 'Nucleic acid amplification tests to detect TB and DR‐TB' from 6 to 10 May 2024.

Data, code and other materials

As part of the published Cochrane review, the review authors have made the following available for download for users of the Cochrane Library: full search strategies for each database (Supplementary material 1); full citations of each unique report for all studies included (Supplementary material 2) or excluded at the full‐text screen (Supplementary material 3), in the final review; consensus risk of bias assessments using QUADAS‐2 (Supplementary material 8) and QUADAS‐C criteria Supplementary material 9; data extraction and reporting forms (Supplementary material 6; Supplementary material 7) and analysis plan (Supplementary material 10). Review authors conducted analysis and data management with Excel and R.

The data package does not contain the data (Supplementary material 5); rather, the summary data is available at open science framework (OSF): https://osf.io/6g84x/files/osfstorage.

The R code is available at OSF: https://osf.io/pwv6j.

Appropriate permissions have been obtained for such use.

History

Protocol first published: Issue 5, 2024