We included studies that conducted an economic evaluation of anti-TB medications or care strategies for active TB. To include active TB patients, studies that included at least one arm of patients treated with anti-TB medication were included. Studies in which subjects were only animals or patients with non-active TB (e.g., latent TB) were excluded. If an article included active and latent TB patients simultaneously with separately presented results of active and latent TB, the article was selected. Studies evaluating the screening and diagnosis of and vaccination for TB were excluded. Articles published before 2010 or abstracts without full text were excluded.

We searched for studies published between January 1, 2010, and January 29, 2020, that economically evaluated active TB treatment, using the PubMed, EMBASE, and Cochrane Library databases. Considering each database system, we combined terms related to the participants of interests and study types using Medical Subject Headings and title and abstract (tiab) as search fields. The search terms were as follows: the Medical Subject Headings term of “tuberculosis,” “Cost-Benefit Analysis,” and the text string of “cost,” “effective*,” “utility,” and “model.” The complete search strategy used for each database is included in Supplementary Table S1. The process of selecting or excluding studies was carried out using Endnote software (Clarivate, London, United Kingdom) as a citation manager. After excluding duplicated articles, two reviewers (JY, MJ) screened the titles and abstracts of all the remaining articles according to the eligibility criteria. Two other reviewers (SH, HL) reviewed the full texts of the retrieved articles to determine whether they met the eligibility criteria. In cases of disagreement between reviewers, a consensus was achieved through discussion.

Using an Excel sheet, two reviewers (JY, MJ) independently extracted information from selected articles and discussed discrepancies during cross-checking. Data on the following characteristics of the included studies were extracted: target population, type of intervention, country, primary outcome, model type, time horizon, and health states included in the model. In addition, we extracted data on the key components of each economic evaluation using a predefined extraction format. The extracted components were intervention/comparator, outcome type, outcome values, cost-effectiveness thresholds, cost-effectiveness status of the outcomes, and funding information. We organized the status of cost-effectiveness of each outcome using stated thresholds or the author’s statements on cost-effectiveness. Two other reviewers (SH and HL) screened all the extracted data to confirm it was complete and accurate. Each cost-related outcome, based on various currencies depending on the country of the study, was converted into US dollars using the exchange rate of July 2020.

To assess the methodological quality of all the selected studies, we used the Quality of Health Economic Studies (QHES) instrument, which has been validated for appraising the methodological quality of economic evaluations (Joshua et al., 2003). The QHES instrument includes 16 criteria, and each criterion has a weighted point value, ranging from 1 to 9, allowing users to estimate total values up to 100. Two reviewers (JY, MJ) evaluated all the included studies using the QHES instrument and resolved any disagreements through discussion and consensus. We allowed an intermediate interpretation; we evaluated performance on a criterion as “partial” when part of the criterion was fulfilled but another was unfulfilled.

A total of 971 citations were identified through a database search. After excluding 309 duplicates, we screened 662 articles by title and abstract, leaving 38 articles for full-text screening. Through full-text screening, 17 articles were eligible for inclusion. A flow chart of the study selection process is presented in Figure 1. The basic characteristics of the 17 included studies are summarized in Table 1. Two studies were carried out in more than one country (Gomez et al., 2016; Lu et al., 2017); however, most (15 out of 17) were carried out in one country. Ten studies included results from low- and middle-income countries (LMICs) according to World Bank income classification and 7 of them (Manabe et al., 2012; Knight et al., 2015; Gomez et al., 2016; Lu et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; Loveday et al., 2018) were conducted in sub-Saharan Africa. Lu et al. (2017) conducted a cost-effectiveness analysis of seven countries specified as having a high burden of TB in the World Health Organization Global Tuberculosis Report 2014 (WHO, 2014), which include a high-income country (HIC) (Estonia) and LMICs simultaneously, to examine the health outcomes of novel treatment. Gomez et al. (2016) estimated the cost and cost-effectiveness of TB treatments in four different countries, two of which were middle-income countries, and the other two were low-income countries.

Five studies conducted cost-effectiveness analyses using various scenarios or settings, including different treatment costs, and durations (Owens et al., 2013), populations with different percentages of drug resistance using several types of outcomes (Law et al., 2014), different countries with different levels of adherence to guidelines (Gomez et al., 2016) or outcome types (Lu et al., 2017), and different TB cohorts and perspectives (Nsengiyumva et al., 2018).

Figure 2 presents a network diagram of the interventions and study populations. In terms of the study population, 7 studies (Hunchangsith et al., 2012; Manabe et al., 2012; Owens et al., 2013; Law et al., 2014; Knight et al., 2015; Gomez et al., 2016; Nsengiyumva et al., 2018) were conducted with drug-susceptible or treatment-naïve TB (hereafter treatment-naïve TB) patients who initiated TB treatment for the first time. Eleven studies were conducted with MDR-TB patients (Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; John et al., 2018; Loveday et al., 2018; Nsengiyumva et al., 2018; Fan et al., 2019). Nsengiyumva et al. (2018) conducted a study with treatment-naïve and MDR-TB patients using two distinct models and included latent TB patients. Six out of 8 studies which were conducted in HICs were dealing with MDR-TB patients (Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Fan et al., 2019), whereas studies from LMICs were relatively balanced with MDR-TB (6 studies) (Lu et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; John et al., 2018; Loveday et al., 2018; Nsengiyumva et al., 2018) and treatment-naïve TB (5 studies) (Hunchangsith et al., 2012; Manabe et al., 2012; Knight et al., 2015; Gomez et al., 2016; Nsengiyumva et al., 2018).

In terms of intervention types, most of the selected articles (13 out of 17) examined interventions of medications, comparing the standard 6-month regimen with longer (8-month) (Manabe et al., 2012) or shorter regimens (2 or 4 months) (Owens et al., 2013; Knight et al., 2015; Gomez et al., 2016) in treatment-naïve TB patients, or compared background regimens with new regimens that contain new drugs, such as bedaquiline, in MDR-TB patients (Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; Fan et al., 2019). One study (Law et al., 2014) compared standardized TB treatments with varying drug resistance in treatment-naïve patients with TB.

The other four articles examined “care” types, which were generally related to new methods of improving patient medication adherence, including comparing self-administered treatment or directly observed treatment (DOT) with other types of care in treatment naïve TB patients (Hunchangsith et al., 2012; Nsengiyumva et al., 2018) or comparing hospital-based centralized care with decentralized care in MDR-TB patients (John et al., 2018; Loveday et al., 2018). In a simulation study, Nsengiyumva et al. (2018) compared DOT with various monitoring strategies in treatment-naïve and MDR-TB patients.

Table 1 presents the components of the economic evaluation model. Nine out of 17 studies used a Markov model (Law et al., 2014; Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; Nsengiyumva et al., 2018; Fan et al., 2019), and 6 studies used a decision tree model (Hunchangsith et al., 2012; Manabe et al., 2012; Owens et al., 2013; John et al., 2018; Nsengiyumva et al., 2018; Fan et al., 2019). Fan et al. (2019) conducted an economic evaluation using a decision tree combined with a Markov model. Distinctively, Knight et al. (2015) used a transmission model, and Gomez et al. (2016) used an individual-based decision-analytic model. Loveday et al. (2018) conducted a study without stating an economic model. Three out of four studies that compared care types used a decision tree model (Hunchangsith et al., 2012; John et al., 2018; Nsengiyumva et al., 2018), while most studies (9 out of 13) that compared drugs or regimens used a Markov model (Law et al., 2014; Wolfson et al., 2015; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; Fan et al., 2019).

The most frequently used primary outcome was disability-adjusted life years (DALYs), used in 10 studies (Hunchangsith et al., 2012; Owens et al., 2013; Law et al., 2014; Knight et al., 2015; Wolfson et al., 2015; Gomez et al., 2016; Lu et al., 2017; Schnippel et al., 2018a; Schnippel et al., 2018b; Nsengiyumva et al., 2018). This was followed by quality-adjusted life years (QALYs), used in five studies (Wolfson et al., 2015; Park et al., 2016; Wirth et al., 2017; John et al., 2018; Fan et al., 2019). Wolfson et al. (2015) used DALYs and QALYs. Other outcomes included death (3 studies) (Manabe et al., 2012; Law et al., 2014; Knight et al., 2015), life-years gained (LYGs) (2 studies) (Park et al., 2016; Codecasa et al., 2017) and new TB cases (2 studies) (Law et al., 2014; Knight et al., 2015).

The time horizon varied across studies, from 2 years (John et al., 2018) to a lifetime (Hunchangsith et al., 2012; Owens et al., 2013; Gomez et al., 2016). Studies that used the Markov model tended to have longer time horizons (e.g., 10 or 20 years) compared with studies that used decision tree models (e.g., 2, 4, or 5 years), although two studies that used decision tree models had time horizons of a lifetime (Hunchangsith et al., 2012; Owens et al., 2013).

Among the health states included in the models for each TB population, studies generally included health states related to cure/fail, lost to follow-up (defaulted), treatment completion, and retreatment for MDR-TB and treatment naïve TB. However, health states related to sputum conversion, end-of-life/palliative care, comorbidity of human immunodeficiency virus (HIV), and surgery were found in models for MDR-TB but rarely found in models for treatment-naïve TB patients.

Six out of 7 studies which were conducted in sub-Saharan Africa were reflecting HIV status in their economic evaluation models. The 6 economic evaluation models used treatment outcomes according to HIV status and included cost for HIV treatment. Schnippel et al. (2018a; 2018b) included health states of HIV treatment such as “initiate HIV treatment” and “continue HIV treatment” in their Markov models. Loveday et al. (2018) adjusted for baseline variables including HIV and antiretroviral therapy (ART) status between 5 cohorts using propensity score weighting with a prospective cohort design. Manabe et al. (2012) divided the patients into HIV-positive and HIV-negative arm in decision tree model to reflect high co-infection of HIV and TB in Uganda. Gomez et al. (2016) used ART coverage level of five different countries as a scenario parameter using individual-based decision analytic model. The Characteristics of TB such as duration and smear positive rates differed by HIV status in Knight et al. (2015) which used transmission model.

To establish an economic evaluation model that is generally simulated by a hypothetical cohort, most of the included studies earned treatment outcome for each intervention and comparator mainly from published data of clinical trials, but Loveday et al. (2018) earned the treatment outcome from prospective cohort analysis, by observing the outcome from 5 cohorts which were allocated with 5 care strategies. In most included studies, safety data of the drugs was reflected in the economic evaluation models since data from clinical trials implicit safety data. Particularly, five studies mentioned how they reflected safety data on their economic evaluation models. Law et al. (2014) included blind health state in their Markov model to reflect ocular toxicity of ethambutol. Wolfson et al. (2015) modeled differences in safety between two strategies in their Markov model using data from randomized trial. Schnippel et al. (2018a) analyzed cost needed for screening adverse reaction. Schnippel et al. (2018b) compared cost-effectiveness with or without considering adverse reactions. Wirth et al. (2017) included cost for adverse events with detailed incidence and cost for each adverse reaction of medications.

Table 2 shows the results of the comparison of the regimens for drug-susceptible/treatment-naïve TB patients. In three studies which were conducted in LMICs, shortened regimens were more cost-effective than regimens with longer duration. The 6-month standard regimen was more cost-effective than an 8-month regimen (Manabe et al., 2012). A hypothetical 4-month regimen was more cost-effective than 6-month standard regimen in most (7 out of 9) comparisons (Knight et al., 2015; Gomez et al., 2016), but in Bangladesh, it was not more cost-effective than the 6-month standard regimen (Gomez et al., 2016). The cost-effectiveness of other new regimens was assessed in HICs. A strengthened retreatment regimen (regimen with second-line drugs for patients who failed or relapsed after cure or default) was more cost-effective than 6-month standard regimen in all comparisons (Law et al., 2014). A moxifloxacin-based 4-month regimen was more cost-effective than a 6-month standard regimen, whereas a moxifloxacin-based 2-month regimen was less costly and more effective in settings with high treatment costs, but it was not possible to assess cost-effectiveness in settings with moderate/low treatment costs because the exact amount of threshold was not suggested (Owens et al., 2013). Ethambutol-added regimen (2 months of isoniazid, rifampicin, pyrazinamide, and ethambutol, followed by 4 months of isoniazid, rifampicin, and ethambutol) was more cost-effective than the 6-month standard regimen when MDR case and TB death were used as outcomes but less effective when DALY was used as the outcome. Standardized MDR treatment (regimen received by patients who failed initial treatment for 24 months) was cost-effective when the outcome was DALY, but cost-effectiveness could not be assessed when MDR case and TB death were used as outcomes due to lack of threshold (Law et al., 2014).

The cost-effectiveness of regimens for MDR-TB patients is presented in Table 3. Three studies included comparison between bedaquiline-based regimen and background regimen in LMICs. For MDR-TB patients who failed the background regimen or for XDR-TB patients, bedaquiline-based regimen was cost-effective (Lu et al., 2017; Schnippel et al., 2018a), but for MDR-TB patients, the cost-effectiveness of bedaquiline-based regimens differed according to the toxicity profile. Bedaquiline-based regimen was less costly and more effective than injection-based regimens when adjusted for toxicity profile in terms of adverse drug reactions but costlier and less effective when adverse drug reactions were not considered (Schnippel et al., 2018b). Six studies assessed regimens with new drugs in HICs. For both MDR- and XDR-TB patients, bedaquiline-based regimens were more cost-effective than background regimen in 6 HICs (Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017; Fan et al., 2019). In Germany, bedaquiline-based regimen was more cost-effective than delamanid or linezolid-based regimens, whereas delamanid-based regimens were more cost-effective than linezolid-based regimens (Wirth et al., 2017). Bedaquiline-based regimen was more effective and less costly (Lu et al., 2017; Schnippel et al., 2018a) or had an incremental cost-effectiveness ratio (ICER) lower than the threshold (Park et al., 2016; Codecasa et al., 2017; Schnippel et al., 2018a) when compared to the standard regimen. Linezolid or delamanid-based regimens are not cost-effective compared to the background regimen (Wirth et al., 2017; Fan et al., 2019). Lu et al. (2017) included cost-effectiveness results from both HIC and LMICs. As a HIC, Estonia had the largest thresholds and probability of bedaquiline plus background being cost-effective was highest when highest price was applied. Regarding concomitant HIV status, Schnippel et al. (2018a) conducted sensitivity analysis by changing proportion of HIV positive patients, showing higher ICER with higher proportion. Gomez et al. (2016) showed lower ICER with higher survival in ART in South Africa, Bangladesh, and Tanzania, but the opposite result in Brazil.

Table 4 presents the results of the comparisons between care types for active TB patients. All the four studies were conducted in different LMICs. For treatment-naïve TB patients, DOT and mobile phone-based care were compared to self-administered treatment in Thailand, and it was not possible to assess their cost-effectiveness because all the values had uncertainty ranges that crossed zero, which indicates statistical insignificance (Hunchangsith et al., 2012). For MDR-TB patients, decentralized care provided from home (John et al., 2018) or the hospital (Loveday et al., 2018) was more cost-effective than hospital-based centralized care in India and South Africa, respectively. In particular, community-based care using mobile devices was more cost-effective than hospital-based decentralized care or community-based care in clinics (Loveday et al., 2018). For treatment-naïve and MDR-TB patients, video-observed therapy or medication monitoring with devices or blister packs were more cost-effective than DOT in Brazil (Nsengiyumva et al., 2018).

Results of the cost-effectiveness analysis are presented in Supplementary Table S2 for each study.

The results of the quality assessment are presented in Supplementary Table S3. In general, most items were fulfilled in the selected 17 articles, but data was not available for all 17 studies for Q4, the criterion regarding pre-specified subgroup analysis. For Q16, the criteria regarding the disclosure of funding, all the included articles except for one (Schnippel et al., 2018b) disclosed the source of funding. Five studies were funded by the pharmaceutical industry. For all five studies, the intervention, including bedaquiline, was more cost-effective than the comparators (Wolfson et al., 2015; Park et al., 2016; Codecasa et al., 2017; Lu et al., 2017; Wirth et al., 2017). Three (Schnippel et al., 2018a; John et al., 2018; Fan et al., 2019) declared no funding, and one (Schnippel et al., 2018b) did not clarify whether it had been funded.

Some studies did not fulfill all of the criteria of the QHES instrument. Manabe et al. (2012) did not conduct an incremental analysis between alternatives (Q6) or indicate the analytic horizon and discount rate (Q8), rating among the lowest of the included 17 studies. Another study by Loveday et al. (2018) also rated among the lowest because it did not explain the economic model of the research, receiving a “No” for Q12, 13, the criteria regarding the economic model of the study. The study by Schnippel et al. (2018b) had the second-lowest scores because it did not clearly disclose the funding source, receiving a “No” for Q16. Hunchangsith et al. (2012) did not state the analytic horizon and discount rate; hence, the study was rated as “partial” for Q8. The remaining studies fulfilled all the QHES items, except for Q4.