Small contribution of gold mines to the ongoing tuberculosis epidemic in South Africa: a modeling-based study

We consider four residency groups in South Africa: gold mine workers, peri-mining residents, labor-sending residents, and other residents of South Africa. Our primary data source for mine workers was data collected during the Thibela TB study [16]. Peri-mining communities were identified based on proximity to Thibela TB study sites, comprising the Lejweleputswa (Free State province), West Rand (Gauteng province), and Dr Kenneth Kaunda (North West province) districts. Labor-sending communities were identified as the OR Tambo and Alfred Nzo (Eastern Cape province) and Ugu and Sisonke (KwaZulu-Natal province) districts. Other residents of South Africa were assumed to comprise all remaining districts. Residents of areas outside of South Africa were not considered.

In the models we accounted for three general categories of parameters: population size, disease natural history, and population mixing. Population sizes were taken from the South Africa Census 2011 [17]. For the mine worker population, we considered both gold mine workers and mine workers of other commodities, representing an upper limit on the at-risk population [18, 19]. The epidemiological characteristics of each population were derived from the literature. Measurements of TB incidence and prevalence at the country level were taken from World Health Organization estimates [20] and at the mine level from the Thibela TB study [21], while HIV prevalence and antiretroviral therapy (ART) coverage were taken from Joint United Nations Programme on HIV/AIDS (UNAIDS)-based measurements [22]. Disease natural history parameters were similar to those found in other TB models and included the rate at which infected individuals progress to active disease as a result of primary disease or reactivation, the effect of HIV on reactivation, and the frequency of different forms of active disease (smear-positive, smear-negative, and extra-pulmonary) and relative infectiousness of each form (Additional file 1: Table S1). Parameters specific to mining included a multiplier for increased Mtb transmission in the mines, the prevalence of silicosis among mine workers, and the effect of silicosis on reactivation (Additional file 1: Table S2). South Africa-specific estimates of healthcare access and treatment effectiveness via directly observed therapy short-course (DOTS) were also included (Additional file 1: Table S3). Population mixing parameters were taken from national tourism surveys [23] as well as data collected during the Thibela TB study (Additional file 1: Table S4).

The static risk model (spreadsheet model) was encoded in Excel and comprises formulas to calculate the FOI (per-susceptible rate of infection) and number of infections occurring in each group (Fig. 1a). A “who acquires infection from whom” (WAIFW) matrix [24] was derived where each element β_ij represents the rate of Mtb transmission from infectives in group j to susceptibles in group i for every community k where contact was assumed possible (spreadsheet matrix 1). Each element of the WAIFW matrix was based on a base rate of transmission β₀ which we defined as the number of new infections generated by each infective case per year averaged over smear-positive, smear-negative, and extra-pulmonary forms of disease. β₀ was multiplied by a community-specific transmission multiplier c_k, which accounts for multiple environmental factors and was calibrated in the individual-based model, and the fractions of each year p_ik and p_jk that individuals from groups i and j, respectively, spend in community k (Fig. 1b). p_ik and p_jk were converted to a frequency of contact between individuals from groups i and j in community k by dividing by the total number of individuals N_k present in community k at any given time, i.e., assuming frequency-dependent transmission [25]. β_ij was then calculated by summing this product over the set A of all communities k where groups i and j spend time.

Additional quantities were derived using the WAIFW matrix:

In the static risk model we also calculated a near-term estimate of TB incidence I_i in each group i (Fig. 1a). In this case near-term refers to 1 year in the future, as prevalence after the first year was not updated to include new cases or losses due to treatment or death. Near-term incidence accounted for TB cases from primary disease resulting from new infections in a given year and TB cases from reactivation of a stable pool of latent infections. Specifically near-term incidence was calculated as the sum of incidence from five sources: (1) primary disease from new infections (k₁L_i), (2) reactivation from non-silicotic, non-HIV-positive latent infections (k₂(1 – S_i)(1 – H_i)P_i), (3) reactivation from non-silicotic, HIV-positive latent infections (k₂m_0,H(1 – S_i)H_iP_i), (4) reactivation from silicotic, non-HIV-positive latent infections (k₂m_s,0S_i(1 – H_i)P_i), and (5) reactivation from silicotic, HIV-positive latent infections (k₂m_s,HS_iH_iP_i). Here k₁ and k₂ represent base rates of primary disease in newly infected and reactivation in latently infected individuals, respectively, and L_i was derived from L_ij above. S_i, H_i, and P_i represent the stable prevalence of silicosis, HIV, and LTBI in group i, and m_s,0, m_0,H, and m_s,H represent multipliers on k₂ for silicotic, HIV-positive, and simultaneously silicotic and HIV-positive individuals, respectively. HIV-positive latent infections were further subdivided into those receiving or not receiving ART, each assigned a separate multiplier on k₂. Because the static risk model did not account for longer-term migration, e.g., for mine workers repatriating to labor-sending areas, the prevalence of silicotics in non-mining areas was assumed to be zero. The resulting values of I_i were compared to published values for mine workers from the Thibela TB study and for all South Africa from WHO estimates (spreadsheet matrix 7).

Monte Carlo simulations were performed where the mine-specific transmission rate and immunity from reinfection were sampled from normal distributions, while all other parameters were held at baseline values. For the mine-specific transmission rate, 95% of the density was assumed to lie within −20% and +20% of the baseline value. This range was selected based on the Gammaitoni-Nucci equation [26], which specifies that the probability of an individual acquiring an infection in a confined space increases exponentially as the inverse of ventilation rate. Therefore, a −20% or +20% difference in transmission rate could result from a +25% or −17% change in ventilation rate, respectively; these were similar to values from the individual-based model calibration (cf. Additional file 1: Figure S2B, S2C). For immunity from reinfection, 95% of the density was assumed to lie within −40% and +40% of the baseline value. One thousand randomly drawn pairs of values for these two parameters were used, and 95% confidence intervals (CIs) were taken from the 0.025 and 0.975 quantiles of the resulting output values. Both the static risk model and the individual-based model as well as input parameter files are available on GitHub (https://​github.​com/​SCTX/​mining_​contribution).

While the static risk model accounts for the current state of the TB epidemic and processes occurring on a sub-annum time scale, particularly population mixing, it does not update the prevalence of different disease states iteratively and does not represent longer-term changes in demographics and risk factors such as HIV. To provide a longer-term representation of the TB epidemic, we developed a dynamic, individual-based model; this model was coded in C++ and based on the TB model available in the EMOD software package [27]. We briefly describe the model here; parameter values and additional details are provided in Additional file 1: Tables S1–S4. Individuals were assigned one of the following disease states: susceptible, latently infected, active pre-symptomatic, active symptomatic, and recovered. Individuals transitioned between these states randomly according to exponentially distributed delays. Birth and non-disease death processes were represented whereby individuals were added to and removed from the simulation at rates consistent for South Africa. A residency status in one of the four groups in the model (mining, peri-mining, labor-sending, and other South Africa) was assigned at birth. Residency status was retained for the lifetime of the individual except for individuals born with labor-sending group status who transitioned to mining status during adulthood and then back to labor-sending status upon retirement. Short-term mixing between groups representing regular visits from mine workers to peri-mining or labor-sending communities was specified by an interaction matrix as in the static risk model. The numbers of discrete agents in the model were scaled in the output to reflect the sizes of the real populations: individuals in mining, peri-mining, and labor-sending groups by approximately 50:1 and in other South Africa by approximately 400:1. These scale factors corresponded to 2011 population sizes of approximately 0.5 million (M), 2.1 M, and 3.4 M in mining, peri-mining, and labor-sending areas, respectively, and 47 M in the remainder of South Africa [17].

Two risk factors were represented in the individual-based model, HIV infection and silicosis. HIV infection was distributed to individuals according to age-specific rates of infection; these were generated from the EMOD HIV model calibrated to South Africa [28] and assumed to be the same across residency groups [29‐31]. The EMOD HIV model assigned a CD4 count to each individual which declined linearly with time in the absence of treatment [29‐31]. ART was distributed to HIV-positive individuals according to eligibility guidelines in South Africa and matched population coverage estimates [22]. ART had the effect of increasing CD4 levels in the model [29‐31]. Silicosis was acquired by mining group individuals at a rate consistent with radiographic observations in mine workers, at approximately 1% per year of employment [32‐34] (Additional file 1: Table S2).

Susceptible individuals in the model were infected at a rate that differed by residency group and depended on the total infectiousness of other groups and the frequency of group interactions. The infectiousness of each group depended on the prevalence of different forms of active TB, where pre-symptomatic, smear-negative, and extra-pulmonary forms of disease were assumed to contribute less infectiousness than smear-positive disease. Data on the frequency of group interactions were specified by a WAIFW matrix as in the static risk model [35].

Infected individuals transitioned to active disease with one of two rates representing primary and reactivation disease. Active disease in the model included a pre-symptomatic period of set duration followed by symptomatic disease of smear-positive, smear-negative, or extra-pulmonary forms. Individuals persisted in symptomatic disease until progressing to self-cure, treatment, or death. HIV and silicosis had the effect of increasing the rate of reactivation, represented as multipliers on the base rate of reactivation [36]. For HIV-positive individuals, the magnitude of the increase varied as the inverse of CD4 level.

Individuals with active symptomatic disease were assumed to seek care at high or low rates corresponding to high- or low-quality access to care, respectively, broadly representing different levels of care in South Africa. Upon accessing care, symptomatic individuals were assumed to receive a sputum smear or GeneXpert test, depending on whether care was sought before or during DOTS availability. The probability of a positive test result corresponded to observed test sensitivities. If a positive test result was obtained, an individual was assumed to undergo treatment, with a rate of disease clearance that depended on whether treatment was given before or during DOTS availability. Following treatment, individuals transitioned to a recovered state that was identical to the susceptible state but assumed to have a reduced probability of reinfection due to immunity.

Several historical population-wide events were simulated in the individual-based model. During each simulation the model was seeded and run for a specified burn-in period. With a burn-in period of 100 simulated years, incidence and mortality were observed to be stable in the different groups in the model, consistent with endemic TB. HIV, DOTS, and ART were introduced at simulated years 1985, 2002, and 2007, respectively, representing country-wide trends.

The model was calibrated to TB incidence in mining areas measured during the Thibela TB study [21] and TB incidence and mortality at the country level for multiple years [37]. Parameters for the transmission rate in mining areas and immunity to reinfection following previous TB exposure were varied during calibration. The transmission rate in mining areas was parameterized as a multiple of the base transmission rate and represented the aggregate environmental factors, e.g., reduced ventilation rates, that may increase Mtb transmission in mines. A likelihood score for each parameter combination was computed using a likelihood function based on a normal distribution where the differences between published high and low estimates of incidence and mortality were taken to represent 95% CIs and the three epidemiological indicators were equally weighted. The joint posterior distribution of the two parameters conditional on the data were estimated via incremental mixture importance sampling (IMIS) [38]. The joint distribution was found to be unimodal and strongly peaked; therefore, parameters for subsequent simulations were set at the maximum a posteriori estimate (joint posterior distribution, Additional file 1: Figure S2A; marginal distributions, Figure S2B, S2C in Additional file 1). For consistency, the values for these parameters were also used in the spreadsheet model. The posterior estimate of the reduction in susceptibility to reinfection was similar to previous estimates of bacille Calmette-Guerin (BCG) protection against active disease, 0.58 (95% CI 0.35–1.01) [39]. Technical details regarding the calibration procedure are available in Additional file 1.

To measure the incidence attributable to mine workers using the individual-based model, we simulated two counterfactual scenarios: first, having no Mtb transmission in the mines and, second, having no Mtb transmission in any area. These scenarios were identical to the baseline scenario until simulated year 2012 when Mtb transmission was stopped in the model. Other processes such as disease progression continued unchanged. The numbers of new cases of active disease between simulated years 2014 and 2019 were counted for each counterfactual scenario and compared to the baseline scenario. This calculation was repeated for each residency group in the model.

We used a static risk model to calculate the FOI and number of transmission events (new infections) in different mining-related communities in South Africa and predict the near-term (following-year) incidence in these communities (Fig. 1a). The model accounted for a number of factors including a higher rate of Mtb transmission due to environmental factors and the amount of time that residents reported spending in their own versus other communities, where mixing was assumed to be proportional to the time spent in each community (Fig. 1b).

As a check on the static risk model, we compared FOI output from the model to data on the annual risk of TB infection (ARTI) in children. Under baseline parameters, we estimated FOI to be 21.2% (95% CI 16.4–26.1%) in mine workers, 4.3% (95% CI 4.3–4.3%, indicating a difference of < 0.05%) in peri-mining residents, 5.8% (95% CI 5.8–5.8%) in labor-sending residents, and 3.5% (95% CI 3.5–3.5%) in other South African residents (Table 1, Additional file 1: Table S5). CIs were derived by sampling parameter values for mine-specific transmission and immunity following previous infection. The FOI estimate for other South African residents in the model was found to be consistent with available ARTI measurements: 2.5–4.2% (across Western Cape, 2005, [40]), 3.8–4.5% (in Cape Town, 2005, [41]), 3.9–4.8% (in Cape Town, 2009, [42]), and 2.1–5.2% (in Johannesburg, 2013, [43]).

Using the static risk model, we then calculated the fraction of FOI in each community attributable to each residency group. We estimated that the majority of each community’s FOI was attributable to local residents: 93.9% (95% CI 92.4–95.1%), 91.5% (95% CI 91.4–91.5%), 94.7% (95% CI 94.6–94.7%), and 98.8% (95% CI 98.8–98.8%) in mining, peri-mining, labor-sending, and other SA communities, respectively (Table 1, Additional file 1: Table S5). Despite the amount of time mine workers were assumed to spend in other areas (up to 20% per annum), the FOI in peri-mining, labor-sending, and other SA communities attributable to mine workers was estimated to be 5.8% (95% CI 5.8–5.8%), 3.6% (95% CI 3.5–3.6%), and 0.1% (95% CI 0.1–0.1%), respectively.

Using the preceding FOI and data on susceptible individuals, i.e., either uninfected or latently infected but susceptible to reinfection, we estimated the number of new infections expected to occur in each community and compared these results to published TB incidence for different communities. For mine workers and the overall population, we estimated TB incidence to be 2963 (95% CI 2208–3858) and 989 (95% CI 980–1000) per 100,000 individuals, respectively. These were similar to published values for these communities, 2957 (in control cluster mines during the Thibela TB study, between 2006 and 2010 [21]) and 977 (717–1276) (in South Africa, 2008 [20, 44]) per 100,000 (Additional file 1: Table S6).

Using the static risk model, we also estimated the fraction of infections occurring each year attributable to each residency group. Out of the overall number of new infections occurring in South Africa per annum, we estimated that 4.0% (95% CI 2.6–5.8%), 5.0% (95% CI 4.5–5.5%), and 9.0% (95% CI 8.8–9.1%) were attributable to mining, peri-mining, and labor-sending residents, respectively (Table 2, Additional file 1: Table S7). When scaled to the fraction of the overall population in South Africa that each group represents, mine workers, peri-mining residents, and labor-sending residents contributed 4.32 (95% CI 2.77–6.15), 1.21 (95% CI 1.09–1.34), and 1.39 (95% CI 1.36–1.41) times as many infections as South Africans as a whole (Table 2, Additional file 1: Table S7). Similarly, when scaled to the fraction of the overall number of prevalent cases in South Africa found in each group, mine workers, peri-mining residents, and labor-sending residents contributed 1.62 (95% CI 1.04–2.30), 1.14 (95% CI 1.02–1.25), and 1.07 (95% CI 1.05–1.09) times as many infections as South Africans as a whole (Table 2, Additional file 1: Table S7). Therefore, while mine workers contribute a larger number of infections on a per-capita or per-prevalent case basis than other South Africans, the majority of these infections occur among mine workers themselves.

To measure the impact of Mtb transmission in the mines on incidence, we used a dynamic, individual-based model of TB in South Africa. In this model we accounted for longer-term demographic changes and additional pathways leading to active disease including reactivation from transmission occurring over a longer time window. We calibrated the model to several epidemiological indicators including incidence and mortality over multiple years [21, 37]. Model estimates of TB incidence and mortality overlapped published ranges, both at the country level (Fig. 2a, b; Additional file 1: Figure S3A, S3B) and at the mining level (Fig. 2c, d; Additional file 1: Figure S3C, S3D). In particular, model incidence reproduced measurements from the Thibela TB study, showing a threefold higher incidence in the mines compared to South Africa overall (Fig. 2a, c). Model incidence in the mines preceding the Thibela TB study was exceeded 4000 per 100,000 (Fig. 2c), consistent with previous studies on mine workers [45]. Model mortality due to TB among mine workers was approximately 1% per annum (Fig. 2d), which was consistent with a range that includes the 0.9% all-cause mortality rate and 4.3% all-cause mortality-plus-medically boarded rate observed as a secondary outcome of the Thibela TB study [21]. As an additional test of the model, including South Africa-specific parameters derived from calibration, we used the model to simulate the Thibela TB study intervention of widely available preventive therapy. Following a cessation of the intervention, we observed a rebound in model incidence similar to the rebound observed during the Thibela TB study (Additional file 1: Figure S4A).

To estimate the fraction of incident cases attributable to gold mines, we simulated a counterfactual scenario of stoppage of Mtb transmission in mining areas and measured the subsequent change in incidence over several years (Fig. 2e). By doing so, we estimated that recent Mtb transmission in the mines contributed 58.2% (95% CI 57.8–58.9%), 4.8% (95% CI 4.3–5.2%), and 4.9% (4.4–5.2%) of the TB incidence in mining, peri-mining, and labor-sending residents, respectively (Table 3). Among other residents of South Africa, the counterfactual scenario had a smaller effect that resulted in a time course that overlapped the baseline scenario at all time points, i.e., within the stochastic noise of the simulation (Table 3). In South Africa as a whole, the fraction of TB incidence due to recent Mtb transmission in the mines was estimated to be 2.4% (95% CI 1.4–3.3%) (Table 3). To test the robustness of this measurement, we performed a series of one-way sensitivity analyses based on varying the infectiousness in each community separately (mining, peri-mining, labor-sending, and other South Africa). In all cases, the resulting fraction of TB incidence due to recent Mtb transmission in the mines was found to vary maximally between 1 and 4% (Additional file 1: Figure S5).

As a second counterfactual scenario, we simulated stoppage of transmission in all areas of South Africa and calculated the fraction of incident cases attributable to recent transmission from any source. Among mining residents, 63.0% (95% CI 58.5–67.7%) of the incident cases were predicted to result from recent transmission, 92.5% (95% CI 92.1–92.9%) of which were attributable to recent transmission in mining areas (Table 3). In contrast, among all South African residents, 37.4% (95% CI 35.2–40.0%) of the incident cases were predicted to result from recent transmission, 3.7% (95% CI 3.0–4.4%) of which were attributable to recent transmission in mining areas (Table 3). These figures were consistent with local Mtb transmission in mines being the source for the majority of incident TB cases in the mines but only a small fraction of the incident TB cases in the remainder of the country.