New tuberculosis vaccines in India: Modelling the potential health and economic impacts of adolescent/adult vaccination with M72/AS01E and BCG-revaccination

Background India had an estimated 2.9 million tuberculosis cases and 506 thousand deaths in 2021. Novel vaccines effective in adolescents and adults could reduce this burden. M72/AS01E and BCG-revaccination have recently completed Phase IIb trials and estimates of their population-level impact are needed. We estimated the potential health and economic impact of M72/AS01E and BCG-revaccination in India and investigated the impact of variation in vaccine characteristics and delivery strategies. Methods We developed an age-stratified compartmental tuberculosis transmission model for India calibrated to country-specific epidemiology. We projected baseline epidemiology to 2050 assuming no-new-vaccine introduction, and M72/AS01E and BCG-revaccination scenarios over 2025–2050 exploring uncertainty in product characteristics (vaccine efficacy, mechanism of effect, infection status required for vaccine efficacy, duration of protection) and implementation (achieved vaccine coverage and ages targeted). We estimated reductions in tuberculosis cases and deaths by each scenario compared to no-new-vaccine introduction, as well as costs and cost-effectiveness from health-system and societal perspectives. Results M72/AS01E scenarios were predicted to avert 40% more tuberculosis cases and deaths by 2050 compared to BCG-revaccination scenarios. Cost-effectiveness ratios for M72/AS01E vaccines were around seven times higher than BCG-revaccination, but nearly all scenarios were cost-effective. The estimated average incremental cost was US$190 million for M72/AS01E and US$23 million for BCG-revaccination per year. Sources of uncertainty included whether M72/AS01E was efficacious in uninfected individuals at vaccination, and if BCG-revaccination could prevent disease. Conclusions M72/AS01E and BCG-revaccination could be impactful and cost-effective in India. However, there is great uncertainty in impact, especially given unknowns surrounding mechanism of effect and infection status required for vaccine efficacy. Greater investment in vaccine development and delivery is needed to resolve these unknowns in vaccine product characteristics.


Model structure and equations
We created an age-stratified compartmental differential equation model of tuberculosis in India, including dimensions for age, tuberculosis natural history, vaccination, and access-to-care. The age and access-to-care structures are identical to those included in Clark et al. 1 Minor modifications from the Clark et al. natural history structure are described below. The vaccination structure is in section 4.3.

1.1
Natural history model structure A natural history structure with eight compartments in Figure S1.1 was created by adapting features of previous models and has been described previously. 1 The latency structure in this model demonstrates a progressive loss of ability to reactivate, with the reactivation rate in the Latent-Fast compartment greater than in Latent-Slow and greater still than in Latent-Zero, where we assume the rate of reactivation is 0. We do not explicitly have a self-clearance compartment. We assume that those in Latent-Fast can only fast progress to subclinical disease or continue to remain latent and transition to Latent-Slow. There is no direct transition between Latent-Fast and Latent-Zero.

Treatment initiation and outcomes
Steps for calculating treatment initiation, treatment completion, non-completion, and mortality rates are described in the Supplementary Material for Clark et. al. 1 We assume the SFR is the ratio between treatment completions to the sum of treatment completions and non-completions. In India, . The data used to calculate the ontreatment outcomes was obtained from the WHO. However, as the private sector accounts for a substantial portion of treatments in India, and not all of the treatments conducted in the private sector are reported, we make adjustments to the on-treatment completion and non-completion fractions from Table S2.3 as described below and in Table S2.4. Table S2. 3 Calculating treatment outcome parameter values for adults and children

Parameter Adults Children
On-treatment mortality fraction Sample from On-treatment completion fraction

On-treatment non-completion fraction
We assume that the total number of treatments is composed of the treatments that are reported and the treatments that are not reported. We assume that the on-treatment mortality fraction is the same in the public and private sector, but want to adjust the treatment completion and non-completion rates to account for differences between those reported and those not reported as in Table S2.4. We assume that 60% of the total treatment occurs in the public sector and the remaining 40% occurs in the private sector. We assume that all treatments not reported are from the private sector, that the treatment completion rate in the private sector is 40%, and that there is no reporting bias (in that they were equally likely to not report treatment completions or non-completions or deaths). Before 2012, only the treatment conducted in the public sector was reported, but since then, treatment in the private sector has begun to be reported. 19

Model simulation
We specified a system of ordinary differential equations defining the derivatives with respect to time of a set of state variables, to simulate the country-specific tuberculosis epidemic between 1900 and 2050. We initialised the simulation by distributing the population between the eight tuberculosis natural history states using a fitted parameter representing the proportion of the population uninfected at the start of the simulation. For each year of the simulation (1900-2050), our models are designed to exactly match the age and country-specific UN population estimates and projections. Forty percent of the population was assigned to the low access-to-care stratum and the remaining sixty percent of the population was assigned to the high access-to-care stratum.

Model calibration
For this India modelling analysis, we followed the same modelling approach as in Clark et al. 1 Broadly, this was as follows: 1. Construct a mechanistic model 2. Calibrate the model by identifying areas of the input parameter space where the output of the mechanistic model was consistent with the historical epidemiologic data 3. Use the calibrated model to simulate and predict future tuberculosis epidemiology and model new vaccines In the context of this analysis, step 1 was achieved by creating the compartment differential equation model as specified in Section 1. For step 2, we independently calibrated a model by identifying areas of the parameter space that made the output of the model match the corresponding calibration targets (Table S3.1 below). Further details on the sources for the calibration targets and any additional modifications are in the subsequent sections.
The model was fitted to the calibration targets using history matching with emulation, a method that allows us to explore high-dimensional parameter spaces efficiently and robustly. [20][21][22][23] History matching progresses as a series of iterations, called waves, where implausible areas of the parameter space, i.e., areas that are unable to give a match between the model output (e.g., the predicted incidence rate by the model) and the empirical data (e.g., the incidence rate calibration target from the WHO data), are found and discarded. In order to identify implausible parameter sets, emulators, which are statistical approximations of model outputs that are built using a modest number of model runs, are used. Emulators provide an estimate of the value of the model at any parameter set of interest, with the advantage that they are orders of magnitude faster than the model.
History matching with emulation, implemented through the hmer package in R, 24,25 considerably reduced the size of the parameter space to investigate. Rejection sampling was then performed on the reduced space to identify at least 1000 parameter sets that matched all targets.
If we were unable to find at least 1000 fully fitted parameter sets using history matching with emulation, we subsequently used an Approximate Bayesian Computation using Markov Chain Monte Carlo method (ABC-MCMC). ABC-MCMC was conducted using the easyABC package in R, modified by the Sebastian Funk, Gwenan Knight, and the Tuberculosis Modelling group at LSHTM for adaptive sampling and to accept seeded parameter values. 26,27 We used parameter sets with the maximum number of targets fitted using history matching with emulation as starting seeds for multiple MCMC chains per country, with the ABC-MCMC algorithm continuously adapting using the last 1000 points, a burn in of 1000 samples, and the noise factor set to 0.0001.
Once we had obtained 1000 parameter sets that produced output consistent with the calibration targets, we used those parameter sets with the mechanistic model to simulate the future (step 3) for each country as the Status Quo no-newvaccine baseline, where we assumed that current trends and quality of non-vaccine tuberculosis services continued into the future at the same rate.
As an alternative future, we calibrated a Strengthened Current Interventions no-new-vaccine baseline. This baseline assumed a scale up in other non-vaccine tuberculosis interventions between 2021 and 2035 in order to meet the target of a 50% reduction in tuberculosis incidence in 2035 compared to the 2015 estimates (an incidence rate of 108.5 (64-164 per 100,000 population). This scale-up was introduced in the model by introducing parameters (sampled between 0 and 1) which act as multipliers on the rate of progression to disease and in the force of infection equation.
The process of generating fits for the Strengthened Current Interventions no-new-vaccine baseline while capturing uncertainty was as follows: 1. Obtain 1000 full fits from the Status Quo baseline 2. Subset the 1000 Status Quo full fits to 100 by: a. Ranking the 1000 Status Quo fits from smallest to largest tuberculosis incidence rate in 2020 b. Retaining every 10th parameter set 3. Use emulation on each of the 100 parameter sets 4. Obtain 100 "groups" of fully fitting parameter sets (one group for each original parameter) 5. Subset each group of fully fitting parameter sets to 10 by: a. Ranking the parameter sets in each group from smallest to largest tuberculosis incidence rate in 2035 b. Retaining every nth parameter set to obtain 10 across the range 6. Obtain 1000 full fits for the Strengthened Current Interventions baseline by combining the 10 parameter sets from each of the 100 emulation sets

Incorporating the COVID-19 pandemic
It will be a number of years before the full implications of disruptions to tuberculosis prevention and care during the COVID-19 pandemic are realised. The WHO provided estimates for the impact on the tuberculosis incidence and mortality rates between 2020-2025 relative to January 2020, shown in Figure S3. 2. 29 In order to ensure that the model is appropriately representing the future trends in incidence and mortality, we calibrated to the projected incidence and mortality for 2025, which is estimated as a 10% increase in both mortality and incidence in 2025 compared to January 2020. To implement this, we calculated a 10% increase to the incidence and mortality rates estimated by the WHO in 2019, shown in Table S3.2, and calibrated to both for 2025.

i. Adjusting the 2015 target bounds
We obtained an estimate for the tuberculosis prevalence in 2015 from "Estimating tuberculosis incidence from primary survey data: a mathematical modelling approach" by Pandey et al 2017. 30 In it, they estimate the prevalence of smearpositive cases across all ages in India, as well as the proportion of cases that are smear-positive. The mean and 95% confidence intervals for the estimates of these are 159. 38 (122.9-196.59) and 0.63 (0.43-0.93) respectively. 30 Cited sources within the paper suggest that these quantities have been modelled as lognormal (smear-positive prevalence) and beta (smear-positive proportion) distributed. 30 The total prevalence, therefore, can be determined as If we assume that the mean estimate for the proportion of cases that are smear-positive is accurate, then we simply quotient the smear-positive bounds by this value (0.63). This gives Since we have confidence intervals and a knowledge of the underlying distributions, we can attempt to determine the hyperparameters of the distributions. Once we have these, we can sample repeatedly from the quotient of the two distributions to get an estimate for its confidence interval. We sample from the numerator's distribution, sample from the denominator's distribution, and quotient them to represent a sample from the (unknown) prevalence distribution. Given enough samples, we can obtain a reasonable estimate of the confidence interval. Since the lognormal distribution has a closed-form, we can simply solve for the hyperparameters.
The beta-distribution is less straightforward, but we can use maximum likelihood estimation to find feasible parameter values. Doing so gives Then we perform monte-carlo sampling to generate a representative sample from our quotient distribution, from which we obtain a 95% confidence interval.

ii.
Adjusting for extrapulmonary tuberculosis In our model we are representing everyone with tuberculosis, which includes both pulmonary (PTB) and extrapulmonary tuberculosis (EPTB). EPTB is not infectious but is included in the WHO estimates of yearly incidence and mortality rates. The 2021 prevalence estimates from the National Tuberculosis Prevalence Survey did not adjust for EPTB in the estimate provided for adults, and neither did the 2015 study which estimated the tuberculosis prevalence from subnational surveys. Therefore, we want to adjust the PTB prevalence estimates and range by the amount of EPTB in order to estimate the total TB prevalence. To estimate the proportion of EPTB, we used the average of the proportion of incident extrapulmonary tuberculosis cases from 2013-2020 (Table S3.3).  32 Averaging the proportion of incident EPTB cases column in Table S3.3 and dividing the estimates and bounds on the pulmonary tuberculosis prevalence estimates by (1-average of proportion of incident EPTB cases) we obtain the following as the estimates of the tuberculosis prevalence per 100,000 population: iii.
Adjusting the 2021 target bounds The National Tuberculosis Prevalence Survey India 2019-2021 reports estimates for the prevalence of all forms of tuberculosis among all age groups in India (312.0 [286.0-337.0] per 100,000 population) and the prevalence of microbiologically confirmed pulmonary tuberculosis among adults aged ≥ 15 years in India (316.0 [290.0-342.0] per 100,000 population). As described in the previous section, we adjusted the estimate of the prevalence of pulmonary tuberculosis in adults for EPTB, giving a revised estimate for the prevalence of all forms of tuberculosis disease in adults of 393.6 (361.2-426.0). We subsequently increased the upper and lower bounds on the all age and adult targets by 30%, leading to estimates of 312.0 (218.4-405.6) and 393.6 (275.5-511.7) respectively. Rationale for adjusting the bounds on the targets is described below.

Rationale 1: Impact of the Covid-19 pandemic
Some state groups started and completed the survey before the COVID-19 pandemic, others during, others after the major pandemic waves had completed. 32 Depending on the impact of COVID-19 measures on tuberculosis, this could bias the estimates of the region either up or down, and bias the overall estimate of the tuberculosis prevalence for India, particularly as Delhi (the region with the highest estimated tuberculosis prevalence) started and completed the survey before the pandemic. 32

Rationale 2: Differences between planned surveyed clusters and actual surveyed clusters
The National Tuberculosis Prevalence Survey India 2019-2021 compares the number of pulmonary tuberculosis cases notified at the state group level in 2019, 2020 and 2021 between those clusters who were surveyed and those who were not surveyed. 32 Although no statistically significant differences were observed between the surveyed and notsurveyed clusters, there are qualitative differences between the number of notifications of pulmonary tuberculosis between groups, where non-surveyed clusters consistently have a lower number of notifications. 32

i. Adjusting the 2021 target bounds
The National Tuberculosis Prevalence Survey India 2019-2021 reports an estimate for the prevalence of tuberculosis infection in India among adults of 0.314 (0.272-0.353). We adjusted the bounds to give a revised estimate of 0.314 (0.114-0.514), with rationale described below.

Rationale 1: Oversampling from Gujarat with no adjustment
Of the 55 clusters where IGRA testing was done, 31 were in Gujarat and 24 were in the remaining 19 state groups. Gujarat had the lowest estimated tuberculosis prevalence per 100,000 population. 32 If we assume that prevalence of tuberculosis infection is correlated with prevalence of tuberculosis disease, then we would anticipate that the tuberculosis infection prevalence estimates from Gujarat would be commensurately low. As more than half of the clusters were from Gujarat, and there is no indication of adjustment for oversampling from this region, it is possible that the reported country-level tuberculosis infection prevalence is an underestimate. If our assumption that prevalence of infection correlates with prevalence of disease was incorrect, the tuberculosis infection prevalence estimates may actually be overestimated. As such, we have adjusted the bounds to account for oversampling with no adjustment, but retained the central estimate, resulting in a calibration target of 0.314 (0.114-0.514).

Tuberculosis case notifications
i. Adjusting to account for the private sector contribution to reported case notifications Treatment in India can occur in the public or private sector. While this varies by state, it is estimated that 60% of treatment is performed in the public sector, and the remaining 40% in the private sector. According to the WHO Global TB Report 2019, reported case notifications only included notifications from the public sector before 2013. 19 From 2013-2020, reported case notifications began to include the private sector ( Figure S3.4). By 2020, approximately 31% of the total reported notifications were from the private sector. 19 Figure S3. 4 Contribution of the private sector to reported case notifications from WHO Global TB Report 2021 29 The model represents case notifications as the number of tuberculosis treatment initiations. We want to calibrate the model to the true number of treatment initiations, as this is what the model will represent. Therefore, this involves adjusting the WHO reported case notifications to reflect underreporting from the private sector. To do this, we must calculate the fraction of total cases notifications (treatment initiations) that are reported, while accounting for both the private and public sector.
Using the percent contribution of the private sector to the reported treatments, and the assumption that all treatments occurring in the public sector are reported, we can calculate the fraction of total notifications that are actually reported.

Note: This calculation is valid for
We want to calculate Using the derived equation, we can calculate the fraction of total notifications reported from 2013-2020 (Table  S3.4). To adjust the WHO reported case notification estimates for underreporting, we divide the estimates by the fraction of total treatments reported ( ), and assume 20% upper and lower uncertainty bounds. The reported and adjusted estimates of case notifications are provided in Table S3.5 and assume 20% upper and lower uncertainty bounds. The reported and adjusted estimates of case notifications are provided in Table S3.5.

Proportion of previously treated incident cases
i. Adjusting the proportion retreated bounds The proportion retreated target is included to ensure that the disease tuberculosis incidence is derived from the correct source (i.e., to ensure that we do not overestimate the amount of incidence that is coming from fast-progression or reactivation without treatment). By dividing the number of notifications who are people who have been previously treated (2) by the total number of notifications (1), we get the proportion of notifications that have been previously treated.
We assume that at equilibrium, the proportion of notifications who have been previously treated will be equal to the proportion of incident disease cases who have been treated previously.
The estimate of the proportion of notifications that have been previously treated for India from the WHO dataset is 10.0% (4.3-14.7). However, country specific estimates may be subject to recall bias as they rely on patients to accurately report previous treatment. Additionally, studies have shown that approximately 11% of patients recorded as "new" have had some form of previous tuberculosis treatment. 35 Therefore, we adjusted the estimates from the WHO dataset, and calibrated to a target of 19.1% (13.9-24.1).

ii. Calculating the proportion retreated target in the model
The subsequent pages describe the methods used to calculate the proportion retreated target.

Definition 1:
The number of notifications, per year is the flow from Dc and T = Definition 2: Being "previously treated" implies that an individual arrived in the R compartment via the T compartment.

Definition 3:
For an individual to count as a notification of a person who was previously treated (Definition 2), they must flow from T → R → Ds → Dc → T Looking at the total number of notifications broken down to their origins, we see that: The "Notifications of people who were in R" term is further broken down into: -People who entered R from T, per year -People who entered R from Dc, per year -People who entered R from Ds, per year We can rewrite the total number of notifications per year equation as: Recall now what we are looking to calibrate to: The denominator is directly available from the model: the total number of notifications ( ). Using definitions 2 and 3 above, the "number of notifications who are people who have been previously treated per year" = "notifications of people who were in R having entered R from T per year". Therefore, we can redefine our calibration target as: We do not have notifications disaggregated by source, but we do have incidence disaggregated by source. Incident cases are defined as the flow into Ds, which can be from R, from Ls, or from Lf.
We obtain output on all of these flows, so we can calculate the proportion of incident cases from each pathway (Lf, Ls, and R) easily by just dividing the total number of incident cases of people from Lf, Ls or R by the total number of incident cases.
The proportions of incidence from each pathway are:

RD = flow from R to Ds = total number of incident cases of people from R, per year LsD = flow from Ls to Ds = total number of incident cases of people from Ls, per year LfD = flow from Lf to Ds = total number of incident cases of people from Lf, per year RD + LsD + LfD = total number of incident cases, per year
Similarly, we can disaggregate the flow from R to Ds further into how the people in R entered R.
Again, we don't have information on the disaggregated numbers of incident cases from R based on how they entered R, but we do have information on the entry to R.
The proportion of the total flow into R from each of T, Ds, and Dc per year is: TR = flow from T to R = total number entering R from T per year DcR = flow from Dc to R = total number entering R from Dc per year DsR = flow from Ds to R = total number entering R from Ds per year TR + DsR + DcR = total number entering R per year If we assume that the flows INTO R from each of T, Ds, and Dc are in the same proportions as the flows OUT of R, then we can disaggregate the outflow from R (which is the number of incident cases of people from R, per year, we called RD in the equation above) into incident cases of people from R who entered R from each of T, Ds, and Dc, per year by multiplying RD by the proportion from each of T, Ds, and Dc Number of incident cases of people from R who entered R from T, per year = (Number of incident cases of people from R, per year)(Proportion of flow out of R that is from people who entered R from T, per year) etc.
We can rewrite RD in terms of the disaggregated pathways from T, Ds, and Dc: Subbing in the expression for RD above into the equation for the proportion of incident cases from R, we obtain: Factor, simplify and rewrite: Proportion of incident cases from R = (Proportion of incident cases from R who entered R from T) + (Proportion of incident cases from R who entered R from Dc) + (Proportion of incident cases from R who entered R from Ds) The assumption we make here is that (at equilibrium) these proportions of incident cases will be equivalent for flows entering Ds (incident cases), entering Dc (progression from subclinical to clinical disease) and entering T (treatment initiation / case notifications).
Therefore, the proportion of notifications of people who were in R having entered R from T will be the same as the proportion of incidence from people who were in R having entered R from T.
Going back to the calibration target once again: Although we do not know the number of notifications of people who were in R having entered R from T per year, this is equal to the proportion of notifications of people who were in R having entered R from T multiplied by the total number of notifications per year We can cancel out the total number of notifications as it is in both the numerator and denominator.
This value is calculated as the proportion of notifications of people who were in R multiplied by the proportion of the entry into R that came from T However, there may be some people who recently entered R from Dc or Ds, but who had also previously had treatment. Therefore, the previous equation is revised as: We assume that the proportion of those in (DcR + DsR) who have been treated previously is the same as the proportion of those in DsR who have been treated previously. We can then set the value: If we substitute in the term, we can see that the same term is repeated again and again. Let Then we can rewrite the above as: Let and , substitute and expand: Let and , substitute and expand: Substituting back in for , , , , and we obtain:

No-new-vaccine baseline
The primary no-new-vaccine simulated was the no-new-vaccine baseline, which assumed non-vaccine tuberculosis interventions continue at current levels into the future. As reported country-level data includes the high coverage levels of neonatal BCG vaccination, this was not explicitly modelled. We assumed that BCG vaccination would not be discontinued over the model time horizon.

Vaccine delivery scenarios
Two recently completed phase 2 trials have demonstrated encouraging efficacy results. The M72/AS01E candidate vaccine is a subunit vaccine for which results from a completed Phase IIb trial were published at the end of 2019. 36 After three years of follow-up, the efficacy of M72/AS01E at preventing disease in latently infected adults from South Africa, Zambia, and Kenya was estimated at 49.7% (95% confidence interval = 2.1-74.2). 36 To confirm this finding, a larger, Phase III follow-up study is needed, which includes participants who are uninfected, adolescents, as well as those living with HIV to assess safety and immunogenicity in these populations. This is being planned.
BCG-revaccination (administering a second dose of BCG to those who were vaccinated neonatally) was previously implemented in many countries, however evidence did not support the effectiveness of this practice. Interest in BCGrevaccination has recently been renewed following results from a trial for the vaccine candidate, H4:IC31. BCGrevaccination was assessed as a third parallel arm alongside H4:IC31 and a placebo in a pre-infection population in South Africa, and although neither vaccine appeared efficacious at preventing infection, BCG-revaccination appeared efficacious at preventing sustained infection (defined as three consecutive positive tests after day 84 of the trial) with an efficacy of 45.4% (6.4-68.1). 37 A larger trial of BCG-revaccination versus placebo in 1800 healthy adolescents from across South Africa is now underway to verify this finding.
We evaluated introducing vaccines with M72/AS01E and BCG-revaccination characteristics compared to the no-newvaccine baseline as described in the subsequent sections.

Classifying tuberculosis vaccines
Before describing the specific characteristics for the vaccine scenarios that we investigated, we provide a brief overview on classifying tuberculosis vaccines (descriptions from Clark et al. 1 ).
Tuberculosis vaccines are characterised on four key characteristics: the vaccine efficacy, the host infection status at the time of vaccination required for the vaccine to be efficacious, the mechanism of effect, and the duration of protection. Vaccine efficacy defines the magnitude of protection induced by the vaccine. Vaccine efficacy is assumed to be either "all or nothing", where the vaccine offers full protection to a subset of individuals (equal to the vaccine efficacy) who were vaccinated, or "degree", where the vaccine offers partial protection to all individuals who received the vaccine. The vaccine mechanism of effect type determines how the vaccine will offer protection. A prevention of infection (POI) vaccine protects individuals from initial or re-infection with Mtb, whereas a prevention of disease (POD) vaccine functions by preventing individuals who may be uninfected or infected with Mtb from progressing to active disease. A prevention of infection and disease vaccine (POI&D) prevents both infection and disease. Finally, the duration of protection represents the length of time following vaccination that individuals are protected.

M72/AS01E and BCG-revaccination scenarios
For each vaccine product, we established one "Basecase" vaccine scenario based on clinical trial data and expert opinion. We then varied vaccine product and delivery scenarios as univariate scenario analyses from the Basecase scenario as described in Table S4.1.

Vaccine eligible population
In our modelling, we assume that there is no pre-vaccination infection testing. Therefore, even if a vaccine is only effective when delivered to uninfected individuals at the time of vaccination, we assume that both uninfected and infected individuals will receive the vaccine, and only the uninfected individuals will receive protection. Our model structure allows for counting and tracking individuals who received the vaccine but do not receive any protection from it.

Efficacy
From trial data, the efficacy of M72/AS01E at preventing disease in latently infected adults was estimated at 49.7% (2.1-74.2). 36 Therefore, our Basecase vaccine efficacy was set at 50%, and based on expert opinion we evaluated 60% and 70% as scenario analyses. BCG-revaccination appeared efficacious at preventing sustained infection with an efficacy of 45.4% (6.4-68.1). 37 The Basecase efficacy was set to 45%, and 70% was evaluated in a scenario analysis.

Protection from repeat vaccinations
In the event that an individual who is currently protected with a vaccine receives another course, after consultation with an immunologist we have made some assumptions on the resulting level of vaccine protection: BCG-revaccination: Based on expert advice, we assume that no additional protection is afforded if a second or third vaccine is administered while the individual is currently protected from the first.
M72/AS01E: Based on expert advice, we assume that overall vaccine protection increases if a second vaccine is administered while the individual is currently protected by a first vaccine. We assume that this protection increases by (1-current protection) times vaccine efficacy, as in Table S4.2. Note that the number of vaccine courses refers to the number of vaccine courses that the individual is currently protected by, not that they have ever a) received, or b) been protected by. For example, if someone receives one vaccine, then wanes, then receives another one, they would only be currently protected by one, not two, vaccines, and so the efficacy would be either 50%, 60%, or 70% depending on the scenario.

Mechanism of effect
We assume that a vaccine that protects against infection will work by reducing the rate of infection for both initial and re-infection, and that a vaccine that protects against progression to disease will work by reducing the rate of progression to subclinical disease. If the vaccine protects against both infection and disease we assume that it has the same efficacy against preventing disease as it does infection. For example, if the vaccine is defined as a prevention of infection and disease vaccine with 50% efficacy, it reduces the rate of infection by 50% and the rate of progression to disease by 50%.

Introduction year
The Basecase introduction years, 2025 and 2030 for BCG-revaccination and M72/AS01E respectively, were determined based on considering when new trial data would become available, as well as incorporating time for licensure and policy change. The introduction year considered in scenario analyses, 2031 and 2036 for BCGrevaccination and M72/AS01E respectively, was based on applying IAVI/Full Value Assessment of Tuberculosis Vaccines analyses from Shelly Malhotra and expert advice to the earliest possible introduction year. 1

Age targeting
The Basecase age was informed by ages of trial participants and expert advice. Additional scenarios were informed by work conducted by Pelzer et. al and expert advice. 38

Vaccine model structure
Depending on the host infection status required at the time of vaccination for the vaccine to be efficacious, we implemented a different vaccine structure in the model to account for differences in Vaccinated Protected, Vaccinated Not Protected, and Vaccinated Waned. Each compartment in the vaccine structure is replicated for all tuberculosis natural history compartments, access-to-care strata, and ages.

No Current Infection vaccines
A No Current Infection (NCI) vaccine requires an individual to be uninfected at the time of vaccination in order for the vaccine to be efficacious. Implementation in the TBVax model of an NCI vaccine with the possibility of two repeat vaccine courses is provided in Figure S4.1. For our purposes, we assume that the level of protection remains the same regardless of the number of vaccine courses received (i.e. level of protection in "Vaccinated Protected (one vaccine course)" is equal to "Vaccinated Protected (two vaccine courses)" etc.). Additionally, because the vaccine is only efficacious for NCI, and in this model once you leave UN (the state where the vaccine is effective) you never return, once you enter a "Vaccinated Not Protected" state you never have the opportunity to become "Vaccinated Protected" again.

Current Infection vaccines
A Current Infection (CI) vaccine requires an individual to be infected at the time of vaccination in order for the vaccine to be efficacious. Implementation in the TBVax model of an CI vaccine with the possibility of two repeat vaccine courses is provided in Figure S4.2. For our purposes, we assume that the level of protection builds with each vaccine course, with efficacy values as in Table S4.2.

Figure S4.2
Vaccine structure for a CI vaccine (where protection builds with each vaccine course)

Any Infection vaccines
An Any Current Infection (AI) vaccine will be efficacious with any infection status (aside from current active disease) at the time of vaccination. The "Vaccine Not Protected" compartments remain as we assume that individuals with subclinical disease may be accidentally vaccinated and would not receive protection from the vaccine. However, we do want to keep track of the number of vaccinations for cost purposes.

AI-1 vaccines:
With each vaccine course the level of protection remains the same (Figure S4.3). Waning occurs from any of the Vaccinated Protected compartments to the Waned Protection compartment.

AI-2 vaccines: With each vaccine course the level of protection builds if the recipient is currently in a Vaccinated
Protected compartment ( Figure S4.4). This is the same structure as the CI vaccine with protection building ( Figure  S4.2). Waning occurs from any of the Vaccinated Protected compartments to the Vaccinated Protected compartment one level below, or to the Waned Protection compartment for those with only one course of protection.

Figure S4.4
Vaccine structure for an AI vaccine (where protection builds with each vaccine course)

Economic analysis methods
Before undertaking this work, we established an economic analysis plan, involving stakeholders and government officials to ensure we had incorporated all necessary information and planned to report on all key outcomes, to outline the methods used in this work. This is summarised below.

Calculation of disability-adjusted life years
We calculated the difference in total disability-adjusted life years (DALYs) from vaccine introduction to 2050 for each scenario compared to the no-new-vaccine baseline. We used the disability weight for tuberculosis disease from the Global Burden of Disease 2019 study, 39 and country-and age-specific life expectancy estimates from the United Nations Development Programme. 40 To incorporate parameter uncertainty in years lost due to disability (YLD) weight estimates, we made 1000 draws from disability weight uncertainty ranges.

5.2
Tuberculosis-related cost model We estimated health system unit costs, patient costs and productivity losses based on a scoping review of published literature. For the tuberculosis programme, we obtained unit costs for drug-susceptible (DS) and drug-resistant (DR) tuberculosis treatment and diagnostic costs. Uncertainty in cost estimates is characterised through gamma distributions around plausible unit cost estimates in a probabilistic sensitivity analysis. There was considerable uncertainty in the cost of delivering a vaccine, including the price of vaccine compounds and programmatic delivery among adolescents. Based on expert opinion from funders, for the M72/AS01E vaccine we assume a $2.50 per-dose vaccination price with two doses per course assumed in the Basecase. Based on the average estimated BCG price from 2020-2023 from UNICEF, 43 the vaccine price per dose for BCG-revaccination was set at $0.17, with one dose assumed per course.

Vaccine introduction
All cost inputs are given in Table S5.1.
Due to uncertainty in unit costs of vaccine supply and introduction among populations who may not typically receive large-scale mass vaccination, we make several assumptions around costs to supply and introduction of vaccines. Uncertainty in cost estimates is characterised through gamma distributions. For the vaccination cost from the societal perspective, the patient time cost of vaccination was added as a multiplier to the number of doses, and therefore included in the equation along with vaccine price, vaccine supply costs, and the cost of delivery.

Cost-effectiveness analysis and willingness-to-pay thresholds
We calculated the incremental cost effectiveness ratio as the ratio between the incremental benefit, in DALYs averted, and the incremental cost, in USD, for each run across vaccination and baseline scenario. Both costs and benefits were discounted to 2025 (when vaccination began) at 3% per year, per guidelines. 48 We measured cost-effectiveness by 2050 against three India specific cost thresholds: 1x gross domestic product (GDP) per-capita (US$1,927.71), 47

5.5
Total costs from the health-system and societal perspectives The following costs are included in the health-system perspective: -Vaccine costs: One-time vaccine introduction costs, recurring vaccine delivery costs, vaccine price per dose, and supply costs -Cost of testing and diagnosis for drug-susceptible and drug-resistant cases -Cost of treatment for drug-susceptible and drug-resistant cases In addition to the costs from the health-system perspective, costs from the societal perspective include: -Vaccine costs: Patient time cost for vaccination -Non-medical patient costs (including transportation) for drug-susceptible and drug-resistant cases -Indirect patient costs for drug-susceptible and drug-resistant cases

Health impact outcomes
The following measures were calculated for each vaccine scenario as the median and 95% uncertainty range -Percent incidence rate reduction in 2050 for each vaccine scenario compared to the estimated value in

Figure S7.2 Tuberculosis infection prevalence, proportion retreated, access-to-care ratio and ratio of subclinical tuberculosis to total tuberculosis trends from 2000-2050 for all ages
The black trend line indicates the median modelled output with 95% uncertainty in shaded grey. The black dot and vertical line is the calibration target from Table S3.1.

Figure S7.3 Tuberculosis incidence and mortality rate trends from 2000-2050 by age group
The black trend line indicates the median modelled output with 95% uncertainty in shaded grey. The black dot and vertical line is the calibration target from Table S3.1.

Figure S7.4 Tuberculosis disease and infection prevalence trends from 2000-2050 by age group
The black trend line indicates the median modelled output with 95% uncertainty in shaded grey. The black dot and vertical line is the calibration target from Table S3.1.

Figure S7.5 Tuberculosis case notification and proportion retreated trends from 2000-2050 by age group
The black trend line indicates the median modelled output with 95% uncertainty in shaded grey. The black dot and vertical line is the calibration target from   Abbreviations: IRR = incidence rate reduction, MRR = mortality rate reduction.          The Basecase BCG-revaccination scenario assumes a 45% efficacy POI vaccine efficacious with no current infection at the time of vaccination, with 10 years duration of protection and reaching 80% coverage. Each BCG-revaccination scenario is delivered routinely to those aged 10 and as a campaign for those aged [11][12][13][14][15][16][17][18]. The scenarios on the figure are labelled with the difference in product characteristics for that scenario compared to the Basecase.