We adapted a previously published,[5] stochastic individual-based simulation model of the transmission dynamics of M. tuberculosis (M.tb.) and fitted the model to current data on TB and HIV epidemiology in South Africa (see S1 File for full details). Briefly, the model includes TB and HIV natural history at an individual-level, with population level incidence sustained by M.tb. transmission. Latent infection with M.tb. (which confers partial protection against progression to disease following re-infection) develops into active (sputum smear negative or positive pulmonary) TB disease at a rate that is higher shortly after (re-)infection and varies with age. The model incorporated data on age-structured social mixing patterns, [6] as well as historic and projected trends in demography [7]. Timing of development of TB disease after (re-)infection is influenced by HIV status and CD4 count; HIV status and age influence the probability of TB disease being smear-positive. The model is built using C. All parameters are found in Table A in S1 File.

In terms of the natural history of TB, individuals are born without M.tb infection; can acquire an M.tb infection which may, or may not develop into active TB disease; and individuals may, or may not have their active TB disease detected and initiate treatment. The force-of-infection for TB is proportional to the prevalence of active TB, weighted by the infectiousness of each prevalent TB case. We assume age-dependent mixing between different age groups, parameterized with data from a social contact study carried out in South Africa as part of ZAMSTAR [6]. Upon infection with M.tb, individuals’ time-since-infection is tracked and they are subject to a hazard of developing pulmonary TB disease per unit time. This hazard is higher in the first 2 years following (re-)infection, with an age-dependence following that used in [8]. In subsequent years, the hazard of activation assumes a lower constant value. Individuals with an M.tb infection have a partial protection against reinfection [9], however, a successful re-infection is treated identically to an initial infection.

Upon activation to pulmonary TB, 65% of cases in 20 year olds were assumed smear positive among those without HIV infection, following the age pattern reported in [8], smear negative TB assumed 23% as infectious as smear positive TB [10,11]. Active TB without treatment results in death (70% for smear-positive disease; 30% for smear-negative disease) or else self-cure over a timescale of 3 years based on [12]. These alternatives are modelled as proportional hazards with a Weibull-distributed time-to-event. Self-cure leaves individuals with an M.tb infection, and acts like a new infection in terms of activation risks.

Detection and initiation of treatment in those with active TB disease is assumed to occur with a probability taken as the WHO estimate of the case-detection rate [13]. The timing of detection and treatment initiation occurs at a certain fraction of an individual’s time-to-outcome without detection and treatment (i.e. more rapidly progressing TB disease is detected proportionately more rapidly).

For PLHIV, CD4+ T-cell counts of cells/mm³ (CD4) and TB risk were explicitly modelled, as was ART initiation and its effect on CD4 count and life expectancy by age. A proportion of PLHIV can be “linked to care” prior to qualifying for ART i.e. PLHIV who know their HIV status but have a high CD4 count. This group start ART at the national guideline threshold. The CD4 count at ART start for those detected as HIV positive with a CD4 count under the guideline threshold was taken from a distribution determined by data from a large set of ART cohorts [14]. This distribution has a median CD4 value at ART start of 152 cells/mm³. HIV incidence for each scenario was taken from a deterministic model of HIV transmission (see S1 File) [15]. The model was calibrated to the situation in South Africa by varying seven key parameters (see S1 File).

The model basecase scenario assumes continued implementation of control methods as used in South Africa up to mid-2013. This includes IPT to PLHIV at 10% coverage and ART to those PLHIV with active TB at 15% coverage by 2014. Coverage was assumed to rise to 80% coverage by 2015 to reflect the high adherence levels to ART guidelines [16]. In 2010, ART guidelines were revised from initiating those with a CD4 count of less than 200 cells/mm³ (2004 guideline), to < 350 cells/mm³ with coverage to this group reaching 80% by 2012 [16]. The percentage of TB patients that were detected and started onto treatment was 69% [17]. This was independent of HIV status and was composed of both probability of detection and pre-treatment loss to follow-up. Introduction and scale-up of the GeneXpert System (Cepheid, Sunnyvale, CA) was implicitly included in the basecase scenario and we assumed this had no impact on underlying parameters such as the rate at which patients start treatment as reported in the most recent South African data [18,19]. Parameter values are given in the Table A in the S1 File. The mean of 15,000 simulations was used for each outcome for each scenario. This number of simulations was used so that stochastic error was then reduced below that of the precision presented in the results and we could be confident in impact size to within 1%.

Interventions were assumed to have started together on the 1^st of January 2014 and to have had short scale up times (1 year) to factor in time for preparations,[20] whilst leaving a maximum amount of time to achieve impact before the beginning of 2016. The interventions under consideration involve three separate control measures: ART, IPT, and improved TB case management. These were grouped into four portfolios of interventions for investigation, which are outlined below and in Table 1 (full details in S1 File). The portfolios were chosen to represent groups of existing interventions that are recommended in different guidelines (NSP and WHO), or not yet in any (Novel Strategies). We then chose the most effective individual interventions and combined these into one portfolio that was predicted to give maximal impact with current tools (Optimised).

We evaluated the robustness of our conclusions by exploring the “NSP” portfolio impact using a sensitivity analysis of key model parameters. Full details are in the S1 File.

In line with recent data,[26] in the basecase scenario TB incidence and mortality were falling in 2012. The decrease past 2014 was due to declining HIV incidence via maintenance of high ART coverage. Both incidence and mortality flattened from 2020 onwards (Fig 1).

Our results suggest, that the “Novel Strategies” and “WHO” portfolios could result in 32% and 22% reductions in TB incidence by 2016 (Fig 1 and Table 2). Both portfolios peak in impact before 2032 and neither achieved greater than 40% reductions in TB incidence or mortality at any time point. The other two portfolios (“NSP” and “Optimised”) had a much larger impact by 2016 (Fig 1 and Table 2). By 2016, the “NSP” portfolio resulted in the prevention of a large number of TB deaths (57%) but only a 32% reduction TB incidence.

The final, “Optimised”, portfolio nearly achieved the 2016 NSP goals with a 46% and 71% reduction in incidence and mortality respectively (Fig 1 and Table 2). By 2050, both TB incidence and mortality could be reduced by more than 75% using the “Optimised” portfolio. However, TB incidence and mortality remained greater than the required zero new cases and deaths NSP target for 2032.

When the portfolios were broken down into their contributing interventions, we found that for the high impact portfolios (“NSP” and “Optimised”) most of the impact (on both incidence and mortality) was attributed to reductions in pre-treatment loss to follow up and to periodic (annual) active case finding. In the “NSP” portfolio ~40% of the impact was attributed to the former and ~40% to the latter at low coverage (Fig 2, see S1 File for other portfolio breakdowns).

The only single intervention that reduces both TB incidence and deaths by more than 50% by 2050 was annual active case finding in the general population (Fig 2 and S1 File). By 2032, the high sensitivity, high coverage scenario reduced TB incidence by 48% and the number of TB deaths by 58%. The intervention with the next highest impact was a 50% decrease in the pre-treatment loss to follow up, with a ~30% and 52% reduction in TB incidence and deaths respectively by 2032. Despite the interacting nature of the HIV and TB epidemics, the ART interventions in general have only a modest impact, except for UTT, which had a long-term impact on TB incidence of ~20%. The IPT interventions have little long-term impact on TB incidence or deaths. Providing IPT to HIV negatives gave the biggest impact with a 13% and 20% decrease in TB incidence and deaths respectively by 2050.

The model predicts that even if the “Optimised” portfolio had been immediately implemented in 2014, the South African 2016 aims will not be achieved (Fig 1). The 2032 goals of zero new cases and deaths are similarly unlikely to be achieved, with only two of the portfolios investigated here (“Optimised” and “NSP”) achieving greater than a 50% reduction in incidence and mortality by 2032. Indeed, both these portfolios achieved this reduction before 2016, thus partially fulfilling the NSP 2016 targets. Little difference was seen here in the impact of the portfolios between 2032 and 2050.

Thus, our results suggest that high levels of TB control could be achieved using scale up of current tools, as reductions of over 50% in the burden of TB could be achieved by 2032.

Varying key model parameters has little influence on the impact of the primary “NSP” portfolio (S1 File). The parameters chosen were the probability that TB on ART is like that in those without HIV, the protection from re-infection if latently infected, the probability of death if TB is not cleared and the ratio of case detection rates by smear status in those without HIV. The only parameter which generated a greater than 3% difference in the impact of the “NSP” portfolio was the effect of ART on the risk of acquiring TB in PLHIV.

Under all parameter ranges explored, the overall difference in impact of the “NSP” portfolio changed within the range of -6.5% to 1% (see S1 File for more details). This suggests that our conclusion that the “NSP” portfolio would not achieve either of the NSP TB targets is robust.

Although current control technologies are unlikely to achieve either of the South Africa targets for TB control, this study predicts that interventions such as active case finding and a reduction in pre-treatment loss to follow up could have a large impact. Our modelling provides evidence on the likely differential relative impact of interventions, needed for future strategy implementation decisions and highlights two important methods that could have a significant impact on future TB burden. We predict that even with the best deployment of the tools we currently have available, new technologies will be needed to reach the 2032 South African TB targets for control; however TB incidence could still be dramatically reduced using only existing control measures.