In this large UK population based cohort of patients with incident diabetes we found only an overall modest 1.3 fold increased risk of tuberculosis. We found no evidence for higher relative increases in TB rates amongst diabetes patients of different age groups or ethnicities, longer duration of disease, those using insulin or with worse glycaemic control. There was strong evidence for differences amongst diabetes patients with different health care utilisation patterns. The highest risk of TB disease was amongst the group least accessing chronic disease health care.
Comparison with other studies and explanation of findings
Our study is the largest cohort study to date exploring the association between diabetes and TB with individual-level adjustment for important demographic and lifestyle factors. The finding of an overall increased risk of TB in those with diabetes is in agreement with previous published studies and reviews [12–29, 4, 30, 3, 31]. However, in contrast we find only an overall small relative effect from diabetes in our UK population. The most recent systematic review and meta-analysis  included eight case-control studies with odds ratios ranging from 1.16 to 7.83 and a random effects analysis of the three included cohort studies showed a three-fold increased risk of TB with diabetes (relative risk 3.11, 95 % CI 2.27 to 4.26). All three cohort studies included in the meta-analysis were conducted in high TB incidence countries and two used cohorts of renal transplant patients.
Since the most recent systematic review, we are aware of seven more published analytical studies in humans looking at the question of TB risk associated with diabetes, including two case-control [26, 27] and five cohort studies [25, 28, 29, 4, 32] summarised in Additional file 5. Our finding of only a modest increase in risk of TB with diabetes is in agreement with these more recent studies. As study size increases there is a decrease in the estimate for the association between diabetes and TB seen for both cohort and case-control designs, even in higher incidence countries. These differences could be due to publication bias in earlier studies and/or the adjustment for more confounding in later studies. Leung et al.  used a cohort from a Hong Kong community based health program for ≥65 year olds and were able to adjust for demographic and lifestyle factors giving an overall adjusted hazard ratio of 1.77 (1.41 to 2.24). Dobler et al.  used a whole population cohort for Australian citizens and adjusted for demographic and indigenous status and TB incidence in country of birth using census aggregate data for the unexposed general population cohort. They found a 1.4 fold increased risk of TB in those with diabetes (RR 1.48, 1.04 to 2.10). The largest matched case-control study by Leegaard et al.  in Denmark found no overall association between diabetes and TB after adjusting for a range of chronic disease and demographic indicators (OR 1.18, 0.96 to 1.45).
We report a median length of follow-up of 4.4 years which is similar to that in previous studies and on the whole, most of the patients included had reasonably well-controlled diabetes. The largest cohort studies previously reported, Kim et al.  and Dobler et al. , which were unable to adjust for individual confounding, had study periods of two and six years, respectively. It is possible that the risk of TB associated with well-controlled diabetes only becomes manifest over much longer periods and we may possibly have under-estimated long-term risks.
The relatively small effect estimate for TB risk from diabetes found in our study could be due to the ability to adequately control for important individual level confounding from lifestyle and demographic risk factors using the CPRD. Our findings might also be a reflection of a successful primary care service with good chronic disease management and, therefore, reduction in attendant infection complications from diabetes. In support of this hypothesis, we see increased risks for TB disease from diabetes in those with the highest and lowest chronic disease management. Diabetes patients most frequently accessing chronic disease care could represent a group with more uncontrolled disease where general practice teams are seeking to improve diabetes management. The group with the least health care utilisation from that incentivised in UK General Practice for chronic disease management, may have more uncontrolled diabetes and have limited access to primary health care. The latter group might form part of a wider hard to reach group who are at increased risk of TB not only from diabetes but from multiple social risk factors . Of note, our findings show diabetes patients receiving standard rates of chronic disease care, which in part will reflect good diabetes control, are at no increased risk of TB compared against patients without diabetes.
We did not find evidence for an effect of age, duration of diabetes or ethnicity for TB risk in diabetes patients although this could be due to type 2 error as we only had very small numbers of TB cases in some diabetes subgroups and had missing data on important confounders. Although some studies have found evidence for increasing TB risk in younger diabetes patients [12, 15], other authors have not [27, 29]. No previous studies have explored the effect of duration of diabetes diagnosis on risk of TB disease but if diabetes acts to increase the risk of TB infection or increases reactivation of latent disease, it would seem probable that cumulative exposure to diabetes would potentiate these risks.
The previous literature, using a variety of different markers, show mixed results for the effect of diabetes severity on the risk for TB. Leung et al.  stratified by glycaemic status and found those with Hba1c <7 % had no increased risk of TB compared with those without diabetes, in contrast to the subjects with Hba1c ≥7 % who were at 2.5 fold risk (HR 2.56, 1.95 to 3.35). Baker et al.  used the number of diabetes complications to explore the effect of diabetes severity and found that those with treated diabetes and ≥2 complications compared against a group without diabetes had a greater risk of TB (RR 3.45, 1.59 to 7.90). Dobler et al.  explored the effects of insulin use as a marker of severity and found that those using insulin had 2.3-fold risk compared against a general population cohort (RR 2.27, 1.41 to 3.66). In contrast, and coherent with our own data, Leegaard et al.  found no evidence for an association between TB risk and dysglycaemia. Again, this might reflect that diabetes patients managed in UK Primary Care have very well controlled disease, not completely captured by mean Hba1c measurements and, therefore, reduced attendant risks from infection.
We aimed to compare the different types of diabetes for risk of TB. The underlying hypothesis being that type 1 diabetes represents a more severe form of diabetes and thus we might expect that it poses a greater risk for TB infection if the relationship between diabetes and TB risk is causal. Only the previous study by Leegaard et al  defined and explored the risk of TB for a group with type 1 diabetes. They classified patients <30-years old using insulin monotherapy and never using oral antidiabetes medications as having type 1 diabetes. Our classification differed in that we defined our type 1 cohort using incentivised diagnostic codes and additional demographic factors to age and insulin prescriptions. Leegaard et al had very small numbers of patients classified with type 1 diabetes, only three amongst their TB cases and the adjusted TB risk estimate reflected the imprecision (OR 2.59, 0.44–15.29). Similarly, we found only one case of TB amongst our group of patients classified with type 1 diabetes and, thus, we were unable to explore the effects of type 1 diabetes further. Under-ascertainment of TB in type 1 diabetes patients within CPRD is a possible cause of our finding only one case of TB in this group. This might be due to those with type 1 diabetes mainly receiving their care in hospital out-patients clinics and notification of TB diagnoses not being returned to general practice. If a large number of cases of TB in the UK are due to reactivation of latent disease from those born in high TB burden countries [2, 1], it might be that incidence of type 1 diabetes in these populations is low, as supported by global incidence studies  or that these patients suffer competing risks before possible reactivation of TB infection.
Current UK guidelines advise considering treatment for latent TB infection in certain groups of adults where active disease has been ruled out but they show signs of TB infection with Mantoux positivity (≥6 mm) and without prior Bacillus Calmette-Guérin (BCG) vaccination, or strong Mantoux positivity (≥15 mm) or interferon-gamma release assay (IGRA) positive and with prior BCG vaccination . There is no specific guidance for patients with diabetes at present.
Strengths and limitations
To our knowledge, this is the largest cohort study to date exploring the association between diabetes and TB in a general population which was able to adjust for important individual level confounding demographic, socioeconomic and lifestyle factors. By using time-updated exposure status, where previous unexposed patients could later develop incident diabetes and join the exposed cohort, we could study time-related phenomena but our design also allowed comparison between more similar groups ensuring we could explore the role of diabetes with reduced confounding from unmeasured social and health-related risk factors. Previous literature describe excluding people with prior diagnoses of tuberculosis but we included this group with a suitable time-lapse as they will be amongst the highest risk groups for TB disease in the UK so our findings are more applicable to the population of interest.
Our study explores effects of important patient characteristics such as age and ethnicity and aspects of the risk factor of diabetes such as duration and severity, which have not been previously explored within one cohort. As far as we are aware this is the only study to look at how TB risk for diabetes patients varies with consultation patterns including receipt of chronic disease health care and shows how UK General Practice systems are able to identify different risk groups. The study is based within the General Practice population using routinely collected clinical data so reducing the likelihood of significant selection bias. In the UK the majority of diabetes care occurs in the primary care setting making our cohort of diabetes patients very inclusive. Where patients receive specialist diabetes care in the face of more challenging disease control, their primary care record is equally important and contemporaneous as primary care co-ordinates ongoing management, such as retinal screening and immunisation. Our study is pertinent to any UK policy seeking to identify high-risk groups in primary care suitable for latent TB infection screening using such routine electronic health data.
There is the potential for misclassification of diabetes and TB within our study as we used routinely collected clinical data without validation from consultation free text or hospital correspondence. We expect misclassification to be minimal as diabetes is an indicator condition within UK Primary Care; thus, practices are incentivised to maintain accurate diabetes patient registers. Seventy percent of our cohort with diabetes had confirmatory prescription data for anti-diabetic medications and the 30 % having no specific therapy but other recorded indicators is in agreement with previous studies . We generated a specific TB Read code list to try to avoid the inclusion of non-mycobacterial disease. Although TB can present non-specifically initially we would expect patients to seek medical attention in a setting with free access to health care. Diagnoses of TB made in secondary care are highly likely to be communicated to the General Practitioner due to the public health risk of this communicable disease, the risk of serious side-effects from antituberculous therapy and possible treatment interactions with medications prescribed in primary care. We found an incidence of TB in our unexposed cohort that was equal to that reported as the UK TB incidence for 2012 (13.9 cases per 100,000 population) , therefore supporting that we reliably identified cases of TB. Any misclassification of our outcome is likely to be non-differential as unexposed patients could join the exposed group during follow-up and “inactive” controls were excluded who are more likely to be misclassified as unexposed. Furthermore, there is no current UK guideline advising screening of patients with diabetes for TB and broader awareness for the association between the two conditions was only recently raised , thus any bias will tend towards underestimating associations.
There is a risk of reverse causality with diagnoses made close together and with a condition such as TB which may initially present with non-specific symptoms and can itself cause a reactive hyperglycaemia associated with acute infection . TB diagnosis and treatment delays can be as long as three months in high-income settings [39, 40] but potential delays in diagnosis of diabetes far exceed this with estimates between four to seven years . Thus, it is far more likely in the situation where a diagnosis of TB closely follows a diagnosis of diabetes that the condition of diabetes or chronic hyperglycaemia has been ongoing for some time previous. Using similar considerations, we did not introduce a minimum follow-up time after diabetes diagnosis. There is no current evidence-base to guide us in considering a mechanistically plausible minimum length of follow-up within which TB risk increases and introducing an arbitrary follow-up time will only produce survivor bias.
Information on country of origin is not routinely recorded within the primary care clinical record and may confound the association between diabetes and TB. Lifestyle factors predisposing a person to diabetes associated with the country of origin are likely superseded by those associated with residence in the UK and remaining lifestyle factors or genetic predisposition could be crudely captured by ethnicity descriptors. However, there is likely residual confounding especially as ethnicity was only broadly described within our study. We did not identify and adjust for HIV status within our study as this is recognised to be under-reported in General Practice records . HIV and its treatment have been associated with metabolic derangement including diabetes  and HIV is a strong risk factor for TB disease [44, 45]. We did not identify and exclude patients with recorded HIV as this may have differentially affected TB patients with diabetes as we expected these patients might be more likely to share a diagnosis of HIV with their General Practitioner due to the risk of multiple drug interactions between their HIV and diabetes medications. We did not adjust for BCG vaccination status. The UK immunisation program recommends childhood BCG immunisation for those at higher risk of TB . The protection offered by BCG immunisation wanes with time , and is likely to offer little protection to the age groups predominantly at risk of TB associated with diabetes.
We performed multiple imputation for missing data on important covariates but there is likely residual confounding from non-differential misclassification of these data. Our main findings were robust to the inclusion of imputed missing data. There were fewer missing data for patients with diabetes, which is to be expected as these represent a group that are better characterised within routine primary care data. Missing data produce loss of study power, despite multiple imputation methods, which calculate standard errors reflecting the increased uncertainty. The study also had insufficient power, risking type 2 error, for the investigation of some interactions as we had small numbers of outcome data in some of our subgroup analyses. Thus, larger well-characterised cohorts in higher TB incidence settings would be needed to study these possible interactions.