Infectious and clinical tuberculosis trajectories: Bayesian modeling with case finding implications

Significance

Up to a quarter of people with prevalent tuberculosis (TB) lack clinical symptoms yet have sufficiently high bacterial burden for detection by sputum smear microscopy. Most current mass screening programs focus on symptoms and do not detect subclinical, smear-positive TB; however, we find that this form of TB contributes the most to future transmission. Therefore, replacing symptom screening with diagnostics that can feasibly screen for the most infectious forms of TB in low-resource settings could improve the impact and efficiency of TB active case finding. The development of assays with relaxed targets for sensitivity but ambitious targets for rapidity, portability, and cost—designed specifically for widespread screening—should be more highly prioritized in TB diagnostic research and development efforts.

Abstract

The importance of finding people with undiagnosed tuberculosis (TB) hinges on their future disease trajectories. Assays for systematic screening should be optimized to find those whose TB will contribute most to future transmission or morbidity. In this study, we constructed a mathematical model that tracks the future trajectories of individuals with TB at a cross-sectional timepoint (“baseline”), classifying them by bacterial burden (smear positive/negative) and symptom status (symptomatic/subclinical). We used Bayesian methods to calibrate this model to targets derived from historical survival data and notification, mortality, and prevalence data from five countries. We combined resulting disease trajectories with evidence on infectiousness to estimate each baseline TB state’s contribution to future transmission. For a person with smear-negative subclinical TB at baseline, the expected future duration of disease was short (mean 4.8 [95% uncertainty interval 3.3 to 8.4] mo); nearly all disease courses ended in spontaneous resolution, not treatment. In contrast, people with baseline smear-positive subclinical TB had longer undiagnosed disease durations (15.9 [11.1 to 23.5] mo); nearly all eventually developed symptoms and ended in treatment or death. Despite accounting for only 11 to 19% of prevalent disease, smear-positive subclinical TB accounted for 35 to 51% of future transmission—a greater contribution than symptomatic or smear-negative TB. Subclinical TB with a high bacterial burden accounts for a disproportionate share of future transmission. Priority should be given to developing inexpensive, easy-to-use assays for screening both symptomatic and asymptomatic individuals at scale—akin to rapid antigen tests for other diseases—even if these assays lack the sensitivity to detect paucibacillary disease.

Tuberculosis (TB) remains among the leading causes of global mortality. One likely driver of TB’s persistent global impact is the high burden of undiagnosed TB. Prevalence surveys reveal multiple person-years lived with TB for each person who is officially diagnosed (1). This undiagnosed TB varies in its infectiousness and symptoms (2). In particular, a large fraction of prevalent TB is subclinical: 36 to 80% of people with TB will not be classified as symptomatic by a typical symptom-screening questionnaire (3). Even among people whose TB has a high enough bacterial burden to be detectable by sputum smear microscopy (a minority of prevalent disease, and an indicator of high transmissibility), 34 to 68% lack symptoms (4). Thus, a substantial portion of prevalent TB may be highly infectious but would not be detected by the routine health system or by symptom-based screening approaches.

Systematic screening, or active case finding, has been recommended by global bodies in response to the high burden of undiagnosed TB (5, 6). However, the tests currently available for TB screening have important limitations. Current screening tests recommended by World Health Organization guidelines either miss subclinical TB (in the case of symptom screening) or are resource-intensive in ways that limit accessibility (5). Most screening tests also fall short of the target sensitivity of 90% set by existing target product profiles (5, 7). Rethinking priorities for screening tests, and considering less sensitive tests that are optimized on other characteristics, might have logistical or cost advantages.

The importance of finding TB in people with no current symptoms depends on both their current infectiousness and their future course of infectiousness, symptoms, and care-seeking. For people whose TB will soon be treated, resolve without treatment, or remain minimally infectious and asymptomatic, early diagnosis may have little impact. But for people whose TB could make a large cumulative contribution to transmission—due to the combination of their infectiousness and the total time they will spend with undiagnosed TB—the benefits of early detection could be substantial.

Because TB is typically treated once diagnosed, there are few direct data on how disease progresses over time. Epidemiological modeling can address this limitation by synthesizing evidence across data sources to draw longitudinal inferences. We sought to use such a modeling approach to describe the future clinical and infectious trajectories of prevalent TB and thus identify efficient screening approaches. Focusing on trajectories of TB disease in HIV-negative adults for the sake of tractability and data alignment, we used Bayesian calibration methods to integrate a wide variety of data sources—including survival data from historical cohorts and recent prevalence surveys from five countries (Bangladesh, Cambodia, Nepal, the Philippines, and Vietnam)—in a model of the bidirectional evolution of both TB symptoms (subclinical versus symptomatic) and sputum bacterial burden (smear-positive versus smear-negative). We then used the calibrated model to simulate the future disease trajectories and contribution to TB transmission of individuals with TB at a cross-sectional point in time (“baseline”), classified by their baseline symptom and smear status. By estimating how much different forms of prevalent TB will contribute to transmission and adverse clinical outcomes in the future, the results of this model can inform research and development for TB screening tools.

Results

Calibration.

In nearly all calibrated simulations, bacterial status was more stable than symptom status (Fig. 1 and SI Appendix, Table S6). For example, the estimated monthly probabilities of smear progression (ranging from mean 1.9% [95% uncertainty interval 1.0 to 2.9%] in Vietnam to 3.9% [2.6 to 5.3%] in Bangladesh) and regression (from 0.8% [0.1 to 2.0%] in Cambodia to 1.4% [0.3 to 3.6%] in Nepal) for those with subclinical TB were substantially lower than the corresponding probabilities of symptom progression (from 7% [4 to 12%] in the Philippines to 17% [13 to 20%] in Bangladesh) and regression (from 22% [10 to 40%] in the Philippines to 41% [26 to 50%] in Nepal) among those with smear-negative TB. Among those with smear-positive symptomatic TB, diagnosis and treatment occurred much more frequently (≥7% per month in 95% of parameter sets for all countries) than smear regression (<2% per month in 95% of parameter sets); symptom regression among this group was less constrained by the calibration process and ranged from <1% to 43% per month. The probability of spontaneous resolution from the smear-negative subclinical state was >15% per month in all countries (in >95% of calibrated simulations). Calibrated models fit closely to both historical and contemporary data ( SI Appendix, Figs. S3 and S4), and the posterior distributions of most parameters were similar across countries (Fig. 1). The complete posterior parameter sets for each country are available at github.com/rycktessman/tb-natural-history.

Fig. 1.

Results of model calibration: posterior parameter distributions. Gray regions show the prior distributions for each parameter, whereas the colored curves show the best-fitting (posterior) distributions after model calibration for each country. Values on the x-axis represent monthly probabilities, except for the four relative risk parameters. Progression relative risk refers to the increased risk of smear or symptom progression when one is symptomatic or smear-positive, compared to being subclinical or smear-negative, respectively. Regression relative risk refers to the decreased risk of smear or symptom regression when one is symptomatic or smear-positive, compared to being subclinical or smear-negative, respectively. Treatment and TB mortality transitions are only shown by smear status because these were assumed only to occur in the presence of symptoms. All priors were sampled from uniform distributions, but exclusion of infeasible parameter sets (for which the transitions out of a state exceeded 100%) resulted in some loss of uniformity ( SI Appendix ). The shift from relatively flat prior (gray) distributions to narrower posterior (colored) distributions indicates that only certain parameter values were consistent with the data used to inform the calibration process.

Individual Clinical Trajectories.

Among people with prevalent TB pooled across the five countries, those with smear-positive subclinical TB at baseline had the longest projected future duration of disease: mean 15.9 mo [11.1 to 23.5], most of which would be spent smear-positive (15.5 mo [10.7 to 23.1]) and over half of which would be spent smear-positive and subclinical (9.6 mo [5.9 to 15.3]) (Fig. 2). In addition, 96% [85 to 100%] of these individuals were projected to eventually develop symptoms; ultimately, 71% [49 to 87%] would be treated, and 17% [8 to 28%] would die of TB. In contrast, 87% [79 to 92%] of those with smear-negative subclinical TB at baseline were projected to resolve, with an average duration of only 4.8 [3.3 to 8.4] mo with TB, and with only 11% [6 to 17%] of individuals ever becoming smear-positive. Although the expected future disease duration of those with smear-positive symptomatic TB at baseline (11.0 mo [6.2 to 16.0]) was also longer than that of the smear-negative states due to less spontaneous resolution, it was shorter than that of smear-positive subclinical TB because of earlier case detection. Projected clinical trajectories were generally similar across countries ( SI Appendix, Figs. S5 and S6).

Fig. 2.

Individual TB natural history trajectory characteristics over 5 y. Panel A shows the average cumulative future time (in months) spent in each specified state (rows), conditional on being in each starting state (columns) at time 0. Panel B shows the proportion of those in each starting state at time 0 (columns) ever reaching other states/outcomes over the subsequent 5 y (rows). Uncertainty intervals for time spent in each state show the 2.5th and 97.5th percentiles across mean values from 50,000 posterior parameter sets for each country pooled together. Uncertainty intervals for the proportions ever reaching a state show the 2.5th and 97.5th percentiles across 50,000 posterior parameter sets for each country pooled together. In both panels, “Total Population” indicates the average times/proportions in a population consistent with each country’s prevalence survey (pooled across the five countries).

Trajectories of Infectiousness.

After incorporating estimates of the relative infectiousness of each state per unit time, we estimated that 6.2 [3.4 to 10.3] times as many secondary infections would arise over the next 5 y from the average person with smear-positive subclinical TB at baseline, compared to the average person with smear-negative subclinical TB (Fig. 3A and SI Appendix, Table S7). Corresponding estimates of other baseline states’ relative contributions to future transmission (continuing to use smear-negative subclinical TB as a reference) were 1.8 [1.3 to 2.4] times higher for smear-negative symptomatic TB and 4.6 [2.4 to 7.5] times higher for smear-positive symptomatic TB. Compared to baseline smear-positive symptomatic TB, the total number of future infections arising from an individual with smear-positive subclinical TB at baseline was projected to be 1.4 [1.1 to 2.0] times greater. This finding—that more future infections were estimated to arise from a person with baseline (prevalent) subclinical smear-positive TB than from a person with prevalent symptomatic smear-positive TB—was robust across all countries ( SI Appendix, p. 21 and Table S7) and as long as subclinical TB was at least 9% as infectious as symptomatic TB (Fig. 3B ).

Fig. 3.

Relative number of future transmission events from individuals with undiagnosed prevalent TB in different symptom and smear states. Both panel A and panel B show the relative cumulative infections generated over 5 y by a single individual in a given TB state at the start of that 5-y period. Panel A shows the full distribution of relative cumulative infections, with vertical lines representing means and 2.5th and 97.5th percentiles across posterior parameter sets. In the left side of panel A, the comparator is someone who initially has smear-negative subclinical TB. In the right side of panel A and in panel B, the comparator is someone who initially has smear-positive symptomatic TB. In panel B, mean relative cumulative infections are compared over a range of values for the per-time relative infectiousness of subclinical (vs. symptomatic) TB, while the per-time relative infectiousness of smear-negative (vs. smear-positive) TB is held fixed at its mean of 0.35. Panels C and D show population-level, rather than individual-level, outcomes: the average contribution of each of the four initial TB states to prevalent TB at time 0 (panel C) and to the cumulative transmission generated over the 5 subsequent years by those with prevalent TB at time 0 (panel D).

We calculated population-level contributions to transmission by combining these estimates with the relative prevalence of each state from prevalence surveys. These estimates thus vary by country, and we report the lowest and highest means (and uncertainty intervals) across the five countries. Of all infections generated in the subsequent 5 y by a cross-sectional sample of people with prevalent TB in the five countries evaluated (i.e., of all transmission potentially avertible by active case finding at a given moment), we projected that between 35% [26 to 43%] and 51% [43 to 57%] would arise from people with prevalent smear-positive subclinical TB, even though they represented only 11 to 19% of the population with prevalent TB (Fig. 3 C and D ). By comparison, people with prevalent smear-positive symptomatic TB (also 11 to 19% of prevalent TB) would generate 27% [19 to 34%] to 35% [31 to 38%] of infections, those with smear-negative symptomatic TB (12 to 24% of prevalent TB) would generate 12% [9 to 16%] to 23% [16 to 30%], and those with smear-negative subclinical TB (42 to 55% of prevalent TB) would generate the remaining 9% [7 to 13%] to 16% [9 to 25%].

Sensitivity Analyses.

In all sensitivity analyses, more secondary infections were estimated to arise from people with prevalent smear-positive subclinical TB than from any other state ( SI Appendix, Table S8). Other results differed modestly in some but not all sensitivity analyses ( SI Appendix, p. 28 and Figs. S7–S10).

Discussion

This modeling study synthesized historical and contemporary data to determine the future clinical and infectious trajectories of individuals with prevalent undiagnosed TB, according to symptom status and bacterial burden. We found that the expected contribution of prevalent subclinical disease to future transmission is strongly tied to baseline bacterial burden—because of corresponding differences not only in baseline infectiousness, but also in future disease course. We projected that individuals with smear-positive subclinical TB at baseline—who accounted for only 11 to 19% of TB in national prevalence surveys—would remain infectious for an additional 14 to 19 mo and generate 35 to 51% of future transmission events, depending on the country. By contrast, although smear-negative subclinical disease accounts for a larger proportion of prevalent TB (42 to 55%), these individuals generated only 9 to 16% of future transmission events, due to a short average TB duration (5 mo) that most often resulted in spontaneous resolution (86 to 89%). These results suggest that less than 50% of future transmissions arising from current prevalent cases can be averted by active case finding that focuses only on people with symptoms. But if all individuals with high bacterial burdens at baseline could be detected regardless of symptoms, then 62 to 78% of transmission could be averted even if those with lower bacterial burdens were missed.

The disproportionate contribution of prevalent smear-positive subclinical TB to future transmission has important programmatic implications. Systematic population screening based on symptoms will miss these individuals—but bacteriologic assays need not be highly sensitive to identify them. A target product profile for an assay to identify these individuals—the characteristics of which would resemble rapid antigen tests for COVID-19—currently does not exist in the TB space. The use case for these assays would be mass screening, not clinical diagnosis. As such, they would differ substantially from current rapid molecular tests, which achieve high sensitivity at the expense of time, complexity, and a cost of $15 to 30 per test that is unaffordable for mass screening in most low- and middle-income settings (8 –10). For an infectivity-focused screening test, sensitivity targets could be relaxed as long as high bacterial burdens were detected—and in turn, more ambitious targets could be set for portability, ease of use, throughput, and cost, with the goal of making widespread systematic screening truly feasible. More research is therefore needed to develop such assays, ideally using accessible specimens such as exhaled breath, oral fluid, or urine.

A major strength of our model is the integration of a sufficient variety of data sources to simultaneously track symptom and smear status, and to incorporate both progression and regression of each, thus allowing for inference about the trajectories of individuals with prevalent undiagnosed TB. Our conclusions were consistent across all sensitivity analyses and five different countries with different TB epidemics and data limitations, increasing generalizability.

Our data sources were better able to inform the values of some parameters than others. In particular, smear transition probabilities were forced to be small by historical evidence demonstrating markedly different cumulative mortality by baseline smear status, while a lack of similar data by baseline symptom status makes higher symptom transition probabilities possible. We found that the frequency of back-and-forth transition in symptom status did not affect the transmission impact of one-time case finding but could, however, affect the optimal frequency of symptom-based screening.

Compared to other recent TB modeling analyses, we estimated a similar duration of symptomatic TB but a shorter overall mean duration of TB than Ku et al. (11), likely because modeling spontaneous resolution only from the smear-negative subclinical state allowed such resolution to occur frequently. By contrast, Richards et al. constrained TB disease to last approximately 2 y (12). Relative to Ragonnet et al. (13), we found a similar rate of smear-negative TB mortality but lower rate of smear-positive mortality; this is likely because the contemporary and historical mortality data were sometimes in conflict with each other, leading to underestimates of historical 5-y smear-positive mortality in some countries and overestimates of present-day mortality in others ( SI Appendix, Fig. S3). Our finding that smear status is relatively stable suggests that models that treat smear status as fixed (13, 14) are reasonable approximations, at least over shorter time horizons.

As with any modeling analysis, this study has certain limitations. First, our Markov model structure did not allow “memory” of an individual’s prior states; in reality, probabilities of disease progression may differ between those who have new incident TB and those who have had more advanced forms previously. Second, we dichotomized symptoms and bacterial burden to correspond with available data, but this dichotomization hides a spectrum. For example, it is possible that the subset of smear-negative TB that is detectable by rapid molecular tests may be more persistent than the average smear-negative TB case, such that detecting Xpert-positive smear-negative TB may yield more benefits than implied here. Furthermore, different prevalence surveys use different definitions of symptomatic, and the average severity of symptoms may be milder (with corresponding lower mortality) among all symptomatic prevalent TB than among patients who sought care and entered historical cohorts. Third, we did not incorporate prevalence survey design effects in our parameterization of the likelihood function; however, survey design is likely to influence the prevalence estimate itself more than the proportion of prevalent cases that are smear-/symptom-positive. Fourth, we only included cohort studies of patients with smear-negative TB if they had been smear-positive (or “bacillary”) prior to enrollment, to increase our confidence that the patients being described truly had TB, but it is possible that some of these patients had already experienced spontaneous resolution. Fifth, evidence on the relationship between smear and symptom status and transmissibility comes from a limited number of household contact and molecular epidemiology studies that are not directly translatable into estimates of instantaneous relative infectiousness. Examples of potential biases include those stemming from behavior change (i.e., people with symptoms may limit contact outside their household), contact saturation (i.e., even low levels of infectiousness may result in infection of all household members), temporal distortions (i.e., less infectious individuals with a longer duration of disease may generate more transmission), and assumptions of stability (i.e., someone who is asymptomatic when diagnosed may have had symptoms previously). We attempted to reflect this uncertainty in our model, and our finding that people with baseline subclinical, smear-positive TB contributed substantially to future transmission was consistent over a wide range of relative infectiousness estimates. Finally, our model is not suitable for representing the population-level dynamics of prevalent TB in high-HIV-burden settings, due to the higher rates of symptom progression and mortality associated with HIV coinfection. Similar limitations may exist regarding generalizability to settings where a large proportion of TB is multidrug-resistant (or previously treated). However, our results about individual disease trajectories are expected to apply to HIV-negative adults with drug-susceptible TB in settings with both high and low HIV prevalence and drug resistance—insofar as those settings have similar passive case detection practices as the countries we modeled and have not yet begun to scale up active case-finding activities.

In summary, a disproportionate amount of future TB transmission likely arises from individuals with undiagnosed prevalent smear-positive subclinical TB. To be effective, active case finding strategies must identify this subset of individuals and thus should not rely on symptom screening for triage. By contrast, an estimated 87% of prevalent smear-negative subclinical TB may spontaneously resolve before being diagnosed and treated, usually with a short disease course and only minimal contribution to ongoing transmission. Novel tools—for example, lower-sensitivity, high-specificity, low-cost, portable diagnostics—to efficiently identify people with smear-positive subclinical TB should therefore be prioritized.

Materials and Methods

We developed a Markov state-transition model to characterize the expected trajectories of prevalent bacteriologically positive TB disease at a population level. For simplicity and consistency across data sources, we focused on HIV-negative adults with pulmonary TB and no previous history of TB treatment. We categorized TB disease into four stages, based on a time-dependent classification of whether an individual would endorse TB symptoms on a standard symptom screen (i.e., subclinical versus symptomatic) and whether microscopy of an expectorated sputum specimen would be positive for acid-fast bacilli if performed (i.e., smear-negative versus smear-positive; selected as a proxy for low versus high bacterial burden that is frequently measured in prevalence surveys) (Fig. 4). TB was assumed to begin from a smear-negative, subclinical state; that is, individuals transition through a smear-negative state before becoming smear-positive, and through a subclinical state before developing symptoms. Similarly, spontaneous resolution of TB (to a non-infectious, asymptomatic, bacteriologically negative state) was assumed to occur only from the smear-negative, subclinical state. We assumed that only those with symptoms can be offered TB treatment (under the standard of care, in which people must first seek care for symptoms) or experience death due to TB.

Fig. 4.

TB natural history model. Shown is the Markov state-transition structure used to evaluate clinical and infectious trajectories of TB disease. Individuals with TB were grouped according to symptom status (subclinical versus symptomatic) and sputum smear status (smear-positive versus smear-negative), as described in the main text. Boxes represent different TB-related states, while arrows represent transitions between states and in (incident TB cases; modeled in calibration to present-day targets only, in order to reach a steady state) and out (TB and non-TB deaths) of the model.

The model progresses in discrete monthly time steps, with monthly probabilities of symptom progression, symptom regression, smear status progression, smear status regression, spontaneous resolution, treatment, non-TB (“background”) mortality, and TB-specific mortality. Individuals who resolve spontaneously leave the model (though we evaluated in sensitivity analysis the possibility of recrudescence from the spontaneously resolved state), as do individuals who initiate treatment (as they are no longer treatment-naïve). To induce correlations between smear and symptom status, we assumed in the main analysis that those with smear-positive subclinical TB are at greater risk of developing symptoms than those with smear-negative subclinical TB, those with smear-negative symptomatic TB are at greater risk of progressing to smear positivity than those with smear-negative subclinical TB, and those with smear-positive symptomatic TB are less likely to regress in symptom or smear status than those with smear-negative symptomatic TB or smear-positive subclinical TB, respectively. We evaluated additional constraints in sensitivity analysis. All details and model equations are presented in SI Appendix, pp. 4–6, and reproducible model code is available at github.com/rycktessman/tb-natural-history. Analyses were performed in R version 4.0.2 (R Foundation for Statistical Computing, Vienna, Austria).

Data and Setting.

All transition probabilities in the model (other than background mortality) were calibrated to data and quantities estimated from the scientific literature (Table 1, full details in SI Appendix, pp. 8–19). These calibration targets were based on historical evidence on TB survival in the pre-antibiotic era and present-day data on TB prevalence, notifications, and mortality in Bangladesh, Cambodia, Nepal, the Philippines, and Vietnam (15 –20), with adjustments made to reported data as appropriate (e.g., to exclude previously treated individuals and clinical overdiagnoses). These countries were selected because they have: (a) high TB incidence (>100 per 100,000 per year according to WHO estimates); (b) recent TB prevalence surveys with reporting of smear and symptom status; and (c) low prevalence of HIV coinfection and TB drug resistance (factors likely to alter disease trajectories but not differentiated in our model). We expect inferences about the natural course of disease to generalize to HIV-negative adults in other settings as well. In our model, cross-sectional prevalence data correspond to the distribution of TB across smear and symptom states under current diagnosis and treatment practices. Estimates derived from notification and mortality data correspond to transitions out of the model via treatment or death (we also model unobserved exits from the model via spontaneous resolution or unnotified treatments). Historical survival data consisted of reviews of TB cohort studies from the pre-antibiotic era (13, 21), from which we extracted 5- and 10-y survival probabilities after being diagnosed with symptomatic TB, stratified by smear status at the time of diagnosis and pooled across cohorts using Poisson meta-regression ( SI Appendix, pp. 14–16).

Table 1.

Model parameters, prior distributions, and calibration targets

Calibration targets (Philippines)	Value	Source
Percent of prevalent TB that is smear-negative subclinical	51% [46 to 56%]	Philippines 2016 Prevalence Survey (16)
Percent of prevalent TB that is smear-positive subclinical	19% [16 to 24%]	Philippines 2016 Prevalence Survey (16)
Percent of prevalent TB that is smear-negative symptomatic	12% [9 to 15%]	Philippines 2016 Prevalence Survey (16)
Percent of prevalent TB that is smear-positive symptomatic	17% [13 to 21%]	Philippines 2016 Prevalence Survey (16)
Ratio of smear-positive prevalence to annual smear-positive notifications	1.95 [1.05 to 3.10]	Estimated – details in SI Appendix
Ratio of annual mortality to prevalence	2.0% [0.1 to 3.1%]	Estimated – details in SI Appendix
Percent of notified true TB that would be smear-positive if tested	45% [32 to 62%]	Estimated – details in SI Appendix
Calibration targets (Vietnam)	Value	Source
Percent of prevalent TB that is smear-negative subclinical	53% [47 to 59%]	Vietnam 2018–19 Prevalence Survey (20)
Percent of prevalent TB that is smear-positive subclinical	11% [7 to 16%]	Vietnam 2018–19 Prevalence Survey (20)
Percent of prevalent TB that is smear-negative symptomatic	24% [19 to 30%]	Vietnam 2018–19 Prevalence Survey (20)
Percent of prevalent TB that is smear-positive symptomatic	11% [7 to 16%]	Vietnam 2018–19 Prevalence Survey (20)
Ratio of bacteriologically positive prevalence to annual true-TB notifications	3.9 [3.0 to 4.8]	Estimated – details in SI Appendix
Ratio of annual mortality to prevalence	4.4% [1.9 to 7.1%]	Estimated – details in SI Appendix
Percent of notified true TB that would be smear-positive if tested	68% [54 to 82%]	Estimated – details in SI Appendix
Calibration targets (Cambodia)	Value	Source
Percent of prevalent TB that is smear-negative subclinical	52% [46 to 57%]	Cambodia 2011 Prevalence Survey (18)
Percent of prevalent TB that is smear-positive subclinical	18% [14 to 23%]	Cambodia 2011 Prevalence Survey (18)
Percent of prevalent TB that is smear-negative symptomatic	15% [11 to 19%]	Cambodia 2011 Prevalence Survey (18)
Percent of prevalent TB that is smear-positive symptomatic	14% [11 to 18%]	Cambodia 2011 Prevalence Survey (18)
Ratio of smear-positive prevalence to annual smear-positive notifications	1.3 [0.9 to 1.7]	Estimated – details in SI Appendix
Ratio of annual mortality to prevalence	3.7% [1.3 to 6.0%]	Estimated – details in SI Appendix
Percent of notified true TB that would be smear-positive if tested	89% [78 to 95%]	Estimated – details in SI Appendix
Calibration targets (Nepal)	Value	Source
Percent of prevalent TB that is smear-negative subclinical	55% [48 to 61%]	Nepal 2018–19 Prevalence Survey (19)
Percent of prevalent TB that is smear-positive subclinical	19% [14 to 24%]	Nepal 2018–19 Prevalence Survey (19)
Percent of prevalent TB that is smear-negative symptomatic	15% [11 to 20%]	Nepal 2018–19 Prevalence Survey (19)
Percent of prevalent TB that is smear-positive symptomatic	12% [8 to 16%]	Nepal 2018–19 Prevalence Survey (19)
Ratio of bacteriologically positive prevalence to annual true-TB notifications	4.7 [3.5 to 6.1]	Estimated – details in SI Appendix
Ratio of annual mortality to prevalence	19.2% [8.1 to 32.1%]	Estimated – details in SI Appendix
Percent of notified true TB that would be smear-positive if tested	73% [55 to 89%]	Estimated – details in SI Appendix
Calibration targets (Bangladesh)	Value	Source
Percent of prevalent TB that is smear-negative subclinical	42% [37 to 48%]	Bangladesh 2015–16 Prevalence Survey (15)
Percent of prevalent TB that is smear-positive subclinical	19% [15 to 24%]	Bangladesh 2015–16 Prevalence Survey (15)
Percent of prevalent TB that is smear-negative symptomatic	19% [14 to 23%]	Bangladesh 2015–16 Prevalence Survey (15)
Calibration targets (Bangladesh)	Value	Source
Percent of prevalent TB that is smear-positive symptomatic	19% [15 to 24%]	Bangladesh 2015–16 Prevalence Survey (15)
Ratio of bacteriologically positive prevalence to annual true-TB notifications	3.9 [3.2 to 4.5]	Estimated – details in SI Appendix
Ratio of annual mortality to prevalence	16.7% [7.2 to 26.2%]	Estimated – details in SI Appendix
Percent of notified true TB that would be smear-positive if tested	92% [79 to 98%]	Estimated – details in SI Appendix
Historical calibration targets (all countries)	Value	Source
Five-year mortality of smear-negative symptomatic TB	13% [9 to 16%]	Pooled estimate of studies in refs. 13 and 21
Ten-year mortality of smear-negative symptomatic TB	21% [15 to 26%]	Pooled estimate of studies in refs. 13 and 21
Five-year mortality of smear-positive symptomatic TB	58% [51 to 64%]	Pooled estimate of studies in refs. 13 and 21
Ten-year mortality of smear-positive symptomatic TB	71% [65 to 77%]	Pooled estimate of studies in refs. 13 and 21
Prior parameter distributions (all countries)	Distribution [lower bound, upper bound]	Notes/rationale
Monthly progression probability from smear-negative subclinical to smear-positive subclinical	Uniform [0%, 20%]	Minimally informative *
Monthly progression probability from smear-negative subclinical to smear-negative symptomatic	Uniform [0%, 20%]	Minimally informative *
Relative risk of smear progression if already symptomatic	Uniform [1, 10]	Constrained to be ≥ 1
Relative risk of symptom progression if already smear-positive	Uniform [1, 10]	Constrained to be ≥ 1
Monthly regression probability from smear-positive symptomatic to smear-negative symptomatic	Uniform [0%, 20%]	Minimally informative *
Monthly regression probability from smear-positive symptomatic to smear-positive subclinical	Uniform [0%, 50%]	Upper bound of 50% corresponds to a symptom duration of ≥ 2 wk in the average person with symptomatic TB
Relative risk of smear regression if subclinical	Uniform [0, 1]	Constrained to be ≤ 1
Relative risk of symptom regression if smear-negative	Uniform [0, 1]	Constrained to be ≤ 1
Monthly probability of spontaneous cure from smear-negative subclinical	Uniform [0%, 50%]	Upper bound of 50% was selected to align with that of symptom regression
Monthly probability of treatment from smear-negative symptomatic	Uniform [0%, 25%]	Minimally informative *
Relative risk of being treated if smear-positive symptomatic (vs. smear-negative symptomatic)	Uniform [1, 20]	Constrained to be ≥ 1
Monthly probability of TB mortality from smear-negative symptomatic	Uniform [0%, 10%]	Minimally informative*
Relative risk of TB mortality if smear-positive symptomatic (vs. smear-negative symptomatic)	Uniform [1, 20]	Constrained to be ≥ 1
Non-calibrated transition probabilities	Value	Source
Annual all-cause mortality (present-day, Philippines)	0.5%	WHO life tables for adults in each country (22)
Annual all-cause mortality (present-day, Vietnam)	0.4%
Annual all-cause mortality (present-day, Cambodia)	0.7%
Annual all-cause mortality (present-day, Nepal)	0.5%
Annual all-cause mortality (present-day, Bangladesh)	0.4%
Annual all-cause mortality (historical)	1%		Based on refs. 13 and (23)
Other non-calibrated parameters	Value		Source
Relative infectiousness of smear-negative TB (vs. smear-positive with the same symptom status)	35% [20 to 55%]	Based on refs. (24 –26); details in SI Appendix
Relative infectiousness of subclinical TB (vs. symptomatic with the same smear status)	70% [50 to 100%]	Based on refs. (27 –35); details in SI Appendix

Brackets show 95% CIs around the calibration target means. All targets and estimated values are restricted to treatment-naïve individuals, as detailed in SI Appendix .

*The specified upper bound was selected for efficiency, after initial calibration runs were not found to prefer values above the specified range.

Calibration.

We calibrated the model to historical calibration targets and each country’s present-day calibration targets using Bayesian Incremental Mixture Importance Sampling (36) with minimally informative, uniformly distributed priors (Table 1). Separate historical and present-day simulations, with shared natural history parameters but distinct treatment assumptions and calibration targets, contributed to a single likelihood for each country ( SI Appendix, pp. 17–18).

Estimation of Future Clinical Trajectories.

After calibrating model parameter values for each country, we simulated individual trajectories for a cohort of 50,000 people with prevalent TB at a cross-sectional point in time (baseline). To replicate the future trajectories of individuals with prevalent TB at that point in time—e.g., trajectories that might be altered by a case-finding campaign—the cohort’s initial state corresponded to the distribution of symptoms and smear status in each prevalence survey. For each set of parameter values in the calibrated posterior distribution, we ran the model forward in time for 5 y. We characterized the resulting future trajectories in terms of (a) the proportion of individuals in each baseline symptom/smear state who would ever reach other specified states and (b) the mean cumulative time after baseline spent in each state (and with active TB of any kind), averaged across all individuals in a given state at baseline. Both outcomes are presented as means, with uncertainty intervals corresponding to the 2.5th and 97.5th percentiles, across all posterior parameter sets. Means and uncertainty intervals were calculated for each country separately (to generate country-specific estimates of outcomes) and for all countries combined (to generate pooled estimates of outcomes, details on SI Appendix, p. 19).

Estimation of Future Infections.

The number of future secondary infections expected to arise from a person with TB (i.e., that might be averted by diagnosing and treating that person at a given point in time) depends not merely on the infectiousness of their current state but on both the infectiousness and the duration of all TB states they will inhabit from that time until resolution, cure, or death. For example, individuals currently in a state that is less infectious, but that is likely to persist undetected for a long time or progress to greater infectiousness, could generate more secondary infections than those in a highly infectious state that will quickly lead to diagnosis and/or mortality. Therefore, we calculated the relative contribution to future transmission from individuals in each of the four prevalent TB states at baseline in our model. This calculation combined the time spent in each future TB state in the simulated disease trajectories with evidence from the literature (e.g., household contact studies) on the relative infectiousness of each state per unit time. Specifically, we estimated that smear-negative TB is 35% [95% uncertainty interval 20 to 55%] as infectious as smear-positive TB when controlling for symptoms, and that subclinical TB is 70% [50 to 100%] as infectious as symptomatic TB when controlling for smear status (Table 1; details in SI Appendix, pp. 19–20). Combining this per-time relative infectiousness with average future disease trajectories for each baseline state allowed us to estimate the relative benefits of identifying and successfully treating one additional person with prevalent TB in each state. We weighted these results by the relative size of the population in each state (from prevalence survey data) to describe the country-specific fractions of future transmission likely to arise from individuals in each of the four states at baseline.

Sensitivity Analysis.

We tested the sensitivity of results from all five countries to various structural and parameter assumptions (details in SI Appendix, p. 7). Additionally, given the wide variation in relative infectiousness estimates reported in the literature and potential biases in their interpretation, we assessed the sensitivity of the simulation results to the relative infectiousness parameters.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Previously published data were used for this work (all data used are cited in the study's references and include: Ragonnet et al. (13); Bangladesh National Tuberculosis Control Programme (15); Philippines Department of Health (16); Cambodia National Tuberculosis Control Program (18); Nepal National Tuberculosis Control Center (19); Nguyen et al. (20); WHO Global Tuberculosis Programme (17); and Tiemersma et al. (21)).