Introduction Despite the availability of effective therapy, tuberculosis (TB) continues to infect an estimated one-third of the world's population, to cause disease in 8.8 million people per year, and to kill 1.6 million of those afflicted [1]. Current TB control measures focus on the prompt detection and treatment of those with infectious forms of the disease to prevent further transmission of the organism. Despite the enormous success of this strategy in TB control, the persistence of TB in many parts of the world suggests the need to expand control efforts to identify and address the individual and social determinants of the disease. Since the early part of the 20th century, clinicians have observed an association between diabetes mellitus (DM) and TB, although they were often unable to determine whether DM caused TB or whether TB led to the clinical manifestations of DM [2–6]. Furthermore, these reports did not address the issues of confounding and selection bias. More recently, multiple rigorous epidemiological studies investigating the relationship have demonstrated that DM is indeed positively associated with TB [7–11]. While the investigators suggested that the association reflects the effect of DM on TB, some controversy over the directionality of the association remains due to observations that TB disease induces temporary hyperglycemia, which resolves with treatment [12,13]. A causal link between DM and TB does not bode well for the future, as the global burden of DM is expected to rise from an estimated 180 million prevalent cases currently to a predicted 366 million by 2030 [14]. Experts have raised concerns about the merging epidemics of DM and TB [15–17], especially in low- to middle-income countries, such as India and China, that are experiencing the fastest increase in DM prevalence [18] and the highest burden of TB in the world [19]. Given the public health implications of a causal link between DM and TB, there is a clear need for a systematic assessment of the association in the medical literature. We undertook a systematic review to qualitatively and quantitatively summarize the existing evidence for the association between DM and TB, to examine the heterogeneity underlying the different studies, and to evaluate the methodological quality of the studies. As our aim was to summarize the effect of DM on TB, we did not include studies that investigated the reverse association. Methods We conducted our systematic review according to the guidelines set forth by the Meta-analysis of Observational Studies in Epidemiology (MOOSE) group for reporting of systematic reviews of observational studies (see Text S2 for the MOOSE Checklist) [20]. Search Strategy and Selection Criteria We searched the PubMed database from 1965 to March 2007 and the EMBASE database from 1974 to March 2007 for studies of the association between DM and TB disease; our search strategy is detailed in Box 1. We also hand-searched bibliographies of retrieved papers for additional references and contacted experts in the field for any unpublished studies. Since we speculated that studies that examined the association between DM and TB may not have referred to the term “diabetes” in the title or abstract, we also searched for studies that examined any risk factors for active TB. We restricted our analysis to human studies, and placed no restrictions on language. We included studies if they were peer-reviewed reports of cohort, case-control, or cross-sectional studies that either presented or allowed computation of a quantitative effect estimate of the relationship between DM and active TB and that controlled for possible confounding by age or age groups. We also included studies that compared prevalence or incidence of DM or TB of an observed population to a general population as long as they had performed stratification or standardization by age groups. We excluded studies if they were any of the following: case studies and reviews; studies among children; studies that did not provide effect estimates in odds ratios, rate ratios, or risk ratios, or did not allow the computation of such; studies that did not adjust for age; studies that employed different methods for assessing TB among individuals with and without DM or for assessing DM among TB patients and controls; studies that investigated the reverse association of the impact of TB disease or TB treatment on DM; anonymous reports; and duplicate reports on previously published studies. Box 1. Search Strategy to Identify Observational Studies on the Association of Diabetes and Active Tuberculosis PubMed MeSH terms 1. “tuberculosis” 2. “diabetes mellitus” 3. “cohort studies” OR “case-control studies” OR “cross-sectional studies” OR “epidemiologic studies” OR “follow-up studies” OR “longitudinal studies” OR “prospective studies” OR “retrospective studies” Text terms 1. “tuberculosis” 2. “diabetes” OR “glucose intolerance” OR “glucose tolerance” OR “insulin resistance” 3. “chronic disease(s)” OR “non-communicable disease(s)” 4. “risk factor(s)” Search strings (all inclusive): 1. 1 AND 2 2. 1 AND 3 AND 5 3. 1 AND 3 AND 6 4. 1 AND 3 AND 7 5. 4 AND 5 (for the year preceding 3/2/07 for articles that may not have been assigned MeSH terms) EMBASE Text terms 1. “tuberculosis”, major subject 2. “diabetes mellitus” 3. “risk factor(s)” 4. “observational study” OR “longitudinal study” OR “prospective study” OR “case-control study” OR “cross-sectional study” Search strings (all inclusive): 1. 1 AND 2 2. 1 AND 3 3. 1 AND 4 Data Extraction The two investigators (CJ, MM) independently read the papers and extracted information on the year and country of the study, background TB incidence, study population, study design, number of exposed/unexposed people or cases/controls, definitions and assessment of DM and TB, statistical methods, effect estimates and their standard errors, adjustment and stratification factors, response rates, the timing of diagnosis of DM relative to that of TB, and the potential duplication of data on the same individuals. Differences were resolved by consensus. For the studies that did not directly report the background TB incidence, we obtained data for the closest matching year and state (or country) made available by public databases (WHO global tuberculosis database, http://www.who.int/globalatlas/dataQuery/; CDC Wonder, http://wonder.cdc.gov/TB-v2005.html). Data Analysis We separated the studies by study design and assessed heterogeneity of effect estimates within each group of studies using the Cochrane Q test for heterogeneity [21] and the I2 statistic described by Higgins et al. [22]. We determined the 95% confidence intervals (CIs) for the I2 values using the test-based methods [22]. We performed meta-analysis for computation of a summary estimate only for the study design (i.e., cohort) that did not show significant heterogeneity. Effect estimates of other study designs were not summarized due to significant heterogeneity. For those studies that reported age, sex, race, or region stratum-specific effects, we calculated an overall adjusted effect estimate for the study using the inverse-variance weighting method, then included this summary estimate in the meta-analyses and sensitivity analyses. We decided a priori to use the Dersimonian and Laird random effects method to pool the effect estimates across studies for the meta-analyses, because the underlying true effect of DM would be expected to vary with regard to underlying TB susceptibility and the severity of DM, and because it would yield conservative 95% confidence intervals [23]. In order to identify possible sources of heterogeneity and to assess the effect of study quality on the reported effect estimates, we performed sensitivity analyses in which we compared pooled effect estimates for subgroups categorized by background TB incidence, geographical region, underlying medical conditions of the population under study, and the following quality-associated variables: time of assessment of DM in relation to TB diagnosis, method of DM assessment (self-report or medical records versus laboratory tests), method of TB assessment (microbiologically confirmed versus other), adjustment for important potential confounders, and the potential duplication of data on the same individuals. To determine whether the effect estimates varied significantly by the above-mentioned factors, we performed univariate meta-regressions, in which we regressed the study-specific log-transformed relative risks (RRs) by the variables representing the study characteristics, weighting the studies by the inverse of the sum of within-study and between-study variance for all studies within the comparison. For background TB incidence, we created an ordinal variable, 1 representing < 10/100,000 person-years to 3 representing ≥ 100/100,000 person-years. Coefficients of meta-regression represent differences in log-transformed RRs between the subgroups; we tested the significance of these coefficients by Student t-test, and significance was set at p < 0.10. We considered studies to be of higher quality if they specified that DM be diagnosed prior to the time of TB diagnosis; used blood glucose tests for diagnosis of DM; used a microbiological definition of TB; adjusted for at least age and sex; were cohort, nested case-control, or population-based case-control studies; or did not have the potential for duplication of data. As the average background incidence rate of TB did not exceed 2 per 100 person-years in any of the of the case-control studies that had not employed incidence density sampling, we assumed TB to be sufficiently rare that the odds ratios would estimate the risk ratios [24], and that it would therefore be valid to compute summary RR in the sensitivity analyses regardless of the measure of association and design of the study. We explored possible effect modification by age by examining the three studies that reported results by age groups [7,9,25]. For this analysis, we graphed the stratum-specific estimates in a forest plot, and tested for heterogeneity of the effects within each study by the Q-test and I2 value. We also performed meta-regression within each study in which we regressed the log-transformed RRs by the mid-points of the age-bands. For the unbound age group, ≥ 60 y, we added half the range of the neighboring age-band, or 5 y, to the cutoff. We computed the factor reduction in RR with 10 y increases in age, and reported the p-value for significance of trend. We assessed publication bias using the Begg test and Egger test [26,27]. Statistical procedures were carried out using R version 2.5.1 [28]. 95% CI of the I2 value was computed using the “heterogi” module in STATA version 10 [29]. Results We identified and screened 3,701 papers by titles and abstracts; of these, 3,378 were excluded because they did not study risk factors for TB, were studies among children, were case reports, reviews, or studies of TB treatment outcome (Figure 1). Of the remaining 323 articles, 232 studies were excluded because they did not report on the association between DM and TB, and 56 studies were excluded because they were review articles (12) or ecological studies (2); studied the clinical manifestations of TB in people with diabetes (11); studied the association of DM and TB treatment outcome (6); assessed latent, relapsed, clustered, or drug-resistant TB as the outcome (6); studied the reverse association of the effect of TB on DM (5); had no comparison group (5); were case reports (3); did not give a quantitative effect estimate (3); had collapsed DM and other chronic diseases into a single covariate (2); or was a study that had been reported elsewhere (1). We contacted the authors of four papers that reported including DM in a multivariate analysis but that did not provide the adjusted effect estimate for DM; we included the papers of the two authors who responded and provided these adjusted estimates [30,31]. Further exclusion of studies that did not adjust for age (11), studies that used a general population as the comparison group for TB incidence or DM prevalence without standardization by age (9), and studies that used different methods for ascertaining TB in the people with diabetes and control group (2), left 13 eligible studies. These included three prospective cohort studies [7,30,32], eight case-control studies [8,11,31–37], and two studies for which study design could not be classified as either cohort or case control, as TB case accrual occurred prospectively while the distribution of diabetes in the population was assessed during a different time period after baseline [9,25]. The studies were set in Canada (1), India (1), Mexico (1), Russia (1), South Korea (1), Taiwan (1), the UK (1), and the US (6), and were all reported in English and conducted in the last 15 y. Two of the cohort studies were among renal transplant patients [30,32], and three of the case-control studies were hospital-based or based on discharge records [8,11,35]. The studies are summarized in Table 1. Figure 1 Flow Chart of Literature Search for Studies on the Association between Diabetes Mellitus and Active Tuberculosis Table 1 Summary of the 13 Observational Studies of Association between Diabetes and Active Tuberculosis Included in the Meta-analysis Table 1 Extended. Figure 2 summarizes the adjusted effect estimates of the 13 studies categorized by the study design. We found substantial heterogeneity of effect estimates from studies within each study design; between-study variance accounted for 39% of the total variance among cohort studies, 68% of the total variance among case-control studies, and 99% of the total variance in the remaining two studies. Despite this heterogeneity, the forest plot shows that DM is positively associated with TB regardless of study design, with the exception of the study by Dyck et al. [25]. DM was associated with a 3.11-fold (95% CI 2.27–4.26) increased risk of TB in the cohort studies. Of note, the study conducted within a nontransplant population provided greater weight (63%) to the summary estimate than the other two cohort studies combined. The effect estimates in the remaining studies were heterogeneous and varied from a RR of 0.99 to 7.83. Figure 2 Forest Plot of the 13 Studies That Quantitatively Assessed the Association between Diabetes and Active Tuberculosis by Study Designs Size of the square is proportional to the precision of the study-specific effect estimates, and the bars indicate the corresponding 95% CIs. Arrows indicate that the bars are truncated to fit the plot. The diamond is centered on the summary RR of the cohorts studies, and the width indicates the corresponding 95% CI. *Other: The studies by Ponce-de-Leon et al. [7] and Dyck et al. [25] were not specified as prospective cohort or case-control. TB case accrual occurred prospectively, while the underlying distribution of diabetes was determined during a different time period after baseline. Table 2 shows that there is an increased risk of active TB among people with diabetes regardless of background incidence, study region, or underlying medical conditions in the cohort. In the sensitivity analyses, we noticed that the strength of association increased from a RR of 1.87 to a RR of 3.32 as background TB incidence of the study population increased from < 10/100,000 person-years to ≥ 100/100,000 person-years, but the trend was not significant (trend p = 0.229). Effect estimates were heterogeneous within each category of background TB incidence (I2 = 60%, 98%, and 76% from highest to lowest background TB incidence category). Table 2 Results of Sensitivity Analyses to Identify Sources of Heterogeneity in the Magnitudes of the Association between Diabetes and Active Tuberculosis We also found that the associations of DM and TB in the study populations from Central America [9], Europe [33,37], and Asia [7,30,32] (RRCentralAm = 6.00, RREurope=4.40, RRAsia = 3.11) were higher than those of North American studies [8,11,33,34–36] (RRNA = 1.46) (meta-regression p CentralAm = 0.006, p Europe = 0.004, p Asia = 0.03). Among North American studies, the pooled estimate of the relative risks for Hispanics from two studies [8,11] was higher (RR = 2.69) than that of non-Hispanics from the same study [8] and other North American studies (RR = 1.23) (meta-regression p = 0.060) (Table 2). In general, stratification of the studies by quality-associated variables did not reduce the heterogeneity of effect estimates. Nonetheless, DM remained positively associated with TB in all strata. Studies that explicitly reported that DM was diagnosed prior to TB showed stronger associations (RR = 2.73) [7,31–34] than those that did not establish the temporal order of DM and TB diagnosis (RR = 2.10) [8,9,11,25,30,35–37], although the difference was not significant (meta-regression p = 0.483). Associations were stronger in studies that classified DM exposure through empirical testing (RR = 3.89) [7,9,32,34] rather than medical records (RR = 1.61) (meta-regression p = 0.051) [8,11,25,30,31,33]; and in those that confirmed TB status using microbiological diagnosis (RR = 4.91) [7,9,35,37] than in the studies that did not confirm by microbiological tests (RR = 1.66) (meta-regression p = 0.015) [8,11,25,30–34,36]. Among case-control studies, those that were nested in a clearly identifiable population or were population-based also reported stronger associations (RR = 3.36) [31,33,34,37] than those that used hospital based controls (RR = 1.62) [8,11,37], but the difference was not significant (meta-regression p = 0.321). Studies that had adjusted for smoking showed stronger associations (RR = 4.40) [33,37], while studies in which an individual may have contributed more than one observation to the data revealed weaker associations (RR = 1.62) [8,11]. Although these results suggest that higher-quality studies gave stronger estimates of association, we also found that the association was weaker in studies that adjusted for socioeconomic status (RR = 1.66) (Table 2) [8,11,37]. Figure 3 presents the summary measures of the association between DM and TB by age group based on the data from the three studies that presented age-stratified RRs. The plots from Kim et al. [7] and Ponce-de-Leon et al. [9] demonstrate stronger associations of DM and TB under the age of 40 y and declining RR with increasing age in age groups over 40 y (trend p Kim = 0.014, p Ponce-de-Leon = 0.184). Each 10 y increase in age was associated with a 0.6-fold reduction in magnitude of association in the study by Kim et al. [7]. This trend was not apparent in the study by Dyck et al. (Figure 3) [25]. Figure 3 Forest Plot of Age-Specific Association between Diabetes and Active Tuberculosis from Kim et al. [7], Ponce-de-Leon et al. [9], and Dyck et al. [25] Size of the square is proportional to the precision of the study-specific effect estimates, and the bars indicate 95% CI of the effect estimates. Arrows indicate that the bars are truncated to fit the plot. *Meta-regression: Factor reduction in RR with 10 y increase in age; p-values are given for test of linear trend. HR, hazard ratio. Both the Egger test and Begg test for publication bias were insignificant (p = 0.37, p = 0.14). Discussion Summary of Findings Our meta-analysis shows that DM increases the risk of TB, regardless of different study designs, background TB incidence, or geographic region of the study. The cohort studies reveal that compared with people who do not have diabetes, people with diabetes have an approximately 3-fold risk of developing active TB. Higher increases in risk were seen among younger people, in populations with high background TB incidence, and in non-North American populations. Heterogeneity of strengths of association may reflect true geographic/ethnic differences in severity of DM, transmission dynamics of TB, and the distribution of effect modifiers such as age, or it may be due to differences in study methodology or rigor. Given this heterogeneity of the RR estimates and the fact that all the cohort studies were conducted in Asia, we note that the summary estimate may not be applicable to other populations and study types. While the included studies covered a relatively broad range of geographic areas, there were none from Africa, where TB incidence is high. Nonetheless, a positive association of DM and TB was noted in two African studies [38,39] and several other studies that we excluded from the meta-analysis [10,40–42], as well as in a previous narrative review [43] of the association of DM and TB. Unlike the previous review, our systematic review identified five additional studies that had examined the association of DM and TB, computed a pooled summary estimate among the cohort studies, and determined important sources of heterogeneity through rigorous sensitivity analyses. Public Health Implication With an estimated 180 million people who have diabetes, a figure expected to double by year 2030, it is clear that DM constitutes a substantial contributor to the current and future global burdens of TB. For example, if we assume a RR of 3 and a prevalence of DM in Mexico of 6%, we can conclude that DM accounts for 67% of active TB cases among people with diabetes, and 11% of cases among the entire Mexican population (see Text S1 for the calculation) [44]. The contribution of DM to the burden of TB may be even higher in countries such as India and China where the incidence TB is greater and mean age is lower. In fact, a recent study by Stevenson et al. determined that DM accounts for 80.5% of incident pulmonary TB among people with diabetes, and 14.8% of incident TB in the total population in India [16]. The population-attributable risk for diabetes is comparable to that of HIV/AIDS; while HIV/AIDS is strong risk factor for TB (RRHIV = 6.5–26 [45], approximately 2–9 times greater than the RRDM estimated in this study), it is a less prevalent medical condition (33 million people infected in 2007 [46], approximately 5–6 times less prevalent than DM). Given these figures it may be puzzling to observe a decrease in TB in those areas that have experienced a growing burden of DM. We attribute this observation to negative confounding by factors such as improved nutrition and TB control measures in the areas of increasing DM such as India and China. Were these other factors to remain the same, we would expect to see a TB incidence trend reflecting that of DM in accordance with the positive association. Biological Plausibility Numerous studies have presented convincing biological evidence in support of the causal relationship between DM and impaired host immunity to TB. Studies in animal models have demonstrated that diabetic mice experimentally infected with M. tuberculosis have higher bacterial loads compared to euglycemic mice, regardless of the route of inoculation of M. tuberculosis [47,48]. Compared to euglycemic mice, chronically diabetic mice also had significantly lower production of interferon-γ (IFN-γ) and interleukin-12 (IL-12) and fewer M. tuberculosis antigen (ESAT-6)-responsive T cells early in the course of M. tuberculosis infection, marking a diminished T helper 1 (Th1) adaptive immunity, which plays a crucial role in controlling TB infection [48]. In experimental studies of human plasma cells, high levels of insulin have been shown to promote a decrease in Th1 immunity through a reduction in the Th1 cell to Th2 cell ratio and IFN-γ to IL-4 ratio [49]. Additionally, an ex vivo comparison study of production of Th1 cytokines showed that nonspecific IFN-γ levels were significantly reduced in people with diabetes compared to controls without diabetes [50]. Another study indicated a dose–response relationship; levels of IFN-γ were negatively correlated with levels of HbA1c (a measure of serum glucose levels over time in humans) [51]. Furthermore, neutrophils from people with diabetes had reduced chemotaxis and oxidative killing potential than those of nondiabetic controls [52], and leukocyte bactericidal activity was reduced in people with diabetes, especially those with poor glucose control [53]. Taken together, these studies strongly support the hypothesis that DM directly impairs the innate and adaptive immune responses necessary to counter the proliferation of TB. Limitations There are several potential limitations to this study. Our analysis was based on estimates derived from observational studies that are vulnerable to confounding by variables associated with both DM and TB. To address the issue of potential confounding, we performed a sensitivity analysis in which we reported separate summary estimates for the studies that adjusted for important potential confounders and those that did not. Studies that controlled for socioeconomic status in a multivariable model found that the adjusted effect of DM was reduced, but not eliminated. Crude effect estimates were not provided in two of the larger studies that adjusted for socioeconomic status, thus the direction of bias cannot be determined. The three studies that did report both crude and the adjusted estimates [33,34,37] found that the adjusted RRs for DM were higher. Although we could not exclude the possibility of residual confounding by unmeasured confounders in these observational studies, such as other chronic diseases that often coexist with diabetes, we found that the effect of DM on TB risk persisted even after adjustment for multiple potential confounders that are likely to be correlated with unmeasured factors. Eight of the studies included in this meta-analysis were case-control studies. Control selection strategies included sampling from hospitals, discharge records, department of health records, the general population, and the cohort in which the study was nested. Sampling controls from hospital or discharge records may have introduced a Berkson bias—a selection bias that can occur when both the exposure and the outcome are associated with attendance at a health-care facility from which cases and controls are recruited [54]. Since DM can lead to multiple health problems, the prevalence of DM is likely to be higher among persons attending clinics or being admitted to hospitals than it is in the general population. This bias would be expected to result in an underestimation of the effect of DM on TB, an expectation that was consistent with our finding that studies using hospital-based controls reported lower effect estimates [54]. Other sources of potential bias include misclassification of either exposure or outcome, such as may have occurred in studies that did not employ laboratory tests to diagnose DM or TB. When we restricted our analysis to studies that used glucose tests to determine DM status, we found that effect estimates were higher than in the studies that relied on less-rigorous methods, consistent with our expectation of a bias toward the null among studies that nondifferentially misclassify the exposure. Studies that utilized glucose tests to classify the exposure may also have reported higher RRs of TB among people with diabetes, since they may have identified undiagnosed people with diabetes who remained untreated and therefore may have had higher glycemic levels that those who self-reported their status. Those studies that confirmed TB through microbiological diagnosis also reported stronger associations, suggesting that diabetes may have a stronger impact on smear-positive and thus transmissible forms of TB. Our result underscores the conclusion by Stevenson et al. that DM accounts for a greater proportion of smear-positive TB than of other forms [17]. In short, we found that magnitudes of association varied by the quality of the studies; at the same time, variations may have been influenced by differences in population characteristics that are correlated with quality-associated variables. Another important limitation of our systematic review is that most of the studies we included failed to examine age as an effect modifier of the relationship between DM and TB. The studies by Kim et al. [7] and Ponce-de-Leon et al. [9] found that estimates varied markedly by age, with substantially higher estimates among younger people. This finding may be explained by heterogeneity of the individuals without diabetes between the age groups. Because baseline glucose tolerance is lower in older persons without diabetes, elderly controls may have had an elevated risk of TB compared to younger ones [55], thus reducing the apparent effect of DM. It is possible that younger people with diabetes might have had type I diabetes, a more severe form of diabetes with a juvenile onset; however, because most studies did not distinguish between type I and type II diabetes we cannot conclude whether the effect modification by age would have been due to differences in types of diabetes. Notably, the study by Dyck et al. [25] did not demonstrate this trend in the age-specificity of the effect of DM on TB and in fact showed a negative association among the elderly. The authors of the study note that results may have been biased by differential mortality in the elderly since individuals with diabetes who would have been most at risk for TB would have already died. Moreover, this study also differed from the others in that it relied on medical records rather than laboratory tests to determine DM status, and it had not included DM occurring in the last six of the 16 y during which TB case accrual occurred. Conclusions In summary, we found consistent evidence for an increased risk of TB among people with diabetes despite heterogeneity in study design, geographic area, underlying burden of TB, assessment of exposure and outcome, and control of potential confounders. Data from these human studies are consistent with emerging information on the biological mechanisms by which hyperglycemia may affect the host immune response to TB. Our findings suggest that TB controls programs should consider targeting patients with diabetes for interventions such as active case finding and the treatment of latent TB and, conversely, that efforts to diagnose, detect, and treat DM may have a beneficial impact on TB control. We also recommend further studies investigating how TB risk varies by type, duration, and severity of DM, for a more thorough understanding of the association that could be translated to a clear public health message. Supporting Information Text S1 Calculation of Attributable Risk Fraction of TB among Patients with Diabetes and Population-Attributable Risk Fraction of TB Due to Diabetes (17 KB DOC) Click here for additional data file. Text S2 MOOSE Checklist (61 KB DOC) Click here for additional data file.