Model structure

We developed a deterministic compartmental model of TB transmission amongst adults (>15 years old) in South Africa, incorporating the role of HIV in driving TB dynamics. The overall model structure is illustrated schematically in Fig 1. We did not aim to model the dynamics of HIV separately as we were interested in the effect of LAM tests on the TB epidemic, and not the HIV epidemic; thus, for the purpose of this analysis, we took the incidence of HIV, the proportion of people living with HIV (PLHIV) with and without TB, their CD4 cell counts, and the coverage of antiretroviral therapy (ART) over time as given inputs for the model. Doing so allows us to capture the role of projected future ART coverage, and HIV trends, in the future trajectory of the TB epidemic in South Africa. We also developed a similar model of TB transmission in Kenya, a country with a lower HIV burden.

Download:

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

Fig 1. Schematic illustration of the model structure.

(A) represents tuberculosis (TB) transmission dynamics in the general community (centre) and hospitalised patients living with HIV, represented by superscript ‘inp’ (right). Model compartments are as follows: uninfected with TB (U), latent infection (L), active TB disease (I), presented for care and awaiting diagnosis (Dx), on TB treatment (Tx), temporarily dropped out of the TB care cascade due to missed diagnosis or pre-treatment loss to follow-up (E), recovered from TB (R), and inpatient admissions (A). Each compartment is stratified further by HIV status, h, defined in Table B in S1 Text. (B) illustrates the transitions between the different HIV states, including antiretroviral therapy (ART) state and CD4 count level (cells/μl). The following states are also further stratified into pulmonary TB and extrapulmonary TB: I, Dx, Tx, E, R^TC, R^SC_, Dx^out, I^inp, A^inp, Tx^inp, E^inp, R^TC_inp, and R^SC_inp. Arrows represent transitions between states, at the per capita rates listed in Table 2. Coloured-dash arrows in (A) illustrate movement of individuals with active TB disease into outpatient status (orange), out of outpatient status (green), and into inpatient status (blue). See S1 Text for further technical details, including model equations and calibration.

https://doi.org/10.1371/journal.pmed.1003466.g001

Download:

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

Table 2. List of model parameters and assumptions.

https://doi.org/10.1371/journal.pmed.1003466.t002

Patients with extrapulmonary TB are often misdiagnosed if tested with sputum-based diagnostics (Table 2). To capture the advantages of LAM testing for diagnosing TB amongst these patients, in the model we distinguished extrapulmonary and pulmonary TB, while assuming that only the latter contribute to transmission. The sensitivity of LAM tests depends on the extent of HIV infection, and in particular the CD4 cell count [12]. Accordingly, amongst those with HIV, we modelled 3 different CD4 cell count strata: those with a CD4 count > 200 cells/μl, those with a CD4 count between 100 and 200 cells/μl, and those with a CD4 count < 100 cells/μl. The model captures the rate at which those with HIV progress through declining CD4 counts, during the course of infection. We also incorporated HIV-associated hospitalisation, assuming CD4-dependent hazard rates of admission into hospital, and further assuming that upon admission, any ART-naïve patients are initiated on ART. We also captured the provision of HIV care in outpatient settings, assuming CD4-dependent rates of ART initiation in these settings. The model does not explicitly capture rifampicin-resistant or multi-drug-resistant TB, as these forms account only for 3%–4% of overall TB burden in South Africa.

Data sources

We drew from WHO data for estimates of TB incidence and mortality in South Africa, along with reported notifications and treatment outcomes [2]. For past HIV trends, we drew estimates from UNAIDS [37] for annual HIV incidence, the proportion of HIV cases being initiated on ART each year, and the proportion of those on ART being virally suppressed. For future projections, we drew from the Thembisa model [42], an HIV modelling framework that is the source of UNAIDs estimates (S5 and S6 Figs). Table 3 summarises the data for South Africa from these sources.

Download:

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

Table 3. Calibration targets for South Africa used to estimate model parameters.

https://doi.org/10.1371/journal.pmed.1003466.t003

We performed a literature search to identify the proportions of PLHIV in the 3 different CD4 strata described above, stratifying by 3 different population types, in line with WHO recommendations: (i) those initiating ART upon admission to hospital, (ii) those initiating ART as outpatients, and (iii) those who are not on ART. S2 Fig summarises the proportions thus extracted, and the sources of data used. These data inform model estimates for the timeliness of HIV treatment as follows: Early treatment yields a patient population with higher CD4 counts at the point of treatment initiation, and, conversely, late treatment is associated with a patient population having lower CD4 counts (Table 3; S2 Fig). Therefore, by fitting the model to simultaneously capture CD4 progression and the distribution of CD4 counts at treatment initiation, we estimated the rate of treatment initiation at different CD4 counts, both in outpatient and inpatient settings (Table 3; S2 Fig). We additionally collected data from the literature for other parameters: for example, for the proportion of HIV cases that are hospitalised per year [39] and for the current standard of TB care amongst HIV inpatients (e.g., the proportion of hospital admissions receiving a TB test in routine practice) [10].

Using adaptive Markov chain Monte Carlo (MCMC), in particular the adaptive algorithm first proposed by Haario et al. [54], we incorporated uncertainty in model inputs (Table 2), propagating this input uncertainty into uncertainty in model projections. We drew 5,000 samples from the posterior distribution. For any model projections based on these samples, we estimated uncertainty intervals using the 2.5th and 97.5th percentiles, referring to this estimate as the Bayesian credible interval (CrI). Further details on the model structure and calibration are given in S1 Text.

Intervention scenarios

We distinguished ‘current’ and potential ‘future’ LAM tests. For the performance of the former, we drew from a systematic review of LF-LAM [12]. For the latter, we took SILVAMP-LAM as an illustrative example, drawing from a recent study that estimated the performance of this test in HIV-negative patients [14]. This same study highlighted the potential of future LAM tests to have improved performance compared to SILVAMP-LAM: Our parameters for future LAM tests could therefore be interpreted as a lower bound for their performance [9] (Table 4). We assumed that differences in test performance between inpatients and outpatients are driven primarily by variations in CD4 distributions between these populations, variations that are captured by the model. Accordingly, we concentrated on study findings stratified by CD4 status, rather than by inpatient or outpatient setting.

Download:

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

Table 4. Test sensitivities.

https://doi.org/10.1371/journal.pmed.1003466.t004

A test’s epidemiological impact (i.e., on incidence and mortality) is driven by its sensitivity, or the proportion of true TB cases it can detect. A test’s specificity, or the proportion of those without TB that it correctly diagnoses as negative, has no direct bearing on its epidemiological impact, and instead is more relevant for the number of unnecessary TB treatments incurred as a result of false-positive diagnosis [55]. As the focus of the current work is on epidemiological impact, we concentrated on test sensitivity and not specificity. Our model therefore does not address the adverse consequences of false-positive results, including additional costs and adverse treatment side effects.

We modelled the deployment of current and future LAM tests in 2 intervention scenarios reflecting models of use in the updated WHO guidelines [13] (Table 1). In ‘PLHIV inpatients only’ (scenario i), LAM testing is conducted in PLHIV inpatients with signs and symptoms of TB and in all PLHIV inpatients with CD4 < 200 cells/μl, independent of symptoms. In ‘PLHIV inpatients and outpatients’ (scenario ii), LAM testing is conducted in PLHIV inpatients and outpatients prior to initiating ART treatment and with signs and symptoms of TB, in all PLHIV inpatients with CD4 < 200 cells/μl, and in all PLHIV outpatients with CD4 < 100 cells/μl, independent of symptoms. Additionally, for future LAM tests alone, we modelled a hypothetical future scenario: ‘universal for all TB presumptive patients’ (scenario iii), in which a future LAM test is deployed as part of routine TB diagnosis in patients presenting with symptoms of TB to a healthcare provider, regardless of HIV status. The impact of this scenario derives from the diagnosis of HIV-negative TB, and from the diagnosis of TB amongst those with undiagnosed HIV (and who may miss the opportunity for urine-based testing under current LAM testing guidelines). To assess which of these factors are most influential for impact, we additionally simulated scenario iii under a hypothetical condition where SILVAMP-LAM has 0 sensitivity for TB in HIV-negative individuals. This artificial scenario is thus deliberately constructed so that the only incremental cases being diagnosed, relative to scenario ii, are those with HIV and not on ART.

We modelled each of these intervention scenarios, assuming them to be initiated in 2020; we assumed that the proportion of the target population accessing these tests increases linearly, until the whole population in South Africa is covered by 2023. We simulated the model forward to 2035, simulating incidence and mortality over this time.

We considered the impact of LAM testing with respect to 2 comparators: (i) a ‘status quo’ comparator scenario, where the current standard of TB care continues indefinitely (where a proportion of patients presenting to care with signs and symptoms of TB are offered an Xpert test by a healthcare provider) and (ii) an ‘Xpert scale-up’ comparator scenario, involving the scale-up of sputum-based Xpert across the country, to diagnose TB both amongst those receiving ART and in routine TB services. We assumed that all patients presumed to have TB based on symptoms receive Xpert testing by 2023 (in inpatient, outpatient, and routine care settings).

Consistent with diagnostic yields reported in the literature (Table 2), we assumed that only 50% of patients with HIV (in both inpatient and outpatient settings) are able to provide a sputum sample for Xpert testing, while 99% of patients are able to provide a urine sample for LAM testing. We also considered clinical diagnosis that occurs after a negative Xpert test result. We assumed that 20% and 30% of patients with negative Xpert test results are clinically diagnosed and offered treatment amongst HIV-negative and HIV-positive patients, respectively [32]. Using comparator and intervention scenarios, we projected estimates for the numbers of TB cases and deaths that would be averted under the intervention scenarios described above.

Finally, as a sensitivity analysis to examine the applicability of our results to other countries with a generalised HIV epidemic, we extended our model to capture epidemiological conditions consistent with Kenya, where an estimated 27% of TB is in HIV-coinfected patients, compared with 59% for South Africa [2]. There is insufficient data from Kenya to calibrate all model parameters; we therefore sought only to capture gross epidemiological indicators consistent with Kenya (TB incidence, mortality, proportion of HIV coinfection, etc.), while assuming the same values as derived for South Africa for all parameters specific to CD4 counts (rates of ART initiation, hospitalisation, etc.; see S2 Table for further details). We then simulated the intervention scenarios described above, in this Kenya-like setting.

Summary

New non-sputum-based diagnostic technologies could raise new opportunities to accelerate current declines in TB incidence [56–58]. In the current work we have addressed one example of such technologies: lateral flow LAM assays, which offer the potential to diagnose TB through urine. Given that LAM tests perform best in PLHIV [12], it was reasonable to hypothesise that they would have a substantial epidemiological impact in settings such as South Africa, where the majority of TB cases are HIV coinfected. However, our results suggest a more nuanced message: For future LAM tests to achieve population-level reductions in TB burden, they would need to be offered more widely than amongst those receiving HIV care, and in routine TB care. Under this scenario, the patients who would benefit are not just HIV-negative cases, but also those with HIV who have not yet been engaged in HIV care. The latter population could arise from a variety of factors, including the rate of HIV diagnosis or gaps in linkage to ART initiation after a diagnosis has been made. In practice, the feasibility of such wide deployment of LAM tests will depend critically on the performance characteristics of the tests concerned.

While much of the existing literature on LF-LAM addresses its diagnostic performance and implementation in defined clinical settings [9,10,12,14], our work complements this evidence basis by addressing the population level in countries with a high HIV and TB burden. To our knowledge, this is the first modelling study to examine the potential incidence and mortality impact that LF-LAM tests may offer.

Limitations and key areas for future work

Although we expect similar results to apply in other settings with a generalised HIV epidemic, our analysis does not address high-TB-burden countries where HIV is not a driving factor, such as India. The potential use of LF-LAM in such settings is an important area for further work. In these settings, we expect that the value of a future LAM test might be driven not by its ability to detect TB amongst those with HIV, but rather by its potential to be used more widely than current diagnostic tools needing laboratory capacity: for example, in primary care and in peripheral healthcare settings.

Our work also does not address cost or cost-effectiveness, another important topic for future work. Recent analysis indicated that LF-LAM and SILVAMP-LAM would be cost-effective when deployed in inpatient settings in both Malawi and South Africa [59]; future work, especially in the context of improved-performance tests, could benefit from incorporating a transmission framework. Relatedly, we have also not addressed the staffing and health system capacity that would be needed to facilitate the expansion of LAM testing in South Africa [60], nor the costs associated with undertaking CD4 cell count measurements. It should be noted that CD4 cell count measurement is increasingly being replaced by viral load monitoring. In addition, we have not addressed potential implementation challenges in the use of future LAM tests. For example, SILVAMP-LAM involves an additional step compared to LF-LAM; findings from ongoing trials will be valuable in determining whether this procedure reduces the performance of this technology to any appreciable extent.

Amongst additional model simplifications, we have not covered paediatric TB, an important and underaddressed part of TB burden. We have also taken a country-level perspective, despite wide subnational variation in TB/HIV burden within South Africa [61,62]. For simplicity we have ignored drug resistance: Despite the potential benefits of LAM tests, one notable limitation of these tests is that they cannot determine drug sensitivity. In the model calibration, we have captured temporal trends and projections in major features of the HIV epidemic in South Africa, such as HIV incidence and the proportion on ART; however, we have modelled other features as static, for example the percentage of HIV cases that are hospitalised. Further data on how this proportion has changed over time should allow our model to better capture these dynamics, potentially affecting our estimates for deaths averted amongst those being hospitalised. However, we have shown that this impact accounts for only a small proportion of overall TB burden in South Africa; we therefore do not expect these changes to affect our overall qualitative findings on the importance of widened eligibility for epidemiological impact.

Given our focus on epidemiological impact, our analysis of test performance is limited to test sensitivity. However, for any future LAM test that is intended for use in routine TB services, test specificity will be a critical performance characteristic, in order to minimise the number of unnecessary TB treatments incurred [55]. As an illustrative example, a test having 95% specificity and 90% sensitivity would—for the 10% TB prevalence that is typical amongst symptomatic individuals presenting to routine TB services—result in 1 false-positive TB diagnosis for every 2 successful TB diagnoses, an unacceptably high rate. Evidence suggests that both LF-LAM and SILVAMP-LAM may have reduced specificity at lower CD4 counts [9,13]. Estimating this risk is an important area for future work: however, an important first step in this direction is to address the real uncertainties in quantifying specificity where the accuracy of the reference standard is unclear. An example of when a reference standard may be unclear includes the use of a sputum-based microbiological reference standard amongst patients with solely extrapulmonary TB, and potentially also among patients with HIV, who are more likely to produce paucibacillary sputum [6,63]. Although a LAM test may correctly diagnose such patients, this diagnosis would be deemed incorrect by a sputum-based microbiological reference standard. For future quantitative analysis, therefore, there is a need for more systematic estimates of specificity that take account of such shortcomings of any given reference standard. Combined use of microbiological and composite reference standards (i.e., including clinical diagnosis) may help in this regard.