Estimating the contribution of subclinical tuberculosis disease to transmission: An individual patient data analysis from prevalence surveys

Decision letter after peer review:

Thank you for submitting your article "Estimating the contribution of subclinical tuberculosis disease to transmission: an individual patient data analysis from prevalence surveys" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Bavesh Kana as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1. The authors should run a sensitivity analysis on:

– The assumption that smear and symptom status are constant over time since this is almost certainly not the case. This could be accomplished, for example, by using the model described in sensitivity analysis #2 – by assigning estimates of per–unit–time relative infectiousness to each state in that model and then calibrating those estimates to the cumulative hazard ratios that are estimated in this study.

– The assumption that the background hazard of infection is the same across all households in a given survey/study. For example, do results change much if households with an index case have a 50% higher background hazard of infection?

– The durations of smear–positive and smear–negative TB are the same. For example, do results change much if smear–positive TB lasts twice as long as smear–negative TB?

2. Did the authors consider fitting a single cumulative hazard model for all 5 countries, that would allow only Λ_B (background cumulative hazard) to vary by country, rather than fitting 5 separate models and then pooling the estimates for the other lambdas after fitting? Does the former approach allow for estimation of the other λ's with more precision – unless the authors have reason to believe the λ's will vary by setting?

3. How were households with multiple TB cases treated in the analysis? If they were excluded, what effect is this likely to have on the analysis? This should be described somewhere (at least in the supplement) and the likely impact on results discussed.

4. Given how much the Philippines 1997 prevalence survey seems to be driving results (in terms of the larger sample size of household contacts), it would be useful to show how the percent of transmission results in Figure 3 look when applied to the Philippines 1997 survey data. Currently only one of the four surveys/studies used to derive the relative infectiousness estimates (Vietnam 2007) is used in the percent of transmission estimates.

Other issues that should be addressed:

1. The finding that subclinical TB is likely to be more transmissible than symptomatic TB is counterintuitive and warrants much more discussion.

2. It may be useful to clarify to readers how Mtb infection differs from subclinical TB (or TB more generally).

3. Figure 1 presents the odds ratios for infection of smear–positive vs. smear–negative (on the left) and clinical vs. subclinical TB (on the right). Figure 2 presents relative per–unit–time infectiousness estimates in the exact opposite way (smear–negative vs. smear–positive and subclinical vs. clinical, and in the opposite order). It would be clearer for readers if the presentation was more consistent across figures (and accompanying text).

4. Presenting figure 1 in terms of the relative cumulative hazard rather than (or in addition to) odds ratios would also help with the continuity between figures 1 and 2 (since the estimates of per–unit–time relative infectiousness used the relative cumulative hazards, not odds ratios).

5. Methods text should contain a bit more information on how the estimated durations of subclinical and clinical TB were used to convert the infection odds ratios into per–unit–time relative infectiousness estimates – this is clear in the appendix but not in the main text.

6. In the abstract, the term "individual population data" is a bit unclear.

7. Supplementary figure 2 may be mislabeled – for Bangladesh, it seemed that only smear–positive cases were included but the graph makes it look like only subclinical cases were included. Are the other panels similarly mislabeled?

8. p5 (line 7–9) Sentence could be made clearer. What does it mean there is a "disconnect"? Is this a pathological link? or causality? lack of evidence?

9. Results: The estimations of per unit time infectiousness of subclinical TB should be brought up front in the results, before the meta–analysis.

10. The discussion around the limitations of household contact and transmission assessment is very limited. This study inherits the limitations of the source studies, a more robust discussion around this issue would be beneficial.

11. Sensitivity of symptom screening algorithms is not discussed: TB disease, although split it into two categories is more likely a continuum.

12. Discussion page 15 (L11–L12): Is it relevant to mention SARS–CoV and Malaria here? Bringing asymptomatic transmission as a transversal subject across diseases without providing more epidemiological insights seems gratuitous. This should be removed.

13. The implications on policy and directly on recommendations around symptom screening in WHO plans are not commented on enough, this is where these results could be more useful.

14. page 16, line 32: The sentence "This is in keeping with increasing data showing that symptoms, in particular the classic TB symptom of cough, are not closely correlated to the amount of Mtb exhaled [28,29] " This is not completely consistent with Ref [28] by Theron et al. Nature Med 2020. The authors actually found that cough alone was a strong indicator for the measured transmissibility supplementary table 8. OR for CASS positivity for cough: 3.38 (1.90–5.99) despite the overall symptom score being negatively associated with transmissibility.

15. Supplementary Table 1: would be useful to add information on the definition of symptoms used to define subclinical against clinical disease.

16. Is there evidence from the literature regarding the subclinical transmission using results from whole genome sequencing? For example, Xu et al. (plos medicine) found in a study from Spain that 5/14 (35.7%) cases likely transmitted TB well before symptom onset. This can be discussed.

Reviewer #1 (Recommendations for the authors):

Slow declines in the global burden of TB are driven by a large gap between TB prevalence and the number of people with TB that are diagnosed and treated. Subclinical TB contributes to this "prevalence–notification gap"; a substantial (around 50%) of people with TB disease do not present with TB symptoms (or have fewer or milder symptoms that would not screen positive on a typical symptom screening questionnaire). The extent to which detecting and curing people with subclinical TB is important to reducing population–level TB transmission (and thus TB burden) is unclear, and depends both on infectiousness (how infectious people with subclinical TB are at any given point in time) and duration (how long they are expected to have TB before being detected or resolving on their own; and their associated future trajectory of infectiousness). Shedding light on this question could have a substantial impact on the global strategy for ending TB, including whether it is important for mass screening efforts (and other strategies to reduce the prevalence–notification gap) to detect those with subclinical TB and, if so, how to prioritize future diagnostics that best accomplish this (as the authors discuss in the last paragraph of the Discussion section).

In this study, the authors aimed to assess the infectiousness of people with subclinical TB (relative to symptomatic TB) at any given point in time; i.e., adjusting for the duration. To do this, they used individual– and household–level data from 4 cross–sectional surveys/studies that collected information on the prevalence of symptomatic TB, subclinical TB, and TB infection in households. The cumulative relative infectiousness of subclinical vs. symptomatic TB can be estimated by assessing how much higher the prevalence of TB infection is in households with a subclinical vs. symptomatic TB case. However, this estimate of cumulative relative infectiousness combines both pre–unit–time infectiousness and duration. To adjust for the duration and estimate per–unit–time relative infectiousness, the authors employed a cumulative hazard model that they then adjusted for estimates of the duration of symptomatic and subclinical TB; these duration estimates came from a mathematical model of TB natural history that is described in a different paper by Ku et al.

Strengths of the analysis include the separation of the duration of different forms of TB to estimate the per–unit–time relative infectiousness, as well as the Bayesian framework used to infer the cumulative relative hazards of infection from different forms of TB. Weaknesses include a number of assumptions made in the analysis that are unlikely to hold true and could bias results: the assumption that household index cases have continually had the same form of TB that they had when they were detected in a prevalence survey (or active case finding study), that smear–positive and smear–negative forms of TB are of the same duration, and that the background hazard of infection is the same for all households in a given survey/study. It would be useful for the authors to conduct a sensitivity analysis on these three assumptions. For example, in sensitivity analysis they estimate the durations of subclinical and symptomatic TB using a model that allows populations to transition between these states – could that model not be used to infer the per–unit–time relative infectiousness estimates? Additionally, historical data from the pre–antibiotic era (and a recent modeling study based on this data https://doi.org/10.1073/pnas.2211045119) suggests differential disease duration by smear status. In countries where a lot of TB is still diagnosed using smears (or was at the time of these prevalence surveys), differential detection/treatment probabilities could also lead to differential disease duration. This could have important implications for the analysis, given that symptom and smear status seem to be somewhat correlated in prevalence surveys (subclinical cases are much more likely to be smear–negative than smear–positive) and that smear status is more strongly predictive of infectiousness.

In all three cases, it is difficult to predict the direction and magnitude of potential bias on results, which is why sensitivity analysis is important and could strengthen the extent to which the authors' conclusions are justified by the analysis (currently this is somewhat unclear). Furthermore, the conclusion that subclinical TB is actually more transmissible than symptomatic TB is counterintuitive and warrants more discussion than is currently included in the study.

Despite these limitations, the broad conclusion that a substantial portion of TB transmission comes from subclinical cases will probably continue to hold even with adjustments to the analysis and additional sensitivity analyses. Even at the lower bound of the wide current estimated prediction interval, 27% of transmission coming from TB is substantial. This is an important finding, and thus additional analysis/revisions that can strengthen the robustness of this finding would be valuable.

Suggestions on analysis

– Suggest running sensitivity analysis on the assumption that smear and symptom status are constant over time since this is almost certainly not the case. This could be accomplished, for example, by using the model described in sensitivity analysis #2 – by assigning estimates of per–unit–time relative infectiousness to each state in that model and then calibrating those estimates to the cumulative hazard ratios that are estimated in this study.

– Suggest running sensitivity analysis on the assumption that the background hazard of infection is the same across all households in a given survey/study. For example, do results change much if households with an index case have a 50% higher background hazard of infection?

– Suggest running sensitivity analysis on the assumption that the durations of smear–positive and smear–negative TB are the same. For example, do results change much if smear–positive TB lasts twice as long as smear–negative TB?

– Did the authors consider fitting a single cumulative hazard model for all 5 countries, which would allow only Λ_B (background cumulative hazard) to vary by country, rather than fitting 5 separate models and then pooling the estimates for the other lambdas after fitting? I wonder if the former approach could allow for the estimation of the other lambdas with more precision – unless the authors have reason to believe the lambdas will vary by setting.

– How were households with multiple TB cases treated in the analysis? If they were excluded, what effect is this likely to have on the analysis? The effect could be to exclude those households that had the highest hazard of infection. This should be described somewhere (at least in the supplement) and the likely impact on results discussed.

– Given how much the Philippines 1997 prevalence survey seems to be driving results (in terms of the larger sample size of household contacts), it would be useful to show how the percent of transmission results in Figure 3 look when applied to the Philippines 1997 survey data. Currently only one of the four surveys/studies used to derive the relative infectiousness estimates (Vietnam 2007) is used in the percent of transmission estimates.

Suggestions on presentation

– The finding that subclinical TB is likely to be more transmissible than symptomatic TB is counterintuitive and warrants much more discussion (I don't believe this is currently discussed at all)

– Depending on the intended audience, it may be useful to clarify to readers how Mtb infection differs from subclinical TB (or TB more generally).

– Figure 1 presents the odds ratios for infection of smear–positive vs. smear–negative (on the left) and clinical vs. subclinical TB (on the right). Figure 2 presents relative per–unit–time infectiousness estimates in the exact opposite way (smear–negative vs. smear–positive and subclinical vs. clinical, and in the opposite order). It would be clearer for readers if the presentation was more consistent across figures (and accompanying text).

– Presenting figure 1 in terms of the relative cumulative hazard rather than (or in addition to) odds ratios would also help with the continuity between figures 1 and 2 (since the estimates of per–unit–time relative infectiousness used the relative cumulative hazards, not odds ratios).

– Methods text should contain a bit more information (like 1 sentence) on how the estimated durations of subclinical and clinical TB were used to convert the infection odds ratios into per–unit–time relative infectiousness estimates – this is clear in the appendix but not in the main text.

Reviewer #2 (Recommendations for the authors):

Emery et al., present for the first time quantitative estimates of the relative infectiousness of subclinical TB cases and the overall contribution of these to the global and country–specific (selected countries) ongoing TB transmission. The authors review and aggregate individual–level data from studies where household contacts and their index cases were assessed for their bacteriological and symptomatic status in order to reclassify and reanalyse household transmission accordingly. They produce aggregated estimates of the Odds Ratio of TB infection for households of subclinical TB index cases vs clinical index cases, and more importantly, they use a novel approach that uses cumulative hazard models to estimate the per unit infectiousness of subclinical TB. This approach accounts for important considerations like background TB transmission and duration of the infectious period. Their results (for ORs and infectiousness) show no significant advantage in the clinical index cases for infection. In light of these results, one can conclude that subclinical TB cases have no transmission disadvantage compared to symptomatic TB cases, and given its high prevalence in high–burden settings might be silently contributing disproportionately to the overall burden of TB. The implications of these results are that current approaches to active case finding and household investigation interventions put most of their efforts to identify symptomatic TB cases, potentially missing important sources of infection.

The main observable limitations of this work are "inherited" from the sources of data, namely:

1. Household contacts of index cases are a very indirect (although useful) way of assessing TB transmission and infectiousness: The direction of transmission (index case to contact) can never truly be asserted, not even can it be ruled out to have occurred from external sources. The authors only briefly mention this but it is the main source of uncertainty when discussing TB household transmission investigation. The approach here proposed does not address these points, and thus it remains a gap and a challenge in the research field.

2. Assessment of TB symptoms is in itself a type of diagnostic tool with a measurable level of performance (sensitivity/specificity). This can drastically alter the estimation of relative infectiousness by clinical/subclinical status. The authors don't mention this and it is not immediately clear how this was done in the source studies.

I have enjoyed having the opportunity of reviewing this manuscript, which I think can be a widely used resource for policymakers, mathematical modellers, and epidemiologists working on TB. The main advantages and solid points I found are summarised in my public review and assessment. Here I add some points that I think need to be addressed to improve the presentation and conveyance of the message.

– p5 (line 7–9) Sentence could be made clearer. What does it mean there is a "disconnect"? Is this a pathological link? or causality? lack of evidence?

– Results: In my view, the estimations of per unit time infectiousness of subclinical TB should be brought up front in the results, before the meta–analysis.

– The discussion around the limitations of household contact and transmission assessment is very limited. As I said before, this study inherits the limitations of the source studies but a more robust discussion around this issue would feel healthier (Kendall editorial note, etc..)

– Sensitivity of symptom screening algorithms is not discussed: TB disease, although we split it into two categories is more likely a continuum, and how good we get at identifying those signs determines how much we will trust these results in the future.

– Discussion page 15 (L11–L12): I wonder how relevant is to mention SARS–CoV and Malaria here. Bringing asymptomatic transmission as a transversal subject across diseases without providing more epidemiological insights seems gratuitous. I would suggest either removing or expanding.

– The implications on policy and directly on recommendations around symptom screening in WHO plans are not commented on enough, and I believe this is where these results could be more useful.

Reviewer #3 (Recommendations for the authors):

The investigators applied a statistical framework to synthesize information from prevalence surveys of active TB disease (regardless of symptoms) and latent tuberculosis infection among household contacts of index cases from tuberculosis surveys. They estimated an odds ratio of 1.2 (95% CI: 0.6–2.3) for infection in a household with a clinical versus subclinical index case (not considering the duration of the disease). They estimated that 68% (27–92%) of global tuberculosis transmission was from subclinical TB.

The major strength of the study is the ability to collect relevant information from prevalence surveys and integrate it using a sophisticated statistical framework. The result is novel in the sense that no such estimation has been provided before.

One major limitation of the study, as acknowledged by the authors, is the very wide confidence interval (27–92%) for the estimated global proportion of transmission contributed by subclinical disease. The wide 95% CI was consistent with a range from modest to very large impact of subclinical TB and therefore has different policy implications for tuberculosis control. It will be useful if the authors could indicate the major source of uncertainty behind this confidence interval. In other words, where and how should we collect more data to improve the precision of this estimation procedure?

Another limitation of the study is the lack of consideration of the conversion between subclinical and clinical TB status (i.e., not a one–way street) in the cumulative hazard model. Its impact on the main finding (the transmission contribution from subclinical disease) remains unclear. Considering the natural history of TB is usually from subclinical to clinical (with conversion and reversion along the way), it is anticipated that the cumulative latent infection among household contacts of a clinical TB index case would likely be contributed by both the clinical period as well as the subclinical period. This "misclassification" of clinical vs. subclinical TB seems to bias the relative infectiousness (clinical vs. subclinical) towards the null value of OR=1, under the hypothesis that clinical TB is more infectious than subclinical TB. In other words, it might therefore underestimate the relative transmissibility of clinical TB.

One of the major results of the study was that an estimated 68% (27–92%, 95% PrI) of global transmission was from subclinical TB. Given that there is substantial between–country variation in the % of subclinical TB among prevalent cases (Figure 3A) and variation in the duration of subclinical TB (see. Ku C–C et al. 2021, citation #32 of this paper), the country–level heterogeneity of the % transmission from subclinical cases should be acknowledged when making inference on the global estimate, accounting for the burden of disease in each geography. The current global estimate was based on the pooled parameter values from 14 countries (not including major countries like India and China) without formal consideration of the burden of disease. I, therefore, found the claim of global estimate a bit too strong.

Despite all these limitations, the present research is considered novel and will contribute to the field of TB epidemiology. It will generate more discussion (and debate) on the importance of detecting and treating subclinical TB. Better data and further refinement of the analytic framework will generate more precise information which will prove to be useful for the revision of TB control strategies.