ABSTRACT
Background Emerging evidence suggests that shortened, simplified treatment regimens for rifampicin-resistant tuberculosis (RR-TB) can achieve comparable end-of-treatment outcomes to longer regimens. We compared a 6-month regimen containing bedaquiline, pretomanid, linezolid, and moxifloxacin (BPaLM) to a standard of care strategy using a 9- or 18-month regimen depending on whether fluoroquinolone resistance (FQ-R) is detected on Drug Susceptibility Testing (DST).
Methods and Findings Genomic and associated demographic data were used to parameterize a mathematical model estimating long-term health outcomes and costs (2022 USD) for each treatment strategy for patients 15 years and older diagnosed with pulmonary RR-TB in Moldova, a country with a high burden of TB drug resistance. In this model individuals were followed over their lifetime, simulating the natural history of TB and associated treatment outcomes, as well as the process of acquiring resistance to each of 12 anti-TB drugs. Compared to the standard of care, 6 months of BPaLM was estimated to reduce lifetime costs by $3433 (95% Uncertainty Interval (UI): 1480, 5771) per individual, with a small non-significant reduction in quality adjusted life expectancy of 0.06 QALYs (95% UI: −0.33, 0.45). For those stopping moxifloxacin under the BPaLM regimen, continuing with BPaLC provided more QALYs at lower cost than continuing with BPaL alone. Such a regimen (6 months of BPaLM, where clofazimine is added in the event of moxifloxacin discontinuation) had a 92% chance of being cost-effective. With the exception of pretomanid and delamanid, 6 months of BPaLM either reduced or resulted in no significant change in the cumulative incidence of resistance to each drug. Sensitivity analyses showed 6 months of BPaLM to be cost-effective across a broad range of values for the relative effectiveness of the drug regimens and the proportion of the cohort with FQ-R.
Conclusions Compared to the standard of care, the implementation of short-course regimens like BPaLM could improve the cost-effectiveness of care for individuals diagnosed with RR-TB, particularly in settings where current long-course regimens are challenging to implement and afford. Further research may be warranted to explore the suitability of 6 months of BPaLM in specific national settings, including locations where DST capacity is limited.
INTRODUCTION
Treatment for rifampicin-resistant tuberculosis (RR-TB) is complex, involving combinations of several drugs—many of which have substantial potential for toxicity—over a prolonged course of therapy. The 2022 WHO Guidelines for the treatment of drug-resistant tuberculosis recommend a shorter, 6-month regimen composed of bedaquiline, pretomanid, linezolid, and moxifloxacin (BPaLM) to treat rifampicin-resistant tuberculosis (RR-TB).1 These guidelines updated earlier 2020 WHO Guidelines that recommended several treatment regimens, each comprising 4-7 drugs for 9-18 months or longer.2
The evidence base for shorter regimens for RR-TB has been broadly positive, including results from observational studies,3, 4 single-arm clinical trials,5, 6 mathematical modeling analyses,7 and the recent multicenter open-label randomized controlled trial TB-PRACTECAL.8 Although trial recruitment was stopped early on the recommendation of a planned, interim review by the study monitoring committee, the analysis suggested that 6 months of BPaLM was non-inferior to the standard of care with respect to treatment outcome (a composite of death, treatment failure, treatment discontinuation, loss to follow-up, or recurrence) and was beneficial with respect to safety.8 These early results point to some clear potential benefits in terms of shorter, simpler regimens for RR-TB, but the absence of larger, confirmatory trials led to a conditional recommendation by the WHO.
The 2020 WHO Guidelines represent the existing standard of care in many settings. In addition to higher prices and supply constraints for newer drugs,9, 10 it is expected that the rollout of the BPaLM regimen as part of the newer 2022 Guidelines may be delayed by concerns about comparative effectiveness and cost-effectiveness9–13 as well as the emergence of drug resistance, particularly in settings with limited capacity to detect resistance to newer antitubercular agents such as bedaquiline, pretomanid, and linezolid.14 The decision to implement the new 6-month BPaLM regimen will depend on setting-specific tradeoffs between regimen effectiveness, cost, the complexity of treatment decisions, and existing levels of resistance to anti-TB drugs in the population. Decision analysis provides a framework to analyze these tradeoffs, and a recent cost-effectiveness study using evidence from TB-PRACTECAL found that 6-months BPaLM may reduce cost and improve health relative to the standard of care in several countries.15 Our analysis builds on this work by focusing on longer term outcomes that are difficult to measure in a trial setting and by examining a wider range of testing and treatment approaches, including whether patients who must stop moxifloxacin (Mfx)—due to side effects or acquired resistance—should continue on BPaL alone, or BPaL plus clofazimine (BPaLC).14, 15
In this study, we investigated the health impact and cost-effectiveness a 6-month BPaLM regimen for the treatment of adults with pulmonary RR-TB, as compared to the standard of care. We considered a range of treatment strategies incorporating these two approaches, varying the timing and frequency of drug susceptibility testing (DST) as well as how regimens would be modified for individuals developing fluoroquinolone resistance. To estimate outcomes, we used a Markov microsimulation model parameterized with detailed genomic sequencing data describing specific patterns of initial drug resistance, and calculated the effect of each treatment strategy on length and quality of life, accounting for regimen effectiveness, risks of severe adverse events (SAEs) due to drug toxicity, and acquisition of resistance. We conducted the analysis for the setting of Moldova, a country with a high proportion of TB strains resistant to first- and second-line drugs,16, 17 and explored the generalizability of findings to settings with a different prevalence of initial fluoroquinolone resistance.
METHODS
Strategies
We compared eight treatment strategies, each reflecting a different approach to drug regimen choice and timing of DST (Table 1). Two strategies adopted drug regimens aligned with the standard of care as defined by the 2020 WHO Guidelines,2 with all individuals started on a WHO longer regimen while awaiting the results of second-line DST by MGIT to fluoroquinolones and injectables. Fluoroquinolone resistance (FQ-R) identified via MGIT was assumed to result in the continuation of an 18-month WHO longer regimen, with refinements as necessary based on DST. If fluoroquinolone susceptibility (FQ-S) was detected, treatment was switched to a 9-month regimen (Figure S1). Under one standard of care strategy (strategy (7)), we modelled the minimum guideline-recommended frequency of second-line DST–every 4 months, and in another (strategy (8)) we increased this to a monthly frequency. While the 2020 WHO Guidelines did not prescribe exactly one combination of drugs for each scenario, we adopted a single combination of drugs for each situation for tractability, based on our best interpretation of the guideline’s hierarchy of group A, B, and C drugs (Figure S1).
The remaining six strategies were modeled on the 2022 WHO Guidelines18 with 6-month BPaLM-based regimens. In three of these strategies, individuals having to stop Moxifloxacin (because of a SAE or because resistance was detected on DST) were continued on BPaL alone, as recommended by the 2022 Guidelines. In the remaining three, they continued on BPaLC. The remaining differences between these six strategies depended on the prescribed schedule of DST to second-line drugs; in two of these strategies, we explored the effects of omitting routine second-line DST at treatment initiation (Table 1).
Population and data
We modeled a cohort of individuals aged 15 years and older diagnosed with RR-TB in Moldova. For each individual, their age and the resistance profile of the strain of M. tuberculosis causing infection were informed by publicly available genomic sequencing data from Moldova.19 These data were collected in 2018–2019, and a full description has been provided by Yang and colleagues.17 We assumed that there was only one strain of M. tuberculosis per individual (i.e., that there were no mixed infections) and that a mutation associated with resistance conferred full resistance to that drug. M. tuberculosis strains lacking relevant resistance mutations were assumed to be fully susceptible to the respective drugs. We excluded data for rifampicin susceptible strains (Figure S3) leaving 674 distinct samples. The proportion of isolates with resistance to each drug is shown in Figure S5. This analysis used publicly available data, and did not require ethical approval.
Model
We used a Markov microsimulation model to simulate lifetime outcomes for a cohort of 10,000 individuals. Individuals in the model were simulated by random draws from the genomic sequencing dataset, with replacement. They were each assigned a drug regimen based on the modeled strategy (Table 1). Individuals then were assumed to transition between four health states: (1) Receiving TB treatment, (2) TB disease – not receiving treatment, (3) Cured post-treatment, and (4) Dead (Figure S2). Within each Markov state, individual events were tracked including true cure, the occurrence of SAEs, second-line DST, changes to the drug regimen, loss to follow-up, relapse, death, and the evolution of drug resistance for that individual’s strain of M. tuberculosis. Extensions to the treatment regimen were implemented for those not observed to have successfully completed treatment.
While the range of SAEs resulting from TB treatment are of many varying durations and degrees of impact on quality of life, we accounted for these events in a simplified way by modeling the risk of a grade 4-5 SAE during the first three months of exposure to each drug, with each SAE conferring a small but lifelong deduction in quality of life (Table S1). SAEs and diagnosed resistance constituted lifetime contraindications to the relevant drug, and replacements were made according to the modelled strategy (Table 1).
Each month, we tracked the drug regimen and the true resistance profile of each individual’s strain of M. tuberculosis. The number of effective drugs in a regimen was defined as the sum of all drugs being received, minus those drugs to which the strain of M. tuberculosis was resistant. The number of effective drugs was used to calculate the rate of cure in the current month (higher with more effective drugs), and the rate of acquisition of new resistance to the remaining effective drugs (lower with more effective drugs; Figure S4 and Table S1). DST to second-line drugs was performed at a frequency informed by the strategy (Table 1), with sensitivity and specificity for diagnosing resistance incorporated (Table S1). Additional detail on model structure is provided in the Supplement.
The effect estimate for cure in BPaLM-based strategies as compared to the standard of care was modeled as the trial estimate for sputum culture conversion from TB-PRACTECAL, conditional on the number of effective drugs in the regimen, up to a maximum of four (i.e., four effective drugs confer a faster cure rate than three, but five or more effective drugs do not confer a faster cure rate than four).8 We varied this parameter in sensitivity analysis. Figure S4 displays the modeled rate of acquisition of new resistance to each drug conditional on the number of effective drugs, also to a maximum of four. Table S1 details the derivation and values for these and all other model parameters.
Outcomes
We measured health outcomes in quality-adjusted life years (QALYs). For each modeled individual, this measure sums the years of life over the individual’s remaining lifetime, weighted by the health-related quality of life experienced in each month.20
We measured the impact on drug resistance by summing for each individual, and for each of 12 anti-TB drugs, the number of months they experienced TB disease with resistance to that drug. We then calculated three summary measures for the impact on drug resistance. In the first, we calculated the mean duration with resistance to each drug for the entire cohort by aggregating the time with resistance across the whole cohort for each drug, then dividing by the size of the starting cohort. Second, we calculated the mean duration of untreated TB disease with resistance to each drug by summing the time with resistance only among those individuals in Markov state (2)—TB disease no longer receiving treatment—and again averaging across the starting cohort. These measures were designed to reflect the relevance of the policies for transmission of drug resistance. We calculated both because—for individuals no longer receiving treatment—there could be a higher risk that M. tuberculosis will transmit to another host, compared to the cohort as a whole. Third, we calculated the lifetime cumulative incidence of acquiring resistance to each drug, per individual in the cohort.
As a set of secondary health outcomes, we calculated the number of severe adverse events (SAEs) experienced per patient to each of the drugs, total life years (LYs), and the proportion of individuals experiencing the end-of-treatment outcomes of Success, Failed by Treatment, Lost to Follow-up (LTFU), and Dead, as would typically be reported programmatically to the WHO.
We measured the total costs under each strategy from the societal perspective in 2022 United States dollars ($) as the sum of direct healthcare, direct non-healthcare, and indirect costs accruing in each period. Direct healthcare costs (i.e., arising directly from the consumption of healthcare goods and services) were calculated by adding the costs of the drugs received, DST to second-line drugs, and a baseline healthcare resource utilization in the form of inpatient and outpatient services. Each SAE was accompanied by a utilization cost for inpatient and outpatient services. Direct non-healthcare costs (e.g., transportation) and indirect costs (e.g., productivity losses) accrued for every additional month on treatment. The indirect costs also accrued for those LTFU prior to cure. Productivity losses secondary to early mortality were not included in total costs, and were calculated separately.
Undiscounted values were calculated at for all outcomes. For QALYs and total costs only, discounted values were also calculated using an annual discount rate of 3%.
Cost-effectiveness analysis
First, we ruled out dominated strategies (i.e., those strategies that were both more expensive and provided fewer QALYs on average than a linear combination of other strategies. We then calculated the incremental cost-effectiveness ratio (ICER; a measure of the additional cost required to produce one additional QALY, as compared to the next cheapest, non-dominated strategy) and identified the cost-effective strategy as that with the greatest health gains subject to the constraint that— in order to provide value for money—the ICER be below the willingness-to-pay (WTP) threshold.21, 22 Lower ($4700 per QALY) and higher ($7021 per QALY) benchmarks for these thresholds in Moldova were based on published estimates using an opportunity cost approach,23 updated to 2022 USD (Table S1). As ICERs may be difficult to interpret in some cases,24 we also calculated the Net Health Benefit (NHB) of each strategy (see Supplement B), with the cost-effective strategy identified as that with the highest NHB.21 This is mathematically equivalent to the ICER approach. The 2022 CHEERS checklist is included in Supplement A.25
Statistical analysis
We estimated results via individual-level microsimulation, with lifetime outcomes for each of 10,000 individuals simulated for each of the diagnostic and treatment strategies described above.
Sensitivity analyses
Probabilistic Sensitivity Analysis (PSA) was conducted to account for uncertainty by constructing distributions for model input parameters (Table S1). In a second-order Monte Carlo simulation, we drew 1,000 parameters sets from the distributions. For each parameter set, the 10,000 individuals were simulated through each strategy, and a set of results was calculated. Finally, point estimates for each outcome were calculated as the mean of these 1,000 second-order simulations, and 95% uncertainty intervals (UIs) were constructed using the 2.5th and 97.5th centiles.26
Some important model parameters have substantial uncertainty. We performed one-way sensitivity analyses on two of these key inputs to understand the relationship with study outcomes. First, we varied the main effect estimate for cure across a uniform distribution fit to the published 95% confidence interval (1.18, 2.14).8 Next, we varied the prevalence of fluoroquinolone resistance among diagnosed RR-TB across the uniform distribution (0%, 40%) to aid the generalization of results to settings with a different prevalence of fluoroquinolone resistance.
Validation
We validated the modeled end-of-treatment (EOT) outcomes to estimates reported to WHO over the period 2010-2019. Further detail is provided in the Supplement.
RESULTS
Health effects, costs, and cost-effectiveness
Health effects, costs, and cost-effectiveness results for all strategies are presented in Table 2 and Figure 1. Among the 6-month BPaLM strategies, the highest health benefits were achieved by Strategy (1) (BPaLC if Mfx stopped, second-line DST upfront, then repeated at 4 monthly intervals), with undiscounted QALYs of 14.91 (95% UI: 12.98, 16.81). Among the standard of care strategies, Strategy (7) (DST every 4 months) had the best performance, producing 14.99 (95% UI: 13.05, 16.80) undiscounted QALYs, an additional 0.08 QALYs (95% UI: −0.52, 0.68) compared to Strategy (1).
Cost-effectiveness plane and acceptability curve
Lifetime costs by category
Strategy (5) (6-months BPaLM, second-line DST at 4 months and then every 4 months, BPaLC if Mfx stopped) had the lowest undiscounted lifetime total costs ($8176, 95% UI: 6245, 10379), followed by Strategy (1) and Strategy (2) (Table 1).
Compared to 6-month BPaLM-based strategies where BPaLC was used if Mfx had to be stopped, strategies continuing only the three-drug regimen BPaL (Strategies (3), (4), and (6)) resulted in worse overall health and additional lifetime total costs. The frequency of second-line DST did not lead to large differences in health or cost outcomes (Figure 1).
We compared cost-effectiveness ratios (ICERs) to current cost-effectiveness criteria for Moldova, with the willingness-to-pay for health improvements assumed to fall between $4700 and $7021 per QALY gained. According to this approach Strategy (1) (6-months BPaLM, DST upfront then every 4 months, BPaLC if Mfx stopped) was the cost-effective strategy with an ICER of $1181 per QALY. Strategy (7) was potentially cost-effective, but only with a willingness to pay over $60,622 per additional QALY. Figure 1B shows the probability of each strategy being cost-effective at a given cost-effectiveness threshold. Within the range of cost-effectiveness thresholds for Moldova, Strategy (1) had the highest probability of being cost-effective (37-42%, depending on the threshold). Taken together, the probability that one of Strategies (1), (2) and (5) (i.e., strategies that use 6-months BPaLM and continued BPaLC for those stopping Mfx) would be cost-effective was greater than 92% across the range of cost-effectiveness thresholds for Moldova. In subsequent sections, we make comparisons between the best performing (i.e., cost-effective) 6-month BPaLM-based and standard of care-based strategies: Strategy (1) and Strategy (7), respectively.
Drug resistance
Compared to Strategy (7), Strategy (1) increased the mean duration with resistance to pretomanid and delamanid among the entire cohort, and also when counting time only among those no longer receiving treatment. For all other drugs, however, Strategy (1) was estimated either to produce no statistically-significant change (i.e., 95% UI for the difference included zero) or to produce statistically significant reductions in the duration with resistance (Figure 3, Figure S8). Compared to Strategy (7), Strategy (1) was estimated to result in either no statistically-significant difference, or a lower lifetime cumulative incidence of resistance for all drugs, except for pretomanid and delamanid (Figure S9).
The impact of the BPaLM regimen on drug resistance
Secondary outcomes
Under Strategy (1), the mean number of SAEs ever experienced per individual was 0.238 (95% UI: 0.197, 0.284). Strategy (7) resulted in a mean number of SAEs of 0.265 (95% UI: 0.235, 0.297), an increase of 0.028 (95% UI: −0.011, 0.059) compared to Strategy (1). Figure 4 displays the proportion ever experiencing an SAE to each drug; the estimates were lower for Strategy (1) than for Strategy (7) for all drugs except linezolid and pretomanid. When health benefits were measured using life years instead of QALYs, Strategy (7) was again estimated to provide a small, non-significant health benefit over Strategy (1) (Table S3). The lowest life expectancy was estimated for Strategies (3), (4), and (6) (BPaLM-based strategies where BPaL was continued in the event of Mfx being stopped). End of treatment outcomes are shown in Figure S10.
Severe Adverse Events
Sensitivity Analyses
Figure 5 shows how cost-effectiveness results change for different values of the hazard rate ratio (HRR) of cure, and the initial prevalence of fluoroquinolone resistance, for Strategy (1) as compared to Strategy (7). In these results, Strategy (1) was estimated to be cost-effective (i.e., had a positive Net Health Benefit) compared to Strategy (7) across the range of values used for these parameters. Similarly, total costs were lower for Strategy (1) compared to Strategy (7) across the range of values assessed. Health outcomes were sensitive to the value of the HRR for cure for the BPaLM regimen as compared to standard of care regimens. For low values of the HRR (HRR = 1.2), Strategy (1) was estimated to lead to a mean 0.55 reduction in QALYs. For high values (HRR = 2), Strategy (1) would lead to a mean 0.35 gain in QALYs.
Sensitivity analyses on the hazard rate ratio of cure and the prevalence of fluoroquinolone resistance
DISCUSSION
In this study we assessed the potential health impact and cost effectiveness of a 6-month BPaLM regimen for treating RR-TB in a setting with a high prevalence of TB drug resistance. Compared to a strategy using 9-18 month regimens based on the 2020 WHO treatment guidelines for drug-resistant TB, we found the 6-month BPaLM regimen would be cost-effective across a range of WTP thresholds, with substantial reductions in the duration and cost of treatment, but little expected change in health outcomes. If implementing a 6-month BPaLM regimen, our analysis suggests that individuals stopping moxifloxain— because of an SAE or a resistant DST result—would be better off with the addition of clofazimine than continuing on BPaL alone. Holding the drug regimen constant, the frequency of second-line DST (to fluoroquinolones and injectables) did not result in substantial differences to health or cost outcomes.
These findings are in line with cost-effectiveness analyses of the TB-PRACTECAL randomized controlled trial,15 which drew similar conclusions for populations across South Africa, Belarus and Uzbekistan. Belarus also has a high proportion of RR-TB among newly diagnosed TB cases,2 but we do not know whether the joint distribution of resistance to other important drugs would differ between Belarus and Moldova. Although South Africa and Uzbekistan have a lower prevalence of resistance to many drugs, we found that 6 months of BPaLM remained cost-effective when the proportion of RR-TB patients with FQ-R was varied across the wide range of 0-40% (compared to Moldova at 28%). Our analysis builds on the aforementioned cost-effectiveness analysis by explicitly modelling the acquisition of drug resistance, with the initial cohort resistance profile informed by genetic sequencing data from Moldova. We also investigated the potential effects of a larger number of policy implementation scenarios, including the frequency of DST, and whether patients having to stop Mfx under BPaLM should continue on BPaL alone or continue on an alternative four-drug regimen.
When modeling the effectiveness estimate for BPaLM as compared to the standard of care, we assumed that the treatment effect for true cure in the model was approximated by the treatment effect for sputum culture conversion from the TB-PRACTECAL trial.8 Although the trial measured clinical outcomes, its primary composite outcome measure combined treatment failure, discontinuation, LTFU, death and recurrence, outcomes that are important to distinguish to calculate long-term health outcomes. The numbers of individuals experiencing each of the long-term outcomes of greatest clinical interest were very small. Even if the effect on true cure is not the same as on culture conversion, we found that 6-months BPaLM remained the cost-effective strategy when the HRR (point estimate 1.59) was varied over the published 95% confidence interval (1.18, 2.14).
While both regimens perform best at lower levels of resistance, sensitivity analyses showed that 6 months of BPaLM may result in a small, non-significant reduction in total QALYs as compared to the standard of care at lower levels of initial FQ-R, or if the BPaLM regimen has lower comparative effectiveness than estimated in the TB-PRACTECAL trial, even while it provides overall value for money. Although some policymakers may be uncomfortable adopting interventions that reduce health benefit on expectation, this difference is small and highly uncertain, and adopting the new regimen would bring substantial benefits in the form of reduced regimen duration, and freeing up funding to spend on other health interventions.
In this analysis we found that 6 months of BPaLM improved or resulted in no change in the duration of disease with resistant strains of M. tuberculosis as well as the cumulative incidence of resistance for all 12 anti-TB drugs, except pretomanid and delamanid. Both measures were influenced by the starting profile of resistance as informed by the WGS data, the rate of acquisition of new resistance to each drug under each modeled drug regimen, and the rate of cure. Changes in the rate of acquisition of resistance are important for individuals undergoing treatment today (some of the effects of this are captured in the QALYs estimated under each strategy) but preventing new resistance is also important for the health outcomes of those who will be diagnosed with RR-TB in the future.
This analysis had several limitations. The modeled strains of M. tuberculosis were based on WGS data from culture positive sputum specimens in 2018-19 in Moldova, and so may not accurately describe current resistance patterns in Moldova or resistance elsewhere, although we hope the sensitivity analysis on the prevalence of FQ-R aids in the generalization of findings. The hazard rate ratio for cure was based on the outcome of sputum culture conversion from TB-PRACTECAL; while culture conversion is a prognostic marker in TB,37 it is not a perfect substitute to quantify the rate of true cure, which is unobservable. Further, real-world outcomes with 6-months BPaLM are likely to be less favorable than in the high-fidelity environment of a randomized controlled trial—for example, there may have been a higher frequency of follow-up in the trial—and the status quo may differ between settings. We did not explicitly model the differences in adherence that may exist between regimens, and we made the simplifying assumption that increasing the number of effective drugs increases the rate of cure and reduces the rate of acquiring resistance. This was based on a previously applied approach7 and is likely to hold qualitatively, but we did not account for the all the differences that may exist between drugs and the interactions between them. For example, the effectiveness of BPaLC vs. BPaL may not be the same as the effectiveness of BPaLM vs. BPaL, yet—SAEs aside—the modelling approach was agnostic to this, conditional on the resistance profile of the strain of M. tuberculosis. Although the probability of an SAE was modelled separately for each drug, we did not incorporate the real-world variation in the duration and consequences of each type of SAE. Finally, we did not formally estimate the longer-term implications of resistance; for this, it would be necessary to model the transmission dynamics of M. tuberculosis. Instead, we estimated the cumulative incidence and duration of resistance as surrogates for the long-term health outcomes they may affect, insofar as lower incidence and fewer months of resistant disease would each result in less transmission of resistant strains.
This study was conducted in the setting of Moldova, a country with a high proportion of RR-TB with resistance to second-line drugs. Through sensitivity analysis on the proportion with FQ-R, we aimed to aid the generalization of findings to other settings. Many of the health-related model parameters are also generalizable beyond Moldova, as TB outcomes under the standard of care were informed by multi-national meta-analyses, and the estimate for comparative effectiveness was from a multi-national trial (Table S1). However, many of the cost parameters were from Moldova and Georgia (GDP per capita of $5,563 and $6,628 in 2022, respectively),38 and so there are likely limitations in the generalization of incremental costs of 6 months BPaLM compared to the standard of care, especially to countries with very different income levels.
To optimize clinical care for RR-TB, decision makers must take account of important health and economic effects for affected individuals as well as society at large. In this study, we estimated favorable outcomes under the 6-month BPaLM regimen in settings with a high burden of drug resistance, conditional on BPaLC being used in the event of moxifloxacin being contraindicated, rather than BPaL alone. The schedule of second-line DST did not appear to affect health outcomes or costs to a great degree across the finite number of DST schedules we explored, and further analyses may be warranted to explore the optimal testing frequency in Moldova and other settings—especially where second-line DST capacity is limited or unavailable39—and to explore additional technologies beyond MGIT for identifying resistance to fluoroquinolones and injectables. RR-TB treatment policy would further be enhanced by continuing to strengthen the empirical evidence base for the comparative effectiveness of different treatment regimens, with sufficient numbers of patients to quantify long-term health outcomes across multiple settings.
Funding Acknowledgment and Disclaimer
This publication was made possible by Grant Numbers T32 AI007433 and R01 AI146555-02 from the National Institute of Allergy and Infectious Diseases. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the NIH.
Potential Conflicts of Interest
JF has received grant funding from the Stop TB Partnership’s Global Drug Facility to support the roll out of child-friendly formulations of second-line TB drugs.
SUPPLEMENTAL INFORMATION
A. CHEERS 2022 Checklist
B. Additional Detail on Microsimulation Model
1. Markov Health States
The model mechanisms included Markov health states and individual patient- and drug-level trackers. The Markov health states were:
Receiving TB treatment
TB disease not receiving treatment
Cured post-treatment
Dead
2. Transitions between Markov Health States
The Markov cycle length was one month. All individuals started in state (1), and stayed there until the earliest of: death, LTFU, or discharge. While in state (1), an individual could be cured of their TB disease. We assumed that all truly cured patients were correctly identified as such; these individuals were discharged after the completion of their planned regimen and transitioned to state (3). When those individuals not truly cured were assessed at the end of treatment, 90% (Table S1) received a 6 month treatment extension, and the remainder were incorrectly assessed to have successfully completed treatment and were discharged. Patients receiving treatment extensions were assessed in the same way at the end of each extension. Individuals discharged before true cure were moved into state (2) and faced a monthly probability of relapsing and returning to treatment.
Individuals also faced a risk each month of becoming LTFU; if this occurred before they were cured, they transitioned to state (2) with a monthly probability of returning to treatment, and if they had already been truly cured they moved into state (3). Self-cure was possible from state (2), at a lower rate than cure on treatment. All simulated individuals faced a risk of death each month, with the mortality rate lower following cure. Prior to cure, the rate of death was higher for those who had stopped treatment. Quality of life was lowest at the start of treatment, and improved as the treatment course progressed, until it reached the maximum value upon true cure.
3. Individual patient- and drug-level trackers
For each simulated individual we tracked whether they had been truly cured during treatment (i.e., before they reached the end of the prescribed regimen and moved into state (3)), their current drug regimen, the duration of treatment with each drug, whether the strain of M. tuberculosis was resistant to each drug, the duration of resistance to each drug since treatment initiation, the results of the most recent DST performed, whether they had experienced a SAE to each drug, and the observed EOT outcomes (Figure S6). This set of trackers collectively informed the event probabilities and health state utility weights (Table S1). In each month, these trackers were also used to calculate the number of effective drugs, by matching the drugs being used in the treatment regimen with the (true) resistance profile to that drug. This variable influenced the probability of cure and the probability that each individual’s strain of M. tuberculosis would acquire resistance to any new drugs (Table S1, Figure S4). Differential adherence by strategy was not modeled explicitly, and we assumed that the effects of imperfect adherence were reflected in published effectiveness estimates.
4. Outcomes
4.1 Calculating Net Health Benefit
NHB was calculated according to convention:21 total discounted costs were converted into QALYs of equivalent value using the exchange rate of WTP for gains in health, itself measured in $ per QALY. These were then subtracted from the total discounted QALYs to produce NHB, measured in units of QALYs.
4.2 End of Treatment Outcomes
We recorded the observed EOT outcome for each simulated individual in the model. We then validated these modeled outcomes against data reported to the WHO from Moldova, presented separately for RR-TB (i.e., MDR/RR-TB) overall and for the subset of patients with XDR-TB. In this analysis, the definition of XDR-TB used is the older definition used by WHO,2 meaning TB that is resistant to any fluroquinolone and to at least one of three second-line injectable drugs, in addition to isoniazid and rifampicin. We used the older definition to allow validation of the modelled EOT outcomes against WHO data. The classification structure for recording EOT outcomes in the model is shown in Figure S6, along with our best interpretation of the EOT structure in the empirical data.
5. Modeled events
5.1 Death
Patients were exposed to a monthly risk of mortality, incorporating the risk from background causes and disease-specific mortality from TB. The TB-specific mortality risk was highest for individuals with TB disease that was not currently being treated. Treatment lowered the TB-specific mortality risk. After an individual was cured, the TB-specific mortality risk decreased but was still greater than zero to account for post-TB sequelae (Table S1).
5.2 Cure
For the first month of treatment and for untreated TB disease, an individual was allowed to self-cure. A higher cure rate was assumed from month two of treatment onwards (conditional on the treatment strategy and the number of effective drugs each month, as previously mentioned).
5.3 Loss to follow-up (LTFU)
Each month, an individual receiving treatment could become LTFU, with a decreasing monthly probability over time (Table S2). Those LTFU before true cure entered the “2) TB Disease – no longer receiving treatment” state and were allowed to subsequently recommence treatment. Those LTFU after true cure, but while still receiving treatment, entered the “3) Cured post-treatment” state and did not return to treatment. The probability of return to treatment is described in Table S1.
5.4 Drug resistance acquisition
The resistance status of each individual’s strain of M. tuberculosis to each drug was assumed to be binary (susceptible or resistant). The probability of developing resistance in a given month to each drug was a function of the number of effective drugs in the regimen. Conditional on the number of effective drugs, the probability of acquiring resistance was independent and identically distributed for all drugs to which M. tuberculosis was exposed in any month, with three exceptions: pretomanid resistance was assumed to confer immediate delamanid resistance and vice versa,40 bedaquiline resistance was assumed to confer immediate clofazimine resistance (but not vice versa),41 and the rate of acquiring resistance to linezolid was assumed to be half the rate as to other drugs.42, 43
We assumed that if an individual had resistance to a drug at a given time, the M. tuberculosis strain could not later revert to being susceptible. While resistance status was tracked for all patients with TB disease regardless of whether they were being treated, the probability of acquiring resistance fell to zero for all drugs in the month following treatment cessation.
5.5 Drug Susceptibility Testing
Individuals received DST by MGIT for moxifloxacin and amikacin according to the frequency prescribed by the strategy (Table 1). We assumed that individuals who had been truly cured would not be able to produce an adequate sputum sample, and as such did not undergo DST. We also assumed full adherence to the prescribed DST regimen. In the event of detecting new drug resistance, the respective drug was discontinued for that individual with a lifetime contraindication, and was replaced immediately. The only exception to this rule of replacement was for moxifloxacin under variations of the 6 months of BPaLM strategy, where some strategies continued BPaL only, rather than replacing moxifloxacin (Table 1).
Sensitivity and specificity values for DST were estimated from the literature (Table S1).
5.6 Severe Adverse Events (SAEs)
We modeled a generalized severe (i.e., grade 4-5) treatment-related SAE, which we assumed took place in any of the first three months of treatment with each drug, but not after. An SAE resulted in the responsible drug being discontinued that month, with a lifelong contraindication and small lifelong decrement to that individual’s health-related quality of life (Table S1). Each additional SAE was assumed to confer the same incremental reduction in quality of life and healthcare resource utilization.
6. Missing Data
In the genomic sequencing dataset, for observations with missing age (n = 12; 1.8%) we imputed the mean age of the cohort (42 years). There was no missing drug resistance data.
7. Probabilistic Sensitivity Analysis
For each input parameter, distributions were fit to the published measure of dispersion (95% confidence intervals or standard deviation) where available. Where there was no accompanying measure of dispersion, we assumed a standard deviation equal to one third of the mean. Information on all distributions is provided in Table S1.
C. TABLES S1–S3
D. FIGURES S1-S10
Schematic of the initial workup phase for the standard of care
Markov state-transition diagram
Genomic sequencing data exclusion criteria
The rate of acquiring drug resistance
Drug resistance at treatment initiation
Assumptions on the classification of End of Treatment Outcomes
Sensitivity analyses on the main effect estimate and the prevalence of fluoroquinolone resistance for the outcome of Incremental Life Years
Duration of TB disease with resistance to each drug
Lifetime cumulative incidence of drug resistance
Validating End of Treatment (EOT) outcomes
E. CALCULATED MODEL PARAMETERS
(following page)
Calculated Model Parameters
Note on terminology
As in most of the modeling literature, but not always the wider public health literature, we use the term rate and probability/proportion to mean different things. The former is in the sense of an instantaneous rate of change. Unlike a probability/proportion, it can be greater than 1. To convert between the two when the rate is constant over time, we use P (X) = 1 − e−r∗t, where P (X) is the probability of an event X occuring in a time period of length t, r is the event rate per unit of time t, and 
.
There are several model parameters that are not well approximated by the available data. In some cases, we adjusted published estimates to better fit the model’s processes. Details for some of these parameter calculations are summarized in Table S1, but here we provide some additional notes on calculations for some parameters.
1. The probability of being correctly identified as not cured (i.e., non-success), conditional on being truly not cured
“True” cure here refers to the elimination of M. tuberculosis in the host, such that it is no longer capable of causing disease. This is contrasted with an observed cure, where the patient clinically appears to have successfully completed treatment, but may subsequently have recurrent disease. “TB” here refers to all RR-TB specifically.
This is an unobserved (i.e. latent) process and as such cannot be estimated from empirical observation.
Instead, we use an estimate from Blöndal et al 2012 [50]. This paper reported that among 129 patients observed to be successfully treated at the end of their TB regimen, 11 had recurrent disease over a median follow-up time of 98.1 months.
The paper also demonstrates that approximately 90% of the cohort had event-free survival by the end of follow-up (Fig. 4 shows Kaplan-Meier curves separately for: a) non-XDR but MDR-TB, and b) XDR-TB. Though there was varying duration of follow-up, we also apply the 98.1 month median follow-up duration here for calculations.
For simplicity, we make the following assumptions with regard to this study:
we assume that all recurrences in this paper are relapses returning to treatment, not reinfection
we assume that all patients who are truly cured are correctly identified as such, and thus do not receive a treatment extension
we assume that there is no self-cure applied to this group
The following diagram represents some of the processes of interest here.
The quantities of interest in the figure are as follows:
98 months, the median follow-up time in the study, adopted as the timeframe for this process
a, the proportion who are not truly cured among those who complete a TB treatment regimen,
b = 1 − a be the proportion who are truly cured among those who complete a treatment regimen,
p1, the proportion of those not truly cured who are undetected/misclassified. These individuals are observed to be successfully treated, and thus are discharged from care.
p2, the proportion of those undetected, not truly cured individuals who eventually return to treatment by 98 months.
1 − p2, the proportion of those undetected, not truly cured individuals who die by 98 months, individuals who are not truly cured and are correctly detected receive a treatment extension as described elsewhere,
The estimate from the study can be represented as follows:
Pr(return to treatment|observed successfully treated) = 
Defined for this population, (i.e., those observed to have been successfully treated):
Let C be the event that such an individual has truly been cured.
Let R be the event that such an individual returns to treatment.
Let D be the event that such an individual dies.
We start therefore with
Because an individual would only be recorded as returning to treatment or dead, not both,
which implies
and thus,
Applying LOTP to P (R),
Because no individual who has truly cured will relapse, this simplifies to
Let us assume that, by 98 months, almost all (let us say 95%) of individuals who were not truly cured at the end of treatment will have either returned to treatment or died.
Let the monthly rate of relapse be μR.
Let the monthly mortality rate of TB on treatment be μD on Rx. The one month probability of death while on treatment is 0.004,
To determine P (D|C),
And so,
assuming a constant relapse rate over the course of 98 months.
Let us also use this estimate to inform the proportion of this cohort not truly cured. Recall
Because of the competing events of death and relapse in this problem where only one is recorded, I simulated a Markov model over 98 monthly cycles using the rates calculated above. After 98 months, 82.5% of the not truly cured had returned to treatment, and 12.5% had died without returning to treatment. As such,
So in this population of individuals who had been observed to successfully complete treatment, we estimate that approximately 10.3% were not truly cured of the disease.
Now let’s obtain the estimate for the proportion misclassified.
Referring back to the diagram and associated notation, we can express P (C) = 1 − P (C) as
While a and b = 1 − a will change based on the cohort characteristics, treatment and other factors, we model p1 as a variable independent of the overall proportion who truly have disease, i.e. it is a test characteristic, 1-sensitivity of a clinical assessment “test” to detect those with the “disease” of being truly cured.
The proportion with observed success in the study by Blondal et al. [50] is 129/211 = 0.611 We note that this is approximately the same as the overall proportion with observed success in the Bastos et al. [47] review of 0.64.
This can be represented as the denominator of the expression above,
This can now be substituted into the equation above to solve for p1.
We obtain an estimate that 13.9% of patients who are truly not cured are misclassified as cured at the end of treatment (i.e. clinical assessment is 84.1% sensitive in being able to detect those not truly cured).
2. Applying the proportion misclassified to the cure rate and proportion cured
We use the value for p1 above to calculate the proportion truly cured from the proportion of observed successes reported.
We assume that the “test characteristics” for diagnosing cure are the same across time, and as such p1 can be applied to determine the true cure at each point clinical assessment takes place in the model.
To distinguish between the notation above for the Blondal et al 2012 paper [50] and the events in our model,
Let c be the event that an individual is truly cured. Let t be the event that an individual is reported as cured (and is thus observed so in the data).
For example, given the reported proportion cured at the end of treatment in the Bastos et al. review [47]
Thus, while 64% were reported to be cured, we would estimate that just under 60% were truly cured, and the rest (4%) were truly not cured and had been misclassified as cured.
3. Calculating different mortality rates for those cured and not cured, while on treatment
There are estimates from the literature [47] for the proportion who are dead at the End Of Treatment (EOT). Using these estimates:
Let p(dead)o,EOT,MDR be the probability that an MDR-TB patient (here, exclusive of XDR) will be observed as dead at the EOT.
p(dead)o,EOT,MDR = 0.08, 95% CI: (0.07, 0.09).
Let p(dead)o,EOT,XDR be the probability that an XDR-TB patient will be observed as dead at the EOT.
p(dead)o,EOT,XDR = 0.21, 95% CI: (0.18, 0.25).
Based on the model mechanisms, we allow for truly cured individuals to have a lower mortality rate, which is a multiplier of background mortality. As a result, the overall mortality estimates from the literature are an average: at any given time point, the observed mortality rate is assumed to be a mean mortality rate of a two groups, one (lower) mortality rate in the cured group, and another (higher) mortality rate in the non-cured group. The challenge is that the proportion cured varies over time. The following details the approach we have taken, which is applied to the MDR-only and XDR population separately.
Let μo,T = −ln(1 − p(dead)o,EOT) be the observed mortality rate in the population over the whole treatment course, converted from the observed proportion dead at the EOT.
Let μC be the monthly mortality rate among those truly cured. This does not vary by month.
Let μNC be the monthly mortality rate among those not truly cured. This does not vary by month.
Let t ∈ {1, 2, …, 21} denote the month of the regimen. We assume for these calculations that EOT outcomes are recorded at 21 months, the modal regimen duration in the studies comprising the Bastos et al. 2017 review.
Let p(cured)t be the proportion of alive patients truly cured at the end of month t.
Among alive patients, true cure and true non-cure are mutually exclusive and collectively exhaustive. As such, p(cured)t + p(cured)t = 1, ∀t
Then,
We are left with a formula with two unknowns, μNC and p(cured)t.
The next step is to provide an approximation for p(cured)t, the proportion of alive patients who are cured at time t. Let p(cured)0 = 0.
Let p(cured)1 = 0.019, allowing only for the mechanism of self-cure in month 1.
Let p(cured)21 be the proportion of patients cured among those alive from the Bastos et al 2017 review. As such, for MDR-TB patients, this will be approximated by a numerator of the observed proportion of treatment successes at end of treatment, and a denominator of the proportion alive at the end of treatment:
For XDR-TB patients, this is approximated in the same way:
To obtain the proportion cured among those alive for months 2-20 in these calculations, we make the simplifying assumption of a constant rate - among the alive non-cured - of converting to cure. We also ignore the dual mechanisms of self-cure and treatment-related cure, along with any differences in duration between MDR- and XDR-TB regimens:
As such, we obtain a monthly cure rate for alive MDR-TB patients, μcure,MDR as follows:
Similarly, we obtain the monthly cure rate for alive XDR-TB patients, μcure,XDR as follows,
From these, we can obtain the cumulative proportion of alive patients cured during each month (not shown). We plug these into formula (1),
μC, the monthly mortality rate among the cured, is a function of background (ASR) mortality. The median age was 35 in the Bastos review, and the modal country of included studies was South Africa, which is also approximately middle of the pack for the income level of the meta-analysis. As such, we adopt the monthly mortality rate for a 35 year old individual in South Africa, and the mortality rate ratio (MRR) for MDR-TB from the parameter table.
μASR = 0.00069
MRR = 3.07
μC = μASR * MRR
μC = 0.00069 * 3.07
μC = 0.00211
The mortality rate among those cured is assumed for now to be the same among those with MDR-TB and XDR-TB, although this could be edited going forward.
From the Excel table “Estimating mort rates”, we have the following values too:
Putting all of this together, for MDR-TB:
And for XDR-TB,
General Acknowledgments
For helpful feedback on research-in-progress presentations, LPJ would like to thank current and former students, postdocs and faculty affiliated with: the Center for Health Decision Science, the PhD Program in Health Policy, and the Center for AIDS Research, all at Harvard University, Cambridge, MA, USA; the Decision Science Methods Group at Erasmus University, Rotterdam, the Netherlands.
REFERENCES
Subject Area
Reviews and Context
0
Comment
0
TRIP Peer Reviews
0
Community Reviews
0
Automated Services
0
Blogs/Media
Author Videos