Quantifying Mycobacterium tuberculosis Transmission Dynamics Across Global Settings: A Systematic Analysis

Abstract

The degree to which individual heterogeneity in the production of secondary cases (“superspreading”) affects tuberculosis (TB) transmission has not been systematically studied. We searched for population-based or surveillance studies in which whole genome sequencing was used to estimate TB transmission and in which the size distributions of putative TB transmission clusters were enumerated. We fitted cluster-size–distribution data to a negative binomial branching process model to jointly infer the transmission parameters (the reproduction number) and the dispersion parameter, , which quantifies the propensity of superspreading in a population (generally, lower values of () suggest increased heterogeneity). Of 4,796 citations identified in our initial search, 9 studies from 8 global settings met the inclusion criteria (n = 5 studies of all TB; n = 4 studies of drug-resistant TB). Estimated values (range, 0.10–0.73) were below 1.0, consistent with declining epidemics in the included settings; estimated values were well below 1.0 (range, 0.02–0.48), indicating the presence of substantial individual-level heterogeneity in transmission across all settings. We estimated that a minority of cases (range, 2%–31%) drive the majority (80%) of ongoing TB transmission at the population level. Identifying sources of heterogeneity and accounting for them in TB control may have a considerable impact on mitigating TB transmission.

Abbreviations

SNP

single nucleotide polymorphism

TB

tuberculosis

WGS

whole genome sequencing

Tuberculosis (TB), an airborne infectious disease caused by Mycobacterium tuberculosis, is an urgent public health issue resulting in an estimated 10 million new cases and 1.5 million deaths globally in 2020, the latter being an increase over the previous year (1). While global TB incidence has declined 11% since 2015, the rate of decline is insufficient to achieve the World Health Organization’s goal of reducing the incidence of new TB cases by 80% by the year 2030 (1, 2). An improved understanding of the patterns of M. tuberculosis transmission may help guide more focused strategies for accelerating TB control.

Individual-level heterogeneity in disease transmission, or differences in the number of secondary cases produced by each index case, varies between pathogens and is fundamental to the understanding of transmission dynamics (3, 4). As heterogeneity increases, a smaller proportion of cases are responsible for a larger fraction of transmission events. For pathogens characterized by extensive variation (“superspreading”), prevention measures targeting high-risk settings or individuals have been shown to disproportionately reduce transmission (3, 5, 6). Hence, accurately characterizing the extent of such heterogeneity may be important for understanding the value of targeted interventions (4, 7).

The degree of individual heterogeneity in M. tuberculosis transmission remains poorly understood. A limited body of evidence suggests marked variability in transmission between index cases, including several outbreak investigations that have identified large clusters of genetically similar TB cases linked to a single or small number of index cases (8–11). While these studies suggest increased heterogeneity, outbreak investigations are not representative of population-level transmission dynamics. A smaller number of studies have used surveillance data to address this question and have estimated similarly high degrees of transmission heterogeneity, yet these studies have been limited to low-incidence settings and may not be generalizable to the global TB epidemic (12–16). A systematic understanding of this phenomenon and its role in shaping population-level TB epidemiology across a range of global settings is needed (17–19).

Whole genome sequencing (WGS) of M. tuberculosis isolates has increasingly been used to identify the sizes of putative TB transmission clusters, defined as the sum of all cases in a given chain of recent transmission, in TB surveillance. In this study, we systematically reviewed the literature for surveillance reports or population-based studies which included WGS data sufficient to estimate the distribution of TB transmission cluster sizes. We used these distributions to estimate the degree of individual-level transmission heterogeneity across settings.

METHODS

Study design and search strategy

We conducted a systematic review of studies that identified the distribution of TB transmission clusters in a defined population. The review was conducted in accordance with the criteria outlined by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (20) and was registered with the International Prospective Register of Systematic Reviews (PROSPERO) (21). Our search strategy was informed by prior literature reviews, through consultation with surveillance experts, and in consultation with science librarians who have expertise in systematic database searches on public health issues at the Harvey Cushing/John Hay Whitney Medical Library at Yale University School of Medicine (New Haven, Connecticut) and the Woodruff Health Sciences Center Library at Emory University (Atlanta, Georgia). Briefly, an initial search for all English-language peer-reviewed studies examining TB transmission via WGS that had been published from 1998 to the present was conducted on July 7, 2021. A follow-up search using the same criteria was conducted on September 23, 2021. Medical Subject Headings and keywords were used to search PubMed (National Library of Medicine, Bethesda, Maryland), and analogous methods were used in the Web of Science (Clarivate Analytics PLC, Philadelphia, Pennsylvania) and Excerpta Medica (EMBASE; Elsevier BV, Amsterdam, the Netherlands) databases (see Web Appendix 1, available at https://doi.org/10.1093/aje/kwac181). After completion of the systematic search, we manually searched references of relevant papers and contacted leading TB genotypic surveillance experts to identify any remaining data sources that met the inclusion criteria. Corresponding authors of potentially included studies were contacted for additional clarification if a study met eligibility criteria but the full distribution of cluster sizes could not be abstracted.

Inclusion and exclusion criteria

Studies were included if they satisfied the following criteria: 1) they were either studies of TB surveillance data or had a population-based study design; 2) they used WGS to identify putative TB transmission clusters, with or without the use of other genotypic or epidemiologic techniques; 3) the investigators explicitly stated their definition of TB transmission clusters; and 4) the full distribution of transmission clusters in the study population could be enumerated, including assumed isolated cases (i.e., cases that could not be linked to any others in the study). If multiple studies conducted analyses on the same data set or substantially overlapping data sets, only 1 study was included based on relevance to the aims of this study and the strength of its design/analysis. Studies that used WGS to analyze transmission only for an outbreak investigation were excluded. After removal of duplicates, article abstracts were screened at least twice for relevance and potential inclusion in the study (J.P.S.). Full articles of relevant studies identified for potential inclusion were adjudicated by the authors (T.C., S.S., A.N.H., N.R.G., D.D.) to determine a short list of potential studies. Short-listed studies were subsequently discussed by all authors, and final included studies were confirmed by consensus.

Data abstraction

Standardized data abstraction forms were developed prior to conducting the search. Extracted study characteristics included the sampling time frame, the specific population from which the surveillance was performed, the estimated sampling fraction, and whether the study included all TB cases or focused exclusively on drug-resistant TB cases. Transmission cluster distribution data were independently abstracted per the authors’ definition, which included the number of single nucleotide polymorphisms (SNPs; a variation in a single M. tuberculosis nucleotide) and additional epidemiologic methods used (i.e., social network analysis) to identify putative transmission clusters. We abstracted TB incidence data from studies if authors cited population-specific incidence; if unavailable, we obtained estimated country-level data on TB incidence from the World Health Organization using the median year of the sampling time frame as a proxy (1). We obtained data on per-capita gross domestic product, in US dollars, from the World Bank and used standardized definitions from the World Bank to classify populations based on country income (22).

We included studies reporting on the surveillance of both drug-susceptible and drug-resistant TB strains. We assumed that each transmission cluster originated with a single introduction event (i.e., a single initial index case) and that introduction events occurred through either importation (i.e., immigration of an infectious individual into the population) or reactivation of latent TB infection acquired in the distant past; drug-resistant clusters may also be a consequence of resistance acquired in a circulating strain. Given this and other potential differences between transmission of drug-susceptible and drug-resistant strains (i.e., infectiousness, infectious periods, etc.), we present results for studies that only reported data on drug-resistant TB surveillance separately.

Parameter inference

Following previous studies, we used a branching process framework to infer transmission parameters from abstracted TB transmission cluster data sets (12, 23–25). This approach assumes that the number of secondary cases produced from each infectious case is distributed according to a negative binomial distribution with mean R (the basic reproduction number) and dispersion parameter k (see Web Appendix 2). The parameter k quantifies the degree of individual heterogeneity by measuring overdispersion in the distribution (e.g., higher-than-expected variation). For a given , smaller values of () suggest increased heterogeneity in the number of secondary cases between individual index cases.

We fitted abstracted cluster data to a negative binomial branching process model and used maximum likelihood estimation to jointly infer R and k (25, 26). Corresponding 95% confidence intervals were obtained using profile likelihood (27). We then used inferred parameters to calculate the expected proportion of infectious cases responsible for 50%, 80%, and 90% of all secondary cases and the probability of observing an outbreak of at least size Y (i.e., ) in a large population of susceptible individuals. We arbitrarily chose values of Y as 5, 10, 25, and 50 to sufficiently capture a range of outbreak sizes. Detailed methods are presented in Web Appendix 3.

Simulation study and sensitivity analysis

We conducted a simulation study to explore 4 mechanisms by which the observed cluster size distributions may differ from the true distribution (Web Appendix 4, Web Table 1). Briefly, we considered surveillance as a 2-step process. First, we considered the probability that each case was observed through routine (passive) genomic surveillance, denoted . This represents the probability that a given case is reported, is culture-confirmed, and yields interpretable WGS results and was applied to all individuals independently. Second, a notified case of TB may trigger additional active case-finding measures (i.e., contact tracing) to identify cases in the chain of transmission missed by routine surveillance. After evaluation with , we considered the probability that a case was observed through subsequent active case-finding measures, denoted . This probability was applied to all cases remaining unobserved after routine surveillance, conditional on at least 1 other case in the chain of transmission being observed through passive surveillance (i.e., the sentinel case required to trigger case-finding).

Third, after evaluation of and , we considered the position of remaining missing cases in the chain of transmission. If an unobserved case was the sole link between previous and future generations of spread, it may appear to “break” the cluster into multiple, smaller pseudoclusters (Web Appendix 4, Web Figure 1). Chains meeting this condition were separated, and each pseudocluster was considered a unique observed cluster. Finally, some unknown proportion of cases will be ongoing at the end of the study time frame (censored; Web Appendix 4, Web Figure 1). We modeled this by randomly designating a proportion of clusters as censored with probability . In addition, we evaluated the influence of the number of clusters in each surveillance system (i.e., the sample size for calculations) by simulating systems with 2,000, 1,000, 500, and 100 clusters.

To explore the effects of these observation mechanisms on parameter inference, we first simulated 1,000 true underlying systems, each containing 2,000 chains of transmission (“perfect surveillance”), using uniformly sampled values of R and k between the minimum and maximum inferred values from included studies. We first applied the previously described observation mechanisms across a range of values for , and to individually to evaluate the impact of each type of bias in sampling. We then conducted an additional scenario analysis in which all 4 observation mechanisms exist; we assumed that 60% of cases were observed through routine surveillance (), 15% of otherwise undiagnosed cases were ascertained through active case-finding (), and 10% of clusters were censored (). These values were conservatively chosen from both assessment of the sampling fraction among included studies (median, 81%) and published reports regarding passive and active case ascertainment in similar settings (28–32).

Lastly, we examined the assumption that the offspring distribution followed a zero-inflated negative binomial assumption (Web Appendix 5). We compared the fits of the zero-inflated negative binomial model and a negative binomial model using Akaike’s information criterion (Web Appendix 5, Web Table 2).

RESULTS

Of the 4,796 studies identified in our electronic and manual searches, 9 studies met our inclusion criteria and were included in our analysis (33–41). Five studies investigated transmission for TB cases irrespective of drug resistance status, while 4 focused only on drug-resistant cases (Figure 1). The studies were from diverse populations: Investigators in 4 studies reported data from high-income countries (Canada (33), the United Kingdom (37), Singapore (38), and Portugal (39)), 4 studies came from upper- or lower-middle-income countries (Ghana (34), Brazil (36), and 2 different provinces in China (40, 41)), and 1 was based in a low-income country (Malawi (35)) (Table 1). Five studies came from relatively high-incidence settings (>50 cases/100,000 population), with the highest reported incidence in Iqaluit, Canada (33) (205 cases/100,000 population; Table 1).

Figure 1.

Identification and selection of studies for inclusion in an analysis of Mycobacterium tuberculosis transmission dynamics, 2021. Studies could meet multiple exclusion criteria; only 1 reason is listed. TB, tuberculosis.

Table 1.

Characteristics of Studies Included in an Analysis of Mycobacterium tuberculosis Transmission Dynamics, 2021

First Author, Year (Reference No.)	Study Location/Setting	World Bank Income Classification (Country Income Level)	Study Type	Time Frame of Isolate Collection	Type of TB Surveillance	TB Incidence (No. of Cases per 100,000 Population)  a
All TB Surveillance
Alvarez, 2021 (33)	Iqaluit, Nunavut, Canada	High-income	Public health surveillance	2009–2015 (6 years)	All TB cases	205b
Asare, 2020 (34)	Accra and East Mamprusi, Ghana	Lower middle-income	Population-based survey	2012–2016 (4 years)	All TB cases	165b
Guerra-Assunção, 2015 (35)	Karonga District, Malawi	Low-income	Population-based survey	1995–2010 (16 years)	All TB cases	106c
Verza, 2020 (36)	Santa Catarina, Brazil	Upper middle-income	Public health surveillance	2014–2016 (3 years)	All TB cases	23.7
Walker, 2014 (37)	Oxfordshire, United Kingdom	High-income	Population-based survey	2007–2012 (6 years)	All TB cases	8.4
Drug-Resistant TB Surveillance
Chee, 2021 (38)	Singapore	High-income	Public health surveillance	2006–2018 (13 years)	Drug-resistant TB	40
Jiang, 2019 (40)	Shenzhen, China	Upper middle-income	Population-based survey	2013–2017 (5 years)	Drug-resistant TB	65b
Macedo, 2019 (39)	Portugal	High-income	Public health surveillance	2013–2017 (5 years)	Drug-resistant TB	23b
Yang, 2017 (41)	Shanghai, China	Upper middle-income	Population-based survey	2009–2012 (4 years)	Drug-resistant TB	65b

Abbreviation: TB, tuberculosis.

^a Incidence data are setting-specific and are taken directly from the cited studies unless otherwise noted.

^b Incidence data were not reported; country-level incidence data from the World Health Organization (1) using the median calendar year of the study’s time frame were taken as a proxy.

^c Midpoint of author-reported range of annual incidence (reported 87–124 cases per 100,000 population).

The median time frame for isolate collection was 5 years (range, 3–16) (Table 1), with a median of 290 genotyped isolates (range, 80–2,309) in each data set (Table 2). Six studies used a genetic distance of 10–12 SNPs as the threshold for defining transmission links; investigators in the 3 remaining studies chose 6 SNPs or fewer (33, 36, 39) (Table 2). In the final definition of a transmission cluster, 8 studies included the use of additional epidemiologic techniques (i.e., social network analysis) in their definition; only 1 study used WGS alone (35). While median cluster size was similar for all studies, there was substantial variation in the percentage of isolated cases (range, 34%–85%) and the maximum cluster size (range, 6–78) across studies (Table 2).

Table 2.

Tuberculosis Cluster Definitions, Assumptions, and Characteristics of Studies Included in an Analysis of Mycobacterium tuberculosis Transmission Dynamics, 2021

First Author, Year (Reference No.)	Definition of a TB Transmission Cluster	SNP Cutoff (Source)	Total No. of Observed TB Cases in Population  a	Cases Available for Genotyping		Cases With Genotyping Data		Unique Isolates		Cluster Size
				No.	%	No.	%	No.	%	Median (IQR)	Maximum
All TB Surveillance
Alvarez, 2021 (33)	1) SNP difference of 3 or less, 2) epidemiologic links, 3) biological plausibility, and 4) only 1 likely source case	3 (own data)	178	147	83	140	95	72	51	2 (2–8)	18
Asare, 2020 (34)	SNP differences for clusters of >5 cases and MIRU-VNTR for clusters of ≤5 cases	10 (Guerra-Assunção et al. (35))	3,303	2,604	79	2,309	89	1,611	70	2 (2–3)	78
Guerra-Assunção, 2015 (35)	SNP differences	10 (own data)	2,332	2,332	79	1,687	72	672	40	2 (2–3)	38
Verza, 2020 (36)	SNP differences and epidemiologic links	5 (own data)	Unknown	N/A	N/A	151	N/A	85	56	2 (2–3)	12
Walker, 2014 (37)	SNP differences and epidemiologic links	12 (own data)	390	269	69	247	92	222	85	2 (2–3)	7
Drug-Resistant TB Surveillance
Chee, 2021 (38)	SNP differences and epidemiologic links	10 (own data)	Unknown	290	N/A	290	100	247	85	2 (2–3)	16
Macedo, 2019 (39)	SNP differences and epidemiologic links	6 (own data)	96	83	86	80	96	27	34	3 (2–8)	21
Jiang, 2019 (40)	SNP differences and social network analysis	12 (own data)	450	450	100	417	93	312	75	2 (2–3)	6
Yang, 2017 (41)	SNP differences and epidemiologic links	12 (own data)	367	324	88	324	100	221	68	2 (2–3)	8

Abbreviations: IQR, interquartile range; MIRU, mycobacterial interspersed repetitive unit; N/A, not applicable; SNP, single nucleotide polymorphism; TB, tuberculosis; VNTR, variable number of tandem repeats.

^a May include clinically diagnosed and extrapulmonary TB cases where microbial culture was unavailable.

Using abstracted cluster data, we jointly estimated the reproduction number, R, and the dispersion parameter, k, along with their 95% confidence intervals (Figure 2, Table 3, Web Figure 2). Maximum likelihood estimates of R across all data sets were below the threshold value of 1.0, with a median of (range, 0.10–0.73). Maximum likelihood estimation values of k across all data sets were suggestive of extensive individual heterogeneity (all ), with a median of (range, 0.02–0.48). There was no apparent correlation between SNP thresholds and (Web Figure 3). Among the studies included in this analysis, heterogeneity tended to increase as population TB incidence or per-capita gross domestic product increased (Web Figure 4).

Figure 2.

Full confidence regions for tuberculosis (TB) transmission parameters from studies included in an analysis of Mycobacterium tuberculosis transmission dynamics, 2021. The graph shows maximum likelihood estimates (MLEs) and 95% confidence regions (CRs) for and . Points indicate MLEs, and lines encompass the 95% CRs. A) All TB surveillance; B) drug-resistant TB surveillance. Exact MLEs for R and k and 95% confidence intervals are presented in Table 3.

Table 3.

Estimates of the Basic Reproduction Number (R) and the Dispersion Parameter (k) for Studies Included in an Analysis of Mycobacterium tuberculosis Transmission Dynamics, 2021^a

First Author, Year (Reference No.)
	MLE	95% CI	MLE	95% CI
All TB Surveillance
Alvarez, 2021 (33)	0.73	0.34, 1.55	0.07	0.03, 0.17
Asare, 2020 (34)	0.21	0.18, 0.23	0.15	0.12, 0.19
Guerra-Assunção, 2015 (35)	0.48	0.43, 0.54	0.24	0.19, 0.33
Verza, 2020 (36)	0.29	0.20, 0.42	0.48	0.16, ∞
Walker, 2014 (37)	0.10	0.06, 0.20	0.06	0.03, 0.16
Drug-Resistant TB Surveillance
Chee, 2021 (38)	0.12	0.05, 0.36	0.02	0.01, 0.04
Jiang, 2019 (40)	0.16	0.12, 0.21	0.24	0.12, 0.81
Macedo, 2019 (39)	0.56	0.31, 1.15	0.19	0.06, 0.99
Yang, 2017 (41)	0.20	0.15, 0.27	0.34	0.15, 1.51

Abbreviations: CI, confidence interval; MLE, maximum likelihood estimate; TB, tuberculosis.

^a Full 95% confidence regions are presented in Figure 2.

Although values in all settings suggested a high degree of individual heterogeneity, we found differences between studies when examining how R and k jointly influenced observed population-level transmission. The expected proportion of infectious cases responsible for 80% of observed transmission ranged from 2.0% to 30.8%, with similar relative differences for 50% and 90% of transmission (Table 4, Figure 3). When examining the relationship between R, k, and the probability of observing a large cluster of size Y, there were substantial differences between settings (Figure 4; Web Table 3). For example, the probability of observing an outbreak of at least 50 cases was over 6 orders of magnitude higher in Iqaluit, Canada (33) than in Oxfordshire, United Kingdom (37) (Web Table 3).

Table 4.

Percentages of Infectious Cases Responsible for 50%, 80%, and 90% of Secondary Transmission in an Analysis of Mycobacterium tuberculosis Transmission Dynamics, 2021^a

	Percentage of Secondary Transmission
	50%		80%		90%
First Author, Year (Reference No.)	% of Infectious Cases Responsible for Transmission	95% UI	% of Infectious Cases Responsible for Transmission	95% UI	% of Infectious Cases Responsible for Transmission	95% UI
All TB Surveillance
Alvarez, 2021 (33)	2.4	1.1, 5.4	7.1	3.2, 15.1	10.9	5.0, 22.6
Asare, 2020 (34)	4.8	4.0, 5.9	13.7	11.4, 16.5	20.6	17.2, 24.5
Guerra-Assunção, 2015 (35)	7.2	5.9, 9.2	19.7	16.5, 24.5	28.9	24.5, 35.4
Verza, 2020 (36)	12.0	5.1, 46.0	30.8	14.4, 77.0	43.5	21.6, 88.0
Walker, 2014 (37)	2.1	1.1, 5.1	6.2	3.2, 14.4	9.5	5.0, 21.6
Drug-Resistant TB Surveillance
Chee, 2021 (38)	0.7	0.3, 1.4	2.1	1.1, 4.2	3.4	1.7, 6.5
Jiang, 2019 (40)	7.2	4.0, 16.6	19.7	11.4, 40.1	28.9	17.2, 54.6
Macedo, 2019 (39)	5.9	2.1, 18.5	16.5	6.2, 43.6	24.5	9.5, 58.5
Yang, 2017 (41)	9.4	4.8, 22.6	25.0	13.7, 50.5	36.0	20.6, 65.8

Abbreviations: TB, tuberculosis; UI, uncertainty interval.

^a The full range of estimates is presented in Figure 3.

Figure 3.

Expected proportion of tuberculosis (TB) transmission attributed to a given proportion of infectious cases in an analysis of Mycobacterium tuberculosis transmission dynamics, 2021. The graph shows the expected proportion of transmission (y-axis) for a given percentage of infectious cases (x-axis). Studies are ordered by value. A) All TB surveillance; B) drug-resistant TB surveillance. Exact values for selected proportions and uncertainty intervals are presented in Table 4.

Figure 4.

Relationship between R, k, and the probability of observing a tuberculosis (TB) outbreak of size Y in an analysis of Mycobacterium tuberculosis transmission dynamics, 2021. Inference of both R and k is needed to properly describe the potential for outbreaks (i.e., “superspreading events”). The probabilities of observing a cluster of at least size Y are denoted by contoured lines (darker shading indicates increasing probability, with values denoted on the contours). The graphs show results for arbitrarily chosen cluster sizes (Y) of A) , B) , C) , and D) . Filled circles represent maximum likelihood estimates (MLEs) from all TB surveillance; filled triangles represent MLEs from drug-resistant TB surveillance. Study names have been removed for clarity (see Web Table 3).

When evaluating the likely effect of imperfect surveillance, our combined situational analysis found that inferred estimates of were generally slightly biased upwards and estimates of were biased downwards (Figure 5). This suggests that reported estimates of heterogeneity may be conservative (biased towards homogeneity) and the average cluster size may be underestimated. This finding concurs with our sensitivity analysis evaluating passive surveillance, active case-finding, and censoring individually; all appeared to be associated with an upward bias towards homogeneity (Web Figures 4–7). Parameter inference was most sensitive to passive surveillance (), though it demonstrated only modest bias despite extreme values of (i.e., only 20% of cases observed; Web Figure 6). Estimates of heterogeneity were less sensitive to active case-finding measures (), though they were slightly biased upwards despite an increase in overall case ascertainment (Web Figure 6). Parameter estimates were robust to censoring, demonstrating only a slight upward bias of (Web Figure 7). The total number of clusters in a simulated surveillance system had a predictable loss of accuracy, though little bias in parameter inference (Web Figure 8).

Figure 5.

Situational analysis of model inference to imperfect tuberculosis (TB) surveillance in an analysis of Mycobacterium tuberculosis transmission dynamics, 2021. We simulated 1,000 perfect and counterfactual imperfect surveillance systems, as described in the Methods section of the text, to evaluate model inference. A) “Perfect surveillance” evaluated our cluster-based approach to parameter inference assuming that 100% of cases were observed and served to validate performance of the procedure. B) “Imperfect surveillance” evaluated the robustness of the model given imperfect case ascertainment. The scenario in panel B assumes that only 60% of cases were observed through routine surveillance (), 15% of otherwise undiagnosed cases were ascertained through active case-finding (), and 10% of clusters were ongoing at the time of data collection (censored; ) (see Methods section of text). Diamonds and solid lines represent the median values and interquartile ranges of maximum likelihood estimates (MLEs) from 1,000 simulations, respectively. Numbers in parentheses indicate reference citation numbers for individual studies included in the analysis. Dotted lines have been added for clarity, to show the direction and magnitude of bias from imperfect surveillance.

DISCUSSION

This analysis leveraged available WGS-defined TB transmission cluster data to quantify individual heterogeneity in TB transmission across various global settings. Despite the airborne route of TB transmission, which might be expected to diminish heterogeneity in disease spread, we found substantial individual-level heterogeneity in TB transmission (). Estimates of k were consistently low across settings and concurred with previous studies investigating TB transmission heterogeneity in low-incidence settings (12, 13).

While multiple studies have sought to quantify the reproduction number (R) in TB transmission (42), relatively few investigators have described any measure of individual transmission heterogeneity. To our knowledge, this is the first study to have explored the combined influence of both R and k on TB transmission. Our joint examination of R and k revealed substantial differences in both the expected proportion of TB transmission attributable to infectious cases and the propensity for a large outbreak, suggesting that incorporating individual-level heterogeneity is critical in the accurate representation of observed population-level TB transmission. Inference of such heterogeneity has been a historical challenge in TB epidemiology due to its unique natural history and the lack of available tools for accurately identifying discrete transmission events between individual cases. Our analysis demonstrated that this limitation can be addressed by leveraging observed transmission cluster data to infer without the need for information on the individual sequence of transmission events. This simplified approach is well-suited to facilitation of future parameter inference as the use of WGS becomes more integrated into routine surveillance and provides an alternate measure for comparing transmission across settings that complements traditional indicators (e.g., incidence rate).

Our findings have several public health implications. We provide evidence that, in the settings contributing data to this review, the median effective reproduction number is below 1 (); however, substantial individual-level heterogeneity in this measure allows that many transmission events may nonetheless occur. This type of emergent heterogeneity in declining epidemics is consistent with that predicted in earlier models (43) and should inform future work examining optimal outbreak response (44, 45). This finding suggests that sustained interventions are required even when local transmission is seemingly under control, as observed with other pathogens characterized by extensive individual heterogeneity (46). The impacts of proposed interventions are often evaluated through modeling studies. The majority of TB modeling studies account for heterogeneity at the group level by ascribing transmission potential based on a priori knowledge of quantifiable factors (i.e., age, CD4 cell count, sputum status). We provide empirical estimates of heterogeneity that can be incorporated at the individual level in future models, affording more accurate prediction of observed TB transmission patterns and improving the evaluation of proposed interventions.

TB burden varies widely between countries and has long been recognized as being inversely related to standard of living and per-capita income (1, 47–49). We explored the hypothesis that individual-level heterogeneity in transmission may mediate this association (i.e., through crowded working and living conditions, lack of available medical care) using TB incidence and country-level gross domestic product as a proxy for these factors. However, we were unable to draw conclusions regarding this association given the limited number of data sets included in this analysis. Furthermore, while several included studies were conducted in localized high-incidence subpopulations, the included studies do not capture places where most of the global TB burden exists. Of the countries accounting for two-thirds of global incident TB, only 1 was represented (China) (1). This may be attributed in part to challenges in the population-wide scaleup of WGS and laboratory capacity in high-burden settings. It is likely that transmission dynamics differ substantially in high-burden countries (50). Scaling up genomic surveillance capacity is critically needed to extend these findings and to examine transmission dynamics in these high-burden settings.

Our models used simplifying assumptions to approximate transmission and assumed that secondary cases were independent and identically distributed according to a negative binomial distribution. We drew transmission cluster distributions from reported literature; thus, the clusters were imperfectly defined. We also assumed that clusters originated with a single index case. This process may have over- or underestimated transmission to some unknown degree. Surveillance data inherently represent an unknown proportion of all cases in each population. Therefore, the inferred parameters should be interpreted as a product of mechanisms that are intrinsic to both TB transmission and the observation process itself; the relative contribution of each is unknown. The sample size of included clusters was variable across studies. In addition to different population sizes across settings, such variation may reflect true differences in TB incidence, a larger number of missing cases in some studies, or some unknown combination of both.

Our sensitivity analysis revealed imperfect surveillance, and variable sample size contributed to a slight overestimation of k and underestimation of R, suggesting our estimates may be slightly conservative (biased towards homogeneity and underestimating the true average cluster size). However, we assumed that cases missed by routine surveillance were missing completely at random, an assumption which cannot be verified by the empirical data. In reality, missing cases may be differential by cluster size based on the characteristics of a given transmission network (i.e., an outbreak in a subpopulation with limited access to the public health system) or by population demography. Notably, pediatric TB is paucibacillary by nature and rarely culture-confirmed, biasing genomic surveillance towards adults. Of the 7 studies in this analysis which reported age statistics in the study population, 5 included pediatric cases and 2 explicitly excluded patients under 15 years of age (see Web Table 4). Since pediatric cases are often a consequence of direct transmission (51), the observed distribution may underestimate R and overestimate heterogeneity to some unknown degree. Recent modeling work suggests that inferences about transmission may be biased in poorly reconstructed TB transmission networks, highlighting a critical need for improved methods to evaluate missing cases (52, 53).

The patterns of individual heterogeneity in M. tuberculosis transmission found in this analysis appear similar to those of other infectious diseases characterized by superspreading events, including coronaviruses (severe acute respiratory syndrome (SARS)-associated coronavirus (SARS-CoV), severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), Middle East respiratory syndrome coronavirus (MERS-CoV)), Ebola virus, and monkeypox virus, among others (3, 54–57). Precisely identifying the specific causal events that contribute to a large outbreak is often a challenge, and the methods used in this analysis precluded the ability to identify the mechanisms driving the observed heterogeneity. Comparable observations in other diseases which have similar modes of transmission and have been globally sustained, such as coronavirus disease 2019 (COVID-19) and influenza, suggest that such variation may be partly considered an intrinsic feature of the pathogen (58). Populations may also have fundamental differences motivating such variation, such as demographic (i.e., age structure, household structure), geographic (i.e., urbanicity (urban/rural areas), informal settlements, climate), and social (i.e., migration, homelessness) characteristics, among others, that vary extensively between settings (59–63). Population-specific characterization of TB incidence, both among infectious cases not resulting in secondary cases and among large observed transmission clusters, may help identify predictive correlates of heterogeneity and further inform targeted intervention strategies.

Quantifying individual-level heterogeneity can help inform the development of targeted control measures seeking to maximize reductions in incidence and provide a clearer picture of TB transmission dynamics in a given population. This analysis illustrates that the simple distribution of TB transmission cluster sizes can be used to infer the degree of individual heterogeneity. Future studies, particularly those in underrepresented and higher-incidence settings, can extend the findings of this review and thus deepen our understanding of heterogeneity in the transmission of M. tuberculosis.