Systematic review, meta-analysis and economic modelling of molecular diagnostic tests for antibiotic resistance in tuberculosis

Conventional culture methods

Conventional culture and drug susceptibility testing (DST) is a slow process, requiring isolation of mycobacteria from clinical specimens, identification of MTBC, and testing of the susceptibility pattern of strains in the presence of anti-TB drugs. Depending on the methodology, culture-based methods can take several months to be completed; however, they remain the ‘gold standard’ for M. tuberculosis DST.

A TB strain is classified as resistant to a drug if ≥ 1% of the bacterial population is able to grow at a ‘critical’ concentration of drug, which inhibits 95% of wild-type strains. 19 DST using agar- or egg-based Löwenstein–Jensen (LJ) solid media is the most commonly utilised method. Standard methods using LJ medium include the absolute concentration method, the resistance ratio method and the proportion method. 19 In the absolute concentration method, a standardised bacterial suspension is inoculated on to drug-free media and media containing graded concentrations of the drug to be tested. The minimum inhibitory concentration (MIC) is the lowest concentration of the drug that inhibits growth. In the resistance ratio method, the growth of the test strain is compared with that of a control susceptible strain. Resistance is expressed as the ratio of the MIC of the test strain to the MIC of the control strain. The proportion method determines the percentage of growth (number of colonies) of a defined inoculum on a drug-free control medium compared with growth on culture media containing the critical concentration (or range of concentrations) of the test drug.

Automated commercial liquid culture systems use the proportion method and are designed to detect growth inhibition in drug-containing media as early as possible. One of the first systems was the BACTEC 460 TB system (Becton Dickinson Diagnostic Instrument Systems, Towson, MD, USA), which uses a radiometric method to measure carbon dioxide production. Subsequently, non-radiometric methods, such as the BACTEC MGIT (Mycobacteria Growth Indicator Tube) 960 system (Becton Dickinson Diagnostic Instrument Systems) have become widely adopted. 20 The BACTEC MGIT 960 system uses fluorimetric detection to monitor oxygen consumption resulting from bacterial growth. Growth in media containing the drug of interest is compared with growth in a drug-free control tube, which is inoculated with a 1 : 100 dilution of the M. tuberculosis strain. Every 60 minutes, the BACTEC MGIT 960 measures the fluorescence emitted from the tube. Once the fluorescence indicates that there are 105–106 colony-forming units (CFUs) per millilitre of medium, the tube is flagged as positive. If there is at least equal growth in a drug-containing tube and the control tube, this indicates that at least 1% of the bacteria present in the sample are resistant to that drug and the sample is classified as resistant. If the tube does not become positive within 6 weeks, the machine marks it as negative. The BACTEC MGIT 960 can reduce the mean time taken to detection of M. tuberculosis to 14.4 days compared with 24.1 days on solid medium. 21

Non-commercial culture systems based on the growth of microcolonies can also facilitate more rapid diagnosis than conventional methods. 22 Microscopic observation of drug susceptibility (MODS) is a liquid culture-based method in which early growth in drug-free and drug-containing media is detected by microscopic examination. 23 Similarly, thin-layer agar (TLA) methods rely on the microscopic detection of growth on solid media. 24

Rapid molecular tests

A significant reduction in the time for diagnosis of clinically significant drug resistance depends on genotypic methods that identify mutations in genes responsible for resistance. Specific nucleotide sequences in processed specimens (crude extracts or treated sputum) can be detected within a few hours, so the total time for the TB detection can be reduced to < 1 day. Improved biosafety is another advantage of molecular assays, as they require high containment only initially. However, molecular methods lack the sensitivity of phenotypic methods, as the genetic basis of resistance is not completely understood for any anti-TB drug. For example, although 95% of rifampicin-resistant M. tuberculosis strains have mutations in the 81 base pair (bp) rifampicin resistance determining region (RRDR) of the gene encoding the RNA polymerase-subunit, rpoB, 5% of strains have no known resistance mutations. 25 Up to 10–25% of low-level isoniazid-resistant M. tuberculosis strains do not have mutations in either katG or the inhA promoter. 26

Molecular tools for the detection of M. tuberculosis and determination of drug resistance that have been developed over the past decade are largely nucleic acid amplification tests (NAATs) to increase sensitivity, combined with highly specific detection systems. The polymerase chain reaction (PCR) is the most common methodology utilised in NAAT; alternatives include real-time PCR, isothermal strand displacement amplification or transcription-mediated amplification, and the ligase chain reaction. 27–30 Although NAAT can theoretically detect a single copy of nucleic acid in a specimen, sensitivity can be significantly compromised by the presence of PCR inhibitors in clinical specimens and loss of nucleic acids during specimen processing, and therefore tends to vary.

Future directions

Whole genome sequencing (WGS) offers a powerful new approach for the diagnosis of M. tuberculosis drug resistance, promising rapid, unambiguous determination of all existing, clinically significant mutations. In the authors’ opinion, the continuing cost reductions of this technology may eventually neutralise arguments over the value of targeted sequencing compared with WGS. Recent studies using WGS to research novel determinants of resistance have revealed that drug resistance may be more multifactorial than previously appreciated which, in some cases, may explain discordance between phenotypes and genotypes. 42–44 As more resistance loci are identified, and the phenotypic effects of multiple mutations and strain background are elucidated, the public health value of routine WGS for diagnosis of drug resistance will increase. In addition, WGS reveals the genetic relatedness of strains, allowing transmission networks to be traced with unparalleled resolution and obviates the need for separate molecular fingerprinting analysis.

Primary objectives

As outlined in Chapter 2, the primary purpose of this review was to produce a critical assessment of studies that report data on the diagnostic accuracy of molecular tests used to detect drug resistance in M. tuberculosis.

Types of studies

All diagnostic studies that compared a molecular (index) test with a gold standard (reference) test were considered in this review. No restrictions on study setting were applied, and studies from all countries were eligible for inclusion. When clearly stated in the paper, the study design was recorded for subgroup analysis.

Types of participants and clinical specimens

Studies of adults or children with suspected MDR-TB or XDR-TB were considered to be eligible for inclusion. This review also considered those studies for which the participant’s details were not clearly stated, and work was carried out on anonymised samples. When categorising the participant population, individuals aged ≥ 16 years were classed as adult participants.

When cited, studies that reported comorbidity with HIV infection or a previous history of TB within the participant population were considered for inclusion. These data were also recorded during the data extraction process for subgroup analysis.

Only studies that were carried out on clinical samples, including pulmonary or extrapulmonary samples, were considered for inclusion. Those studies that used clinical specimens ‘spiked’ with mycobacteria or cultured isolates were excluded from the review. Studies where the test was carried out on both cultured and clinical samples were considered for inclusion only if the diagnostic accuracy data were reported by the type of sample tested. Studies that pooled data from clinical and cultured specimens were excluded.

Index tests

Studies that used a genetic rapid diagnostic method to test for either first- or second-line drug susceptibility were considered for inclusion. Although studies using commercial assays are more commonly present in the literature, data on in-house tests were also considered for inclusion when there was sufficient information to assess their diagnostic accuracy in detecting MDR-TB or XDR-TB.

Reference tests

Studies that compared an appropriate index test against an appropriate reference standard test were potentially eligible for inclusion. This review accepted any TB phenotypic drug susceptibility test that has been endorsed by the WHO. These include culture on solid or liquid media and use of rapid liquid culture systems. If a study compared a rapid culture test against a NAAT with the ability to detect mutations associated with drug susceptibility, it was included in the review.

Search strategy

A standardised search strategy (PROSPERO registration CRD42011001537, see Appendix 1), was designed to generate a comprehensive list of relevant studies from five electronic literature databases: EMBASE, PUBMED, MEDLINE, Bioscience Information Service (BIOSIS) and Web of Science. The strategy design was based upon a previously successful model used by the European Centre for Disease Prevention and Control (ECDC). It was further validated by comparing the citation output against the bibliography of two published diagnostic reviews of rapid diagnostic tests for TB and drug susceptibility. 45,46

The electronic databases were initially searched on 29 October 2012, and the search was repeated in August 2013 to update the review. The search strategy was confined to any paper published between 1 January 2000 to the date of the last search, 15 August 2013. These limits were imposed so that early diagnostic studies of older commercially available tests, such as INNO-LiPA, would not be missed.

Additional sources were checked to ensure that the review included studies that were not missed. These included Cumulative Index to Nursing and Allied Health Literature (CINAHL), NHS Economic Evaluation Database (NHS EED), System for Information on Grey Literature in Europe Social Policy & Practice (SIGLE), diagnostic equipment manufacturer websites and experts within the field. Additional hand-searching was carried out to identify papers using the citation lists of published diagnostic accuracy reviews. 45,46 In addition, when study authors were contacted to confirm details within their papers they were also requested to suggest studies that could be potentially missing from the review. A systematic review database was developed in Microsoft^® Office 2010 (Microsoft Corporation, Redmond, WA, USA) to manage and record the review process.

Study selection

Studies were included in the review if they met the inclusion criteria listed below. Studies that met these criteria were included irrespective of the published language, the country of origin or their current publication status (i.e. grey literature, published or ‘in press’).

The eligibility criteria were as follows:

• Studies that:

  • assessed rapid genetic diagnostic methods to detect drug susceptibility of M. tuberculosis

  • used human clinical samples

  • compared the results of the rapid (index) test with sequencing or a culture-based sequencing DST as a reference standard

  • reported sufficient data to calculate the true positive (TP), true negative (TN), false positive (FP) and false negative (FN) of the rapid diagnostic test

  • reported at least 10 samples susceptible to the drug of interest and 10 samples that were resistant to the drug of interest, identified by the reference standard.

All articles that could potentially meet the eligibility criteria outlined above were selected for initial review. The assessment of study eligibility was not blinded to publication details, such as journal or author names.

Data extraction

Two reviewers independently evaluated titles, abstracts and full-text papers (see Appendix 2). Each reviewer had an independent copy of the review database.

In order to identify valid studies, the initial data extraction from the electronic databases were reviewed by one of the study reviewers (MC). The second reviewer (CT) validated a randomised subset (50%) of the titles for eligibility. The subsequent potentially eligible studies were randomly sorted and abstracts were divided between reviewers (MC, CT, AS), with each reviewer assessing a randomised subset (60%) of the abstracts, with a 10% overlap to check consistency.

The full-text papers of eligible papers were independently reviewed and data were extracted by the first reviewer MC (100%) and by the second reviewers AS (69%) and AB (31%).

Key variables – including patient characteristics, drug resistance(s) investigated, test used, characteristics of the tests (i.e. nature of assay used, method of drug resistance, detection and location) – were recorded. Other characteristics to be recorded – including study quality, publication details, time for analysis, sensitivity and specificity – were extracted for the data set. A sample data collection form is shown in Appendix 3 and the full list of extracted variables is given in Appendix 4.

These data were matched and, when the extracted data did not match, they were re-extracted by both the first and second reviewers MC and AB, respectively. In the event that the TN, FP, FN, TN did not match the reported sensitivity and specificity, the data were re-extracted a third time by the primary reviewer (MC), and the study author was contacted for clarification.

Quality assessment

The quality of each study deemed eligible – either through bias or concerns regarding applicability – was assessed using the Quality Assessment of Diagnostic Accuracy Studies tool version 2 (QUADAS-2). 47 The QUADAS-2 tool was tailored to this study by selecting relevant key questions in each QUADAS-2 domain. It was considered that, although the questions regarding bias when conducting the reference test were applicable, the questions regarding reference test applicability were not, as these were part of the study eligibility criteria.

During data extraction, each study was independently evaluated using QUADAS-2, and disagreement between reviewers’ opinions was resolved by discussion, with reference back to the original article. As recommended by Whiting et al. ,47 if a domain contained one or more question answered as ‘not clear’ or ‘high-bias’ then it was considered an area of potential bias or a concern regarding the applicability of the test. As recommended by Whiting et al. ,47 further quantitative sensitivity analysis was not performed.

Data synthesis

The study extraction data set from each reviewer was cleaned using Microsoft Excel version 2010, Microsoft Access version 2010 and Stata version 13.1 (StataCorp LP, College Station, TX, USA: www.statacorp.com).

Each group of tests, with sufficient studies available, was analysed separately. For each test comparison, the sensitivity, specificity and their exact 95% confidence intervals (CIs) were calculated using Stata. If meta-analysis was not found to be appropriate, because of a small number of studies or clinical heterogeneity, a qualitative narrative synthesis was used.

Statistical heterogeneity of sensitivities and specificities were investigated in any group of studies that had more than four studies with apparent clinical heterogeneity. Statistical analysis of the review data reflects that suggested by Lijmer et al. 48,49 and Higgins and Thompson. 50 Heterogeneity between studies was assessed using the I²-statistic. 50

Accuracy is usually presented in individual studies in terms of sensitivity and specificity, that is, dichotomous data rather than differences in distributions. Standard meta-analytic techniques, which are a simple pooled estimate of sensitivity and another of specificity, may be inappropriate, as these two statistics are likely to be correlated. Therefore, diagnostic accuracy across studies were summarised using a summary receiver operating characteristic (SROC) curve. This figure is a graphical summary of a hierarchical logistic regression model used to explain variability in study diagnostic odds ratios (DORs). In particular, variability across studies as a result of blinding or the use of different thresholds to define positivity were assessed and modelled using this approach. Hierarchical summary receiver operating characteristic (HSROC) curve models were used because they account for both between- and within-study variation in TP and FP rates.

Details of included and excluded studies

A total of 8922 titles and abstracts were identified through database searches and hand-searching. After the first phase of screening, 557 papers were identified as being potentially eligible for the review (Figure 1).

FIGURE 1.

Flow diagram showing study selection process.

A total of 56 studies contained sufficient information on the performance of the rapid diagnostic tests to be included in the review. These studies were divided into five categories, depending on the rapid test investigated: INNO-LiPA studies (n = 9),33,35,51–57 GeneXpert studies (n = 6),29,38,58–61 MTBDRplus studies (n = 18),34,62–75,98–100 MTBDRsl studies (n = 6)37,76–80 and other rapid diagnostic studies (n = 17);81–97 the last category comprised any report with insufficient information about the type and manufacture of the assay tested, or assays that had been designed in-house.

Subgroup analysis was limited by the dearth of information reported on participant age, HIV status and other identified risk factors for heterogeneity in diagnostic studies (Table 1). Among the 56 studies, only one study37 was conducted on samples from HIV-positive patients and two studies58,97 on patients that were not HIV positive. The remaining studies were conducted on samples with either mixed or undisclosed HIV status. The majority of studies did not report any data on participant age (n = 44), and the remaining studies were carried out on adults (n = 8) or a mix of adults and children (n = 4).

GeneXpert

Findings of the review

A total of six29,38,58–61 studies reporting the use of GeneXpert were eligible for inclusion and subsequently analysed (Table 2). The reported sensitivity and specificity of the GeneXpert test varied between 81.3–100.0% and 97.4–100.0%, respectively (see Table 2). The pooled estimates of sensitivity (96.8%, 95% CI 94.2% to 99.4%) and specificity (98.4%, 95% CI 97.8% to 99.0%) suggested a high level of diagnostic accuracy when this test was used to detect rifampicin resistance in clinical samples (Figure 2).

TABLE 2 - GeneXpert sensitivity and specificity for the detection of rifampicin resistance in clinical samples

Study (first author and year)	TP	FP	FN	TN	Sensitivity (%)	95% CI (%)	Specificity (%)	95% CI (%)
Boehme (2010)29	200	10	5	505	97.6	94.4 to 99.2	98.1	96.5 to 99.1
Boehme (2011)38	236	14	14	796	94.4	90.8 to 96.9	98.3	97.1 to 99.1
Naidoo (2010)60	157	8	1	635	99.4	96.5 to 100	98.8	97.6 to 99.5
Kim (2012)58	21	0	0	41	100.0	83.9 to 100	100.0	91.4 to 100
Kurbatova (2013)59	57	1	5	38	91.9	82.2 to 97.3	97.4	86.5 to 99.9
O’Grady (2012)61	13	2	3	78	81.3	54.4 to 96.0	97.5	91.3 to 99.7

FIGURE 2.

Detection of rifampicin resistance by GeneXpert: forest plot of (a) sensitivities and (b) specificities. D + L, DerSimonian and Laird; ES, effect size; I–V, instrumental variables.

The calculation and analysis of the DOR for each study also suggested a high level of diagnostic accuracy when the study data were pooled (DOR = 1209.2, 95% CI 446.3 to 3276.4; p < 0.05) (Figure 3). However, there was marked heterogeneity between studies, indicated by both the wide 95% CI of the DOR, and both tests for heterogeneity using both the chi-squared distribution [11.46, degrees of freedom (df) = 5; p = 0.04] and the I² method (56.4%), which is more robust when evaluating meta-analyses containing a small number of studies.

FIGURE 3.

Detection of rifampicin resistance by GeneXpert: forest plot of DORs. D + L, DerSimonian and Laird; M–H, Mantel–Haenszel.

Subsequent analysis and the construction of a HSROC curve indicated similarly high levels of predicted sensitivity (96.5%, 95% CI 91.4% to 98.5%) and specificity (98.2%, 95% CI 97.6% to 98.8%), with the majority of the studies and the summary point clustered in the top left-hand corner of the plot (Figure 4). However, again there was evidence of heterogeneity between the study estimates in the distended predicted 95% confidence ellipse.

FIGURE 4.

Summary receiver operating characteristic plots of GeneXpert diagnostic accuracy for the detection of rifampicin resistance in clinical samples.

Further investigation into the potential causes of heterogeneity could not be carried out as the small number of studies included in the analysis could not be reasonably broken down into subcategories for subgroup analysis (see Table 1).

Secondary analyses

In order to investigate possible sources of heterogeneity between studies, a qualitative assessment of the studies was carried out using the QUADAS-2 framework. Each study was evaluated using a predefined set of indicator questions to evaluate factors that may bias the study findings or raise concerns about the applicability of the study. The summarised results of this assessment are outlined in Table 3 and Figure 5.

TABLE 3 - Quality Assessment of Diagnostic Accuracy Studies version 2: risk of bias and applicability concerns summary of GeneXpert diagnostic accuracy for the detection of rifampicin resistance in clinical samples, by study

☺, low risk of bias; ☹, high risk of bias; ?, unclear risk of bias.

Study (first author and year)	Risk of bias				Applicability concerns
	Patient selection	Index text	Reference standard	Flow and timing	Patient selection	Index text
Boehme (2010)29	☺	☺	☺	?	☺	☺
Boehme (2011)38	☺	?	?	?	☺	?
Naidoo (2010)60	?	?	?	?	?	☺
Kim (2012)58	☹	☹	☹	?	☺	☺
Kurbatova (2013)59	☺	☹	☹	?	☺	☺
O’Grady (2012)61	?	?	?	☺	☺	?

FIGURE 5.

Risk of bias and applicability concerns summary of GeneXpert diagnostic accuracy for the detection of rifampicin resistance in clinical samples.

In the ‘Flow and timing’ domain, used to assess possible biases introduced by the use of patient samples, there was a marked lack of clarity surrounding the timing and the thresholds of the index test and reference test and the exclusion of samples from the assay.

Concerns regarding a high risk of bias in both the ‘Reference test’ and ‘Index test’ domains were associated with the absence of, or a lack of clarity associated with, blinding within the study. The study identified as high risk in the ‘Patient selection’ domain did not enrol patients either consecutively or randomly, and there was a strong indication that convenience sampling had been used.

When assessing the applicability of the methodology to the research question, the research setting and patient population did not raise concerns regarding the applicability of the assay.

There was no statistically significant evidence of publication bias based on both Egger’s test and Begg’s test (p > 0.05), based on rifampicin resistance.

INNO-LiPA

Findings of the review

A total of nine studies33,35,51–57 reporting the use of INNO-LiPA were eligible for inclusion (Table 4). It was possible to extract data broken down by smear status from only two studies35,52 (see Table 1).

TABLE 4 - INNO-LiPA sensitivity and specificity for the detection of rifampicin resistance in clinical samples

Study (first author and year)	TP	FP	FN	TN	Sensitivity (%)	95% CI (%)	Specificity (%)	95% CI (%)
Pulmonary specimens
Costeira (2007)51	11	2	1	99	91.7	61.5 to 99.8	98.0	93.0 to 99.8
Drobniewski (2000)52	55	3	9	14	85.9	75.0 to 93.4	82.4	56.6 to 96.2
Higuchi (2004)53	22	1	0	19	100.0	84.6 to 100.0	95.0	75.1 to 99.9
Johansen (2003)54	26	0	0	21	100.0	86.8 to 100.0	100.0	83.9 to 100.0
Ogwang (2009)55	13	2	2	13	86.7	59.5 to 98.3	86.7	59.5 to 98.3
Sam (2006)56	31	2	1	623	96.9	83.8 to 99.9	99.7	98.8 to 100.0
Skenders (2005)57	31	2	3	52	91.2	76.3 to 98.1	96.3	87.3 to 99.5
Viveiros (2005)33	31	0	1	310	96.9	83.8 to 99.9	100.0	98.8 to 100.0
Sputum specimens
Smear positive
Drobniewski (2000)52	11	0	4	11	73.3	44.9 to 92.2	100.0	71.5 to 100.0
Seoudi (2012)35	52	5	4	1554	92.9	82.7 to 98.0	99.7	99.3 to 99.9
Smear negative
Seoudi (2012)35	20	0	1	100	95.2	76.2 to 99.9	100.0	96.4 to 100.0

The reported sensitivity and specificity of the INNO-LiPA test varied between 86.7–100.0% and 82.4–100.0%, respectively (see Table 4). The pooled estimates of sensitivity (95.4%, 95% CI 92.2% to 98.3%) and specificity (99.7%, 95% CI 99.5% to 100.0%) suggested a high level of diagnostic accuracy when this test was used to detect rifampicin resistance in clinical samples (Figure 6).

FIGURE 6.

Detection rifampicin resistance by INNO-LiPA: forest plots of (a) sensitivities and (b) specificities. D + L, DerSimonian and Laird; ES, effect size; I–V, instrumental variables. a, All specimens; b, smear-positive specimens.

The calculation and analysis of the DOR for each study suggested a high level of diagnostic accuracy when the study data were pooled (DOR = 770.2, 95% CI 159.6 to 3717.2; p < 0.05) (Figure 7). However, there was some evidence of heterogeneity between the estimates reported by the studies, indicated by both a wide 95% CI of the DOR, and both tests for heterogeneity using both the chi-squared distribution (43.9, df = 9; p < 0.05) and the I² method (79.5%).

FIGURE 7.

Detection of rifampicin resistance by INNO-LiPA: forest plot of DORs. M–H, Mantel–Haenszel. a, All specimens; b, smear-positive specimens.

Subsequent analysis and the construction of a HSROC curve indicated similarly high levels of predicted sensitivity (94.1%, 95% CI 89.5% to 96.7%) and specificity (99.1%, 95% CI 96.5% to 99.8%), with the majority of the studies and the summary point clustered in the top left-hand corner of the plot (Figure 8). However, there was a strong indication of heterogeneity, particularly with regard to specificity of the test, as demonstrated by the shape and orientation of the 95% confidence ellipse.

FIGURE 8.

Summary receiver operating characteristic plot of INNO-LiPA diagnostic accuracy for the detection of rifampicin resistance in clinical samples.

Further investigation into the potential causes of heterogeneity could not be carried out, as the small number of studies included in the analysis could not be broken down into large enough subcategories for subgroup analysis (see Table 1).

Secondary analyses

In order to investigate possible sources of heterogeneity between studies, a qualitative assessment of the studies was carried out using the QUADAS-2 framework. The summarised results of this assessment are outlined in Table 5 and Figure 9.

TABLE 5 - Quality Assessment of Diagnostic Accuracy Studies version 2: risk of bias and applicability concerns summary of INNO-LiPA diagnostic accuracy for the detection of rifampicin resistance in clinical samples, by study

☺, low risk of bias; ☹, high risk of bias; ?, unclear risk of bias.

Study (first author and year)	Risk of bias				Applicability concerns
	Patient selection	Index text	Reference standard	Flow and timing	Patient selection	Index text
Costeira (2007)51	?	?	?	?	☺	☺
Drobniewski (2000)52	☹	?	?	?	☺	☺
Higuchi (2004)53	?	☹	☺	☹	☺	☺
Johansen (2003)54	☺	?	?	?	☺	☺
Ogwang (2009)55	☺	?	?	?	☺	☺
Sam (2006)56	☺	☺	☺	☺	☺	☺
Seoudi (2012)35	☺	?	?	☺	☺	☺
Skenders (2005)57	☺	☹	☹	?	?	?
Viveiros (2005)33	☺	?	?	?	☺	☺

FIGURE 9.

Risk of bias and applicability concerns summary of INNO-LiPA diagnostic accuracy for the detection of rifampicin resistance in clinical samples, by study.

As with the GeneXpert studies, there was a marked lack of clarity surrounding the timing and thresholds of the index reference tests. This, combined with a lack of clarity surrounding the exclusion of samples from the assays, contributes to the high proportion of studies designated to have an unclear risk of bias in the ‘Flow and timing’ domain.

Concerns regarding a high risk of bias in both the ‘Reference test’ and ‘Index test’ domains were associated with the absence of, or a lack of clarity associated with, blinding within the study.

When assessing the applicability of the methodology to the research question the studies were categorised as either low or unclear concern.

There was no statistically significant evidence of publication bias based on both Egger’s test and Begg’s test (p > 0.05), based on rifampicin resistance.

GenoType MTBDRplus

Findings of the review

A total of 18 studies34,62–75,98–100 reporting the use of GenoType MTBDRplus^® (Hain Lifescience, Nehren, Germany) were eligible for inclusion into the meta-analysis (Table 6).

TABLE 6 - GenoType Mycobacterium tuberculosis drug resistance plus assay sensitivity and specificity for the detection of isoniazid and rifampicin resistance in clinical samples

Study (first author and year)	TP	FP	FN	TN	Sensitivity (%)	95% CI (%)	Specificity (%)	95% CI (%)
Rifampicin
Macedo (2009)98	23	0	0	43	100.0	85.2 to 100.0	100.0	91.8 to 100.0
Nikolayevskyy (2009)99	103	4	4	38	96.3	90.7 to 99.0	90.5	77.4 to 97.3
Somoskovi (2006)74	51	0	41	51	55.4	44.7 to 65.8	100.0	93.0 to 100.0
Albert (2010)73	15	4	0	73	100.0	78.2 to 100.0	94.8	87.2 to 98.6
Dorman (2012)68	12	2	2	200	85.7	57.2 to 98.2	99.0	96.5 to 99.9
Mironova (2012)34	323	35	16	311	95.3	92.4 to 97.3	89.9	86.2 to 92.9
Eliseev (2013)66	69	2	0	31	100.0	94.8 to 100.0	93.9	79.8 to 99.3
Barnard (2008)70	94	2	1	357	98.9	94.3 to 100.0	99.4	98.0 to 99.9
Farooqi (2012)65	51	1	4	54	92.7	82.4 to 98.0	98.2	90.3 to 100.0
Asencios (2012)71	22	1	2	75	91.7	73.0 to 99.0	98.7	92.9 to 100
Tukvadze (2012)62	112	4	13	329	89.6	82.9 to 94.3	98.8	97.0 to 99.7
Maschmann (2013)100	23	2	5	32	82.1	63.1 to 93.9	94.1	80.3 to 99.3
Anek-Vorapong (2010)72	19	0	0	145	100.0	82.4 to 100.0	100.0	97.5 to 100.0
Crudu (2012)69	100	2	6	48	94.3	88.1 to 97.9	96.0	86.3 to 99.5
Hillemann (2006)64	15	0	0	27	100.0	78.2 to 100.0	100.0	87.2 to 100
Cauwelaert (2011)67	47	4	1	202	97.9	88.9 to 99.9	98.1	95.1 to 99.5
Lacoma (2008)75	29	1	0	20	100.0	88.1 to 100.0	95.2	76.2 to 99.9
Lyu (2013)63	57	4	2	365	96.6	88.3 to 99.6	98.9	97.2 to 99.7
Isoniazid
Macedo (2009)98	24	0	0	43	100.0	85.8 to 100.0	100.0	91.8 to 100.0
Albert (2010)73	21	0	5	66	80.8	60.6 to 93.4	100.0	94.6 to 100.0
Barnard (2008)70	114	1	7	330	94.2	88.4 to 97.6	99.7	98.3 to 100.0
Farooqi (2012)65	45	0	14	49	76.3	63.4 to 86.4	100.0	92.7 to 100.0
Asencios (2012)71	30	3	1	66	96.8	83.3 to 99.9	95.7	87.8 to 99.1
Tukvadze (2012)62	52	2	125	281	29.4	22.8 to 36.7	99.3	97.5 to 99.9
Anek-Vorapong (2010)72	27	0	2	135	93.1	77.2 to 99.2	100.0	97.3 to 100.0
Hillemann (2006)64	17	0	0	25	100.0	80.5 to 100.0	100.0	86.30 to 100.0
Cauwelaert (2011)67	55	4	14	181	79.7	68.3 to 88.4	97.8	94.6 to 99.4

The reported sensitivity and specificity of the MTBDRplus to detect resistance to rifampicin ranged between 82.1–100.0% and 89.9–100.0%, respectively (see Table 6). The pooled estimates of sensitivity (94.6%, 95% CI 91.6% to 97.6%) and specificity (98.2%, 95% CI 97.2% to 99.3%) suggested a high level of diagnostic accuracy when this test was used to detect rifampicin resistance in clinical samples (Figure 10).

FIGURE 10.

Detection of isoniazid and rifampicin resistance by MTBDRplus: forest plots of (a) sensitivity of isoniazid, (b) specificity of isoniazid, (c) sensitivity of rifampicin and (d) specificity of rifampicin. D + L, DerSimonian and Laird; ES, effect size; I–V, instrumental variables.

The calculation and analysis of the DOR for each study suggested a high level of diagnostic accuracy when the study data were pooled (DOR = 666.0, 95% CI 339.3 to 1307.2; p < 0.05) (Figure 11). However, there was some evidence of heterogeneity between the estimates reported by the studies, indicated by both a wide 95% CI of the DOR and both tests for heterogeneity using both the chi-squared distribution (35.1, df = 17; p < 0.05) and the I² method (51.5%).

FIGURE 11.

Detection of (a) isoniazid and (b) rifampicin resistance by MTBDRplus: forest plot of DORs. D + L, DerSimonian and Laird; M–H, Mantel–Haenszel.

Subsequent analysis and the construction of a HSROC plot indicated similarly high levels of predicted sensitivity (96.1%, 95% CI 91.9% to 98.1%) and specificity (98.1%, 95% CI 96.7% to 98.1%), with the majority of the studies and the summary point clustered in the top left-hand corner of the plot (Figure 12). However, there was a strong indication of heterogeneity, particularly with regard to the sensitivity of the test, as demonstrated by the length and orientation of the 95% confidence ellipse.

FIGURE 12.

Summary receiver operating characteristic plots of MTBDRplus diagnostic accuracy for the detection of (a) isoniazid and (b) rifampicin resistance in clinical samples.

The reported sensitivity and specificity of the MTBDRplus to detect resistance to isoniazid in clinical samples ranged between 29.4–100.0% and 95.7–100.0%, respectively (see Table 6). The pooled estimates of sensitivity (83.4%, 95% CI 66.3% to 100.0%) and specificity (99.6%, 95% CI 99.0% to 100.0%) suggested a high level of diagnostic accuracy when this test was used to detect isoniazid resistance in clinical samples (see Figure 10).

The calculation and analysis of the DOR for each study suggested a high level of diagnostic accuracy when the study data were pooled (DOR = 565.2, 95% CI 175.3 to 1823.0; p < 0.05) (see Figure 11). However, there was some evidence of heterogeneity between the estimates reported by the studies, indicated by both a wide 95% CI of the DOR, and both tests for heterogeneity using both the chi-squared distribution (18.4, df = 8; p < 0.05) and the I² method (56.5%).

Subsequent analysis and the construction of a HSROC plot indicated similarly high levels of predicted sensitivity (90.0%, 95% CI 73.6 to 96.7) and specificity (99.4%, 95% CI 97.9% to 99.8%), with the majority of the studies and the summary point clustered in the top left-hand corner of the plot (see Figure 12). However, there was at least one extreme outlier, Tukvadze et al. ,62 which, potentially, contributed to the substantial heterogeneity apparent in the sensitivity estimate. The authors of this study62 had been unsuccessfully contacted regarding a discrepancy between the diagnostic accuracy data reported and the published calculation of sensitivities and specificities. When the study62 was excluded from the analysis, the heterogeneity within the sample was reduced I² method (36.1%). However, there was only a small change in the predicted estimates of sensitivity (91.8%, 95% CI 82.8% to 96.2%) and specificity (99.4%, 95% CI 97.3% to 99.9%).

Further investigation into the potential causes of heterogeneity could not be carried out, as the small number of studies included in the analysis could not be broken down into large enough subcategories for subgroup analysis (see Table 1).

Secondary analyses

In order to investigate possible sources of heterogeneity between studies, a qualitative assessment of the studies was carried out using the QUADAS-2 framework. The summarised results of this assessment are outlined in Table 7 and Figure 13.

TABLE 7 - Quality Assessment of Diagnostic Accuracy Studies version 2: risk of bias and applicability concerns summary of MTBDRplus for the detection drug susceptibility in clinical samples, by study

☺, low risk of bias; ☹, high risk of bias; ?, unclear risk of bias.

Study (first author and year)	Risk of bias				Applicability concerns
	Patient selection	Index text	Reference standard	Flow and timing	Patient selection	Index text
Macedo (2009)98	☹	☺	☺	☹	☺	☺
Nikolayevskyy (2009)99	?	?	?	?	?	☺
Somoskovi (2006)74	☹	?	?	?	☺	?
Albert (2010)73	☺	☺	☺	?	☺	☺
Dorman (2012)68	☺	?	☺	?	☺	☺
Mironowa (2012)34	☺	?	?	?	☺	?
Eliseev (2013)66	☺	?	?	?	☹	☺
Barnard (2008)70	?	?	?	?	☹	☺
Farooqi (2012)65	?	?	?	?	?	?
Asencios (2012)71	☺	?	?	?	?	☺
Tukvadze (2012)62	☺	?	?	?	☺	☺
Maschmann (2013)100	☹	☹	☹	?	☺	☺
Anek-Vorapong (2010)72	?	?	☺	?	☺	☺
Crudu (2012)69	?	☺	?	?	☺	?
Hillemann (2006)64	?	?	?	?	?	?
Cauwelaert (2011)67	?	?	?	?	?	☺
Lacoma (2008)75	☹	?	?	?	☹	☺
Lyu (2013)63	☹	?	?	?	☹	☹

FIGURE 13.

Quality Assessment of Diagnostic Accuracy Studies version 2: risk of bias and applicability concerns summary of MTBDRplus for the detection drug susceptibility in clinical samples.

When compared with the studies included in the INNO-LiPA and GeneXpert analyses, there appeared to be a marked reduction in the quality of methodological reporting with studies using the MTBDRplus.

The lack of detail regarding timing, thresholds, patient selection and blinding resulted in the majority of studies classified as either ‘high bias’ or ‘unclear bias’ in the four key Quality Assessment of Diagnostic Accuracy Studies (QUADAS) domains. This is particularly marked in the ‘Flow and timing’ domain, for which all studies performed poorly.

There was no statistically significant evidence of publication bias based on both Egger’s test and Begg’s test (p > 0.05) based on rifampicin resistance, although there was strong evidence of publication bias using isoniazid (Egger’s test, p = 0.001).

GenoType MTBDRsl

Other studies

In addition to the studies that reported diagnostic data from the widely used commercial tests, 17 studies81–97 reporting data on other genetic rapid diagnostic tests were identified using the inclusion criteria. These tests were broadly broken into those studies that utilised a variety of TB biochip assays (n = 7, 41.2%) and those that used in-house PCR assays (n = 10, 59.8%).

The majority of studies (n = 13) reported the diagnostic accuracy of these test to detect rifampicin resistance in clinical samples (Table 9). The reported sensitivities and specificities of these studies were high, ranging between 79.8% and 100.0% and 92.2% and 100%, respectively. However, as the manufacture and design of these assays were not consistent it was not possible to conduct a meta-analysis on this data set.

Observation

Selection of days on which to observe the diagnostic tests was made based on convenience. Several biomedical scientists (BMSs) perform each diagnostic test at the NMRL and the observations were not restricted to any particular members of staff. The observer followed the BMS continuously while they were carrying out the diagnostic test.

Data collection

Based on the NMRL standard operating procedures (SOPs) for each diagnostic test, paper forms were created on which the working time involved in each significant task was recorded (see Appendix 7). The wall clocks present in each laboratory were used to time each task and the start and end time of each task was recorded on the form. Timings were rounded to the nearest minute or, if the time taken was significantly < 1 minute, rounded to the nearest 10 seconds. Any relevant additional information (such as tasks carried out while waiting for a sample in the centrifuge) was recorded in the ‘notes section’ next to the relevant task. The data were recorded in a spreadsheet, and the start and end times for each task were used to determine the working time for each task. The sum of the working times for each task equals the total working time. Every task involved in conducting the diagnostic test (such as paperwork, setting up, cleaning up, recording results, etc.) was included in the working time.

If a waiting time (e.g. waiting for a centrifuge to finish a spin) was ≤ 15 minutes, the waiting time was included in the working time, as the BMS could not realistically undertake separate tasks in that time. If the waiting time was > 15 minutes, unless the BMS remained unoccupied, the time was not recorded as working time, as the BMS had enough time to complete other unrelated tasks. If the BMS used the waiting time to complete tasks that were relevant to the test being timed, those tasks were timed and it was noted that they took place while waiting. The time taken for automated machines to conduct stages of an assay (> 15 minutes) or the time cultures spent in incubators was not recorded because no labour on the part of the BMS was involved.

The number of samples being processed during each observation was recorded. In addition, the number of each type of sample was recorded, because some sample types require slightly different methods of preparation. Reference cultures arrive on either solid or liquid medium and the number of each was also recorded.

Data analysis

At the NMRL, samples are processed in batches and, with the exception of the GeneXpert, individual samples are almost never assayed. For many tasks, the time taken to perform it is independent of the number of samples involved, for example preparation of PCR master mixes. The theoretical time taken for one sample was calculated by dividing the time taken by the number of samples. Thus, this is the mean time taken for a sample in an average batch size at the NMRL and does not represent the actual time that would be taken if a single sample was processed individually.

GeneXpert assay

At the NMRL, the GeneXpert system is used to assay respiratory and CSF specimens. As all samples received are subsequently cultured to permit further phenotypic testing, respiratory specimens undergo a preliminary decontamination step before following the manufacturer’s recommendations for sample processing. Occasionally, a large volume (> 2 ml) of CSF is received and it is concentrated by centrifugation prior to processing.

At the NMRL, the GeneXpert assay is performed on demand, on a daily basis, on up to four samples simultaneously. This assay was observed between November 2012 and April 2013: 10 times for respiratory specimens (Table 10) and seven times for CSF specimens (Table 11). The mean theoretical hands-on time taken to process a single CSF specimen was 24 minutes. For respiratory specimens, the equivalent calculated time for a single specimen was 44 minutes.

Line probe assays

The INNO-LiPA assay is run on primary samples received for molecular testing, other than respiratory specimens and CSF (which are run on the GeneXpert). The MTBDRplus assay, which is less sensitive, is subsequently run on isolates determined to be rifampicin resistant by either the INNO-LiPA or GeneXpert, in order to identify isoniazid-resistance mutations. The MTBDRplus assay is also run on cultures received by the reference laboratory.

At the NMRL, the INNO-LiPA and the MTBDRplus assays are conducted on a weekly basis. Typically, there are samples requiring INNO-LiPA testing every week, but the MTBDRplus assay is not always needed. The initial stages for each assay, resulting in amplification-ready DNA, are performed together. Set-up of the first-round PCR is also carried out for both assays together. Typically, the subsequent steps are carried out the following day. These are performed separately for each assay. The INNO-LiPA assay requires a second-round PCR and electrophoresis step to identify reactions containing amplicon. If no tubes contain sample amplicon then the hybridisation stage is not performed. Thus, for this assay, the time taken to identify negative samples is less than that to identify positive samples. For the MTBDRplus assay, all samples and controls are hybridised to strips, and there is no difference between the processing time for positive and negative samples.

Between November 2012 and April 2013, the INNO-LiPA assay performed in isolation was observed seven times and the two LPAs performed together observed a further five times (Table 12). The theoretical time per sample for each stage was calculated and the stages required for each assay summed to estimate the theoretical time for a single sample for each assay performed independently, if processed in a batch of 2–15 samples. The calculated per sample hands-on time for a PCR-positive INNO-LiPA assay was 45 minutes and for the MTBDRplus assay was 50 minutes.

Culture-based drug susceptibility testing

Phenotypic testing is considered the gold standard for determining MTB drug resistance, and samples received at the NMRL for molecular testing are subsequently assayed using culture-based methods. The observations of the culture-based methods took place between April 2013 and June 2013.

Prior to DST, primary specimens are cultured in drug-free media (LJ and MGIT) to provide sufficient replicating bacilli for testing. LJ slopes are examined weekly for growth, and MGIT growth is monitored daily. Purity of the culture is assessed by plating on Columbia Blood Agar plates and acid-fast staining. The theoretical mean time taken to complete all tasks for the culture of one primary specimen was 20 minutes (Table 13).

Testing for resistance to first-line drugs (rifampicin, isoniazid, ethambutol, pyrazinamide and streptomycin) is conducted using solid egg-based LJ medium that is prepared in-house. The resistance ratio method utilised requires three slopes per drug. In addition, four control slants (positive, p-nitrobenzoic acid, thiophene 2-carboxylic acid hydrazide and low temperature) are included for each sample to identify positivelyMTB complex isolates. Thus, a total of 19 slopes are inoculated per sample. The LJ slopes are prepared on demand, in batches corresponding to 30 eggs-worth of medium, which fills approximately 780 bijou bottles. In each batch, slopes containing a single drug are prepared; these are stored for up to 2 months. The theoretical time taken to prepare LJ slopes for testing one sample was estimated at 4 minutes (Table 14). In total, the theoretical hands-on time required to set up the first-line sensitivities for one specimen was 21 minutes (see Table 14). Including the time required for the initial culture in drug-free media, the total hands-on time for one specimen was 41 minutes.

Testing for second-line drugs [aminoglycosides, capreomycin, fluoroquinolones, thioamides and linezolid (Zyvox^®, Pfizer)] is conducted in the BACTEC MGIT 960 system at the NMRL and requires 11 tubes per sample. Media-containing MGITs are purchased commercially and prepared by the addition of supplied supplements plus drugs when required. The theoretical time to prepare tubes sufficient to test one sample was 9 minutes. In total, the hands-on time required for one specimen (for 10 drugs) was 28 minutes (Table 15). Including the time required for initial specimen culture, the total theoretical per-sample time was 49 minutes.

Diagnostic strategies

• The baseline for comparison was the current diagnostic strategy, including smear microscopy, culture for identification of MTB and DST for culture-positive cases. It was assumed that smear microscopy results are available within 1 day of collection of the sputum sample. Culture testing was assumed to be 100% accurate for both TB diagnosis and drug susceptibility and was treated as the reference standard for other tests. The results of culture tests were assumed to take 14 days for diagnosis of MTB and an additional 14 days for DST. Note that times to test results will differ between patients, and that the figures used here are intended as an average across the population.

• The intervention evaluated was the addition of a rapid molecular assay for the detection of TB disease and drug resistance, alongside the current diagnostic strategy. We assumed that molecular test results would be available for both TB status and drug sensitivity 3 days from collection of the sputum sample, on average.

• The currently available molecular assays test for MTB and rifampicin resistance (GeneXpert and INNO-LiPA) and MTBDRplus also tests for isoniazid resistance. In practice, the results of culture-based DST are used to identify sensitivity to other drugs and to derive individualised drug regimens. However, for the modelling we did not distinguish between different patterns of drug resistance. This is a simplification and relies on the assumptions that patients with resistance to a single drug would not incur significant additional risks or costs for isolation or drug treatment compared with patients with DS disease, and that patients with different patterns of multiple drug resistance would face similar risks and costs to one another (this is unlikely to be true for XDR-TB, but this is still a rare phenomenon in the UK). Furthermore, we assumed that the diagnostic accuracy of the molecular assays for detecting ‘true’ MDR-TB is the same as for detecting rifampicin resistance (as reported in Chapter 3). This is justified by the observation that a large majority of MDR patients are resistant to rifampicin. 38

• The analysis focused on the GeneXpert system, the INNO-LiPA and MTBDRplus assay as exemplars for the rapid molecular assays.

• At present, microscopy and mycobacterial culture are most usually performed in local hospital laboratories, but DST for MTB-positive cases is performed at specialist mycobacterial reference laboratories. 112 We assumed that these arrangements would not change with the wider adoption of molecular testing. However, we did investigate alternative scenarios for the location of rapid molecular assays: centralised, with molecular testing conducted only at TB specialist reference laboratories, and localised, with molecular testing distributed at local TB laboratories. It was assumed that centralised testing would reduce the mean cost per test (owing to economies of scale in the laboratory arising from more intensive use of capital) but at the expense of possible delay and an additional cost for transport of samples.

Population and subgroups

• The transmission model adopted a population approach, modelling incidence of TB and the process of diagnosis and treatment in three sections of the community with high incidence of MDR-TB: Black African, South Asian and Eastern European individuals.

• The population subject to intervention was individuals being tested for suspected TB disease. This includes individuals who present symptomatically or as a result of active case finding. Individuals may have been referred to TB services from general practice, emergency departments or other specialties.

• Individuals undergoing diagnosis are categorised into subgroups which are defined by their true disease status, prior risk status perceived by clinicians and test results (see Diagnostic subgroups, below).

Care pathways

• For each diagnostic strategy and subgroup, we defined a care pathway that specified what medical treatments and health services individuals would receive between presentation to the point when they either complete TB treatment successfully, or fail to complete treatment (are lost to follow-up). See Care pathways, below.

• We estimated a total cost for each care pathway, including the cost of consultations and investigations prior to referral to TB services, TB diagnosis, medication, inpatient care and isolation, outpatient care, and contact tracing and associated treatment of any latent TB cases.

• In addition, for each care pathway the health impact is estimated in terms of the mean discounted QALYs lost as a result of TB disease and treatment.

Diagnostic subgroups

Various factors define the unique care pathways that patients can follow after referral for suspected TB. Under the current diagnostic strategy (the baseline comparator), the key factors are as follows.

True clinical status (no tuberculosis, DS tuberculosis, MDR-TB): a patient starting the process of diagnosis for pulmonary TB can be in one of three clinical states: no active TB disease, DS TB, or MDR-TB. As noted above, for simplicity, we did not differentiate between resistance to single or multiple drugs, or between different patterns of drug resistance.

Smear status (positive, negative): the smear status is identified by microscopy, within 1 day of collection of the sputum sample. Smear status indicates infectivity and so determines if the clinician decides to isolate the patient or provides presumptive medication prior to diagnosis.

MDR risk status (high risk, low risk): risk status is determined a priori by the clinician, with patients who are deemed to be at ‘high risk’ of MDR-TB invoking more conservative action. The main risk factors for MDR-TB are history of prior TB drug treatment, birth in a high-incidence country, HIV infection, residence in London, age between 25 and 44 years, and male gender. 108 However, these features are non-specific, and presumptive treatment of every patient with one or more of these risk factors would be unrealistic.

For the purposes of calculation, we defined risk status simply as a percentage of suspected TB cases that a clinician deems to be at sufficiently high risk of MDR-TB to warrant admission to a negative pressure isolation room and presumptive treatment with MDR medications if the person is smear positive. However, the decision to prescribe MDR medication presumptively would also depend on the expected delay in diagnosis. We assumed that presumptive treatment would not be used with molecular testing, as the results are available in a few days compared with 2–4 weeks for culture and DST alone. In the absence of better information, we assumed that clinicians would be relatively ‘well-calibrated’ in their judgement, and that the proportion of patients identified as high risk would be similar to the ratio of MDR cases to all TB cases in the relevant population: 1.6% (13 cases) of Black African patients, 1.4% (31 cases) of South Asian patients, and 24% (13 cases) of Eastern European patients. 113 Based on these figures, we assumed that 2% of patients would be treated as high risk in the Black African and South Asian populations, and 25% in the Eastern European population.

These three factors define the 12 diagnostic subgroups that have distinctive care pathways under the current diagnostic strategy (Table 17).

The introduction of a rapid assay increases the number of possible care pathways, because errors in molecular test results can impact on treatment decisions prior to day 28, when the correct culture results are available. There are 32 permutations of true disease status, smear status, risk status and molecular assay results that lead to distinct care pathways (Table 18).

One final factor that leads to variation in the care pathway, and hence in the cost and QALYs attached to the diagnostic subgroups, is whether or not the individual completes their course of medication. This affects the duration of care the patient receives and therefore the total duration over which they accrue costs and QALYs from treatment, and whether they are cured or re-enter the population as a transmission risk. We assumed that, on average, individuals who fail to complete treatment default half-way through the course of treatment: uncertainty over this proportion was reflected in the PSA by sampling from a beta distribution with a mean of 0.5 and standard error of 0.1.

Care pathways

Each of the above diagnostic subgroups was assumed to follow a distinct care pathway. We considered the care pathway to be broken down into the following stages:

Pre referral This is the period of time from when a patient becomes symptomatic or is first suspected of having TB, to the time when they are referred to a TB or respiratory specialist for investigation. We assumed that the referral can come from one of three channels: from primary care, from an emergency department, or from another hospital specialty.

Diagnosis The duration of the diagnostic period is from the first collection of sputum samples until the definitive diagnosis for TB and drug sensitivity is received. The diagnostic period can overlap with treatment as an inpatient or outpatient.

Inpatient treatment From the point the patient is admitted until they are discharged. For smear-positive patients, inpatient care will include isolation until the patient is no longer considered a transmission risk. If the patient is smear positive and at high risk of MDR-TB, or if they are diagnosed with MDR-TB, we assumed isolation in a negative pressure room.

Outpatient treatment Once a patient is deemed well enough, and is no longer a transmission risk, they will be discharged and continue treatment as an outpatient.

The addition of a rapid molecular test is assumed to impact on the care pathways in two main ways during the diagnostic period: first, it may influence clinical decisions about what presumptive drug treatment is offered, if any, and, second, it may influence decisions about whether or not a patient is admitted and, if so, to what level of isolation. These decisions may also have implications for the total duration of drug treatment and the time spent as an outpatient, which depend on when effective treatment is initiated.

Examples of care pathways

The care pathways described above define the utilisation of health services (tests, medications, inpatient care and outpatient follow-up) and hence the costs incurred for the NHS. They also define the duration of patients’ exposure to TB symptoms and to adverse events of treatment, and hence to their QALY loss over the treatment period. Figures 15 and 16 illustrate how the costs and QALYs are incurred over the care pathway for a patient with smear-positive DS TB, but who is perceived in advance to be at high risk of MDR disease.

FIGURE 15.

Care pathway for high-risk patient with smear-positive DS TB, with culture only. +ve, positive; NP, negative pressure.

FIGURE 16.

Care pathway for high-risk patient with smear-positive DS TB, and correct molecular test result. +ve, positive; isol., isolation; NP, negative pressure.

Figure 15 shows the expected results under the current diagnostic strategy (culture only). Owing to the perceived transmission risk, the patient is admitted to negative pressure isolation on presentation and remains there until DST results become available at day 28, after which he/she is discharged to outpatient follow-up. MDR medication is started on admission as a pre-emptive measure, but the patient is transferred to a standard regimen when culture confirms DS disease at day 28. The overall course of treatment lasts for 6 months (183 days), if completed. The patient has impaired quality of life (utility) as a result of developing symptoms during the prereferral period, more severe symptoms while in hospital, and then some continuing loss of quality of life after discharge while recovering. While on MDR treatment, the patient is at risk of additional side effects (particularly ototoxicity attributable to injectable agents).

Figure 16 shows how the treatment pathway, and associated costs and QALY loss, would be expected to change for this patient with a (correct) molecular test result on day 3. At this time, the patient would be released from negative pressure isolation to a normal isolation room and start a standard drug regimen. Both of these changes would save money for the NHS, and the patient would not be exposed to the risks of the MDR regimen, reducing the risks of toxicity. In this case, the length of hospital stay and total duration of the pathway would not change.

Method for estimating costs

Costs were estimated for each care pathway, including costs (1) of tests and consultations prior to referral for a TB diagnosis, (2) for tracing, testing and treating latent infection in contacts of each index case, (3) of tests during the diagnostic period, (4) of medication, (5) and for inpatient care and outpatient follow-up. The first two of these items (prereferral and public health control costs) were treated as a fixed (deterministic) cost per index case. The other costs differed between the pathways. All costs were estimated from a NHS perspective and were discounted to a present value at an annual rate of 3.5%.

An example of a cost calculation for one of the care pathways is shown in Table 23. This is for a smear-positive patient considered to be at high risk of MDR disease, but who in fact has DS disease and is treated under the current culture-based diagnostic strategy (as in Figure 15). The high cost for this pathway is driven by the (unnecessary) expense of negative pressure isolation, and MDR medications. In reality, only a small proportion of patients presenting with TB symptoms will be considered at such high risk of MDR-TB and incur this high cost.

Method for estimating quality-adjusted life-year loss

The QALY loss associated with each pathway was also estimated. This was estimated relative to expected survival and quality of life (utility) for individuals of the same age but without TB and not undergoing diagnosis or treatment for TB. QALYs were discounted to present values using an annual rate of 3.5% per year. Three potential health impacts were considered: impairments to utility attributable to TB per se, utility impairments attributable to adverse effects of treatment, and the case fatality risk.

Probabilistic sensitivity analysis

Probabilistic sensitivity analysis was used to examine the impact of uncertainty over the input parameters to the transmission model. For each parameter subject to uncertainty, a probability distribution was defined and a set of 10,000 parameter values were generated by random (Monte Carlo) sampling in Excel. Separate streams of random numbers were set for each parameter to ‘seed’ them in such a way that when the distributions of some parameters were changed for alternative decision scenarios (described below), the values of other parameters would not change. This reduces unnecessary variation between scenarios. Matrices of PSA sample values were saved as text CSV (comma-separated values) files and provided to the mathematical modellers as inputs for the transmission model. Alongside the ‘economic data’ (cost and utility estimates) described in this chapter, samples of some other key input parameters were included in the PSA file, including the estimates of diagnostic accuracy obtained from the systematic review presented in Chapter 3.

Distributions were chosen to reflect the appropriate ranges for the parameters of interest. For example, beta distributions were used for probabilities and utility multipliers, as these can take on values of only between zero and one. Costs were sampled from a gamma distribution, as they must be > 1 and usually have a positive skew. Some parameters not subject to significant uncertainty were treated as fixed. For example, the unit costs of most medications are defined nationally, and, although there is some local variation in prices paid by NHS providers as a result of negotiated discounts, no data are available on these discounts. We therefore followed the NICE convention of using NHS list prices for medications. 123 Similarly, the costs of many NHS services (e.g. outpatient consultations and routine tests) are now set in a national tariff that is not subject to local negotiation, and so we treated these costs as deterministic. However, there were some large drivers of costs which were subject to considerable uncertainty: notably the duration of time spent in inpatient care and the cost of negative pressure isolation. These elements were sampled probabilistically. For sampling purposes, when no estimate of uncertainty was available, standard errors were set at 10% or 20% of the mean. For all elements estimated probabilistically, the distribution applied and standard error are reported with their mean values in the tables below.

Decision scenarios

Six sets of PSA samples were taken to reflect different combinations of assumptions over the location of molecular testing (centralised vs. localised) and over the proportion of patients being tested for TB who are judged by clinicians to be at high risk of MDR disease. The location of testing has potential impacts on the cost and timing of the diagnostic process. We assumed that microscopy and initial culture testing is currently conducted locally, but that DST for culture-positive cases is conducted only at specialist TB reference laboratories. These arrangements were assumed not to change with the introduction of molecular testing. Under the ‘centralised’ scenario, the equipment and consumables for molecular testing would be located at the TB reference laboratories and used efficiently, at maximum capacity. However, a cost would be incurred for transport of the sample to the reference laboratory for molecular testing – we assumed an average cost of £6.50 per sample. Under the ‘localised’ scenario, we assumed that molecular tests would be conducted at local laboratories, incurring no additional transport costs compared with current practice. However, we assumed that they would be operating below capacity for molecular tests, losing some economies of scale. Across England, there are 110 local TB laboratories, each processing on average 1030 tests per year. 112 We estimated that this would increase the mean cost per test by approximately £6 compared with centralisation of molecular tests. The net effect of transport and capacity on the cost per molecular test is therefore minimal. There is, however, a potential time delay associated with the transport of samples to a reference laboratory for molecular testing: we assumed 1 day extra, on average.

We produced PSA samples for centralised and localised locations of testing for each of the three populations modelled. For each population, we varied the proportion of patients with suspected TB who clinicians would judge to be at sufficiently high risk of MDR-TB to justify presumptive treatment: 2% for the Black African and South Asian populations and 25% for Eastern Europeans.

Test accuracy parameters

Estimates of the sensitivity and specificity of the rapid molecular tests for detecting drug sensitivity in patients with confirmed TB disease, compared with the reference standard of culture DST, have been presented in Chapter 3. However, many studies lacked the breakdown of diagnostic accuracy by smear status and by test for TB or drug susceptibility. This limited which studies could be used to inform the economic analysis.

It was important to differentiate diagnostic accuracy by patients’ smear status for the economic analysis, as this impacted on decisions about presumptive treatment and isolation. Smear status also had an important impact on transmission probability in the dynamic model (reported in the following chapter). Similarly, the care pathways and transmission risks also depended on the accuracy of molecular tests for early detection of TB disease.

The diagnostic accuracy parameters for the GeneXpert, MTBDRplus and INNO-LiPA assays are shown in Tables 25–27. These estimates were based on studies included in the systematic review. In addition, estimates of the sensitivity and specificity of GeneXpert assays for detection of TB (compared with culture as the reference standard) were obtained from Drobniewski et al. 124

To account for uncertainty, the diagnostic accuracy results were sampled probabilistically. For sampling of the diagnostic accuracy, it is important to retain the correlations between the sensitivity and specificity parameters. To do this, we first calculated the DOR from the mean estimates of the sensitivity and specificity. The DOR is a summary measure of diagnostic accuracy, the ratio of the odds of a positive test result if the patient has the disease over the odds of a positive test result if the patient does not have the disease, and can be expressed as a function of the sensitivity (Se) and specificity (Sp).

We drew random samples for one of the accuracy parameters (sensitivity); as this is simply a proportion, it was sampled from a beta distribution. We then rearranged the equation above and used the DOR and the sampled sensitivity values to calculate the corresponding specificity values. This approach is likely to underestimate the overall uncertainty (as it does not allow for uncertainty over the DOR). Simultaneous sampling of correlated sensitivity and specificity parameters (e.g. with a Dirichlet distribution) would better account for uncertainty. However, the reported data in the literature lacked estimates of covariance or the absolute numbers of TP, TN, FP and FN results for each assay broken down by smear status.

For GeneXpert, we were able to obtain only one estimate of diagnostic accuracy for smear-negative patients. This estimate had a mean of 1.00, so we were unable to resample this value and retain this mean value in our PSA estimates. For this reason, we set the sensitivity as fixed and sampled the specificity directly from a beta distribution.

Unit costs for health-care resources

All costs were considered from a NHS perspective. Some items relate to one-off events, such as the cost of the rapid molecular test or the cost of the culture and DST. Others are time dependent, including the cost of medication, inpatient care (including isolation) and outpatient care. Where possible, unit costs were based on published national tariffs. Otherwise, estimates were derived from local information or from the literature.

Table 28 shows the unit costs for diagnostic tests. Costs for culture and rapid molecular assays were estimated, based on the costs of consumables, capital costs and overheads from the NMRL laboratory finance department. Capital costs were discounted over a period of 10 years and apportioned to each test, assuming that all equipment is used at full capacity. Labour costs of administering the assays were estimated based on the person-hours per test, estimated from a time-and-motion study at the NMRL (see Chapter 4). A transport cost of £6.50 per sample was included in the total cost of administering each assay, including culture. All costs have been inflated to 2011–12 levels. It has been assumed that the NMRL cost profile is indicative of the other reference laboratories in Newcastle and Birmingham. Regional variation in cost of labour and overheads has not been explored here, because of a lack of data.

Unit costs for medications (Table 29) were sourced from the British National Formulary (BNF),128 with one exception: costs for prothionamide are not currently listed in the BNF, and so a mean cost was taken from three hospitals. The unit costs of other health services are listed in Table 30.

Costs for general practice consultations were taken from the Personal Social Services Research Unit (PSSRU) estimates,129 and from the Department of Health reference costs for inpatient and outpatient care. 126 The costs of negative pressure isolation were uprated for inflation from published estimates,52 using the hospital and community services pay and price index. 129 Costs for standard isolation were estimated by the authors from the NHS reference costs for TB requiring non-elective inpatient stay. 126 The lower estimate of an excess bed-day was taken as the cost of a non-isolated inpatient-day and the average cost of a bed-day for pulmonary TB with complications was taken as an estimate of a day in isolation. The cost of hospital admissions was treated probabilistically. Details of the figures and distributions sampled are available in Table 30.

Resource utilisation

Parameters on health-care utilisation per patient were sourced from literature, local audit, and, in some cases, expert clinical advice (Tables 31–33).

Utility and quality-adjusted life-year loss

Estimates of parameters used to estimate QALY losses attributable to TB are shown in Table 34. Patients with TB were assumed to occupy three health states and attract a corresponding utility decrement while in that state: active TB pretreatment, active TB on treatment as inpatient, and active TB on treatment as an outpatient. The utility estimates applied are listed in Table 34. For each care pathway these utility estimates were integrated over the estimated time period spent in each health state to calculate a QALY loss attributable to TB.

Parameter estimates used to estimate QALY loss attributable to MDR treatment are listed in Table 35. Based on available data we were able to estimate the serious adverse event rate for MDR treatment only at an average of 4 months on treatment, with 42% experiencing ototoxicity. 118 On any pathway for which a patient is on MDR treatment for < 4 months, the adverse event rate was adjusted pro rata, based on their time on treatment. On pathways for which a patient is on treatment for > 4 months, it was assumed that they would attract this maximum risk of adverse effects.

Following Veenstra et al.,120 we assumed that, of those patients who experienced a MDR treatment-related hearing impairment, 50% would have a ‘mild’ impairment not requiring any treatment, 25% would have a ‘moderate’ impairment requiring a hearing aid and 25% would have a ‘severe’ impairment that would be appropriate for cochlear implantation. Utility losses attributable to moderate and severe levels of impairment were sourced from the literature. There is good evidence for the validity and responsiveness of the Health Utilities Index Mark 3 (HUI3) as a measure of health-related quality of life in people with hearing impairment. 134 HUI3 utility estimates were obtained from a sample of 609 hearing-impaired adults assessed at four UK audiology clinics between April 2000 and October 2002,131 and for 311 ‘postlingually deafened’ adults before and after they received a cochlear implant. 132 To estimate the utility loss associated with hearing impairment per se, we subtracted the reported HUI3 scores (with treatment) from a population norm, based on a large general population sample (n = 4048) in the USA. 135 We assumed a utility loss owing to mild (untreated) hearing impairment of 0.05, in the absence of an empirical estimate. Lifetime discounted QALY loss attributable to hearing loss was then estimated, based on a mean life expectancy for individuals at TB diagnosis.

Lifetime discounted costs of treatment for treatment-related hearing loss were also estimated. Costs for hearing aids included costs for assessment, purchase of devices, fitting, follow-up, replacement (assumed once per 5 years), and maintenance (one repair per year). Audiology costs were taken from the Department of Health National Tariff. 127 Lifetime costs for patients receiving a cochlear implant were based on an estimate for adults receiving a unilateral implant taken from a published Health Technology Assessment report,133 and uprated for inflation. 129 This estimate included the cost of the implant, assessment, the procedure, treatment for complications, replacement and maintenance.

Analysis

All analysis of the model was done for all three ethnic groups considered in this report. Time and resource constraints meant that not all possible sensitivity analyses could be performed.

Probabilistic sensitivity analysis

Probabilistic sensitivity analysis was performed to evaluate the impact on cost-effectiveness of diagnostic and treatment time delays, diagnosis costs and treatment costs, and associated QALYs. We ran the model simulation with 10,000 parameter sets and report the resulting incremental costs and incremental QALYs.

Deterministic sensitivity analysis

Parameters relating to diagnostic test performance were subject to deterministic sensitivity analysis, in which each parameter was varied individually across its range. Results are presented in tornado plots.

South Asians

There is a marginal effect of introduction of molecular testing on annual numbers of diagnoses (Figure 18) owing to a small reduction in transmission and an increase in FP results. In the INNO-LiPA and MTBDRplus scenarios there are increases in diagnoses owing to the latter effect dominating, whereas in the GeneXpert scenario the former dominates. Local molecular testing has a faster turnaround time than regional molecular testing, so the annual number of diagnoses in the local testing configuration is slightly lower than in the regional testing configuration. However, this effect is tiny, with < 1% fewer diagnoses annually.

FIGURE 18.

Effects of introducing molecular testing on annual numbers of TB diagnoses in the South Asian population. Annual numbers of TB diagnoses in South Asians following the introduction of different rapid molecular tests. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. Note that the vertical axis does not start at zero and the magnitude of the changes is, in fact, small.

There is a significant incremental net benefit (INB) in the GeneXpert scenario, MTBDRplus scenario and INNO-LiPA scenario (Table 42 and Figure 19). These molecular tests reduced net costs and increased QALYs as more individuals were appropriately treated for TB. As the options are cost-saving, results are presented as an INB rather than using incremental cost-effectiveness ratios (ICERs).

FIGURE 19.

Cost-effectiveness plane of different molecular testing interventions compared with current practice in South Asians, over a 10-year time horizon. This figure shows the cost-effectiveness plane resulting from comparing the effect of different rapid molecular tests for TB confirmation and DST with the current practice in South Asians. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. The solid diagonal line indicates the threshold of £30,000 per QALY and the dashed line £20,000 per QALY.

For all tests (GeneXpert scenario, MTBDRplus scenario and INNO-LiPA) the INB for local and regional testing are very similar (see Table 42 and Figure 19). GeneXpert and MTBDRplus scenarios have similar INBs, which are higher than the INNO-LiPA scenario. INNO-LiPA has a higher FP rate, leading to more inappropriate treatment of patients, increasing costs and reducing QALYs. It also has the lowest sensitivity for detection of smear-negative TB cases, leading to fewer TB diagnoses and treatment, further reducing the number of QALYs gained. When comparing over a time horizon of 20 years rather than 10 years, the INBs are greater but the overall picture is broadly similar, although MTBDRplus appears to be relatively more cost-effective than the other two rapid assays (Figure 20 and Table 43).

FIGURE 20.

Cost-effectiveness plane of different molecular testing interventions compared with current practice in South Asians, over a 20-year time horizon. This figure shows the cost-effectiveness plane resulting from comparing the effect of different rapid molecular tests for TB confirmation and DST with the current practice in South Asians. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. The solid diagonal line indicates the threshold of £30,000 per QALY and the dashed line £20,000 per QALY.

GeneXpert only

There is uncertainty in the GeneXpert scenario over the impact on both costs and, particularly, health, and hence there is uncertainty in cost-effectiveness (Figures 21 and 22). However, in all realisations of the PSA there is a net reduction in costs, and net increases in QALYs. When local and regional testing are compared (Figure 23), there are increases in costs in all realisations of the PSA and increases in the QALYs in most realisations when moving from local to regional testing.

FIGURE 21.

Cost-effectiveness plane of PSA samples of local GeneXpert-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing GeneXpert-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold of £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 22.

Cost-effectiveness plane of PSA samples of regional GeneXpert-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing GeneXpert-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert base TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 23.

Cost-effectiveness plane of PSA samples of regional vs. local GeneXpert-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing regionally based GeneXpert-based TB confirmation and DST in South Asians with locally based GeneXpert-based TB confirmation and DST in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert base TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold of £30,000 per QALY and the black dashed line £20,000 per QALY. Note the very small magnitude of the differences.

Figure 24 shows that the most influential test performance characteristics that affect the cost-effectiveness of the GeneXpert scenario compared with current practice are the sensitivity for detection of MDR infection in those in whom TB is detected (as FN results lead to inappropriate treatment for DS infection), and specificity for detection of MDR infection in those in whom TB is detected (as FP results lead to inappropriate treatment for MDR infection, which incurs cost and harms health). Sensitivity and specificity for detection of TB infection and time to obtain culture-based DST results are much less influential.

FIGURE 24.

Tornado plot of the influence on INB in South Asians of GeneXpert-based testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of GeneXpert-based TB confirmation and DST in South Asians compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.

GeneXpert with INNO-LiPA test

In the INNO-LiPA scenario there is considerable uncertainty over the impact on both costs and health. However, in all realisations of the PSA for both local and regional testing there is a reduction in costs, and an increase in QALYs (Figure 25 and 26).

FIGURE 25.

Cost-effectiveness plane of PSA samples of local GeneXpert with INNO-LiPA-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing GeneXpert with INNO-LiPA test-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert with INNO-LiPA test-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 26.

Cost-effectiveness plane of PSA samples of regional GeneXpert with INNO-LiPA-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing GeneXpert with INNO-LiPA test-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert with INNO-LiPA test-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

Figure 27 shows that the most influential test performance characteristics that affect the cost-effectiveness of the GeneXpert with INNO-LiPA scenario compared with current practice are the sensitivity for detection of MDR infection in those in whom TB is detected (as FN results lead to inappropriate treatment for DS infection), and specificity for detection of MDR infection in those in whom TB is detected (as FP results lead to inappropriate treatment for MDR infection, which incurs cost and harms health). Sensitivity and specificity for detection of TB infection and time to obtain culture-based DST results are much less influential.

FIGURE 27.

Tornado plot of the influence on INB in South Asians of GeneXpert with INNO-LiPA-based testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of GeneXpert with INNO-LiPA-based TB confirmation and DST in South Asians compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.

MTBDRplus test

In the MTBDRplus scenario (Figures 28 and 29), there is a wide range of uncertainty with regard to the magnitude of the impact on health and costs, but in all realisations of the PSA there is a net reduction in costs and increases in QALYs.

FIGURE 28.

Cost-effectiveness plane of PSA samples of local MTBDRplus-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing MTBDRplus-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 29.

Cost-effectiveness plane of PSA samples of regional MTBDRplus-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing MTBDRplus-based TB confirmation and DST in South Asians with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

When local and regional testing are compared (Figure 30), there are modest increases in costs in almost all of the realisations of the PSA, and modest increases in QALYs in most PSA samples, but with modest decreases also occurring in some realisations.

FIGURE 30.

Cost-effectiveness plane of PSA samples of regional vs. local MTBDRplus-based testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing regionally based MTBDRplus-based TB confirmation and DST in South Asians with locally based MTBDRplus-based TB confirmation and DST in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY. Note the very small magnitude of the differences.

Figure 31 shows that the most influential test performance characteristics that affect the cost-effectiveness of the MTBDRplus scenario compared with current practice are the specificity for detection of TB infection (as FP results lead to inappropriate treatment, which incurs cost and harms health), sensitivity for detection of TB infection (as FN results delay treatment), and specificity and sensitivity for detection of MDR infection in those infected with TB. Figure 18 shows how moving from current practice to the GeneXpert scenario results in a reduction in annual TB diagnoses, whereas moving from current practice to the MTBDRplus scenario results in an increase in diagnoses; the key difference is the specificity of the tests.

FIGURE 31.

Tornado plot of the influence on INB in South Asians of MTBDRplus-based testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of MTBDRplus-based TB confirmation and DST in South Asians compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.

GeneXpert only compared with MTBDRplus test

Comparing the cost-effectiveness of regionally based testing in the GeneXpert and MTBDRplus scenarios finds that there is considerable uncertainty with regard to both costs and health benefits (Figure 32). In all realisations of the PSA there are reductions in costs, but in a large proportion of samples there are also reductions in QALYs when moving from the MTBDRplus scenario to the GeneXpert scenario.

FIGURE 32.

Cost-effectiveness plane of PSA samples of regional GeneXpert vs. regional MTBDRplus molecular testing compared with current practice over a 10-year time horizon in South Asians. This figure shows the ICER pairs resulting from comparing regionally based GeneXpert-based TB confirmation and DST in South Asians with regionally based MTBDRplus-based TB confirmation and DST-based TB confirmation and DST in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

Black Africans

As with South Asians, there is a marginal effect of introduction of molecular testing on annual numbers of diagnoses (Figure 33) owing to a small reduction in transmission and an increase in FP results. In the INNO-LiPA and MTBDRplus scenarios there are increases in diagnoses as a result of the latter effect dominating, whereas in the GeneXpert scenario the former dominates. Local molecular testing has a faster turnaround time than regional molecular testing, so the annual number of diagnoses in the local testing configuration is slightly lower than in the regional testing configuration. However, this effect is tiny, with < 1% fewer diagnoses annually.

FIGURE 33.

Effects of introducing molecular testing on annual numbers of TB diagnoses in the Black African population. Annual numbers of TB diagnoses in Black Africans following the introduction of different rapid molecular tests. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. Note that the vertical axis does not start at zero and the magnitude of the changes is, in fact, small.

There are significant INBs in the GeneXpert, MTBDRplus and INNO-LiPA scenarios (Table 44 and Figure 34). These molecular tests reduce net costs and achieve an associated QALY gain as more individuals are appropriately treated for TB. As the options are cost saving, results are presented as INB rather than using ICERs.

For all tests (GeneXpert scenario, MTBDRplus scenario and INNO-LiPA) the INBs for local and regional testing are very similar (see Table 44 and Figure 35). The GeneXpert scenario has the highest INB, with the MTBDRplus and INNO-LiPA scenarios having INBs that are similar to each other.

When comparing over a time horizon of 20 years, rather than 10 years, the INBs are greater but the overall picture is broadly similar (see Figure 35 and Table 45).

FIGURE 34.

Cost-effectiveness plane of different molecular testing interventions compared with current practice in Black Africans over a 10-year time horizon. This figure shows the cost-effectiveness plane resulting from comparing the effect of different rapid molecular tests for TB confirmation and DST with the current practice in Black Africans. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. The solid diagonal line indicates the threshold £30,000 per QALY and the dashed line £20,000 per QALY.

FIGURE 35.

Cost-effectiveness plane of different molecular testing interventions compared with current practice in Black Africans over a 20-year time horizon. This figure shows the cost-effectiveness plane resulting from comparing the effect of different rapid molecular tests for TB confirmation and DST with the current practice in Black Africans. Three different molecular tests are considered (GeneXpert only, GeneXpert combined with INNO-LiPA line-probe assay and MTBDRplus) in two locations: either local to the hospital or in a regional laboratory. The solid diagonal line indicates the threshold £30,000 per QALY and the dashed line £20,000 per QALY.

GeneXpert only

As with the South Asian population, there is uncertainty in the GeneXpert scenario over the impact on both costs and, particularly, health, and hence there is uncertainty in cost-effectiveness (Figures 36 and 37). However, in all realisations of the PSA there is a net reduction in costs, and net increases in QALYs. When local and regional testing are compared (Figure 38), there are increases in costs and reductions in QALYs in all realisations of the PSA, although the magnitudes are small particularly for costs.

FIGURE 36.

Cost-effectiveness plane of PSA samples of local GeneXpert-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing GeneXpert-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert base TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 37.

Cost-effectiveness plane of PSA samples of regional GeneXpert-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing GeneXpert-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green dashed line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 38.

Cost-effectiveness plane of PSA samples of regional vs. local GeneXpert-based testing compared with current practice over a 10-year time horizon in Black Africans. This figure shows the ICER pairs resulting from comparing regionally based GeneXpert-based TB confirmation and DST in Black Africans with locally based GeneXpert-based TB confirmation and DST in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY. Note the very small magnitude of the differences.

Figure 39 shows that the most influential test performance characteristics that affect the cost-effectiveness of the GeneXpert scenario compared with current practice are the same as for South Asians: sensitivity for detection of MDR infection in those in whom TB is detected, and specificity of detection of MDR infection in those in whom TB is detected.

FIGURE 39.

Tornado plot of the influence on INB in Black Africans of GeneXpert-based testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of GeneXpert-based TB confirmation and DST in Black Africans compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.

GeneXpert with INNO-LiPA test

In the INNO-LiPA scenario there is considerable uncertainty over the impact on both costs and health. However, in all realisations of the PSA for both local and regional testing there are reductions in costs, and in the large majority of realisations there are increases in QALYs (Figures 40 and 41).

FIGURE 40.

Cost-effectiveness plane of PSA samples of local GeneXpert with INNO-LiPA-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing GeneXpert with INNO-LiPA test-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert with INNO-LiPA test-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 41.

Cost-effectiveness plane of PSA samples of regional GeneXpert with INNO-LiPA-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing GeneXpert with INNO-LiPA test-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of GeneXpert with INNO-LiPA test-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

Figure 42 shows that the most influential test performance characteristics that affect the cost-effectiveness of the GeneXpert scenario compared with current practice are the same as for South Asians: sensitivity for detection of MDR infection in those in whom TB is detected, and specificity of detection of MDR infection in those in whom TB is detected.

FIGURE 42.

Tornado plot of the influence on INB in Black Africans of GeneXpert with INNO-LiPA-based testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of GeneXpert with INNO-LiPA-based TB confirmation and DST in Black Africans compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.

MTBDRplus test

As with South Asians, in the MTBDRplus scenario (Figure 43 and 44), there is a wide range of uncertainty with regard to the impact on health and costs. In all realisations of the PSA there is a net reduction in costs, but the effect on health is uncertain, with the PSA realisations covering a large range of incremental QALYs, both negative and positive.

FIGURE 43.

Cost-effectiveness plane of PSA samples of local MTBDRplus-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing MTBDRplus-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

FIGURE 44.

Cost-effectiveness plane of PSA samples of regional MTBDRplus-based testing compared with current practice in Black Africans over a 10-year time horizon. This figure shows the ICER pairs resulting from comparing MTBDRplus-based TB confirmation and DST in Black Africans with current practice in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY.

When local and regional testing are compared (Figure 45), there are reductions in QALYs and increases in costs in almost all of the realisations of the PSA when moving from local to regional testing.

FIGURE 45.

Cost-effectiveness plane of PSA samples of regional vs. local MTBDRplus-based testing compared with current practice over a 10-year time horizon in Black Africans. This figure shows the ICER pairs resulting from comparing regionally based MTBDRplus-based TB confirmation and DST in Black Africans with locally based MTBDRplus-based TB confirmation and DST in PSA. The stochastic parameters included in the samples were those associated with TB diagnosis time delay, costs and QALYs on the cost-effectiveness of MTBDRplus-based TB confirmation and DST compared with current practice. Parameters were varied in both current practice and the intervention scenario. The green line indicates the threshold £30,000 per QALY and the black dashed line £20,000 per QALY. Note the very small magnitude of the differences.

Figure 46 shows that the most influential test performance characteristic that affects the cost-effectiveness of the MTBDRplus scenario compared with current practice is the sensitivity for detection of TB infection in smear-positive patients. Figure 33 shows how moving from current practice to the GeneXpert scenario results in a reduction in annual TB diagnoses, whereas moving from current practice to the MTBDRplus scenario results in an increase in diagnoses; the key difference is the specificity of the tests.

FIGURE 46.

Tornado plot of the influence on INB in Black Africans of molecular testing compared with current practice, over a 10-year time horizon. This figure shows the effects of varying different test performance parameters on the INB of MTBDRplus-based TB confirmation and DST in Black Africans compared with current practice, if a QALY is valued at £20,000. Individually, parameters are varied from their minimum (green bar) to maximum (black bar) values.