24
views
0
recommends
+1 Recommend
1 collections
8
shares
• Record: found
• Abstract: found
• Article: found
Is Open Access

# Modeling of Novel Diagnostic Strategies for Active Tuberculosis – A Systematic Review: Current Practices and Recommendations

research-article
PLoS ONE
Public Library of Science

Bookmark
There is no author summary for this article yet. Authors can add summaries to their articles on ScienceOpen to make them more accessible to a non-specialist audience.

### Abstract

##### Introduction

The field of diagnostics for active tuberculosis (TB) is rapidly developing. TB diagnostic modeling can help to inform policy makers and support complicated decisions on diagnostic strategy, with important budgetary implications. Demand for TB diagnostic modeling is likely to increase, and an evaluation of current practice is important. We aimed to systematically review all studies employing mathematical modeling to evaluate cost-effectiveness or epidemiological impact of novel diagnostic strategies for active TB.

##### Methods

Pubmed, personal libraries and reference lists were searched to identify eligible papers. We extracted data on a wide variety of model structure, parameter choices, sensitivity analyses and study conclusions, which were discussed during a meeting of content experts.

##### Results & Discussion

From 5619 records a total of 36 papers were included in the analysis. Sixteen papers included population impact/transmission modeling, 5 were health systems models, and 24 included estimates of cost-effectiveness. Transmission and health systems models included specific structure to explore the importance of the diagnostic pathway (n = 4), key determinants of diagnostic delay (n = 5), operational context (n = 5), and the pre-diagnostic infectious period (n = 1). The majority of models implemented sensitivity analysis, although only 18 studies described multi-way sensitivity analysis of more than 2 parameters simultaneously. Among the models used to make cost-effectiveness estimates, most frequent diagnostic assays studied included Xpert MTB/RIF (n = 7), and alternative nucleic acid amplification tests (NAATs) (n = 4). Most (n = 16) of the cost-effectiveness models compared new assays to an existing baseline and generated an incremental cost-effectiveness ratio (ICER).

##### Conclusion

Although models have addressed a small number of important issues, many decisions regarding implementation of TB diagnostics are being made without the full benefits of insight from mathematical models. Further models are needed that address a wider array of diagnostic and epidemiological settings, that explore the inherent uncertainty of models and that include additional epidemiological data on transmission implications of false-negative diagnosis and the pre-diagnostic period.

### Most cited references40

• Record: found
• Abstract: found
• Article: found
Is Open Access

### Rapid Diagnosis of Tuberculosis with the Xpert MTB/RIF Assay in High Burden Countries: A Cost-Effectiveness Analysis

(2011)
Introduction Tuberculosis (TB) control in developing countries is hampered by the inadequate care that can be delivered on the basis of diagnosis by microscopy alone—an issue that is acute in populations with a high prevalence of HIV and/or multidrug resistant (MDR)-TB. It is estimated that 1.7 million people died from TB in 2009, many of them remaining undiagnosed [1]. The Xpert MTB/RIF assay (referred to as Xpert in this article), is a real-time PCR assay that is a design-locked, all-within-cartridge test, and that has demonstrated high performance and could be deployed in a range of low- and middle-income settings [2],[3]. It has recently been endorsed by the World Health Organization (WHO) as a promising new rapid diagnostic technology that has the potential for large-scale roll-out (www.who.int/tb/laboratory). Xpert is commercially produced and sold at concessional prices. However, because the price is considerably higher than that of smear microscopy, there is a concern among TB program managers and policy makers that Xpert may not be cost-effective in low- and middle-income settings. There is little previous research into the cost-effectiveness of TB diagnostics. A study considering a hypothetical TB diagnostic found that cost-effectiveness would be maximized by high-specificity, low-cost tests deployed in regions with poor infrastructure [4]. Other studies have examined the cost-effectiveness of culture, PCR, and novel methods for drug susceptibility testing such as line-probe assays (LPA). These studies all found that these diagnostic tests are potentially cost-effective [5]–[7]. However, because of their technical requirements, mycobacterial culture, PCR, and LPA can only be deployed in specialised laboratories. We present the first (to our knowledge) economic evaluation of the Xpert rapid test for TB. [2]. Methods The aim of this study was to assess whether Xpert is likely to result in an improvement of the cost-effectiveness of TB care in low- and middle-income settings. We did this by estimating the impact of Xpert on the costs and cost-effectiveness of TB care in three countries, using decision analytic modelling, comparing the introduction of Xpert to a base case of sputum microscopy and clinical diagnosis. The model's primary outcome measure is the cost per disability adjusted life year (DALY) averted. Our model followed a cohort of 10,000 individuals suspected of having TB through the diagnostic and treatment pathway, estimating costs and health gains. In the diagnostic pathway, the TB cases among the individuals with suspected TB were either diagnosed as having TB or not, depending on the test sensitivities in the pathway. Similarly, individuals with suspected TB who were not TB cases may have been diagnosed as having TB, depending on the pathway's test specificities. A diagnosis of TB was followed by treatment. Individuals with suspected TB completed the pathway when they were either cured, failed treatment, died, or, if they had no TB from the start, remained without TB. Three different diagnostic scenarios are compared (Figure 1). The base case is defined as two sputum microscopy examinations followed, in smear-negative individuals with suspected TB, by clinical diagnosis that might include chest X-ray and antibiotic trial [8]. The inclusion of an antibiotic trial (empirical treatment with one or more broad-spectrum antibiotics to exclude other infectious causes of pulmonary disease) is no longer part of the WHO diagnostic strategy for HIV-infected patients. However, in the clinics participating in the demonstration study from which the diagnostic performance parameters were sourced [2], an antibiotic trial was still commonly provided during the diagnostic process as an adjunct to the treatment of smear-negative individuals with suspected TB. Antibiotic trial was therefore included in the base case; the model assumed that for each country the use of antibiotic trial and chest X-ray was proportional to the observed use in the demonstration study clinics. In comparison, two alternative pathways involving Xpert were considered: (1) two smear examinations, if negative followed by Xpert on a single sputum specimen (“in addition to”); (2) Xpert instead of smear examination: single sputum specimen tested by Xpert for all individuals with suspected TB (“replacement of”). 10.1371/journal.pmed.1001120.g001 Figure 1 Simplified schematic of model. Each scenario included drug resistance testing of previously treated patients [9], either by conventional drug susceptibility testing (DST) or Xpert. All patients diagnosed with TB were treated using the standard WHO-recommended regimens. Patients awaiting DST results were started on first-line treatment (isoniazid [H], rifampicin [R], pyrazimamide [Z], and ethambutol [E] for 2 mo followed by HR for 4 mo for new patients, and HRZE for 3 mo with streptomycin added during the first 2 mo followed by HRE for 5 mo for patients with a history of previous TB treatment) and switched to second-line treatment when a DST result of rifampicin resistance became available. The second-line treatment regimens differed between the countries but commonly included a fluoroquinolone and an aminoglycoside (kanamycin, amikacin) or capreomycin in addition to one or more first-line drugs and ethionamode, cycloserine, and/or 4-aminosalicylic acid (PAS). If Xpert identified rifampicin resistance, this was confirmed by conventional DST or LPA as practice in the respective countries. LPA, used as a screening test on smear-positive sputum samples in South Africa, detects rifampicin resistance within 24 h by molecular methods. While awaiting this result, the patient was started on second-line treatment, but then switched to first-line treatment if resistance to rifampicin was not confirmed. TB cases that remained undiagnosed were assumed to return to the clinic after 3 mo, with 10% of undiagnosed cases becoming smear-positive within that time. Key model input parameters are shown in Table 1 and further details can be found in Text S1. The model was parameterised for three settings: India (low HIV prevalence, low MDR prevalence), Uganda (high HIV prevalence, low MDR prevalence), and South Africa (high HIV prevalence, high MDR prevalence). In each cohort, TB cases were characterized as: (1) new or previously treated, (2) HIV-negative or HIV-positive, and (3) MDR or drug susceptible. These proportions were sourced from country reports [1],[10],[11]. 10.1371/journal.pmed.1001120.t001 Table 1 Model inputs: cohort composition and diagnostic parameters, by country. Cohort Proportions and Diagnostic Parameters India South Africa Uganda Distribution Source Cohort proportions Smear-positive TB 0.1 0.1 0.1 Beta Model assumption Smear-positive TB among pulmonary TB cases, HIV-negative 0.723 0.723 0.723 Beta Demonstration study, all sites [2] Smear-positive TB among pulmonary TB cases, HIV-positive 0.446 0.446 0.446 Beta Demonstration study, all sites [2] Previous TB treatment among pulmonary TB cases 0.192 0.168 0.073 Beta WHO [1] Multidrug resistance, among new TB cases 0.023 0.066 0.011 Beta WHO [10] Multidrug resistance, among previously treated TB cases 0.172 0.245 0.117 Beta WHO [10], survey [11] HIV infection, among pulmonary TB cases 0.006 0.588 0.593 Beta WHO [1] Diagnostic parameters Sensitivity for diagnosing pulmonary TB (SEM) Xpert MTB RIF, smear-positive TB cases 0.983 (0.005) 0.983 (0.005) 0.983 (0.005) Beta Demonstration study, all sites [2] Xpert MTB RIF, smear-negative TB cases, HIV-negative 0.793 (0.025) 0.793 (0.025) 0.793 (0.025) Beta Demonstration study, all sites [2] Xpert MTB RIF, smear-negative cases, HIV-positive 0.718 (0.040) 0.718 (0.040) 0.718 (0.040) Beta Demonstration study, all sites [2] Smear microscopy (two slides), HIV-positive 0.723 (0.015) 0.723 (0.015) 0.723 (0.015) Beta Demonstration study, all sites [2] Smear microscopy (two slides), HIV-negative 0.446 (0.036) 0.446 (0.036) 0.446 (0.036) Beta Demonstration study, all sites [2] Mycobacterial culture 1 (—) 1 (—) 1 (—) Model assumption Clinical diagnosis 0.160 (0.073) 0.209 (0.039) 0.444 (0.096) Beta Demonstration study [2] Proportion culture-positive individuals with suspected TB who had chest X-ray 0.032 0.262 0.867 Beta Demonstration study [2] Proportion culture-positive individuals with suspected TB who had antibiotic trial 1 0.051 0.241 Beta Demonstration study [2] Specificity for diagnosing pulmonary TB (SEM) Xpert MTB RIF 0.990 (0.002) 0.990 (0.002) 0.990 (0.002) Beta Demonstration study, all sites [2] Smear microscopy (two slides) 1 (—) 1 (—) 1 (—) Model assumption Mycobacterial culture 1 (—) 1 (—) 1 (—) Model assumption Clinical diagnosis 0.942 (0.009) 0.953 (0.007) 0.869 (0.030) Beta Demonstration study [2] Proportion culture-negative individuals with suspected TB who had chest X-ray 0.037 0.059 0.790 Beta Demonstration study [2] Proportion culture-negative individuals with suspected TB who had antibiotic trial 1 0.009 0.887 Beta Demonstration study [2] Sensitivity for detecting rifampicin-resistance (SEM) Xpert MTB RIF 0.944 (0.015) 0.944 (0.015) 0.944 (0.015) Beta Demonstration study, all sites [2] Conventional drug susceptibility testing 1 (—) — 1 (—) — Model assumption Line-probe assay — 1 (—) — — Model assumption Specificity for detecting rifampicin-resistance (SEM) Xpert MTB RIF 0.983 (0.005) 0.983 (0.005) 0.983 (0.005) Beta Demonstration study, all sites [2] Drug susceptibility testing 1 (—) — 1 (—) — Model assumption Line-probe assay — 1 (—) — — Model assumption Cost parameters US$2010 (min, max) First-line category 1 treatment: total 227 (103, 352) 454(306, 602) 185 (146, 224) Triangular WHO-CHOICE [13], literature review [14]–[19] First-line category 2 treatment: total 352 (159, 546) 998 (451, 1546) 287 (130, 445) Triangular WHO-CHOICE [13], literature review [14]–[19] Cotrimoxazol preventive treatment: 1 mo 4, 50 10, 53 3, 25 Triangular WHO-CHOICE [13] Treatment of bacterial infection 3, 66 9, 70 2, 41 Triangular WHO-CHOICE [13] Chest X-ray 11 (9, 13) 16 (14, 18) 3 (2.6, 3.7) Triangular WHO-CHOICE [13], literature review [14]–[19] Second-line treatment total 2,256 (1,463, 3,050) 3,492 (2,068, 4,917) 1,759 (1,285, 2,233) Triangular WHO-CHOICE [13], literature review [14]–[19] DALY parameters: DALYs averted (min, max) HIV positive, sputum smear-negative 9.38 (8.62, 10.39) 10.71 (9.85, 11.90) 11.58 (10.63, 12.90) Triangular See Text S1 HIV negative, sputum smear-negative 13.18 (12.32, 13.96) 13.83 (12.83, 14.72) 18.65 (17.56, 19.61) Triangular See Text S1 HIV positive, sputum smear-positive 9.67 (8.62, 10.39) 11.03 (9.85, 11.90) 11.92 (10.63, 12.90) Triangular See Text S1 HIV negative, sputum smear-positive 16.43 (16.02, 16.79) 17.52 (17.05, 17.93) 22.63 (22.13, 23.07) Triangular See Text S1 The distribution column indicates which probability distribution was specified for each parameter in the Monte Carlo simulations. For triangular distributions the mode, upper and lower limit are given. All beta distributions have boundaries (0, 1). SEM, standard error of the mean. Diagnostic test performance data were sourced from a demonstration study of Xpert use in nine facilities [2]. Sensitivity and specificity parameters for all diagnostic tests and procedures were calculated taking sputum culture as the reference standard. The sensitivity and specificity of Xpert and sputum microscopy (light-emitting diode [LED]) fluorescence microscopy) was based on weighted averages across the sites. Since clinical diagnostic practice of smear negatives in the base case varied considerably between sites, site-specific data were used to estimate performance of the clinical TB diagnosis. A patient was defined as having clinically diagnosed TB if microscopy was negative, but the onset of treatment preceded the availability of the culture result. Estimates of the economic costs of each pathway were made from a health service perspective. All costs were estimated using the ingredient costing approach. This approach identifies all the inputs (and their quantities) required to perform a test or deliver treatment and then values them to arrive at a cost per test/person treated. Diagnostic costs were collected at the demonstration sites. These costs included all building, overhead, staff, equipment and consumables, quality control and maintenance, and calibration inputs. The resource use associated with each test was measured through observations of practice, a review of financial reporting, and interviews with staff in the Xpert demonstration sites. Resource use measurement took into account the allocation of fixed resources between Xpert and any other uses. Estimates of device and test prices, calibration, and training costs were obtained from suppliers. Costs for treatment were estimated using drugs costs from the Global Drug Facility and the MSH International Price Tracker, and unit costs for outpatient visits and hospitalisation sourced from WHO-CHOICE [12]. A review of previous costing studies was used to validate these estimates [13]–[18]. As our constructed estimates were higher than those found in our review, we took the mid-point between our estimate and the lowest estimate found in the literature. All local costs were converted using the average exchange rate for 2010 (imf.statex.imf.org). Where relevant, costs were annualised using a standard discount rate of 3% [19]. All costs are reported in 2010 US$. Treatment outcome probabilities were taken from published meta-analyses of clinical trials, cohort studies, and systematic reviews [20]–[28]. DALYs averted from patients being cured were estimated using the standard formula [19]. Further details can be found in Text S1. Since the Xpert scenarios are most likely to be more costly and more effective than the base case, an incremental cost effectiveness ratio (ICER) was calculated to describe the additional cost for any additional DALYs averted by Xpert over the base case. This ICER was then compared to WHO's suggested country-specific willingness to pay (WTP) threshold, defined as the cost per DALY averted of each country's per capita GDP (US$1,134 for India, US$5,786 for South Africa, and US$490 for Uganda in 2010). If the ICER is below this threshold the intervention is considered cost-effective. In the demonstration study from which our parameter estimates were sourced [2], the probability that an individual with suspected TB was a true TB case varied considerably by location; the proportion with smear-positive TB being 8.9% in India, 14.3% in South Africa, and 32.4% in Uganda. This variation probably reflects the local patterns of (self-) referral, in particular for the extremely high proportion of TB cases among the individuals with suspected TB in Uganda. Therefore to enable generalizability, we assumed a 10% proportion of smear-positive TB in individuals with suspected TB for all three countries as our point estimate with a range of 2.5% to 25% in our uncertainty and sensitivity analyses [29]. A large number of one- and two-way sensitivity analyses were conducted to assess the robustness of our model. These analyses examine the robustness of our results when one or two parameters are varied between the outer limits of their confidence intervals. We examined the sensitivity of our results to the probability that a suspect has TB or MDR-TB or has been previously treated. We examined the impact of varying treatment costs on our results. We tested for different prices of Xpert cartridge. We examined the impact of varying the proportion of individuals with suspected TB who get chest X-ray in addition to Xpert, as physicians may continue clinical diagnosis for smear-negative TB. Similarly we examined the impact of assuming that all HIV-infected individuals with suspected TB who have negative Xpert undergo the clinical diagnosis procedure, with costs based on site-specific use of chest X-rays and antibiotics, and sensitivity and specificity based on site-specific diagnostic performance of clinical diagnosis. We assessed the sensitivity of our results to the performance of the base case in three ways: (1) assuming one instead of two smears; (2) by varying the sensitivity of smear examination; and (3) by replacing the site-specific performance estimates for clinical diagnosis with estimates averaged across the three sites. Recognising that the performance of clinical diagnosis is a trade-off between sensitivity and specificity, we varied the sensitivity and specificity in opposite directions across a plausible range of values. As physicians in the demonstration study were aware that they would receive the results of sputum culture of all individuals with suspected TB, we tested for the effect of deferring treatment decisions until the availability of culture results. For each site culture was costed and assessed on the basis of current practice. We did not include a sensitivity analysis of the use of alternatives to culture such as microscopic observation drug susceptibility test (MODS) [30], as this was not practiced on site, and we found no good source of costing data. We examined the effect of reprogramming Xpert so that no resistance result is obtained. In addition, we conducted a probabilistic sensitivity analysis (Monte Carlo simulation) to explore the effect of uncertainty across our model parameters. This analysis randomly sampled each parameter in our model simultaneously from their probability distribution (Table 1; Text S1), and repeated this 10,000 times to generate confidence intervals around our estimates of incremental cost per DALY averted. The model and the analyses were constructed using TreeAge software. Percentage ranges in the text reflect ranges across countries unless stated otherwise. The demonstration study was endorsed by national TB programmes of participating countries and approved by nine governing institutional review boards (IRBs). The requirement to obtain individual informed consent was waived. The costing and cost-effectiveness assessments were outlined in the study protocol reviewed by the IRBs. Results The cost for the Xpert test (including all costs, such as the cartridge, equipment, salaries) ranges from US$22.63 in India to US$27.55 in Uganda, at an Xpert cartridge price of US$19.40 (including a 25% mark-up for transportation) and US$17,000 per four-module instrument (Tables 2 and 3) [2]. This cost falls to as low as US$14.93 with volume-driven price reductions. As FIND has negotiated a fixed price for Xpert, the difference in costs between sites is primarily determined by the intensity of use of the four-module instrument. Other factors also influence costs, but to a lesser extent; these include local wage levels and the room space used. A single sputum smear examination costs between US$1.13 and US$1.63. Unit costs for culture (Löwenstein–Jensen [LJ] or mycobacteria growth indicator tube [MGIT]) range from US$13.56 to US$18.95. Unit costs for tests that diagnose MDR-TB (where relevant for all first-line drugs) range from US$20.23 for LPA only to US$44.88 for MGIT and LPA. 10.1371/journal.pmed.1001120.t002 Table 2 Cost of diagnostic tests at the study sites (2010 US$). Diagnostic Test Type of Laboratory Costs per Test (2010 US$) India South Africa Uganda AFB Smear (one smear) Peripheral/hospital 1.13 1.58 1.63 Xpert (current pricing) US$19.4 including transport Peripheral/hospital 22.63 25.90 27.55 Xpert (volume>1.5 million/y) US$15.5 including transport Peripheral/hospital 18.73 22.00 23.61 Xpert (volume>3.0 million/y) US$11.7 including transport Peripheral/hospital 14.93 18.20 19.85 Culture (LJ) Reference 13.56 — 15.45 Culture (MGIT) Reference — 15.24 18.95 Culture + DST (LJ) Reference 22.33 — 23.98 Culture + DST (MGIT) Reference — 41.17 44.88 DST (MGIT + LPA) Reference — 33.01 38.82 DST (LPA), on sputum Reference — 20.23 21.84 10.1371/journal.pmed.1001120.t003 Table 3 Cost of Xpert (current pricing) by input type (2010 US$). Input Type Costs per Test (2010 US$) India South Africa Uganda Overhead 0.18 0.88 0.40 Building space 0.02 0.08 0.12 Equipment 2.84 3.50 7.00 Staff 0.11 1.82 0.24 Reagents and chemicals 19.40 19.40 19.40 Consumables 0.07 0.22 0.38 Total 22.63 25.90 27.55 The use of Xpert substantially increases TB case finding in all three settings; from 72%–85% to 95%–99% of the TB suspect cohort (Table 4). When Xpert is deployed “as a replacement of” instead of “in addition to” smear microscopy, the number of TB cases detected is similar—while the number of MDR-TB cases detected increases substantially. When undiagnosed TB patients are assumed not to return for diagnosis, TB case detection increases from 62%–76% in the base case to 86%–94% in the Xpert scenarios. 10.1371/journal.pmed.1001120.t004 Table 4 Cohort, cases detected, total cohort costs, and costs per case detected. Country Scenario Cohort n Individuals among the Cohort Who Have TB Total TB Cases Detected Percent of TB Cases Detected Total MDR Cases Detected Percent of MDR Cases Detected Total Diagnostic Costs (2010 US$) Diagnostic Cost per TB Case Detected, Excluding MDR (US$2010) Additional Diagnostic Cost per MDR Case Detected (2010 US$) Treatment Costs (2010 US$) Treatment Costs Percent of Total Cohort India Base case Tuberculosis (MDR) 72 59 82 38 52 1,077 — — 89,223 19 Tuberculosis (no MDR) 1,318 1,079 82 — — 8,412 — — 268,122 59 No tuberculosis 8,611 — — — — 46,106 — — 100,759 22 Total 10,000 1,138 82 38 — 55,595 49 165 458,103 100 In addition to smear Tuberculosis (MDR) 72 71 99 49 68 2,335 — — 115,932 25 Tuberculosis (no MDR) 1,318 1,300 99 — — 13,831 — — 325,381 70 No tuberculosis 8,611 — — — — 184,298 — — 22,414 5 Total 10,000 1,371 99 49 200,464 146 116 463,727 100 Replacement of smear Tuberculosis (MDR) 72 71 99 67 93 3,038 — — 151,603 30 Tuberculosis (no MDR) 1,318 1,298 99 — — 28,986 — — 328,669 65 No tuberculosis 8,611 — — — — 174,538 — — 22,414 4 Total 10,000 1,369 99 67 206,562 151 24 502,687 100 South Africa Base case Tuberculosis (MDR) 184 131 72 56 31 2,345 — — 230,989 22 Tuberculosis (no MDR) 1,729 1,237 72 — — 13,772 — — 659,365 63 No tuberculosis 8,087 — — — — 22,014 — — 156,213 15 Total 10,000 1,368 72 56 38,131 28 86 1,046,567 100 In addition to smear Tuberculosis (MDR) 184 175 95 112 61 7,131 — — 423,146 31 Tuberculosis (no MDR) 1,729 1,649 95 — — 30,341 — — 882,010 65 No tuberculosis 8,087 — — — — 205,858 — — 45,788 3 Total 10,000 1,824 95 112 243,331 133 57 1,350,945 100 Replacement of smear Tuberculosis (MDR) 184 175 95 165 90 9,504 — — 583,064 39 Tuberculosis (no MDR) 1,729 1,645 95 — — 46,866 — — 880,190 58 No tuberculosis 8,087 — — — — 193,053 — — 45,788 3 Total 10,000 1,820 95 165 249,423 137 30 1,509,043 100 Uganda Base case Tuberculosis (MDR) 36 30 85 14 38 499 — — 26,422 5 Tuberculosis (no MDR) 1,882 1,594 85 — — 11,282 — — 282,928 59 No tuberculosis 8,082 — — — — 51,565 — — 171,803 36 Total 10,000 1,625 85 14 63,345 39 163 481,154 100 In addition to smear Tuberculosis (MDR) 36 34 95 22 63 1,392 — — 41,123 11 Tuberculosis (no MDR) 1,882 1,794 95 — — 34,694 — — 320,685 85 No tuberculosis 8,082 — — — — 230,369 — — 14,908 4 Total 10,000 1,828 95 22 266,455 146 124 376,717 100 Replacement of smear Tuberculosis (MDR) 36 34 95 32 90 1,849 — — 56,488 14 Tuberculosis (no MDR) 1,882 1,790 95 — — 57,204 — — 322,502 82 No tuberculosis 8,082 — — — — 217,185 — — 14,908 4 Total 10,000 1,824 95 32 276,238 151 27 393,899 100 The diagnostic cost (including the costs of testing all individuals with suspected TB) per TB case detected is US$28–US$49 for the base case and increases significantly to US$133–US$146 and US$137–US$151 when Xpert is used “in addition to” and “as a replacement of” smear microscopy, respectively, depending on the setting (Table 4). The resulting change in treatment costs is more moderate, due to a reduction in the numbers of false positives in the base case from clinical diagnosis. For example, in India, the percentage of treatment costs spent on false-positive diagnoses falls from 22% to 4% when Xpert is used “as a replacement of” smear microscopy in comparison to the base case. ICERs for each Xpert scenario are presented in Table 5. The mean ICER for using Xpert “in addition to” smear microscopy compared to the base case ranges from US$41 to US$110 per DALY averted depending on the setting. The mean ICER for using Xpert “as a replacement of” smear microscopy ranges from US$52 to US$138 per DALY averted. The mean ICER for using Xpert as “a replacement of” smear microscopy compared to using Xpert “in addition to” smear microscopy ranges between US$343 and US$650. This higher ICER is due to the fact that the effectiveness gain from using Xpert as “replacement of smear microscopy” is derived from additional MDR-TB cases detected, and the cost-effectiveness of treating MDR-TB is lower than that for drug-susceptible TB. All the ICERs found are well below the WTP threshold. 10.1371/journal.pmed.1001120.t005 Table 5 Cost per DALY (US$ 2010). Country Scenario Total Cost Total DALYS Cost per DALY ICER Compared to Base Case, Mean Monte Carlo Simulation ICER, Median (2.5, 97.5) ICER Compared to in Addition to, Mean Monte Carlo Simulation ICER, Median (2.5, 97.5) India Base case 513,698 17,133 30 — — — — In addition to smear 664,191 19,887 33 55 54 (40, 70) — — Replacement of smear 709,248 20,019 35 68 68 (51, 87) 343 361 (239, 578) South Africa Base case 1,084,698 15,805 69 — — — — In addition to smear 1,594,276 20,420 78 110 109 (88, 133) — — Replacement of smear 1,758,467 20,702 85 138 136 (105, 172) 582 594 (353, 956) Uganda Base case 544,499 22,182 25 — — — — In addition to smear 643,172 24,570 26 41 34 (−3, 69) — — Replacement of smear 670,137 24,611 27 52 37 (0, 73) 650 276 (−1895, 2,406) The results of the probabilistic sensitivity analysis (Monte Carlo simulation) are also shown in Table 5. Aside from the replacement of smear microscopy in Uganda all estimates remain cost-effective. Figure 2 provides an illustration of the cost-effectiveness of Xpert deployed as “a replacement of” smear microscopy in comparison to the “in addition to” scenario for a range of WTP thresholds. This graph, known as an acceptability curve, shows that if the WTP is US$490 in Uganda, there is around a 75% probability that Xpert as a replacement of smear is cost-effective when compared to the “in addition to” scenario. 10.1371/journal.pmed.1001120.g002 Figure 2 Cost-effectiveness acceptability curves. ICER “replacement of smear” compared with “in addition to smear.” Nearly all of our one- and two-way sensitivity analyses did not increase the ICER compared to the base case of either Xpert scenario above the WTP threshold (Table 6). Figure 3 shows ICER variation when parameters for the suspect population and the performance of the base case change. Varying the true proportion of those with TB and MDR-TB in the cohort has little effect on our results, although Xpert ICERs substantially worsen when the proportion of smear-positive TB cases becomes 5% or less (translating into 7%–9% with any type of TB). Varying assumptions on the performance of the base case alters ICERs substantially. Increasing the sensitivity of smear examination reduces the cost-effectiveness of Xpert, but not below the WTP threshold. If clinical diagnosis has a higher specificity and lower sensitivity than in our study sites, Xpert ICERs worsen, but also remain below the WTP threshold. But, if clinical diagnosis has a lower specificity and higher sensitivity than in our study sites, ICERs for Xpert substantially improve. Adding chest X-ray for 50% of the individuals with suspected TB tested by Xpert has limited impact on the cost-effectiveness of Xpert. Adding clinical diagnosis for all HIV-positive individuals with suspected TB with a negative Xpert result has no or limited effect in India and South Africa, but doubles ICERs for Xpert in Uganda (although not above the WTP threshold). This reflects differences in HIV prevalence as well as relatively high cost and low specificity of clinical diagnosis in Uganda owing to more extensive use of X-ray. Incorporating the cost of culture and increasing the proportion of TB diagnosis based on the culture result, has a mixed effect. Xpert remains cost-effective up until the point where 40%–70% of patients receive a culture-based diagnosis. Above proportions of 50%–90%, the base case becomes more effective. If however, culture performance is less than 100%, the base case does not become more effective than the Xpert-based scenarios until nearly 100% of patients receive a culture-based diagnosis (unpublished data). 10.1371/journal.pmed.1001120.g003 Figure 3 Selected sensitivity analyses. Sensitivity of the model for the prevalence of tuberculosis, for the prevalence of multidrug-resistant tuberculosis, and for the accuracy of clinical diagnosis. Patterns of ICERs in 2010 US$ for varying the proportion of individuals with suspected TB in the cohort who have smear-positive TB (A, D, G); for varying the proportion of new patients with TB who have multidrug-resistant TB (MDR-TB, B, E, H); and for varying the specificity of the clinical diagnosis of TB in the base case (C, F, I). (A, B, and C), South Africa; (E,D, and F), India; (G, H, and I), Uganda. Black lines, Xpert assay in addition to sputum smear examination; grey lines, Xpert assay as replacement of sputum smear examination. The proportion of individuals with suspected TB in the cohort who have smear-negative TB varies along with the proportion of individuals with suspected TB in the cohort who have smear-positive TB in a linear manner, depending on the HIV-infection prevalence (A, D, G; see Table 1 and Text S1). Similarly, the proportion of previously treated patients with TB who have MDR-TB varies linearly with the proportion of new patients with TB who have MDR-TB (B, E, H; see Table 1 and Text S1). The sensitivity of clinical diagnosis in the base case varies inversely with the specificity (range, 8%–80%; C, F, I). 10.1371/journal.pmed.1001120.t006 Table 6 Costs per DALY 2010 US$: sensitivity analyses. Assumption ICER Compared to: India South Africa Uganda Base Case In Addition to Smear Replacement of Smear Base Case In Addition to Smear Replacement of Smear Base Case In Addition to Smear Replacement of Smear Primary estimate Base case — 55 68 — 110 138 — 41 52 In addition to smear — 343 — — 582 — — 650 Reprogrammed test: no signal MDR Base case — 50 51 — 87 86 — 37 40 In addition to smear — 107 — — NA — — 289 Clinical diagnosis performance pooled across countries Base case — 62 78 — 89 121 — 53 58 In addition to smear — 342 — — 582 — — 650 Proportion retreatment doubles Base case — 115 119 — 209 220 — 67 73 In addition to smear — 170 — — 334 — — 200 Cartridge cost reduces to US$11.70 Base case — 42 54 — 102 129 — 26 36 In addition to smear — — 318 — 570 — — 561 50% of individuals with suspected TB have X-ray added to Xpert Base case — 73 87 — 126 154 — 47 58 In addition to smear — — 378 — — 606 — — 686 HIV-infected individuals with suspected TB have clinical diagnosis when Xpert is negative Base case — 55 68 — 132 157 — 82 90 In addition to smear — 343 — 610 — 706 Undiagnosed patients with TB do not return for diagnosis Base case — 50 67 — 109 138 — 33 43 In addition to smear — — Dominated by in addition to scenario — — 1,442 — — Dominated by in addition to scenario Single smear examination in base case Base case — 48 58 — 105 130 — 40 48 In addition to smear — — 343 — — 582 — — 650 60% of case receive culture diagnosis Base case — Dominates base case Dominates base case — 67 311 — Base case more cost-effective Base case more cost-effective In addition to smear — — 343 — — 582 — — 650 Sensitivity of smear examination increase by 15% Base case — 106 130 — 131 165 — 59 74 In addition to smear — — 343 — — 582 — — 650 NA, not available. Discussion Our results suggest that Xpert is likely to be more cost-effective than a base case of smear microscopy and clinical diagnosis of smear-negative TB. The extent and type of cost-effectiveness gain from deploying Xpert is dependent on a number of different setting-specific factors. First and foremost of these factors is the performance of current TB diagnostic practice. Where the sensitivity of current practice is low, but specificity high, Xpert has a substantial impact on effectiveness. Where the sensitivity of current practice is high, but specificity low, Xpert will lower treatment costs by reducing the number of false positives. This latter effect is illustrated by the case of Uganda, where the model predicts a reduction in the treatment costs of false positives from US$171,803 to US$14,908, contributing to the overall reduction in treatment costs. Other factors that are likely to determine the extent of cost-effectiveness gain include the proportion of those co-infected with HIV and the proportion of those with MDR-TB, and the numbers of true TB cases in the suspect population. However, our results show that increasing proportions of HIV in the suspect population will not always reduce the ICER of Xpert (Figure 3). This finding is counter-intuitive. One would expect the cost-effectiveness of a diagnostic test that diagnoses smear-negative TB to improve with increases in HIV prevalence. However, as the proportion of individuals co-infected with HIV in the suspect population increases, so the sensitivity of Xpert decreases. Depending on the relative costs and performance of the base case, this counter-effect means that the relationship between HIV prevalence and Xpert's cost-effectiveness is weaker than anticipated. Nor can we conclude on the direction of the relationship between cost-effectiveness gain and the level of prevalence of MDR-TB in the suspect population at this time. Our model demonstrates that when transmission effects are excluded, the ICER of Xpert increases as the MDR-TB prevalence increases (Figure 3). This result occurs because although the effectiveness of Xpert increases with a higher MDR-TB prevalence, the ICER of treating MDR-TB is higher than that of drug susceptible TB, thus countering the gain from increased effectiveness. Unsurprisingly, we also find that higher proportions of TB cases in the suspect population improve the cost-effectiveness of Xpert. The cost per TB case detected will also decrease with increases in TB prevalence. As TB programmes already fund elements of the base case, cost-effectiveness may therefore be initially improved by using existing diagnostic tools, such as X-ray and clinical scores, to screen the TB suspect population prior to Xpert. In the longer run, however, the expansion of X-ray as a permanent approach for suspect screening is unlikely to be cost-effective, and further work examining alternative screening approaches may be required. Moreover, different approaches are likely to be adopted for specific suspect populations, most notably those already known to be HIV infected, those who have already failed treatment, and those at a high risk of MDR-TB. We therefore recommend that further work is conducted to explore the impact on cost-effectiveness of different algorithms when Xpert is applied to more limited suspect groups. A number of factors limit our analysis. Firstly, the assumption of no transmission effects or additional mortality benefit from early diagnosis is a conservative approach and will underestimate the cost-effectiveness of Xpert—particularly where the introduction of Xpert is likely to increase the numbers of drug-resistant patients who are appropriately and rapidly treated. Likewise, we do not factor in patient costs. A full societal evaluation would make all options less cost-effective, but Xpert is likely to fare better than alternatives, as it requires fewer patient visits. In addition, if Xpert can achieve earlier diagnosis, substantial reductions in patient costs prior to treatment may be achieved [31]. The reference standard for the test performance parameters in our model did not include culture-negative TB based on response to treatment, because this diagnostic category will include cases with no TB or extra-pulmonary TB that cannot be diagnosed by sputum-based tests. This situation may have lead to overestimation of the sensitivity and underestimation of the specificity of Xpert. Owing to lack of evidence, we only included one repeat visit for false negatives in our model, to capture those who quickly progress to smear-positive TB. This number may be insufficient and miss both the additional costs and effectiveness of further repeated visits. On the other hand, our assumption that 100% of false negatives still alive and with TB after 3 mo have a repeat visit may be an overestimation, thereby inflating ICERs for the Xpert scenarios. We assumed that a negative Xpert result does not lead to further TB diagnostic procedures. This assumption may not be true in practice, in particular not for HIV-infected individuals with suspected TB [32]. Our sensitivity analyses show that adding clinical diagnostic procedures for this group can substantially reduce cost-effectiveness of Xpert when HIV prevalence is high and X-ray is used extensively. Also because of the lack of data, we have not included a high MDR-TB, but low HIV-prevalence setting. This lack of data restricts our ability to generalise findings to all low- and middle-income settings, particularly the former Soviet states, where this epidemiological pattern is common in suspect populations. Finally, our sensitivity analysis demonstrates that Xpert may not be cost-effective when compared to a base case in which a high proportion of smear-negative TB cases are diagnosed by culture. However, this result is based on our assumption that culture performs at 100% sensitivity and specificity. In addition, we did not include costs of specimen transport, increased risk of false-negative cultures or contamination, reduced sensitivity when only one specimen is cultured, and possible delay effects on mortality and patient drop out. All these simplifications will inflate the cost-effectiveness of a base case that includes culture. As is standard practice, we determine cost-effectiveness in comparison to the WHO WTP threshold. Unfortunately, achieving this threshold does not mean that the resources are available in low- and middle-income countries, merely that Xpert should be afforded [33]. In reality, resourcing for tuberculosis services in low- and middle-income countries is extremely constrained. Countries may therefore need to prioritise. In terms of priorities, suspect populations with a high likelihood of TB, particularly in settings with high HIV and MDR-TB prevalence, are an obvious choice. However, our findings illustrate that it is also important to balance these factors with the current performance of the existing diagnostic pathway. Countries or areas that have the weakest performance in terms of diagnosing smear-negative cases may benefit the most, even when they have relatively low levels of MDR-TB and HIV; although additional investment may be required to strengthen aspects of the health system to ensure that Xpert can be implemented successfully. Funding Xpert may also mean that scarce resources are not made available to other equally deserving areas. Care must therefore be exercised to take into account the broader tuberculosis control and health system needs of any particular setting when funding Xpert. Our model is robust given the current evidence and data available. However, key data in this area—particularly on the characteristics of TB suspect populations, the feasibility of implementing Xpert at scale, and the extent to which clinicians allow diagnostic test results to influence treatment decisions—remain scarce. Moreover, it is likely that there will be costs associated with Xpert scale-up that we cannot predict at this point. Although our model strongly suggests that Xpert will be cost-effective in a wide variety of settings, Xpert scale-up will substantially increase TB diagnostic costs. Given this increase, and the current data paucity, we recommend careful monitoring and evaluation of initial roll-out before full scale-up. Funding should be provided for implementation studies, including pragmatic trials, in a number of countries to accelerate this process. As we did not assess cost-effectiveness in a setting with high MDR but low HIV prevalence, we also recommend additional economic modelling studies before embarking on roll-out in these settings, taking into consideration operational factors that may affect outcomes such as patient drop-out and physician behavior [34]. Finally, although Xpert is a highly promising technology, there is still room for improvement in TB diagnostics. Xpert has incomplete sensitivity for smear-negative TB and rifampicin resistance and does not detect resistance to isoniazid and other drugs. Other promising tests, such as microscopic observation drug susceptibility test (MODS) [35], should be evaluated for their cost-effectiveness, including comparisons with Xpert. Our finding should not discourage investment in other promising new TB diagnostic technologies, particularly those that further improve the diagnostic sensitivity and detection of wider forms of drug resistance and can be implemented at peripheral health care level at low cost. Conclusion Despite the fact that there is considerable concern from policy makers about the costs and affordability of new diagnostic technologies in low- and middle-income countries, our results suggest that Xpert is likely to be a highly cost-effective investment. If demonstrated test performance is maintained at scale, Xpert has the potential to substantially increase TB case detection. Moreover, in the settings modelled, TB treatment costs are not predicted to substantially increase with the introduction of Xpert; instead, treatment is likely to be switched from those who do not benefit from treatment, to those who do. Our results suggest that funding should be provided to initiate the roll-out of Xpert in low- and middle-income countries, as a promising means of enabling access to effective treatment for all those with the disease. We recommend, however, that this roll-out is carefully evaluated to validate our results before full scale-up—to ensure that Xpert implementation is done in a way that does not negatively impact TB programmes, their funding, and the health systems that support them. Supporting Information Text S1 Details of model assumptions, test turnaround times (Table S[A]), treatment outcome probabilities (Table S[B]), probabilities of death and spontaneous recovery with false-negative tuberculosis diagnosis (Table S[C]), and variables used in the DALY calculations (Table S[D]). (DOC) Click here for additional data file.
Bookmark
• Record: found
• Abstract: found

### Epidemiological benefits of more-effective tuberculosis vaccines, drugs, and diagnostics.

(2009)
The Bill and Melinda Gates Foundation supports an ambitious portfolio of novel vaccines, drug regimens, and diagnostic tools for tuberculosis (TB). We elicited the expected efficacies and improvements of the novel interventions in discussions with the foundations managing their development. Using an age-structured mathematical model of TB, we explored the potential benefits of novel interventions under development and those not yet in the portfolio, focusing on the WHO Southeast Asia region. Neonatal vaccination with the portfolio vaccine decreases TB incidence by 39% to 52% by 2050. Drug regimens that shorten treatment duration and are efficacious against drug-resistant strains reduce incidence by 10-27%. New diagnostics reduce incidence by 13-42%. A triple combination of a portfolio vaccine, drug regimen, and diagnostics reduces incidence by 71%. A short mass vaccination catch-up campaign, not yet in the portfolio, to augment the triple combination, accelerates the decrease, preventing >30% more cases by 2050 than just the triple combination. New vaccines and drug regimens targeted at the vast reservoir of latently infected people, not in the portfolio, would reduce incidence by 37% and 82%, respectively. The combination of preventive latent therapy and a 2-month drug treatment regimen reduces incidence by 94%. Novel technologies in the pipeline would achieve substantial reductions in TB incidence, but not the Stop TB Partnership target for elimination. Elimination will require new delivery strategies, such as mass vaccination campaigns, and new products targeted at latently infected people.
Bookmark
• Record: found
• Abstract: found

### Mathematical models in the evaluation of health programmes.

(2011)
Modelling is valuable in the planning and evaluation of interventions, especially when a controlled trial is ethically or logistically impossible. Models are often used to calculate the expected course of events in the absence of more formal assessments. They are also used to derive estimates of rare or future events from recorded intermediate points. When developing models, decisions are needed about the appropriate level of complexity to be represented and about model structure and assumptions. The degree of rigor in model development and assessment can vary greatly, and there is a danger that existing beliefs inappropriately influence judgments about model assumptions and results. Copyright © 2011 Elsevier Ltd. All rights reserved.
Bookmark

### Author and article information

###### Contributors
Role: Editor
###### Journal
PLoS One
PLoS ONE
plos
plosone
PLoS ONE
Public Library of Science (San Francisco, USA )
1932-6203
2014
23 October 2014
: 9
: 10
###### Affiliations
[1 ]Department of Epidemiology, Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland, United States of America
[2 ]TB Modelling Group, TB Centre, London School of Hygiene and Tropical Medicine, London, United Kingdom
[3 ]Department of Global Health and Development, London School of Hygiene and Tropical Medicine, London, United Kingdom
[4 ]Department of Epidemiology, Harvard School of Public Health, Boston, Massachusetts, United States of America
University of Erlangen-Nuremberg, Germany
###### Author notes

Competing Interests: The authors have declared that no competing interests exist.

Conceived and designed the experiments: AZ RGW AV TC DWD RMGJH. Performed the experiments: AZ RMGJH. Analyzed the data: AZ RMGJH. Wrote the paper: AZ RGW AV TC DWD RMGJH.

###### Article
PONE-D-14-03791
10.1371/journal.pone.0110558
4207742
25340701

This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

###### Page count
Pages: 8
###### Funding
AZ is funded through a Canadian Institutes of Health Research post-doctoral fellowship. RGW, AV, DWD, TC, and RMGJH are funded by the TB Modelling and Analysis Consortium (TB MAC) grant (OPP1084276) from the Bill and Melinda Gates Foundation. RGW is also funded by the Medical Research Council (UK) (Methodology Research Fellowship: G0802414 and grant MR/J005088/1) and CDC/PEPFAR via the Aurum Institute (U2GPS0008111). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
###### Categories
Research Article
Biology and Life Sciences
Computational Biology
Population Modeling
Infectious Disease Modeling
Plant Science
Plant Pathology
Infectious Disease Epidemiology
Medicine and Health Sciences
Diagnostic Medicine
Epidemiology
Public and Occupational Health
Global Health
Infectious Diseases
Bacterial Diseases
Tuberculosis
Multi-Drug-Resistant Tuberculosis
Tropical Diseases
Research and Analysis Methods
Research Assessment
Systematic Reviews

Uncategorized
Uncategorized