The impact of HIV and tuberculosis interventions on South African adult tuberculosis trends, 1990-2019: a mathematical modeling analysis

Highlights • We developed a tuberculosis (TB) transmission dynamic model for South Africa.• We quantified TB incidence and mortality attributable to HIV between 1990 and 2019.• We estimated the impact of TB programmatic interventions on TB incidence between 1996 and 2019.• A significant proportion of TB incidence and mortality was attributable to HIV.• TB screening and antiretroviral therapy led to substantial TB incidence reductions.


Introduction
South Africa is ranked among the top 20 high tuberculosis (TB) burden countries ( World Health Organization, 2019 ). The TB epidemic grew rapidly in the early 1990s, primarily driven by HIV ( Abdool Karim et al. , 2009 ). HIV infection is the strongest individual-level TB risk factor, increasing the risk of progression to TB disease and reactivation of latent TB infection, worsening treatment outcomes and increasing mortality . Although the effect of HIV on TB has been established, very few studies have quantified its population-level effect on incidence and * Correspondence to: Associate Professor Leigh F. Johnson mortality over time. South Africa has implemented TB control interventions, including directly observed therapy (DOTS), which was scaled up in 1996 ( World Health Organization, 1994 ). This strategy had multiple components, including directly observed treatment, political commitment, improved microscopy services, surveillance and monitoring, and quality treatment ( World Health Organization, 1994 ). Other interventions, which were scaled up in the mid-20 0 0s, included the provision of isoniazid preventive therapy (IPT) and antiretroviral therapy (ART) for HIV-positive individuals ( Churchyard et al. , 2014b ;Lawn et al. , 2011). Declines in TB notifications and mortality from 2008 have largely been attributed to ART, which was made widely available during the mid-20 0 0s ( Dye and Williams, 2019 ;Loveday et al. , 2019 ;Surie et al. , 2018 ). This is supported by the established individual-level effectiveness of ART in reducing TB incidence and mortality ( Lawn et al. , 2006 ;Ross et al. , 2021 ). In addition, IPT also reduces the risk of developing TB in people living with HIV ( Ayele et al. , 2015 ;Ross et al. ,  Substantial effort has also been invested in identifying TB cases in South Africa. Between 2004 and 2012, the annual number of microbiological TB tests performed doubled ( Nanoo et al. , 2015 ). Xpert MTB/RIF was introduced in 2011 to replace smear microscopy ( Di Tanna et al. , 2019 ;Theron et al. , 2014b ). Although earlier modeling studies anticipated substantial health benefits from Xpert MTB/RIF implementation compared with microscopy ( Menzies et al. , 2012 ), clinical trials have found minimal or no impact on TB mortality ( Di Tanna et al. , 2019 ). To understand these dynamics better, modeling studies with detailed diagnostic algorithms that account for empirical treatment are required.
There have been no formal analyses to quantify the contribution of the abovementioned TB interventions to the declining TB trends. Such evaluations are essential in assessing which South African TB program components are most critical to decreasing TB incidence. Therefore, we sought to (i) describe the South African TB epidemic trends between 1990 and 2019; (ii) assess the burden of TB attributable to HIV; and (iii) assess the impact of TB interventions, including DOTS, increased TB screening, Xpert MTB/RIF as an additional first-line diagnostic tool replacing smear microscopy, IPT, and ART on TB incidence.

Methods
We developed an age-and sex-structured deterministic compartmental model of TB and HIV for the South African adult population (aged 15 years and older). The core TB states are modeled following conventions described by previous studies ( Menzies et al. , 2012 ). Transitions between states include TB infection, progression to TB disease, natural recovery, diagnosis and treatment initiation, death, and treatment cure. Age-and sex-specific relative risks were applied to rates of progression to TB disease to capture age and sex differences in TB risk factors.
After cure by TB treatment, two post-treatment states dependent on time since cure are defined: short term (within 6 months after cure) and long term (6 or more months after cure). In both states, individuals are at risk of reinfection, whereas in the short term, individuals are assumed to have a high likelihood of relapse ( den Boon et al. , 2007 ). The natural history parameters are described in Table 1 . The model structure is shown in Figure 1 . A detailed description of the model is provided in the supplementary material. The risk of infection depends on the mean contact rates, age-and sex-mixing patterns ( Dodd et al. , 2016 ), the proba-bility of transmission per contact, and the prevalence of infectious TB.
The Thembisa HIV model forms the HIV component of the model ( Johnson and Dorrington, 2020 ). Thembisa is a compartmental model of the South African HIV epidemic, designed to answer policy questions relating to HIV prevention and treatment. The model is age-and sex-structured, and the HIV epidemic is simulated dynamically from 1985. Subpopulations who are HIVpositive are further stratified by HIV testing history, CD4 count, and duration since ART initiation. The model also captures changes in the ART guidelines over time ( Johnson and Dorrington, 2020 ) and is calibrated to South African data. HIV is assumed to affect the TB natural history parameters. These HIV effects are modeled as relative risks, which vary by CD4 count and duration since ART initiation ( Table 1 ).
IPT is modeled for individuals with latent tuberculosis infection who are HIV-positive and eligible as per guidelines ( Department of Health, 2013 ). Uptake depends on CD4 count, ART duration, and latent TB status. IPT uptake started in 2010; the number of IPT initiators was obtained from the District Health Information System. Assumptions on IPT duration, completion, and efficacy are in the supplementary material.
The assumed health-seeking patterns in the model are based on South African studies. We assumed different health facility attendance rates for individuals: (i) with TB, attending health facilities due to TB-related symptoms; (ii) without TB, attending due to other health conditions; and (iii) without TB, attending due to TB-like symptoms. We assumed females were more likely to seek care than males ( Horton et al. , 2016 ), and individuals who are HIV-positive had higher health-seeking rates than individuals who are HIV-negative ( Corbett et al. , 2003 ). Individuals who are smear-positive experience more TB symptoms than individuals who are smear-negative ( Onozaki et al. , 2015 ). We consider smear microscopy and Xpert MTB/RIF as the first-line diagnostic tools, accounting for the phased implementation of Xpert MTB/RIF from 2011 onwards. To estimate the numbers of true-and false-positive TB diagnoses, we specified the sensitivity and specificity of these tests. After initial negative test results, we assumed a proportion of individuals were followed up for a second test by culture.
The model also allows treatment initiation in a proportion of individuals who do not have laboratory-confirmed TB (empirical treatment). Health-seeking parameters and rates of TB screening are estimated through calibration by fitting the model to the number of microbiological tests performed ( Nanoo et al. , 2015 ) and number of cases treated. Once individuals start the 6-month TB treatment course, the following outcomes are considered: cure,  failure, discontinuation, and death (with the rates of cure and failure depending on treatment discontinuation rates). Treatment outcome assumptions are based on the electronic TB treatment register (ETR.net) data for drug-susceptible TB and are shown in the supplementary material.
We simulated posterior distributions numerically using incremental mixture importance sampling, that is, using importance sampling to draw a sample of parameter combinations from regions of the parameter space that yield the highest likelihood values ( Raftery and Bao, 2010 ). The means for the model estimates were calculated over 10 0 0 posterior samples, and 95% confidence intervals (CIs) were calculated by taking the 2.5th and the 97.5th percentiles of the posterior sample (supplementary materials).
We performed a sensitivity analysis to assess how the model inputs that varied in the calibration process were correlated with the estimated TB incidence and mortality for 2019. • We assumed the relative rate of treatment discontinuation was 0.48 under DOTS (Pasipanodya and Gumbo, 2013). • Xpert MTB/RIF is assumed to be more sensitive than microscopy, but is associated with reduced empirical treatment. • ART is assumed to reduce TB incidence and mortality, through both direct effects on viral load, and indirect effects on CD4 count ( Table 1 ). • IPT is assumed to reduce TB incidence by 52% in latently-infected adults ( Ayele et al. , 2015 ).

B
To assess the burden of TB attributable to HIV, we simulated a scenario with no HIV infection.
HIV transmission probabilities were set to zero, so that there was no HIV epidemic. C To assess the impact of DOTS, we simulated a scenario without DOTS.
Treatment discontinuation rates held constant (no reduction due to DOTS). D To assess the impact of IPT, we simulated a scenario where no IPT is implemented.
The number of HIV-infected individuals initiated on isoniazid preventive therapy in each year was set to zero. E To assess the impact of ART, we simulated a scenario where there is no ART.
Annual numbers of ART initiations are set to zero. To assess what would have happened without any programmatic changes, we simulated a scenario without any interventions in C to G.
Including all changes described in C-G. ART = antiretroviral therapy; DOTS = directly observed therapy; IPT = isoniazid preventive therapy.

Model experiments to assess the impact of HIV and programmatic interventions over time
To quantify the effects of HIV, DOTS, increased TB screening, Xpert MTB/RIF, ART, and IPT on TB incidence and mortality, we ran the scenarios A-H described in Table 2 . Each of the counterfactual scenarios B-H was compared with baseline scenario A to assess the change in TB incidence and mortality attributable to the relevant factor.

Results
The estimated number of TB deaths for males and females was consistent with the recorded number of deaths. The estimated deaths rapidly increased from 1994, peaked in 2006, followed by a decline to 31,0 0 0 (95% CI 30,0 0 0-33,0 0 0) and 21,0 0 0 (95% CI 21,0 0 0-22,0 0 0) in 2019, in males and females, respectively ( Figures 2 a and 2 b). The model estimates for the numbers of people starting treatment were slightly inconsistent with the data. Before 2010, the model overestimated the number of females initiating treatment. After 2008, the model underestimated the number of males initiating treatment ( Figures 2 c and 2 d). The estimated TB prevalence was reasonably close to the results of the 2018 TB prevalence survey; in 2019, the estimated TB prevalence was 1.02% (95% CI 0.97-1.06%) and 0.6% (95% CI 0.52-0.57%) in males and females, respectively ( Figures 2 e and 2 f).
In the counterfactual scenario, without HIV, the model estimated that TB incidence and mortality would have remained relatively low ( Figure 3 ), although still high enough for South Africa to be classified as a high TB burden country.
Reductions in TB incidence due to DOTS and IPT were small ( < 3%) in all years ( Figures 4 a and 4 b). On the other hand, the reduction in TB incidence due to ART was evident from 2006. The impact of ART increased monotonically until 2019, with a reduction of 20.0% (95% CI 19.2-20.7%) in TB incidence ( Figure 4 c).
Apart from DOTS and IPT, most interventions had a greater impact on TB mortality than on TB incidence (supplementary material, Figure 17). For example, reductions in mortality in 2019 due to interventions were 37.6% (95% CI 35.8-39.5%) for ART, 37.9% (95% CI 37.1-38.6%) for scaled-up TB screening, and 3.2% (95% CI 2.4-  Most of the model input parameters had the expected relationships with the outcomes (shown by correlation coefficients and scatter plots). However, because several parameters were varied simultaneously in the calibration process, some counterintuitive associations need to be interpreted in terms of correlations between model parameters (for further discussion, see section 18 of the Supplementary materials).

Discussion
HIV has had a devastating impact on TB incidence and mortality. Between 1990 and 2019, 8.8 million South Africans developed TB, and 2.1 million lives were lost. HIV caused 55% (4.8 million) of the TB cases and 69% (1.4 million) of TB deaths. We also showed that interventions implemented by the South African TB program have led to notable reductions in TB incidence, with ART and increased screening contributing to most of the decline. Our model also showed that the other interventions-DOTS, IPT, and Xpert MTB/RIF-had modest impacts on TB incidence. For most of the interventions (increased screening, ART, and Xpert MTB/RIF), the impact on TB mortality was proportionately greater than the impact on TB incidence (Supplementary material).
Although HIV is the strongest driver of the TB epidemic, our model estimated that even in the absence of HIV, TB incidence in South Africa would remain high. This was demonstrated in the no-HIV counterfactual scenario, in which there were an estimated 235 cases per 10 0,0 0 0 population in 2019. This rate is much higher than the estimated TB incidence for industrialized regions, such as Europe and America ( World Health Organization, 2019 ). The high TB burden in the HIV-negative population indicates other underlying factors that drive the epidemic (i.e., low rates of diagnosis and risk factors that increase susceptibility to TB disease).
The provision of ART substantially impacted TB incidence; in 2019, it led to a 20% reduction. The benefits of ART in reducing incidence ( Dye and Williams, 2019 ;Nanoo et al. , 2015 ;Surie et al. , 2018 ) depend on CD4 count and duration of ART (Lawn et al. , 2011)-individuals who are HIV-positive who initiate ART earlier at higher CD4 counts and stay on ART for longer experience the greatest benefits of ART. In the model, the effect of ART on reducing TB incidence increased during the mid-20 0 0s, when access to ART expanded in South Africa. Over time, the CD4 count threshold at which individuals can start ART has increased ( Meintjes et al. , 2017 ), and average ART durations have increased, consequently contributing to the substantial reduction in the population-level TB incidence.
Intensified TB screening also led to significant declines in TB incidence. Between 2005 and 2012, South Africa scaled-up effort s to identify TB cases and testing rates doubled ( Nanoo et al. , 2015 ). As a result, there were rapid reductions in TB incidence owing to increased TB screening during this period. In 2019, increased screening led to a 28% reduction in TB incidence.
The reasons for DOTS having minimal impact on TB may include high ongoing Mycobacterium tuberculosis transmission rates, high prevalence of substantial risk factors, such as HIV, which increase progression to disease, and the emergence of resistant TB ( De Cock and Chaisson, 1999 ;Wood et al. , 2011 ). Lastly, we have only considered one component of the DOTS strategy; considering other aspects could have led to a larger impact. Nonetheless, our findings of the minimal impact of DOTS align with studies that suggested that DOTS would have minimal impact in settings with a high HIV burden ( Dowdy and Chaisson, 2009 ;Dye et al. , 1998 ).
The small population-level impact of IPT on TB incidence in individuals who are HIV-positive is consistent with other epidemiologic analyses, which attribute the limited impact to the low implementation of IPT in South Africa ( Dye and Williams, 2019 ). However, other reasons may be because this HIV-positive population has a high risk of progression to disease or because IPT does not necessarily cure latent Mycobacterium tuberculosis infection in individuals who are HIV-positive ( Houben et al. , 2014 ). In addition, there appears to be minimal protection from IPT after IPT discontinuation or completion. Thus, the short duration of IPT protection might explain the relatively modest population-level impact in South Africa ( Churchyard et al. , 2014a ). Another possible reason may be that ART eligibility criteria have changed over time. In recent years, more people have started ART at higher CD4 counts. Those starting ART at higher CD4 counts stand to benefit less from IPT.
Our findings regarding the small effect of Xpert MTB/RIF on TB incidence align with studies that found no significant effect of Xpert MTB/RIF on TB mortality ( Di Tanna et al. , 2019 ). It has been suggested that reductions in empirical treatment offset the positive effect of Xpert MTB/RIF ( Theron et al. , 2014a ). The introduction of Xpert MTB/RIF has increased the number of microbiologically confirmed diagnoses; however, this has not equated to more new diagnoses because many cases were diagnosed empirically before adopting Xpert MTB/RIF. In addition, there was more culture testing in those testing negative under microscopy than under Xpert MTB/RIF ( McCarthy et al. , 2016 ). Another modeling group, which had initially estimated a substantial positive impact of Xpert MTB/RIF on health outcomes, conducted a reanalysis accounting for empirical treatment and the sensitivity and specificity of diagnostic algorithms ( Menzies et al. , 2015 ). The revised analysis found a reduction in the benefits, with 70% fewer disability-adjusted life years averted due to Xpert MTB/RIF ( Menzies et al. , 2015 ).
We implemented a detailed diagnostic algorithm that estimated true-and false-positives from microbiological diagnoses. We also considered the proportions of individuals who initiate treatment empirically, as informed by South African pragmatic trials and op-erational studies. As a result, we showed that Xpert MTB/RIF has indeed led to a reduction in the number of individuals without TB who started treatment (by 56%) and reduced the number of individuals who started treatment on an empirical basis (by 28%). Xpert MTB/RIF possibly has other benefits, such as reducing the time to diagnosis and time to treatment initiation; however, we did not model the effect of Xpert MTB/RIF on these endpoints. Nonetheless, we assumed that loss to follow-up before treatment initiation was lower when Xpert MTB/RIF compared with smear microscopy testing was used (supplementary material). We did not model other benefits of Xpert MTB/RIF, such as the ability to detect drug-resistant TB ( Di Tanna et al. , 2019 ).
This study was subject to several limitations. First, we only considered the adult population (aged 15 years and older). Second, owing to the lack of data on the national rollout of Xpert MTB/RIF, we relied on expert opinion regarding Xpert MTB/RIF implementation. Third, it was difficult to quantify the extent of empirical treatment before introducing Xpert MTB/RIF due to the lack of data and studies to inform our assumptions. Fourth, our model does not distinguish between symptomatic and asymptomatic TB. However, we made assumptions about the prevalence of symptoms for modeling screening algorithms. Fifth, our model did not fit the number of treatment initiations data very well, particularly in the earlier years of the ETR data, before 2010. The model estimates for treatment initiations peaked earlier than the ETR data, for both males and females, and it overestimated the treatment initiations in males. This may be because, in earlier years, there was greater under-reporting in the ETR data. Lastly, this analysis only focused on the past impact of interventions implemented in South Africa. Thus, we did not explore the potential impact of new interventions or improvements to current interventions.
To the best of our knowledge, this is the first comprehensive retrospective assessment of the impact of HIV and multiple TB interventions on the South African TB burden at a national level. This study demonstrated the tremendous effect HIV has had on TB incidence and mortality. However, even in the HIV-negative population, the TB incidence remains unacceptably high. The South African TB program has made notable efforts that have led to a significant reduction in TB incidence and mortality. Further modeling studies are needed to identify the changes to current programs that are required to accelerate these reductions in future.

Author contributions
MK, LJ, and AB contributed to the study conceptualization, analysis, and interpretation of the results. LJ and MK wrote the code for the mathematical model. LJ and AB were the study supervisors. MO curated the electronic tuberculosis register data and contributed to interpretation of results. MK wrote the first manuscript draft, and all authors critically reviewed versions of the manuscript and agreed on the final version to be submitted for publication. All data used in this study were drawn from publicly available data sources, which are described and reported in the supplementary material.

Funding
MK received PhD funding from the Fogarty International Center of the National Institutes of Health (D43 TW010559), the South African Department of Science and Technology/National Research Foundation Centre of Excellence in Epidemiological Modelling and Analysis, and the International Epidemiology Databases to Evaluate AIDS (National Institutes of Health, UO1AI069924). LJ was supported by the Bill and Melinda Gates Foundation (019496). The content of this manuscript is solely the responsibility of the authors and does not necessarily represent the views of the funders.

Ethical approval statement
Not applicable. All data used in this study were drawn from publicly available data sources which are described and reported in the supplementary material.

Declaration of competing interests
The authors have no competing interests to declare.