Start4All protocol for a Bayesian cost-effectiveness model of tuberculosis screening and diagnosis in seven high burden low-income and middle-income countries

Abstract

Introduction

High costs of screening and diagnostic tests remain a major barrier to timely tuberculosis (TB) identification in resource-limited settings. Evidence on the cost-effectiveness of scalable screening algorithms is limited. Start4All is a research project aimed at developing and evaluating algorithmic approaches to TB screening and diagnosis, with the goal of optimising technical and allocative efficiency when expanding diagnostic coverage to primary healthcare and community settings.

Methods and analysis

Five screening and diagnostic tests will be evaluated: a capillary blood-based assay (C-reactive protein (CRP)), sputum-based rapid molecular tests (PCR; individual and pooled Xpert MTB/RIF Ultra assay (Xpert Ultra, Cepheid®, California, USA)), a lateral-flow urine-based test for lipoarabinomannan (LF-LAM), and digital chest X-rays with artificial intelligence-based computer-aided detection (CXR-CAD). A microbiological reference standard of positive culture using the mycobacteria growth indicator tube will be used to confirm TB disease.

We will compare the cost and effectiveness of concurrent and sequential positive serial combinations (screening algorithms) of CRP, CXR-CAD, LF-LAM, individual and pooled Xpert Ultra. Diagnostic performance will be estimated using sensitivity, specificity, predictive values and proportions of positive results, with Bayesian inference used to derive these estimates. The analysis will include adults (15 years and older) only and will be stratified by HIV status and level of care, including facility and community-based case finding. Effectiveness will be assessed based on the number of people with TB detected. Cost analysis will be conducted from the provider perspective, incorporating commodity and implementation costs. A decision tree model will be developed to assess the cost per number of persons with confirmed TB detected across all countries. Probabilistic sensitivity analysis will be conducted to account for uncertainty in model parameters, incorporating willingness-to-pay and willingness-to-accept thresholds.

Ethics and dissemination

WHO ethical review committee approval ERC.0003921. Data will be available on reasonable request to the principal investigator of the consortium.

Trial registration number

NCT05845112.

Strengths and limitations of this study.

• Modelling several screening algorithms across a wide range of possible test combinations and accounting for uncertainty.

• Covers several countries, populations, settings and case finding approaches.

• The proposed model provides a flexible framework that can be adapted or used by countries for decision-making.

• Relies on secondary data and a number of simplifying assumptions.

Introduction

Tuberculosis (TB) is the leading cause of death from a single infectious disease globally. Despite progress in diagnosis and treatment, 2.7 million people who developed TB disease were not diagnosed or notified in 2023, and over 1.25 million died of TB.¹ TB disproportionately impacts vulnerable populations with limited healthcare access, including those living in poverty, men, people living with HIV (PLHIV), rural residents, urban poor, displaced persons, older persons and children. The WHO recently reported that 47% (range 27%–92%) of TB-affected households face catastrophic costs.² Furthermore, funding for TB services in low-income and middle-income countries (LMICs) has been chronically insufficient with current levels at only 26% of the targeted US$ 22 billion/year by 2027.¹

Effective TB screening and diagnostic tests are crucial for improving population health outcomes by minimising delays in diagnosis and linkage to care, and reducing transmission, as well as lowering patient costs and unnecessary treatment expenditures.¹ However, barriers such as limited access to accurate and timely diagnostic tools behove development of cost-effective approaches, particularly for vulnerable populations in LMICs. Since 2011, rapid molecular tests have revolutionised TB diagnosis, replacing the traditional reliance on microscopy and culture methods. While newer molecular tests are recommended by WHO as highly effective, rapid and reliable diagnostic tools for TB,³ their high cost limits widespread use and sputum smear microscopy remains the primary TB diagnostic method for over half of the TB-affected population in high-burden countries.¹ A recent review highlighted that cost-effectiveness (CE) of TB screening and diagnostic tools may be a function of several contextual factors, including population risk and associated prevalence, as well as testing volumes, yet evidence on the cost and CE of alternative TB screening and diagnostic tools is lacking.⁴

WHO defines an algorithm for systematic TB screening as combining one or several screening tests and a separate diagnostic evaluation for TB disease.⁵ Evidence suggests that the introduction of initial screening tests, especially symptom-agnostic tools, such as chest X-rays (CXR) with artificial intelligence (AI)-based computer aided design (CAD) software and pooling of samples, may not only improve diagnostic accuracy, but also lead to large cost savings by reducing the need for more expensive confirmatory tests,⁶,⁸ potentially enabling large-scale TB screening programmes where they are most needed.⁹ Additionally, a marginal reduction in diagnostic accuracy may be considered acceptable if it facilitates the use of decentralised and/or more cost-efficient tests that can increase overall case detection at the population level¹⁰ . The Start4All consortium is a 4-year Unitaid-funded research project to evaluate the optimal algorithmic approaches to TB screening and diagnosis across seven countries (Bangladesh, Brazil, Cameroon, Kenya, Malawi, Nigeria and Vietnam). The aim of Start4All is to determine the most accurate, feasible, acceptable, scalable and cost-effective solutions to expand TB diagnostic coverage to primary healthcare (PHC) and community settings.

This protocol paper outlines the methods used for the economic modelling analysis conducted alongside the Start4All diagnostic evaluation study, the protocol for which is planned to be published in this journal and is summarised in the statistical analysis plan (SAP) published elsewhere.¹¹ This initial analysis presents a modelling study that may support countries in identifying algorithms to be prioritised in their specific contexts—an essential step given the evolving diagnostic landscape, diverse target populations and the varying capacities of health systems across settings.

Our study aims to address a critical lack of evidence on the CE and implementation of multi-step TB screening and diagnostic pathways, particularly across diverse and vulnerable populations, by generating multi-country data that can inform TB policy, optimise screening strategies and guide future research.

Objectives

Our primary objectives are to identify the most cost-effective algorithms relative to individual sputum Xpert Ultra (Cepheid®, California, USA) testing to detect TB through facility-based case finding (FBCF):

1. At PHC facilities across all countries (aggregate analysis) and all individuals, regardless of HIV status.

2. At district hospitals (DH) across all countries (aggregate analysis) and for all individuals, regardless of HIV status.

Our secondary objectives are to identify the most cost-effective algorithms relative to individual sputum Xpert Ultra testing to detect TB through:

1. FBCF at PHC facilities across all countries (aggregate analysis), for PLHIV and people without HIV or not known to have HIV.

2. FBCF at DH across all countries (aggregate analysis), for PLHIV and people without HIV or not known to have HIV.

3. FBCF at PHC facilities in individual countries (Bangladesh, Brazil, Cameroon, Kenya, Malawi and Nigeria) for all individuals, regardless of HIV status.

4. FBCF at DH in individual countries (Bangladesh, Cameroon, Kenya, Malawi, Nigeria and Vietnam) for all individuals, regardless of HIV status.

5. Community-based case finding (CBCF) in informal settlements in Bangladesh for all individuals, regardless of HIV status.

6. CBCF in rural poor in Cameroon for all individuals, regardless of HIV status.

7. CBCF in internally displaced people (IDPs) in Nigeria for all individuals, regardless of HIV status.

8. CBCF in nomads in Nigeria for all individuals, regardless of HIV status.

9. CBCF in urban, low-income populations in Vietnam for all individuals, regardless of HIV status.

Methods and analyses

Study design

The study took place between 1 September 2022 and 28 May 2025. The health economic study is being conducted alongside multi-country (n=7), multi-centric cross-sectional studies (n=20), with different population groups and at different health system levels (see online supplemental appendix S1 for overview). Publication of the protocol for the multi-country studies is planned in this journal, while the SAP has been published elsewhere.¹¹ At stationary healthcare facilities, including PHC facilities and DH, passive case-finding (PCF) was implemented for individuals with presumptive TB (see online supplemental appendix S1). In some cases, intensified case finding (ICF) was adopted for people who were considered by WHO definitions to have a higher risk of TB, for example, PLHIV.⁵ PCF and ICF will be considered jointly in the analysis as FBCF. Individuals were also identified through CBCF in pre-specified population groups, including informal settlements, IDPs or nomads.

Five screening and diagnostic tests were evaluated on each participant in the study (table 1): a capillary blood-based assay (C-reactive protein (CRP)), sputum-based rapid molecular tests (PCR; individual and pooled Xpert Ultra testing), a lateral-flow urine-based test for lipoarabinomannan (LF-LAM), and digital CXRs with AI-based CAD software. Pooled sputum testing involved multiple sputum samples from different individuals being combined and tested together, usually in pools containing four individual samples, by Xpert Ultra. A microbiological reference standard of positive culture using the mycobacteria growth indicator tube was used to confirm TB disease.

Table 1. Summary of tests evaluated in Start4All.

Assay type	Assay	Included in health economic evaluation
1. Point-of-care CRP
Quantitative	POC CRP	Yes
Semi-quantitative	Rapid CRP	No*
2. Urine lateral flow test	LF-LAM	Yes
3. Molecular diagnostic	Pooled Xpert Ultra	Yes
4. Molecular diagnostic	(Single) Xpert Ultra	Yes
5. CAD interpretation of digital CXR	Digital CXR+CAD	Yes

^*Performance metrics will be combined for all CRP tests and the economic evaluation will consider the cost of quantitative CRP only, as it makes up the majority of CRP tests in the study

CAD, computer-aided detection; CRP, C-reactive protein; CXR, chest X-ray; LF-LAM, lateral-flow urine-based test for lipoarabinomannan; POC, Point of Care.

Data from the study will inform estimates of diagnostic performance—including sensitivity, specificity, predictive values and diagnostic yield per test attempted—for individual tests, as well as combinations of CRP, CXR-CAD, LF-LAM and individual or pooled Xpert Ultra tests. The proportion of positive results from each test or test combination will also be calculated.

Given a large number of possible combinations and permutations, a framework was developed to select a set of individual tests, as well as concurrent (simultaneous testing) and positive sequential (stepwise testing following a positive result) test combinations (screening algorithms), for modelling and evaluation (online supplemental appendix S2). In total, 20 algorithms were pre-selected for FBCF and 40 for CBCF. As with individual tests, key performance metrics will be assessed for each selected algorithm.

Bayesian statistical inference will be used to estimate overall diagnostic accuracy metrics for each algorithm.

The CHEERS framework was used as a guide to develop this protocol (online supplemental appendix S3).¹²

Patient and public involvement

Data collection of health outcomes took place in seven high-TB-burden countries, where country partners conducted a variety of community engagement activities to ensure appropriate participation of the public in the study. These activities included the organisation of consensus building meetings at national and local levels, during which feedback on the design and usefulness of the study was received and incorporated. CBCF events in a subset of countries were also accompanied by awareness building and community mobilisation efforts as described in table 3. The consortium constituted and continually engaged a scientific advisory board, who were informed and provided feedback on the progress of the study and its findings. Finally, the study included qualitative research elements that enabled the collection of insights and commentary from key stakeholders such as public health officers, policymakers and key affected groups.

Economic evaluation

The health economic analysis outlined in this protocol will compare the modelled costs and accuracy of selected algorithms with the individual Xpert Ultra test through a CE analysis. The single specimen individual Xpert Ultra was selected as the comparator for the economic evaluation because it is recommended by WHO.

Analysis will be performed at aggregate and individual country level for FBCF (PHC and DH) and at individual country level for CBCF.

The analysis will be performed from the provider perspective, as no primary data was collected from patients. The time horizon is approximately 1 year, though this varies between countries based on duration of the recruitment period. For FBCF, only the costs incurred at the point of testing, including the diagnostic process and any immediate resources used, will be considered, excluding follow-up care or treatment costs. For CBCF, the costs associated with mobilising individuals for screening, including community engagement efforts and communication materials used during the recruitment phase, will be captured.

The primary analysis will focus on aggregated CE results across multiple countries for FBCF. The secondary analysis will provide a more detailed examination of individual country-level CE results, covering both FBCF and CBCF approaches.

Cost analysis

Unit costs were defined as the cost per person tested and will be calculated for each algorithm at aggregated and individual country level. For the aggregated analyses, only commodity costs will be considered. This is because there is significant variability in implementation costs across countries, making the inclusion of health system costs non-comparable and analytically meaningless. Individual country analyses will account for commodity and implementation costs.

Commodity costs will be based on ex-works costs only and be calculated as a function of several parameters, including cost inputs (capital and non-capital items) and a set of underlying assumptions (table 2). Input cost values will be based on available literature and Start4All procurement costs (online supplemental appendix S5). Capital items are defined as a durable resource with a useful life typically exceeding 1 year and value greater than US$ 100, whose costs are depreciated over time. Capital item costs will be annuitised over the lifetime of the item at a 3% discount rate¹³ and apportioned per test, based on testing volumes for each test. Warranty and service level agreement costs will be included. Non-capital items are defined as resources with a useful life of less than 1 year that are consumed during routine service delivery, for example, Xpert Ultra cartridges.

Table 2. Summary of cost input parameters and modelling assumptions.

Input parameters	Disaggregation and level of detail	Data source
Commodity costs*	Per test, capital and non-capital items listed individually	22Pricing information
FBCF implementation cost	Per country, cost per outpatient visit (remains the same for PHC and DH)	17Modelled cost per outpatient visit
CBCF implementation cost	Per country, given as overall cost per month or per campaign	Collected from four countries: Bangladesh, Cameroon, Nigeria and Vietnam
Modelling assumptions
Number of tests per day	Per test and setting (Xpert Ultra)	PHC and DH23: Community: assumption on positivity rate of screened individuals who proceed to confirmatory testing (15%)
Number of tests per day	Per test and setting (CXR-CAD and CRP)	PHC and DH: assumptions on capacity and use Community14,16
Number of pooled samples	Per test (Xpert Ultra with Pooling only)	7Description of pooled sputum testing with Xpert Ultra
Number of people screened	Per country, per month/campaign for CBCF implementation	14,16Description and results of active case finding campaigns
Cost of transporting samples from testing site to facility	Per country, CBCF only	Collected from four countries in which CBCF implemented: Bangladesh, Cameroon, Nigeria and Vietnam
Discount rate	Remains the same for all tests and countries	13Discount rate suggested in the literature
Expected life years of capital equipment	Per capital item type, remains the same for all tests and countries	24Suggested methods for adjusting cost data over time periods

^*Based on ex-works values

CAD, computer-aided detection; CBCF, community-based case finding; CRP, C-reactive protein; CXR, chest X-ray; DH, district hospitals; FBCF, facility-based case finding; PHC, primary healthcare.

Testing volumes for individual Xpert Ultra will be based on assumptions from the literature.¹⁴ At CBCF, volumes will be calculated by applying a 15% rate to the total number of individuals screened daily,¹⁴,¹⁶ representing those who proceed to confirmatory testing with individual Xpert Ultra. For CXR-CAD and CRP at facility level, testing volumes will be determined using capacity and utilisation assumptions. For CBCF, daily screening volumes will be drawn from existing literature.¹⁴,¹⁶

For FBCF, sputum specimen transportation costs will be assumed to be zero, based on the assumption that Xpert Ultra is available onsite at all facilities. For CBCF, transport of specimen will be costed based on information on running CBCF screening camps from the four countries implementing CBCF (Bangladesh, Cameroon, Nigeria and Vietnam).

For FBCF, country-specific implementation costs will be determined by the cost per outpatient visit.¹⁷ The cost per FBCF algorithm will be estimated by summing the cost of individual tests included in the algorithm and assuming one outpatient visit to complete the algorithm.

CBCF implementation costs will be collected directly from the four countries where CBCF activities take place (Bangladesh, Cameroon, Nigeria and Vietnam), based on assumptions regarding the typical implementation of a CBCF programme, excluding research costs (table 3). The number of people screened per month or campaign will be estimated based on literature.¹⁴,¹⁶ The total cost per campaign or per month of running the CBCF programme will be divided by the estimated number of people screened to calculate the cost per person screened in each country. The cost per CBCF algorithm will be determined in a manner similar to facility-based algorithms, but by combining commodity costs with primary programme implementation cost.

Table 3. Summary of costed CBCF activities.

Cost component	Overview of costed activities (number of staff involved)*
	Bangladesh	Cameroon	Nigeria	Vietnam‡
Pre-screening/community mobilisation	Announcers through microphones	Field visit by team	Production and airing of radio jingles in three languages	Provincial and district level consensus-building meetings
	Leaflet printing and distribution	Stakeholder meeting: seven health districts and regional team	Field visit by team (including hotel, per diem, etc)	Invitation letters and loudspeaker announcements
	Staff training (ultra-portable X-ray)	Community health workers (stipend, if relevant)	Stakeholder meeting(s)	Ward level consensus-building meetings
		Radio station fees	Incentive for community volunteers	Informal communication to community about upcoming CBCF
			Training for community volunteers
Screening event	†	Staff allowances and accommodation	Accommodation for teams	Staff allowances
		Water for participants	Per diem/food allowance	Participant allowances
		Non-medical equipment and other items (eg, printer, paper and Internet)	Incentive for participants	Logistics
			Other (non-medical) items	Non-medical equipment and other items (eg, laptops/tables, printer)
Transportation	Vehicle hire (for carrying logistics: X-ray machine and relevant kits)	Purchased screening van for equipment transport	Purchased cars+motorbikes	Fee for sputum transportation to lab
		Fuel and tolls for staff transport	Fuel and vehicle hire for (pre-) screening teams and sputum transportation
HR (screening team)	Project research physician	Driver	Driver	Event supervisor (1)
	Field assistant/health worker	Radiographer (2)	Radiographer (2)	Doctor (1)
	Technical assistant (radiography)	Laboratory technician (2)	Laboratory technician (2)	Nurse (7)
	Medical technologist	Research medical officer (MD)	Research medical officer (MD)	X-ray technician (2)
		Medical research assistant (nurse/lab technician)	Medical research assistant (nurse/lab technician)	Radiologist (1)
				Laboratory technician (2)
Additional laboratory costs	n/a	n/a	Incentives for laboratory technicians	Testing fees
				Laboratory coordination and management

^*The following costs were not included: programme management and coordination, any costs related to culture

^†ACF implementation approach costed is a ‘low cost’ cost, where screening events take place relatively close to healthcare facilities, including laboratories; the Ministry of Health in Bangladesh is also exploring a more costly option where screening activities take place further afield

^‡In Vietnam, screening for other primary care indications is included

ACF, Active Case Finding; CBCF, community-based case finding; HR, Human Resource.

Cost per outpatient visit is available in 2017 international dollars (Int$)¹⁷ and will be adjusted to 2024 dollars using global inflation figures provided by the World Bank.¹⁸ CBCF implementation costs will be converted to 2024 Int$ using Purchasing Power Parity (PPP) conversion factors.¹⁸ For consistency, commodity costs published in 2024 will be used. Project procurement costs are from 2023/2024.

Outcomes (effects)

Effectiveness will be assessed by the number of persons with confirmed TB detected by each algorithm, simulating for a hypothetical cohort of 1000 participants. Data from the study will be used to estimate, for each algorithm, country and population setting under evaluation, the conditional probabilities of both positive results at each testing step, as well as the positive and negative predictive values (PPV and NPV). Each of these parameters is a probability parameter, and as such, the posterior distributions are estimated as beta distributions, the estimated distributional parameters of which are used as input to the CE model.

Aggregate analyses will combine parameters from all countries for PHC and DH populations, disaggregated by HIV status. Individual country analyses will use population-specific and site-specific accuracy parameters.

The number of individuals with a false positive result and the number of missed cases will also be estimated following the methods described above and analysed alongside CE results.

Cost-effectiveness modelling

We will use a Bayesian decision tree model to simulate the cost and persons with TB detected for all algorithms selected a priori. Separate decision trees will be built in Amua V.0.3.1 software (Dr Zachary Jonathan Ward, Center for Health Decision Science at Harvard T.H. Chan School of Public Health, Boston, MA, USA)¹⁹ for FBCF and CBCF, following standard methodology.¹³ An example decision tree is provided in online supplemental appendix S6.

The expected cost of each algorithm will be calculated as a function of the unit cost of individual tests included and the (conditional) probability of positive/negative test results at each screening step and PPV and NPV of each algorithm (detail provided in online supplemental appendix S6).

A total of 22 separate and independent model runs will be performed across a range of different settings, accounting for HIV status for aggregated analyses (table 4). Each time, all algorithms selected a priori for facility-based and CBCF settings will be included.

Table 4. Model runs for economic analysis.

Geographical scope	Setting	Simulation model runs
		Adults†: all	Adults: people without HIV/HIV status not known	Adults: PLHIV	Total
Aggregated analyses*
PHC	FBCF, all	1	1	1	3
DH	FBCF, all	1	1	1	3
Total		2	2	2	6

Individual country	Setting	Adults: all	Adults: people without HIV/HIV status not known	Adults: PLHIV	Total
Bangladesh	FBCF, PHC	1			1
	FBCF, DH (outpatient department)	1			1
	CBCF (informal settlements)	1			1
Brazil	FBCF, PHC	1			1
Cameroon	FBCF, PHC	1			1
	FBCF, DH (outpatient department)	1			1
	CBCF (rural poor)	1			1
Kenya	FBCF, PHC	1			1
	FBCF, DH (outpatient department)	1			1
Malawi	FBCF, PHC	1			1
Nigeria	FBCF, PHC	1			1
	FBCF, DH (outpatient department)	1			1
	CBCF (nomads, IDPs)	2			2
Vietnam	FBCF, PHC	1			1
	CBCF (urban, low-income)	1			1
Total		16	0	0	16
Grand total		18	2	2	22

^*Aggregated analyses will be performed with aggregated accuracy data and commodity costs only

^†Children will not be included in the health economic analysis

CBCF, community-based case finding; DH, district hospitals; FBCF, facility-based case finding; IDP, internally displaced people; PHC, primary healthcare; PLHIV, people living with HIV.

For each model run, point estimates for expected costs, persons with TB detected and the cost per TB case detected will be tabulated and results displayed on a CE plane (figure 1).

Figure 1. Interpretation. Individual Xpert (the comparator) is at the origin. The x axis represents the difference in the number of TB cases between a given algorithm and individual Xpert. The y axis represents the difference in cost. The WTP and WTA threshold lines are drawn in quadrants I and III respectively. Algorithms above the line are not considered cost-effective. Quadrant I: more TB cases are detected at a higher cost. The ICER is calculated for algorithms on or below the WTP threshold line. Quadrant II: more TB cases are detected at a lower cost. Algorithms in this quadrant are dominant and automatically retained. Quadrant III: fewer TB cases are detected at a lower cost. The trade-off between accuracy and cost is explored for algorithms on or below the WTA threshold line. Quadrant IV: fewer TB cases are detected at a higher cost. Algorithms in this quadrant are dominated and automatically ruled out. ICER, Incremental Cost-Effectiveness Ratio; TB, tuberculosis; WTA, willingness-to-accept; WTP, willingness-to-pay.

Handling uncertainty

Three types of uncertainty will be considered in the decision model, following standard guidelines.²⁰

First, heterogeneity of individuals, including varying levels of TB prevalence in the population and differences in cost will be reflected through repeated model runs for different country and population settings.

Second, parameter uncertainty will be addressed through probabilistic sensitivity analysis (PSA), using Monte Carlo simulation (100,000 iterations). Uncertainty of the performance parameters will be propagated into the model using the posterior distributions estimated in the performance analysis described above. Uncertain cost parameter inputs will be included in the model as fitted distribution functions, based on most likely, optimistic and pessimistic scenarios (online supplemental appendix S4). Where no data can be found on varying test prices, ±20% of the price point estimate will be assumed for the optimistic and pessimistic scenario, respectively, except for LF-LAM, where available literature will be used to define a reasonable range.²¹ Xpert Ultra cartridge costs will be varied by±20%. Gamma or Poisson distributions will be fitted using R (V.4.1.0 (The R Foundation for Statistical Computing, Vienna, Austria)), depending on data characteristics. The analysis will use the ‘optimistic’ and ‘pessimistic’ scenarios to set cost distribution parameters. Poisson distributions will be bounded to avoid unrealistic values in the PSA as a result of long tails. All costs in the decision model will be included as a beta distribution function. For aggregate analyses, results will include a point estimate and a 95% credible interval, defined as the 2.5th and 97.5th percentiles of results across these simulations.

Third, structural uncertainty will be partially addressed by modelling FBCF and CBCF separately. Costing assumptions will also be varied across different population settings, adjusting for the number of tests performed a day.

Uncertainty surrounding the cost and number of persons with TB detected will be summarised through scatter plots and CE acceptability curves (CEACs). Decision uncertainty will be reflected by varying the willingness-to-pay (WTP) and willingness-to-accept (WTA) thresholds (figure 1). WTP will be assessed where algorithms fall in quadrant I. WTA will be assessed where algorithms fall in quadrant III. Separate CEACs will be generated for WTA and WTP analysis.

Since there is no recognised WTP or WTA threshold for our outcome indicator (cost per person with TB detected), the estimated cost of treating one person with drug-susceptible TB (DS-TB)²¹ will be used as a minimum threshold value to ground the analysis and varied in the sensitivity analysis (SA).

In the absence of established guidance, a minimum threshold of a 60% probability of being cost-effective will be set as a starting point and varied in the SA.

Discussion

Effective and timely screening algorithms are essential for improving TB case detection and reducing transmission globally. This health economic modelling study compares the cost and number of persons with TB detected across various country and population settings and case-finding strategies, relative to individual Xpert Ultra testing. The goal is to identify a subset of the most promising algorithms for further real-world evaluation by examining trade-offs between cost and accuracy, and potential cost savings through additional screening tests that can reduce the number of individual Xpert Ultra tests required without impairing the number of persons with TB detected.

While the application of a follow-on diagnostic test after a negative screening result has the potential to improve diagnostic accuracy—particularly by reducing false positives—it also increases overall costs and may pose challenges to implementation feasibility, especially in resource-limited settings. Therefore, only positive concurrent and sequential screening algorithms are included since these have potential to exclude individuals who do not have TB early in the pathway and thus save resources. Applying follow-on diagnostic tests after a negative screening result would counteract this objective by limiting the opportunity to reduce confirmatory testing and associated costs.

The analysis will be conducted at two levels. First, aggregated analyses will identify a set of promising algorithms across several countries for FBCF in PHC and DH settings. Second, country-specific and population-specific analyses will generate contextual results to inform national policy and practice. This dual approach provides practical support for decision-making at country, regional and global level.

The analysis will use Bayesian methods to evaluate CE, accounting for uncertainty at different levels. Though our modelling framework will not account for detailed patient pathways or provider delays (eg, turnaround time), it will provide a high-level view of CE to help inform country selection of appropriate diagnostic strategies for their context. Distinct approaches will be applied for FBCF and CBCF, but models will remain generic and applicable across countries. Results will be stratified by level of care (PHC and DH) and HIV status (PLHIV, people without HIV or not known to have HIV, and all individuals, regardless of HIV status) for aggregated analyses. Extensive sensitivity analyses will account for parameter, model and decision uncertainty, ensuring robust findings across different demographic and epidemiological settings.

A limitation of the study is the exclusion of patient-incurred costs, due to the lack of primary data and limited up-to-date WHO TB Patient Cost study coverage in the recruiting countries.² While this restricts patient-level cost inclusion, the number of false positives and missed cases will be reported to support algorithm prioritisation. Another limitation is that costs associated with FBCF will not distinguish between PCF and ICF, as comparing them falls outside the scope of this study. However, this will not impact key outcomes, since no direct comparison is being made between these approaches or across diagnostic settings. As our aggregate analysis seeks to support the evaluation of algorithms rather than provide a nominal estimate of health system costs across all countries, implementation costs will be included for individual country analyses only. A final limitation is that this study is not measuring patient-important outcomes including time to treatment initiation or treatment outcomes, but such analysis is planned for in further studies under Start4All.

Although primary cost data are limited, cost assumptions will vary by care level and diagnostic setting. Commodity and implementation costs will be considered in country analyses and CBCF implementation costs will be estimated based on data obtained from four countries (Bangladesh, Cameroon, Nigeria and Vietnam). Robust cost estimates will be generated using procurement data combined with literature, with PSA accounting for heterogeneity and uncertainty at three levels: costing assumptions, cost per test and cost per algorithm.

Our cost estimates include many uncertainties, such as over-estimating non-capital costs, but underestimating commodity costs overall by using ex-works estimates. However, since the same approach will be applied across all model runs, this limitation is unlikely to affect comparative results. CBCF implementation cost per person will be held constant across algorithms, assuming test-specific cost differences cancel out. Multiple model runs will reflect variations in costs across settings.

The model does not include socioeconomic equity, gender disaggregation, drug-resistance disaggregation or child populations due to scope and data limitations. However, specific analyses will be conducted for CBCF settings in four countries, including informal settlements, nomadic populations and refugees/IDPs. Separate models for PLHIV will help identify algorithms particularly suitable for this group.

WTP and WTA thresholds will be applied based on the position of each algorithm on the CE plane. This allows for a nuanced evaluation of trade-offs between cost and diagnostic performance. Algorithms with lower performance, but substantial cost savings may be attractive, especially in resource-limited settings. The benchmark used for WTP and WTA in this study is the provider cost of treating one person with DS-TB, reflecting the trade-off between early detection and the higher costs of treating advanced or drug-resistant TB. This approach aligns with global policy goals prioritising early detection to reduce transmission and avoid costly treatments.

An initial threshold of 60% probability of CE will guide decision-making, balancing uncertainty with a reasonable level of confidence for policy-making. Sensitivity analyses will explore the impact of lower thresholds to assess the robustness of conclusions under greater uncertainty.

The study does not include treatment costs or model TB transmission effects, which are important limitations. However, WTP and WTA thresholds implicitly account—at least in part—for long-term treatment and transmission costs, offering a practical way to evaluate economic trade-offs. These thresholds allow exploration of policy priorities and budget constraints, making the model relevant across diverse settings.

While our modelling results are intended to inform policy, guide implementation and support further research on the CE of TB diagnostic algorithms, we acknowledge that CE is only one of several important considerations. Factors such as feasibility, scalability and acceptability are equally critical for successful real-world implementation.

Ethics and dissemination

Ethics approval has been granted for phase I only. Oversight is provided by country ethical research committees, LSTM and Stop TB Partnership, as shown in table 5.

Table 5. Master and country ethical research committees’ identification codes.

Body	Number
LSTM REC	22–084
WHO ERC (master protocol)	ERC.0003921
WHO ERC Bangladesh	ERC.0004062
WHO ERC Brazil	ERC.0004061
WHO ERC Cameroon	ERC.0004014
WHO ERC Kenya	ERC.0004079
WHO ERC Malawi	ERC.0004033
WHO ERC Nigeria	ERC.0004062
WHO ERC Vietnam	ERC.0004084

ERC, ethical research committee; REC, Research Ethics Committee.

Informed consent was obtained from all participants, and all data were anonymised prior to being accessed by the study authors.

The results of this study will be disseminated through publication in peer-reviewed open-access journals, presentations at national and international scientific conferences, and summaries shared with relevant stakeholders and participants where appropriate.