Sustained high fatality during TB therapy amid rapid decline in TB mortality at population level: A retrospective cohort and ecological analysis from Shiselweni, Eswatini

Setting

The predominantly rural Shiselweni region (203,376 inhabitants in 2017) in Eswatini comprises Nhlangano, Hlathikulu and Matsanjeni health zones [20]. The region had a high HIV prevalence of 26% in ≥15-year-olds in 2017 [21]. Although regional drug-sensitive and drug-resistant TB notifications declined fivefold from 1341 to 269 cases per 100,000 population between 2009 and 2016, HIV co-infection among TB cases remained high at ~71% [14]. Figure 1 summarises the main programmatic interventions implemented in the region [14, 22, 23]. HIV and TB care initially centralised at three secondary care facilities in 2008, was—with the support of Médecins sans Frontières (MSF)—decentralised to 22 nurse-led and medical doctor-supported primary care clinics between 2009 and 2011. The public sector primarily delivered TB care in the region, while the private sector played a minor role. Laboratory capacity was expanded with the introduction of Xpert MTB/RIF testing in the three health centres in 2009 and TB-LAM testing in all clinics of Nhlangano health zone in 2018. ART eligibility criteria evolved from 2009 onwards, initially targeting HIV-coinfected TB patients and PLHIV who had a CD4 ≤200 cells/mm³ (Figure 1). Eventually, universal ART (treat-all) was piloted in Nhlangano zone in 2014 and extended to the other health zones 2 years later.

Study design, data management and definitions

We analysed patient-level (retrospective cohort analysis, from 2009 to 2019) and population-level (ecological analysis, from 2009 to 2017) trends and predictors of TB-related deaths in first-line TB treatment cases in Shiselweni.

Patients registered in routine paper-based drug-sensitive TB registers at facilities in the region were considered first-line TB treatment cases. The TB-related variables were defined in accordance with WHO recommendations [24]. The date of registration—corresponding to the date of prescription of TB drugs—was considered as time zero (start of observation period). Final TB treatment outcomes were assigned ~1 year after registration. Patients dying during treatment were considered a TB-related death. TB treatment cases that were transferred in were excluded from analyses to avoid duplications. These registers were the only data source for this study.

For the cohort analysis, we used treatment data from Shiselweni from January 2009 to December 2017, as well as from Nhlangano health zone from 2018 to 2019. Patient-level treatment data were unavailable for the other two zones for 2018–2019. For the ecological analysis, we considered TB cases from the entire region for the period from 2009 to 2017 only, lacking sex- and age-stratified population-level data for Nhlangano zone thereafter.

Assumptions behind the relationship of variables for the cohort analysis were summarised in a directed acyclic graph (Figure S1). We identified factors that may be (in)directly associated with timely TB care registration and death during therapy, which were then included in analysis. All available variables were incorporated into the final fitted models for multiple imputation and Poisson regression.

Statistical analysis

Analyses were performed with Stata 16.1. Frequencies and proportions describe baseline and crude patient-level and population-level outcome data overall and by HIV status. Annual treatment outcomes were reported in the same year of TB treatment initiation (cohort specific).

Multiple imputations by chained equation was used to account for missing baseline values, with 10 imputed datasets created and imputation diagnostics being satisfied (Figures S2 and S3). We also accounted for undocumented deaths. Some treatment outcomes (loss to follow-up [LTFU], treatment failure, outcome not evaluated, transfer out) may contain undocumented deaths, possibly resulting in underestimation of TB-related deaths. Therefore, these outcomes (17.7%) were assumed missing, with multiple imputations used to obtain the final binary outcome of treatment success and death (Table 1). Interactions among variables were not evaluated in the process of multiple imputations. However, interactions of HIV status with other variables were assessed for the cohort and ecological analysis.

Ethics

This analysis was approved by the Eswatini Health and Human Research Review Board. It also fulfilled the exemption criteria set by the MSF Ethics Review Board (ERB) for a posteriori analysis of routinely collected clinical data and thus did not require MSF ERB review.

Cohort analysis (2009–2019)

Crude TB treatment outcomes

Treatment success increased from 65.7% to 81.5% between 2009 and 2019 and was highest in 2018 (84.8%) (Figure 2, Table S1). Treatment success tended to be higher for HIV-negative patients in all years; in 2019 it was 86.4% in HIV-negative patients versus 80.5% in PLHIV.

Of 11,883 TB treatment cases, 1302 (11.0%) patients were registered as dead during treatment. By HIV-status, there were 210/2798 (7.5%) deaths in HIV-negative patients, 984/8443 (11.7%) deaths in PLHIV, and 108/642 (16.8%) in patients with unknown HIV status. While the absolute number of TB treatment cases and deaths declined each consecutive year (Table S1), PLHIV sustained higher case fatality ratios (CFRs) in all years (Figure 3). The CFR remained above 10% in most years (other than 2009, 2017 and 2018), with the highest CFR recorded in 2015 (15.3%) and the lowest in 2018 (5.7%). For HIV-negative patients, only the years 2012 and 2013 had reported CFRs at 11.8%–11.2%, while it ranged between 4.5% and 9.2% for the remaining years.

Ecological analyses (2009–2017)

Mortality rates

TB treatment cases (2785 in 2009 vs. 497 in 2017) (Table S1) and annual TB-related mortality rates per 100,000 population decreased in consecutive years (see Table 4, Figure 4). In summary, the peak mortality rate was 174/100,000 population in 2009 and 2010, and it decreased 6.4-fold (−147/100,000 population) to 27/100,000 population in 2017. In all years, men (peak year 2010: 200/100,000 population) had a higher mortality rate than women (peak year 2009: 158/100,000 population). PLHIV sustained a many times higher mortality rate (peak year 2009: 967/100,000 population) versus the HIV-negative population (peak year 2011: 27/100,000 population). The absolute decline between peak years and 2017 was most pronounced in PLHIV (6.9-fold; −826/100,000 population) versus the HIV-negative population (6.8-fold; −23/100,000 population) while the relative decline was of comparable magnitude. By age, mortality rates were highest in the 30 to 39 years age group (peak year 2009: 469/100,000 population) and the ≥60 years group (peak year 2010: 481/100,000 population) while the decline was highest in the ≥60 years group (37.0-fold; −468/100,000 population).

TABLE 4. Annual TB-related mortality rates per 100,000 population by HIV-status and selected demographic factors, and accounting for undocumented deaths, in Shiselweni, from 2009 to 2017.

	Annual mortality rates per 100,000 populationa									Relative decline in mortality rates until 2017		Absolute decline in deaths per 100,000 population until 2017
	2009	2010	2011	2012	2013	2014	2015	2016	2017	From 2009	From peak year	From 2009	From peak year
Overall	174	174	109	90	67	50	43	29	27	6.4	b	147	b
By sex
Women	158	151	88	80	53	32	35	21	20	7.9	b	138	b
Men	193	200	133	102	83	71	52	39	35	5.5	5.7	158	165
By HIV status
HIV positive	967	908	514	417	311	260	211	139	141	6.9	b	826	b
HIV negative	14	25	27	24	18	8	9	7	4	3.5	6.8	10	23
By HIV status and sex
HIV-negative women	9	14	16	20	10	1	5	2	2	4.5	10.0	7	18
HIV-negative men	19	37	38	28	25	14	13	12	6	3.2	6.3	13	32
HIV-positive women	760	704	379	321	226	157	158	98	92	8.3	b	668	b
HIV-positive men	1323	1257	743	579	454	430	299	206	223	5.9	b	1100	b
By age group, years
0–9	30	47	11	15	4	2	4	2	4	7.5	11.8	26	43
10–19	17	25	15	8	12	4	0	0	4	4.3	6.3	13	21
20–29	244	230	126	85	64	70	48	34	37	6.6	b	207	b
30–39	469	356	292	238	132	167	135	87	79	5.9	b	390	b
40–49	398	377	258	216	160	119	125	91	63	6.3	b	335	b
50–59	264	307	203	182	153	98	71	61	89	3	3.4	175	218
≥60	391	481	273	255	243	58	76	43	13	30.1	37.0	378	468

• ^a The peak years of mortality rate are in bold.
• ^b Data not presented as peak year was 2009.

Treatment fatality in treatment cases

The overall crude annual treatment fatality remained high throughout the years. Despite a decline in CFR in HIV-negative patients from 2014 onwards, the overall high CFR was driven and sustained by PLHIV and comparable to previously reported trends from Eswatini and other RLS [15, 30].

Similar to other settings [31-34], older age increased the fatality risk irrespective of HIV status, possibly related to factors increasing biological vulnerability to TB disease and unmeasured competing causes of death (e.g., cardiovascular disease, cancer) in older people. Interestingly, sex did not show an obvious association with case fatality, but estimates may have been distorted by lack of adjustment for other socio-demographic factors or correlations between covariates.

Comparable to other studies [30], case fatality was higher in HIV-negative patients and PLHIV who presented with extra-pulmonary TB and missing bacteriological status, while a negative bacteriological status increased the fatality risk in HIV-negative patients only. Possible explanations are (undiagnosed) underlying infectious (e.g., pneumocystis pneumonia in PLHIV) and other chronic lung diseases (e.g., lung cancer, silicosis) associated with high case fatality, delay in diagnosis and treatment (e.g., ruling out other infectious diseases), unrecognised drug-resistant TB disease, and possibly more disseminated and severe disease in patients with extra-pulmonary TB disease. Importantly, negative bacteriological status—which is more common in immunosuppressed patients and a risk factor for death in advanced HIV disease [30, 35]—was not clearly associated with increased case fatality in PLHIV. In our setting, health workers were capacitated to start TB treatment in bacteriologically negative cases irrespective of HIV status, potentially reducing delays in treatment initiation in those at greatest need for treatment.

Previous TB therapy was clearly associated with higher fatality risk in HIV co-infected patients while the association with HIV-negative patients was not obvious. In Eswatini, retreatment cases and PLHIV are more likely to present with drug-resistant TB disease [36] that is associated with increased case fatality. Accurate and timely diagnosis of drug-resistant TB was compromised in our setting because of suboptimal access to culture-based drug-resistance testing, and the inability of the Xpert and line probe assays to detect the locally prevalent RpoB I491F rifampicin resistance mutation [36-38], thus possibly resulting in misclassification of drug-sensitive and drug-resistant TB disease. These challenges suggest a need for faster and more accurate tests for rifampicin-resistant TB disease as well as more potent first-line TB treatment regimens that could cover undiagnosed drug-resistant TB disease.

Similar to other studies [2-4], higher CD4 cell counts and ART decreased the fatality risk, possibly due to less severe HIV disease and improved immunity. However, the overall impact of programmatic and policy factors mainly targeted at PLHIV remained inconclusive. While the programmatic availability of Xpert testing and the expansion of ART eligibility criteria lacked strong associations, the availability of TB-LAM testing was associated with a lower fatality risk. Although Xpert and TB-LAM testing may improve diagnosis and result in more timely treatment [39-43], the effect of Xpert on case fatality remains inconclusive, as does the population-level impact of TB-LAM [44-46]. Patients only benefit from improved diagnostics if these are available and used at sites where treatment decisions are made, health workers correctly perform and interpret them, and administrative delays are minimised (e.g., reporting back of test results). Although health policies promoting earlier access to ART may reduce time periods of increased risk for (severe) TB disease, a patient-level benefit may only be achieved if patients actually initiate ART at the appropriate time and remain virally suppressed. Notably—and as suggested by a 7% increased fatality risk for each consecutive calendar year in PLHIV—other temporal opposing factors may hide possible patient-level benefits of policy and programmatic interventions—for instance, when they are related to unmeasured temporal changes in access to care (e.g., hard to reach and to treat patients are possibly proportionally overrepresented in recent years). Overall, findings suggest that theoretically valuable programme and policy interventions at population level may not always translate to obvious patient-level health benefits, or that effects of new interventions may be difficult to isolate in observational studies.

TB-related mortality at population level

The absolute crude number of TB-related deaths was highest in 2010 and decreased each consecutive year thereafter. Although mortality rates remained higher in PLHIV, as also reported elsewhere [2-4], the absolute decline was most pronounced in PLHIV and was seen in all age groups and both sexes. This decline coincided with rapidly falling numbers of TB treatment cases in this region [14], suggesting that changes in TB notifications and treatment cases may be an important driver for changes in TB-related mortality at population level. With a lower prevalence of active TB in the population, the absolute number of TB cases resulting in death decreases. This inference is supported by the multivariate analysis, indicating an association between declining TB notifications and a lowered mortality risk. However, this risk reduction may primarily signify a decreased risk of TB infection in the population rather than a risk reduction in mortality alone. Although a spurious decline in TB treatment notifications may have resulted in more undiagnosed TB disease and undocumented deaths, the decline in TB is likely real in our setting due to higher ART coverage in PLHIV as well as increased case finding activities and better TB diagnostics in recent years, in addition to a similar decline reported from other parts of the country and settings in Southern Africa [13, 47-49]. Future decline in TB-related mortality may be achieved by further reducing active TB disease—for instance, by reducing the pool of people with increased susceptibility to TB infection (e.g., reducing HIV infections, early ART) and decreasing the risk of activation of latent TB infection with TB preventive therapy.

Overall, population-level multivariable analysis confirmed crude trends with a rapid mortality risk reduction in PLHIV but less obvious mortality risk reduction in the HIV-negative population when compared with 2009. The crude decline in HIV-negative cases may have been too small to detect an obvious temporal trend in adjusted analysis. In the HIV-negative population, the elderly population had the highest mortality risk, surprisingly followed by the 20–29 years group (vs. 30–39 years). The latter may be a spurious finding or explained by unaccounted HIV infections. For instance, mortality risk in PLHIV was also high in young adults (e.g., 15–34 years) who were known to have high levels of undiagnosed, untreated and virally elevated HIV infections [21, 28].

Notably, the relative mortality risk was comparable between sexes. It may be explained by complex differentiated temporal trends in sex-related and HIV-associated TB factors with regards to epidemiology, behaviour and exposure, biological and genetic determinants, and access to care [50-54], possibly balancing crude differences between sexes.

Limitations and strengths

First, we accessed first-line TB treatment data, and were thus unable to account for undiagnosed TB or pre-treatment mortality (38% in some African settings [53]), and deaths related to drug-resistant TB therapy. Our estimates of TB-related mortality rates were therefore at the lower range (27/100,000 in 2017) compared with WHO estimates (55/100,000 in 2017) [55]. In contrast, all deaths during TB treatment were assumed to be related to TB disease, which may overestimate mortality rates if some deaths were caused by other conditions.

Second, although, population-adjusted estimates of TB mortality depended on the accuracy of disaggregated population-level denominators, overall trends in mortality rates appeared reasonable, with higher rates and risks of death in PLHIV, older people, and the mid-year age group for PLHIV.

Third, our dataset contained missing values. Under the assumption that values were missing at random (MAR), multiple imputation was used to impute missing baseline values as well as treatment outcomes that could clinically not be assigned to the binary outcome of death and treatment success. Multiple imputation allows all observations to be retained, thus increasing power and precision [56]. As patients lost to follow-up and transferred during TB treatment have a risk of death [32], the omission of non-binary outcomes would have disregarded undocumented deaths and likely resulted in an underestimation of TB-related deaths. Importantly, the assumption of MAR may have been violated and TB-related deaths underestimated if, for instance, LTFU and treatment failure have a disproportionally higher risk of undocumented deaths, or if missing categories have a meaning (e.g., sicker patients are less likely to have measurements). Omitting these outcomes from analysis might have resulted in an underestimation of TB-related deaths. The absence of contextualised information about reasons for and the extent of deaths in LTFU prevented informed assumptions and adjustments in multiple imputation. Alternative correction methods, such as confirming undocumented deaths with the National Population Register, were not feasible due to limited access. As a result, our imputation model might have contributed to underestimating TB-related deaths. Finally, despite a high proportion of missing values for CD4 cell count, we retained it in analyses due to the absence of alternative immune-suppression proxy variables. Avoiding additional assumptions about missingness, we primarily attributed it to inadequate documentation, given the availability of point-of-care CD4 testing in clinics since the early years. Anticipated associations with deaths were consistent with findings in comparable studies [2-4].

Fourth, we could not adjust for some unmeasured factors that were identified in the directed acyclic graph (DAG) to be associated with TB treatment initiation and the outcome, thus possibly resulting in biased effect estimates. They included communicable (e.g., lower respiratory infections due to pathogens other than TB) and non-communicable (e.g., diabetes mellitus) comorbidities, disease severity and socio-economic determinants (e.g., employment status, income) as well as factors influencing HIV and TB treatment (e.g., delayed TB diagnosis and treatment, viral load suppression in HIV coinfected patients, other concurrent treatments).

A strength was that that the study was conducted in a routine public sector setting, generalizable to many rural contexts in Southern Africa with an intertwined HIV/TB epidemic. In addition, the long observation period allowed us to explore temporal trends amid the rapid expansion of TB and HIV care.