Equity, cost and disability adjusted life years of tuberculosis treatment supported by digital adherence technologies and differentiated care in Ethiopia: a trial-based distributional cost-effectiveness analysis

Introduction

The 2018 World Health Assembly on Digital Health highlighted the potential of digital technologies to advance universal health coverage, while cautioning that contributions should be carefully assessed to ensure investments do not divert resources from alternative, more cost-effective non-digital approaches [1]. Tuberculosis, a notifiable disease of public health importance, requiring adherence to treatment schedules of 6 months or more, may benefit from innovation to support treatment adherence. In 2022, 7.5 million people globally started treatment for tuberculosis, and with an estimated 1.3 million deaths, it remains one of the deadliest infectious diseases [2]. In Ethiopia the tuberculosis incidence rate is 126 (85-176) cases per 100,000 population compared to the global average of 133. Disease risk factors include undernourishment, HIV, alcohol use disorders, diabetes, and smoking. Of individuals newly started on tuberculosis treatment in 2021, 87% successfully completed. Current treatment options for drug-susceptible tuberculosis (DS-TB) include a six-month course of treatment or a 6-20-month course for rifampicin-resistant tuberculosis (RR-TB). Intermittent dosing of tuberculosis medication as is observed with suboptimal dosing is possible due to the post-antibiotic effect of drugs included in first-line treatment. However, the treatment regimen will be less bacteriostatic when dosing do not coincide with the metabolic activity of semi-dormant persistent tubercle bacilli [3]. Consistent adherence is therefore recommended to reduce the risk of treatment failure, relapse and acquired drug resistance, especially among persons with initial pulmonary cavitation or HIV disease.

Historically, directly observed therapy short course (DOTS) was used to achieve high rates of treatment success, necessitating persons with TB (PWTB) attending the health facility daily to be observed when taking their medication. Digital solutions for adherence monitoring has made community-based, patient-centered treatment adherence support possible. A range of digital adherence technologies (DATs) have been assessed with limited evidence of improvements in end of treatment outcomes [4–6] and cost-effectiveness [7–10]. Interventions range from video-supported therapy and digital pillboxes to SMS-based interventions. The use of these interventions may be dependent on PWTB’s access to digital devices, and concerns have been raised about equity implications [11]. Since lower household socio-economic position (SEP) increases individuals’ risk of contracting tuberculosis disease [12], simultaneously increasing the risk for suboptimal treatment adherence and poor health outcomes, it is important to assess the equity implications of digital interventions. As part of the adherence support coalition to end tuberculosis (ASCENT) project, a pragmatic cluster randomised trial was conducted to assess the effectiveness and cost-effectiveness of digital pillboxes (pillbox) and medication labels (labels) with daily monitoring using a web-based platform to inform differentiated care, in reducing a composite unfavourable outcome, including tuberculosis recurrence in Ethiopia. The trial did not find evidence of a difference in the composite unfavourable outcomes between the labels (adjusted odd ratio [aOR]=1.14; 95%CI:0.83-1.61; p=0.62) or pillbox study arms (aOR=1.04; 95%CI:0.74-1.45; p=0.95) when compared against the standard of care (SOC) [13]. We conducted a trial-based distributional cost-effectiveness analysis to assess the equity impact, costs and cost-effectiveness of digital adherence technologies with differentiated care compared to the SOC in Ethiopia.

Methods

Results

Between May 2021 and August 2022, 3885 adults (≥18 years) starting DS-TB treatment were enrolled, with follow-up visits completed in August 2023 (figure S1). Participants were excluded if their diagnosis changed to “not TB” or if they were diagnosed with RR-TB treatment less than 28 days before treatment start or were re-enrolled leaving 1295 participants in the standard of care (SOC), 1305 in the Labels -, and 1258 in the Pillbox arms of the study in the intention-to-treat population. Participants were analysed in the based on DAT received, leaving 1295 in SOC, 1152 in labels and 1411 in pillbox (table S1). Baseline characteristics of the clusters and trial participants are shown in Table 1.

Baseline characteristics were similar between study arms. 62% (806/1295) of participants in the SOC, compared to 66% (756/1152) in the labels and 66% (936/1411) in the pillbox arms had bacteriologically confirmed tuberculosis. Slightly more participants in the SOC arm had a history of previous tuberculosis disease at 9% (112/1295) compared to 5% (63/1152) in the labels and 6% (81/1411) in the pillbox arms. Considering the variables describing household socio-economic position (SEP), similar proportions of participants in the SOC 21% (277/1295) compared to the labels 23% (266/1152) and pillbox 23% (324/1411) arms of the study reported no individual income. Participants in the SOC arm were less wealthy (20% classified as lowest wealth quintile) compared to the labels and pillbox arms (18% and 19% respectively). Participants in the labels study arm were wealthier (43% (498/1152) in the two wealthiest quintiles) than those in the SOC (35% (450/1295)) and pillbox (36% (506/1411)) study arms. The number of health facility visits per treatment episode compared between study arms are shown in table S4. Compared to the SOC, there was a reduction in the mean number of visits to health facilities in the pillbox and labels study arms throughout the treatment period; labels compared to SOC by -20.2 (95%CI: -26.3; -14.1; p<0.001) and pillbox compared to SOC -20.4 (95%CI: -26.5;-14.3; p<0.001) (table S4). Cohort level patient events during treatment compared between study arms are summarised in table S5. There were more home visits in the labels (77), and pillbox (58) compared to the SOC (5) study arms. Similarly nights in hospital was greater in the labels (648) and pillbox arms (310) compared to the SOC (135). This was partly explained by one healthfacility in the labels arm that admitted more people to hospital, and may also be as a result of more intensive follow-up because of monitoring using the DATs. A total of 22 participants required a second course of tuberculosis treatment in the SOC, compared to 24 in the labels and 31 in the pillbox study arms. With 4 participants in the SOC, 1 in labels and 2 in the pillbox study arms starting rifampicin resistant tuberculosis treatment. The economic costs of treatment are summarised in Table 2.

Intervention costs were $13 (95%CI: 12; 13) per patient in the Labels compared to $16 (95%CI: 15; 16) in the Pillbox arm, with an empirically derived pillbox reuse rate of 2.16 (95%CI: 1.58; 2.85) included in the calculation. Per patient provider costs were much higher in the SOC $491 (95%CI: 443; 539) compared to the Labels $182 (95%CI: 177; 204) and pillbox study arms $178 (95%CI: 156; 200), driven by the larger number of health care visits in the SOC. The cost of medication, diagnostic tests, hospitalisation, and further treatment were similar between study arms. Patient-incurred costs are similar between study arms at $16 (95%CI 15; 16) per person in the SOC, $14 (95%CI: 13; 14) and $12 (95%CI: 11; 13) in the labels and pillbox arms respectively despite the large difference in the number of health facility visits in the intervention arms. The distributions of DALYs, health facility visits, patient – and provider costs by household socio-economic position (SEP) are shown in Figure 1.

• Download figure
• Open in new tab

Figure 1. illness concentration indices for (A) Disability Adjusted Life Years (DALYs); (B) Number of health facility viists; (C) Patient costs in 2023 USD; and (D) Provider costs in 2023 USD.

The concentration index (ICI) assesses if there is a within study arm difference in the outcome by ranked by household socio-economic position (SEP). Clustering was taken into account in estimating 95% confidence intervals (95%CI). A negative concentration index means that the outcome, here a negative consequence, is concentrated among less wealthy households. Conversely, a positive index value means that the outcome is concentrated among relatively wealthier households. We compared the difference in the distribution in outcome between SEP groups, where the p-value represents the likelihood of any observed difference between the groups being by chance.

The equity impact of the interventions in four domains, namely DALYs, health facility visits, patient-incurred costs and provider costs are shown in figure 1, figure s2 and table S6. The illness concentration indices (ICIs) for DALYs were -0.06 (95%CI: -0.24; 0.11) in SOC; -0.02 (95%CI: -0.17; 0.14) in labels and -0.02 (95%CI: -0.16; 0.12) in pillbox study arms, with no evidence of differences in the distribution of DALYs within or between study arms (p=0.920). The number of health facility visits were concentrated among the comparatively poor in the SOC -0.07 (95%CI: -0.10; -0.05) and pillbox - 0.08 (95%CI: -0.09; -0.06) arms but not in the labels arm 0.01 (95%CI: -0.00; 0.02). Patient costs were concentrated among those relatively poor in the SOC arm -0.05 (95%CI: -0.07; -0.02), while concentrated among the wealthy in the labels arm 0.03 (95%CI: 0.01; 0.05) and pillbox 0.01 (95%CI: - 0.01; 0.02) with evidence of differences between study arms (p<0.001). Provider costs were concentrated among the relatively wealthy, suggesting that the wealthy most benefits from public health expenditure, in the SOC 0.13 (95%CI: 0.08; 0.18); labels 0.07 (95%CI: 0.05; 0.08) and pillbox 0.07 (95%CI: 0.04; 0.10) arms, though there was insufficient evidence (p=0.225) of a difference between study arms. The cost-effectiveness and distributional cost-effectiveness of the interventions compared to the SOC are summarised in Table 3.

The difference in societal costs, using multiple imputation, was a reduction by $299 per patient in the Labels and $298 in the Pillbox arms when compared against the SOC. The societal cost per DALY averted was $676 (95%CI 17; 1435) in the Labels and $941 (95%CI -53; 1652) in the Pillbox study arms.

• Download figure
• Open in new tab

Figure 2. Cost-effectiveness planes

Scatter plot of the estimated joint density of incremental costs and incremental DALYs averted, showing uncertainty in the ICER estimate obtained from the analysis models. The dots represent the findings of a probabilistic sensitivity analysis.

When plotted on cost-effectiveness planes, we find that the ICERs are crossing the axes but slightly concentrated in the bottom left hand quadrant suggesting that DATs are less costly and less effective than the standard of care. ICERs in this quadrant are interpreted as cost-effective if they fall below the cost-effectiveness threshold. Using the outcomes of the probabilistic sensitivity analysis, plotted on the cost-effectiveness acceptability frontier, the probability that each of the interventions would be cost-effective compared against a range of cost-effectiveness thresholds (CETs) are shown in figure S3. Using a range of US$5 to US$900 per DALY averted for the CETs, we find that the interventions have a 56% probability (labels) and 49% probability (pillbox) of being cost-effective compared to SOC at a CET of US$100. The results of the distributional cost-effectiveness analysis estimates the distributional cost-effectiveness ratio at 18.68 for labels and 18.96 for pillbox. We compare this finding against a relative health inequality aversion parameter of 5.66. In the absence of empirical research on the level of inequality aversion by policymakers in Ethiopia, we used an estimate previously derived for decisions in Ethiopia [20]. As the value of the inequality aversion parameter increases, it suggests that there is a greater willingness to achieve fewer health improvements to reduce inequality in health or other outcomes. Our distributional cost-effectiveness ratio estimated (18.68 and 18.96) exceeds the parameter used for previous decisions in Ethiopia for vaccines (5.66) suggesting that policy makers would need to be more averse to inequality than in previous decisions, to recommend the implementation of DATs based on health equity.

In the sensitivity analyses (table S7), we found that more visits in the SOC increased potential cost-savings from implementing the interventions. If we assumed an increased cost of the pillbox, for example in setting where reuse rate is low, the cost savings compared to the SOC reduced by 16%. Removing the costs and DALYs associated with recurrence reduces cost savings and has minimal effect on the DALYs averted. Valuing participants using an average wage rate increases the cost savings to patients. Main study findings, however, remained similar under sensitivity analysis.

Discussion

In health facilities implementing DATs, trial participants made 20 fewer visits per treatment episode compared to the standard of care (SOC), substantially reducing the cost to providers. We didn’t find evidence of differences in poor health outcomes (DALYs) between study arms or their distribution between socio-economic position (SEP) groups. However, there was a reduction in patient-incurred costs among the least wealthy households in the intervention compared to SOC study arms. When assessing uncertainty, the 95% confidence intervals of the incremental cost-effectiveness ratios (ICERs) were wide, suggesting substantial uncertainty that DATs would be cost-effective when compared to the SOC. We estimated that the labels and pillbox arms have a 49-56% probability of improving the cost-effectiveness of tuberculosis treatment compared to SOC at a cost-effectiveness threshold of $100 per DALY averted. Our study follows the work of other cluster randomised trials evaluating the effectiveness and costs of Digital Adherence Technologies (DATs) in China [5,10], South Africa [21], Uganda [7] and Ethiopia [22], by including a detailed analysis of human resources, patient costs and equity in addition to costs and cost-effectiveness.

Differences in provider costs between the SOC and intervention arms of the study were primarily driven by the number of healthfacility visits per treatment episode. In a separate analysis, we showed that health care workers spent more time with PWTB per visit in the intervention arms than when they were attending more frequently in the SOC arm [23]. This increase in time spent remained lower than the cumulative time of multiple visits in the SOC. Our evaluation of cost of treatment included above service-level costs. These costs were highly variable, particularly between urban and rural health facilities where fewer people with tuberculosis (PWTB) were accessing services. Given that we had overhead costs from only 12 clusters, single imputation of provider costs has likely resulted in larger variability than if we had collected data for all 78 clusters. Provider costs reported in this study were broadly similar to the costs estimated by Terefe et al’s comprehensive assessment of the cost of tuberculosis care in Ethiopia, though we used a micro-costing approach [17]. Our total cost estimate, includes costs incurred by all participants who received the intervention irrespective of whether they completed treatment (ITT population). Provider costs incurred through diagnostic tests, hospitalisation and subsequent treatment episodes were similar between study arms, and differences in treatment costs were predominantly driven by the reduction in the number of health facility visits.

Patient costs were similar in the intervention - compared to the SOC arms of the study, despite large differences in facility visits. This is explained by lower transport costs per visit reported in the SOC where there are more frequent visits compared to the intervention study arms. Furthermore, we used the human capital approach for valuing time spent seeking care, whereby individual self-reported income is used to value people’s time. This likely does not fully captured the full cost of time spent. Approximately 35% of participants reported no- or irregular income with consequently no cost allocated to their time spent. Furthermore, we collected self-reported income data at enrolment into the trial when people were on treatment and already had suffered disease-associated income loss. A recent meta-analysis of patient costs associated with tuberculosis in Ethiopia, found that pre-diagnosis costs were more than double post-diagnosis costs US$97.6 (56.4-184.3) versus US$45.1 (−119.1-209.37)) [28]. Differences in time spent between study arms due to a reduction in number of visits, would therefore only have a marginal effect on productivity costs. We quantified this potential bias by conducting a sensitivity analysis estimating the productivity costs for each of the study arms by valuing all participants’ time equally.

Our study is the first to empirically demonstrate that DATs could be cost saving by reducing the number of health facility visits. This observed difference is dependent on the frequency of health facility visits in the SOC. At the time of our study, PWTB were attending health facilities on average 30 times per treatment episode, our findings may therefore not hold in settings where less frequent facility visits are already routine practice. Furthermore, the observed difference between study arms would have been more pronounced prior to COVID-19, when visits to health facilities were more frequent [25]. Other studies evaluating the economic costs of DATs found them to be more costly and effect-neutral [7,8,10]. In Uganda, the provider cost of 99DOTS was estimated to be $59 (range: $50-$70) per participant successfully treated [7]. Similarly, in Morocco, the cost-effectiveness of a digital pillbox from a provider perspective was found to be $434 per DALY averted among MDR-TB patients. This was a model-based analysis where the number (and cost) of health facility visits were assumed to be the same between study arms [8]. A study in China, considering the societal perspective, found digital pillbox to be more costly and effect neutral with an ICER of $3,668.59 per DALY averted. Data for assessing number of health facility visits was limited (n=120) and did not fully assess differences in practice between studyarms [10].

Strengths of our study include it being a trial-based cost-effectiveness analysis with detailed data collection of patient events, healthfacility visits and rigourous assessment of trial endpoints embedded within a randomised study where the SOC reflects routine practice in Ethiopia. We furthermore went beyond the traditional cost-effectiveness framework to analyse the equity impact of the interventions. Study limitations include, the trial follow-up period was not long enough to capture all health and economic consequences of the interventions – particularly tuberculosis transmission. This question was addressed in a transmission modelling study following this study [26]. The number of health facility visits were recorded prospectively, and patient event data (including hospitalisations) collected at the 12 month follow-up. Self-reported patient events would not be available for patients who died or became LTFU during treatment episode and we underestimate the cost associated with care-seeking prior to death or LTFU. Given the representativeness of the sample, imputed costs would undervalue true costs. Our DALY estimates are likely to be conservative, because we only include DALYs for the current and identified future treatment episodes and not for any longer-term sequalae of tuberculosis disease [27]. Our estimates are similarly likely to underestimate the full costs and economic losses incurred because our analytical time horizon, restricted to the current tuberculosis event, includes costs incurred once participants started treatment and not during the diagnostic phase. We could only assess the equity distribution impact of the intervention among those already accessing care, and expecting to live in the health facility’s catchment area for the duration of treatment. We may therefore be underestimating the value of the intervention among more marginalised groups. Nevertheless, we expect the findings to be representative of adults who are accessing care at public health clinics in Ethiopia.

We used relative SEP based on multi-dimensional poverty assessment, to incporate equity in the decision framework [29]. Alternative approaches are based on estimates of individual income to assess inequality [30]. Measures of income-based inequality would correctly ascertain immediate ability to respond to a health shock, but would fail to account for resource-sharing within families, saving ability, and non-monetary assets that are important in assessing poverty in rural areas. Multi-dimensional poverty metrics include locally-relevant assets such as land, livestock, social capital and education that individuals may draw on in the event of a health shock. Our evaluation as part of the trial, was not powered to compare outcomes between SEP groups within study arms, but sufficient to evaluate how the concentration of the outcome between those relatively wealthy and relatively poor varies between study arms.

In conclusion, we found the implementation of DATs to be effect-neutral but cost-saving due to a marked reduction in healthfacility visits. These savings would only be fully realised if healthcare providers have other productive tasks to fill their time previously spent on DOTS, or in contexts where routine care is not daily facility-based DOTS. We found limited evidence of differences in DALYs averted between study arms and estimate a 49-56% probability that DATs will improve the cost-effectiveness of tuberculosis treatment at a CET of $100 per DALY averted. The implementation of DATs reduced the inequitable burden of patient-incurred costs.

Authors’ contributions

NF, KLF, AWT and CFM designed this study. The trial was designed by KLF, KvK, JvR and DJ. All authors have contributed to the study design. KLF provided statistical expertise in clinical trial design and analysis. NF conducted the analysis and provided health economics expertise in design and analysis. AWT, MB, MS, AB, TA, NM, KvK and JvR implemented the study. All authors contributed to the interpretation of the findings and revision of the final manuscript. The authors read and approved the final manuscript.