Transmission model

We developed a dynamic transmission model of TB, illustrated in Fig 1 and described in further technical detail in S1 Text. The model captures key features in the natural history of TB, including asymptomatic but infectious disease, smear status, TB mortality, and spontaneous cure. The model focuses on pulmonary TB and disregards extrapulmonary TB, assuming ACF to focus on the former. Amongst those with TB (indicated by the orange box in Fig 1), we assumed that smear-negative pulmonary TB cases are, on average, 20% as infectious as smear-positive ones [9] (see Table 1 for model parameters). We incorporated ‘passive’ TB service delivery, including the split between public and private sectors that is a major feature of the health system in India [12]. Additionally, the model also distinguishes drug-susceptible TB (DS-TB) and drug-resistant TB (DR-TB), the latter including both rifampicin-resistant and multi-drug-resistant forms of TB. Although DR-TB only accounts for approximately 5% of the TB burden in India, it also consumes a disproportionate amount of the TB budget due to the high cost of second-line treatment [1]. In the model, we account for the delay in recognition of DR-TB as a result of missed opportunities for drug sensitivity testing, as well as the potential impact of Xpert-based ACF in reducing this delay.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic illustration of the model.

The tuberculosis (TB) transmission model distinguishes TB by smear status and by symptom status. Upon developing symptoms, symptomatic individuals seek care through either the private or public sectors (‘passive’ TB services) after a certain delay, estimated to match data in Table 2. Although not shown here for clarity, the model captures these sectors separately, including the lower standard of TB care in the private sector (see S1 Text for full model details). Those successfully diagnosed initiate TB treatment; we assume that 15% of diagnoses in the public sector are conducted with Xpert, the remainder by microscopy. All those lost from the TB care cascade, whether because of missed diagnosis, pre-treatment loss to follow-up, or failed treatment, temporarily disengage from care-seeking, before once again seeking care after a given delay. Compartments shown in orange denote the effect of an active case finding (ACF) intervention on this ‘passive’ system; we assume that ACF consists of initial symptom screening, followed by microbiological confirmation. Meanwhile, the separate compartments on the right represent a subset of the general population that may be detected by the ACF intervention (represented by the orange compartments) because they have TB-like symptoms, but without TB: These may include, for example, individuals with chronic obstructive pulmonary disease, bronchitis, and other lung conditions. They incur a cost to the health system for diagnosis and—if they are mistakenly diagnosed with TB—a cost in false-positive treatment. The number incorrectly identified as having TB is dependent on the specificity of both the screening and confirmatory stages. Finally, at any stage individuals may die of natural causes or of TB (in the diseased compartments) or recover spontaneously. For simplicity, these transitions are not shown in the figure (see S1 Text for full model details).

https://doi.org/10.1371/journal.pmed.1003456.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. List of model parameters.

https://doi.org/10.1371/journal.pmed.1003456.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Data used to calibrate the compartmental model.

https://doi.org/10.1371/journal.pmed.1003456.t002

We separately simulated the dynamics of a non-TB symptomatic (NTS) population, i.e., individuals who would be eligible for a TB diagnosis based on the presence of TB-like symptoms, but who do not have TB. This population is necessary for tracking the overall number of diagnostic tests being conducted during the case-finding intervention, as well as the unnecessary treatment of false-positive TB [13].

Model calibration

The model described above was calibrated to annual risk of Mycobacterium tuberculosis infection and prevalence estimates representative of urban slums in India (Table 2). To inform model parameters for symptom onset from asymptomatic stages, and for progression of smear status with developing disease, we drew from a prevalence survey in Chennai [5]. In particular, we calibrated the model to the proportion of prevalent TB cases that are symptomatic and the proportion smear-positive amongst those who had not yet presented for care, stratified by symptom status (Table 2).

Uncertainty was estimated using Latin hypercube sampling to sample model parameters from their respective ranges (10,000 samples) and simulating the model to 2018, as described in S1 Text. Parameter sets were accepted in the sampling if they provided model projections that fell within the calibration target ranges, while the rest were rejected.

Intervention

Using the calibrated model, we simulated an ACF algorithm using symptom screening followed by a confirmatory microbiological test. We assumed that all TB diagnoses through ACF would be microbiologically confirmed, i.e., we assumed that there is no role for presumptive clinical diagnosis in ACF. For the screening stage, we modelled the use of symptom screening, with individuals reporting any symptom suggestive of TB deemed eligible for microbiological confirmation. For the sensitivity and specificity of this approach for TB, we drew from a recent meta-analysis of screening approaches [14]. We also performed sensitivity analyses (described below) under an alternative screening approach: identifying those with prolonged cough, rather than any TB symptom (with prolonged cough screening having higher specificity but lower sensitivity than ‘any TB symptom’).

For the confirmatory test, we distinguished 2 types of strategy: a ‘moderate accuracy’ strategy, using a confirmatory test with performance and cost consistent with smear microscopy, and a ‘high accuracy’ strategy, using a confirmatory test with performance and cost consistent with Xpert (see Table 1 for parameters). The reason for these designations is that—although there exists data from clinical trials on the performance of smear microscopy and Xpert—there is currently little evidence for characteristics such as the specificity of either test in real-world conditions. In particular, test specificity is often assessed against culture as a reference standard and thus does not address how to interpret smear-positive, culture-negative cases that nonetheless show clear signs of having TB [15,16]. We do not address the important question of how reference standards might be improved to meet these challenges. Instead, our analysis casts light on which types of uncertainty are most critical for future studies to address in strategic planning for ACF activities. Consistent with the capabilities of smear microscopy and molecular tests such as Xpert, we assumed that the high-accuracy test can detect rifampicin resistance at the same time as detecting TB, while the moderate-accuracy test cannot. We assumed that all individuals recognised as having rifampicin resistance are immediately initiated on second-line therapy, while those with unrecognised rifampicin resistance are only switched to second-line therapy after failing first-line treatment.

We considered ‘sensitivity’ as the proportion of TB cases that yield a positive microbiological result under a given test, and likewise ‘specificity’ as the proportion of individuals without TB who yield a negative microbiological result. Moreover, we assumed that all smear-positive cases test positive through a smear-like test, and that no smear-negative cases do (i.e., sensitivities of 1 and 0, respectively). This is a simplification for model purposes: We also performed a sensitivity analysis under alternative assumptions, where smear status is associated with a given probability of diagnosis.

We simulated the intervention at different levels of coverage to run from 2020 to 2035, scaled up linearly over the first 4 years (2020–2023), ultimately to screen a given proportion of the population per year. We assumed for simplicity that this proportion is selected at random from the slum population each year. The impact of ACF was measured as the percentage reduction in cumulative incidence between 2020 and 2035. Finally, we calculated the incremental costs of the intervention relative to a baseline of status quo, i.e., current standards of TB care continued indefinitely, without ACF.

Economic evaluation

We focus on the programmatic perspective, i.e., considering only costs borne by the TB programme, and ignoring broader societal costs. The model accounts for ‘passive’ TB services, i.e., those routine services (diagnosis and treatment, both first- and second-line) independent of ACF, and contingent on patients presenting for care. Since these services could be lowered by the rollout of ACF, this will impact the overall incremental costs. Concentrating on programmatic costs, we counted only the costs of diagnosis and treatment in the public sector, and not the private sector. Moreover, new second-line regimens introduced in 2017 present a range of possible costs for DR-TB treatment. However, as the focus of the current study is on ACF, for simplicity we assumed costs and outcomes consistent with currently used regimens (Tables 1 and 3), while performing sensitivity analyses with respect to this assumption. For all unit costs, for simplicity we assumed uncertainty intervals of ±20%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Unit (service) costs used in the analysis.

https://doi.org/10.1371/journal.pmed.1003456.t003

For the ACF costs, we assumed that all patients identified through ACF are linked to treatment in the public sector. We focused for simplicity on service costs related to the numbers of confirmatory tests and patient-months of first- and second-line treatment incurred by the intervention. In doing so, we are neglecting the ‘intervention costs’ needed to facilitate the delivery of these interventions (e.g., human resources). As discussed below, these are important cost components in any exercise aiming to estimate the full costs of ACF [17]. For the purpose of the present study, however, if these intervention costs are similar for moderate- and high-accuracy testing, they may not substantially impact a comparison in cost between the two.

Finally, we tracked the costs of treating false-positive TB diagnoses, as driven by the imperfect specificity of the screening and diagnostic tools involved, and the low prevalence of TB (<5% amongst those with symptoms, even in urban slums). Unnecessary TB treatment carries heavy societal costs, including avoidable stigma [13], as well as the monetary costs to patients and households (e.g., travel costs) involved in completing a regimen of TB treatment [18]. Such factors are outside the scope of our current study: Here we focus on unnecessary programmatic spending on the treatment of false-positive TB, recognising (as discussed below) that this approach represents only one narrow part of the overall adverse effects of false-positive treatment [13]. All unit costs are provided in Table 3. Overall, we caution that our estimates should not be interpreted as representing the overall cost of any ACF intervention, given that they miss these cost components, as well as the implementation costs described above.

Sensitivity analysis

To test model sensitivity to alternative screening strategies, we first simulated the impact on incidence and incremental costs of using prolonged cough for symptom screening, which has higher specificity and lower sensitivity than the symptom screening strategy used in the main analysis, any symptom suggestive of TB. For additional sensitivity analyses, as a focal model output, we calculated the cumulative cases averted between 2020 and 2035 under an assumed budget of US$20 million, calculating the ratio of cases averted between the high-accuracy and moderate accuracy scenarios. This ratio offers an estimate of the relative cost efficiency of the 2 approaches. We examined model sensitivity to individual parameters by conducting a partial rank correlation between this model output and each of the model inputs listed in Table 1. We additionally evaluated the focal model output under different scenarios: (i) assuming that smear microscopy can detect 25% and 75% of smear-negative and -positive cases, respectively (as opposed to 0% and 100% in the main analysis) and (ii) using alternative scenarios for the burden and management of DR-TB, which can consume a disproportionate share of programmatic spending. We modelled a scenario consistent with Mumbai, which has a substantially higher DR-TB burden than the national average. We also modelled a scenario with the adoption of new, more effective, and less costly second-line regimens.

Planning and execution of methods

At the outset of this analysis, the modelling plan focused on the confirmatory test, assuming the use of ‘any symptom suggestive of TB’ as a screening strategy. Subsequently, our main adjustment to this initial plan was in response to constructive reviewer comments, on the need to better understand the influence of the screening algorithm. In response, we incorporated the additional analysis as described above, on the alternative use of prolonged cough as a screening strategy. Also in response to reviewer comments, we incorporated the additional sensitivity analyses described above, for the alternative scenarios for the burden and management of DR-TB. Our initial modelling plan otherwise underwent no data-driven changes.

Quantifying specificity

We note that quantifying ‘true’ specificity is a complex challenge, partly as the culture reference standard (against which specificity is judged) also has imperfect sensitivity, raising the question of how to interpret smear-positive, culture-negative results, particularly amongst those with strong clinical signs of TB. As initial steps in this direction, future ACF implementation research could aim, for example, to supplement microbiological reference standards with ‘composite’ reference standards that additionally incorporate clinical diagnosis and patient assessment on follow-up, including response to anti-TB treatment. Our results also have implications beyond the particular diagnostic tests being examined here. For example, a new generation of molecular diagnostic test, Xpert Ultra, has higher sensitivity than Xpert, but also lower specificity [6]. Its use in ACF would lead to similarly unacceptable numbers of false-positive TB treatments, as those illustrated in Fig 5A.

Role of screening

Our sensitivity analysis shows that the performance of the screening stage can also play an influential role, although whether this widens or narrows the gap between the 2 diagnostic tests (at a given budget) depends on the relative magnitudes of (i) the reduction in false-positive TB treatments arising from a higher-specificity screening algorithm and (ii) the long-term second-line TB treatment costs that can be averted by a high-accuracy test (Fig C and Table A in S1 Text). In our current work, the overall effect is for our qualitative findings to remain unchanged. Strategies such as X-ray screening (not modelled in the current study) can show greater sensitivity and specificity than symptom screening alone [23], while allowing TB detection amongst those not reporting symptoms. Implementation of X-ray screening is more resource-intensive than symptom screening, but could be facilitated by the use of mobile radiography units [24,25], along with newly emerging technology for automated X-ray reads [26,27]. While the present analysis has focused on confirmatory tests, a more systematic exploration of these and other screening strategies will be an important area for future work.

The patient perspective

Focusing on programmatic costs, our analysis does not address the important issue of patient costs associated with TB [28–30]. In the context of routine TB services, the costs of care-seeking and TB treatment can have a substantial impact on productivity and household income, and are an important cause of catastrophic health expenditure [29]. Previous work in India has shown that ACF can bring about substantial reductions in these patient costs [18], essentially by bringing TB services to those in need in a timely way. By neglecting patient costs, our analysis therefore does not capture the societal cost savings that would result from higher-sensitivity testing strategies. Our analysis also does not address the societal costs of false-positive TB diagnosis, including the potentially life-changing impact of stigma [13] and the potential side effects of TB treatment. Our analysis thus does not capture the societal cost savings that would result from higher-specificity ACF approaches. Overall, therefore, we expect that inclusion of societal costs would act to widen the gap between the strategies shown in Fig 3. The extension of our analysis to incorporate these important costs to the patient is an important area for future work.

Key considerations in intervention costs

Our study has focused on the performance of confirmatory tests in ACF but has not addressed the implementation of these tests. For example, the overall cost of an Xpert test will depend on whether Xpert units are readily accessible for use in ACF, or whether they are only available in central laboratories, requiring the additional expense of sample transport and relaying test results to the patient. Previous work has addressed these considerations in the context of routine TB services [31]. Although such considerations are outside the scope of our current analysis, we note that access to Xpert facilities is likely to be a more pressing concern in peripheral health facilities in rural India than in urban slums. India’s current National Strategic Plan includes measures to improve the capacity for Xpert testing nationwide [32]. Moreover, recent developments offer prospects for reduced reliance on laboratory infrastructure, for example with the development of new, more mobile molecular diagnostics such as Xpert Omni [33] and Truenat [34], as well as the deployment of Xpert MTB/RIF through mobile diagnostic vans [35]. All of these developments would tend to reduce the per-test cost of high-accuracy diagnostic tools, by allowing closer proximity of testing to the ACF intervention. For any study aiming systematically to estimate the full cost of ACF, these and other implementation factors will be important to take into account (e.g., as demonstrated in [17,31]). Nonetheless, our findings illustrate an important consideration in any such study: In addition to these important cost components, the specificity of the confirmatory test, along with that of the whole ACF cascade, is likely to be a key driver of the cost/impact ratio of any ACF intervention.

Our analysis has ignored intervention costs directly associated with the rollout of the intervention, such as human resource costs and costs associated with the purchase of equipment. We also take a simple approach of assuming unit costs that are unaffected by the scale of the intervention, thus ignoring the potential for economies of scale, especially at high levels of population coverage. If human resource costs are similar under the 2 strategies, they would have little effect on the relative costs of these strategies. Overall, however, we emphasise that our estimates should not be interpreted as estimates of the actual cost of ACF. An important area for future work would be to incorporate these important elements [17,36], for more comprehensive estimates of the full cost of ACF.

Model limitations

For simplicity, we ignored pre-treatment loss to follow-up. Although commonly observed in practical implementations of ACF [37], we do not expect this simplification to alter our essential findings, as long as it affects the 2 testing strategies roughly equally. Again, for simplicity, we assume that all false-positive TB diagnoses have the same rates of treatment completion as true-positive diagnoses; if their completion rates are substantially lower, this would tend to reduce the treatment cost associated with false-positive TB diagnosis. Finally, we note that there is considerable uncertainty about the potential transmission impact of ACF interventions, owing partly to a lack of direct evidence, particularly in South-Asian settings [4]. In the present analysis, our impact projections are based on a series of assumptions, including perfect implementation of symptom screening and diagnosis, high-quality engagement with and participation by the community, effective linkage to treatment, immediate rapid bacteriological suppression upon treatment (thus interrupting transmission), and assumptions about the degree to which ACF is able to identify TB cases while they still have substantial potential for passing on infection [4]. Lessons from current and future ACF initiatives will be invaluable in identifying which of these assumptions needs most attention.