Serum Bilirubin and Risk of Respiratory Disease and Death FREE

Serum total bilirubin is routinely measured in the primary care setting to identify hepatobiliary and hemolytic diseases. After excluding disease, moderately increased levels (>1 mg/dL) generally indicate benign hereditary hyperbilirubinemia (Gilbert syndrome). This common condition is caused by a deficiency in the liver enzyme uridine diphosphate–glucuronosyltransferase 1 (UGT1A1), which converts insoluble bilirubin to a form suitable for renal and biliary excretion.¹

Although bilirubin levels are highly heritable and genetic variation of UGT1A1 explains a large proportion of the variability,² enzymes involved with bilirubin production from heme, including heme oxygenase, may also exert an influence.³ Elevated heme oxygenase activity, caused by genetic variation or environmental factors, appears beneficial for respiratory health.⁴ Part of this effect is purportedly due to the powerful cytoprotective properties of bilirubin, including antioxidant, anti-inflammatory, and antiproliferative effects.⁵ Experimental studies using animal models support a protective effect of increased bilirubin against respiratory injury by environmental stressors.⁵ The epidemiological relationship between bilirubin level and the risk of respiratory disease is not well characterized.

We therefore examined the association between serum bilirubin levels and the incidence of chronic obstructive pulmonary disease (COPD), lung cancer, and all-cause mortality in a large population-based cohort of primary care patients from the United Kingdom.

We used information from a cohort study of UK primary care data recorded between January 1, 1988, and December 31, 2008. The Health Improvement Network (THIN) database includes anonymized general practice records on more than 7 million patients from the United Kingdom. It is one of the largest national collections of primary care data and is broadly representative of the general practice population in terms of demographics and consultation behavior.⁶ Clinical diagnoses recorded by general practitioners (GPs) have recently been shown to be accurate compared with other reliable sources.⁷ Data on diagnoses, interventions, symptoms, and referrals to secondary care are electronically recorded as Read Codes, a hierarchical coding system used in UK primary care.⁸ All prescriptions are also recorded in therapy records. In addition, the computer system contains information on variables such as height, weight, blood pressure, smoking status, and laboratory test results. The database provides the Townsend score measure of deprivation, a composite measure of social deprivation in quintiles (owner occupation, overcrowding, car ownership, and unemployment). It is based on patient postal code and linked to UK census data from 2001 for approximately 150 households in that postal area.

Bilirubin levels tend to stabilize following adolescence⁹; therefore, we selected patients aged 20 years or older with at least 1 bilirubin test recorded on the same day as a complete blood count (defined as white blood cell, hemoglobin, and platelet counts) and liver enzyme assay (defined as a minimum of alkaline phosphatase and alanine aminotransferase). Details of gamma-glutamyltransferase and aspartate aminotransferase test results were also extracted if simultaneously recorded.

Patients with abnormal liver enzyme assays or blood counts, defined using reference ranges, were excluded (eTable 1). Patients were also excluded if they had Read Codes corresponding to acquired hemolytic disorders or hepatobiliary disease (including malignancy), gallstones, cirrhosis, hepatitis, or alcoholic liver disease before or up to 2 years after the date of the bilirubin measurement. Patients with diagnoses at any date in their medical history of inherited disorders causing abnormally high bilirubin levels, including sickle cell anemia and severe hereditary hyperbilirubinemias (eg, Crigler-Najjar syndrome), were excluded from the cohort. A small number of patients (n = 2319) had bilirubin fractions recorded (direct/indirect) and were excluded if direct fractions exceeded 30% of total bilirubin levels as indicative of hepatobiliary disease.

Exposure . Serum total bilirubin level in patients without evidence of liver or hemolytic disease was the exposure variable. Based on the observed frequency distributions, published reference ranges, and levels observed in healthy individuals with the genotype underlying Gilbert syndrome (approximately 11% of the population),¹⁰ patients with bilirubin values below 0.18 mg/dL (3 μmol/L) for both sexes and exceeding 2.34 mg/dL (40 μmol/L) for men and 1.75 mg/dL (30 μmol/L) for women were excluded. The top cutoffs were selected to approximate the upper standard deviation seen across studies reporting associations between the common UGT1A1 genotypes and bilirubin levels by sex (eTable 2). Intraindividual variability in bilirubin is reported to be low¹; therefore, we used the value from the first recorded test to approximate the exposure throughout adulthood.

Outcomes. Lists of Read Codes for COPD and lung cancer were developed using previously reported guidelines.¹¹ Incident cases were defined using published methods.¹² In brief, disease incidence rates in newly registered patients were plotted over 500 days. Initially, there is a spike in recording rates when GPs are taking disease histories in newly registered patients, and incident cases were defined as those that occurred after the recording rate had stabilized. This was 50 days for lung cancer and 80 days for COPD. All-cause mortality was also selected as an outcome because moderately higher bilirubin could protect against a range of age-related diseases. Mortality was ascertained by the presence of a date of death and all cases were considered to be incident. Patients with the outcomes of interest anytime before or within 90 days of the date of the bilirubin test were excluded from the analyses.

Covariates. Variables previously shown to be associated with bilirubin levels were selected for investigation as potential confounding factors.^13- 15 Hence, we extracted information on sex, social deprivation score, age, smoking status, alcohol intake, body mass index (BMI), and systolic blood pressure for the primary analysis.

THIN data are generated via entry into a single software system (Vision, In Practice Systems, London, England) by primary care physicians. Most smoking data are recorded as “never,” “current,” or “ex.” A recent study found that the proportions of current smokers recorded in THIN data was similar to that in the Health Survey for England, in which self-reported smoking status has been biologically validated using cotinine.¹⁶ Alcohol consumption was grouped into 3 categories: heavy (≥336 g/wk or Read Codes indicating “heavy” or “very heavy” intake), light to moderate (<336 g/wk or Read Codes specifying “moderate,” “occasional,” or “light” drinking), and current nondrinkers. If there were multiple entries per patient for any of these variables, the one recorded closest to the bilirubin test date was used. The median time between the bilirubin measure and recording of the various covariates was as follows: smoking, 128 days (interquartile range [IQR], 16-397 days); systolic blood pressure, 21 days (IQR, 3-163 days); BMI, 184 days (IQR, 16-688 days); and alcohol intake, 347 days (IQR, 77-1634 days).

Each GP practice in THIN has a date that indicates when the practice achieved a standardized mortality rate similar to the UK population after accounting for age and sex characteristics.¹⁷ This “acceptable mortality recording” (AMR) date is also a good proxy for when GPs were using computer systems to record data adequately. Patients entered the cohort at the latest date of when they registered with the GP or at the AMR date. Patients were followed up from this date until they experienced an outcome of interest, died, or transferred to a different GP practice; the practice stopped contributing data to THIN; or the study period ended. Separate censoring was used for each of the 3 outcomes. Patients with less than 6 months of follow-up were excluded from the cohort.

Potential confounders were identified using univariate regression analysis with log-transformed bilirubin levels as the dependent variable. Variables with P < .10 or with strong supporting evidence from the literature were considered to be potential confounders. Regression analyses for COPD, lung cancer, and death were conducted with missing data for potential categorical confounders treated as separate categories. Rather than categorize continuous covariates and include missing categories, which may cause underadjustment,¹⁸ patients with missing data on BMI and systolic blood pressure were excluded from the regression analyses. We used Poisson regression to model the association between bilirubin levels and outcomes. This is a type of survival analysis for count data and implicitly controls for the length of follow-up. Goodness of fit to the Poisson model was assessed using the deviance statistic as a test for overdispersion.¹⁹ Estimates of the association between 0.1-mg/dL (1.7-μmol/L) increases in bilirubin and disease incidence were first calculated assuming a simple linear relationship followed by linear spline interpolation.²⁰ Linear splines are a series of linear functions that intersect at a “knot,” in this case a value of bilirubin. Unlike categorization, this method can flexibly capture dose-response relationships while still using the full information available in data. Most healthy persons with bilirubin levels greater than 1 mg/dL are homozygous for a gene promoter variant termed UGT1A1 *28 (rs8175347) (eTable 2), which causes reduced gene expression and underlies the right-skewed distribution of bilirubin in healthy populations. We introduced a knot at 1 mg/dL to examine whether there was a change in the relationship before or beyond this level. The likelihood ratio test was used to compare linear with linear spline models and identify the model that was the best fit with the data.

The general practice was included in the models as a random effect to account for data clustering. The presence of any interaction effects between smoking status and bilirubin levels was examined by comparing models with and without an interaction term between smoking and bilirubin using the likelihood ratio test. The effect of including smoking intensity and duration in the regression models was investigated in a subgroup of patients for whom these data were recorded or were possible to estimate from smoking history during follow-up.

All analyses were stratified by sex due to differences in both the means and standard deviations of bilirubin levels reported previously and were performed using Stata software, version 11 (Stata Corp, College Station, Texas). P < .05 was considered statistically significant and testing was 2-sided.

The THIN scheme for obtaining and providing anonymous patient data to researchers was approved by the National Health Service South-East Multi-center Research Ethics Committee (MREC) in 2002. The current study was reviewed and approved by the Cambridge MREC.

The total number of adult patients with simultaneous serum bilirubin levels, blood counts, and liver enzymes recorded in THIN was 1 096 700. After excluding patients with evidence of hepatobiliary or hemolytic disease, 504 206 patients from 371 practices were eligible for inclusion, with a median follow-up of 8 years (Table 1). At the end of the study period (December 31, 2008), 23 122 patients (5%) had died and 61 795 (12%) had transferred to a different general practice. For 13 093 patients (3%), the most recent data collection was before January 1, 2008.

Bilirubin levels in this cohort were right-skewed (Figure), with median levels of 0.64 mg/dL (IQR, 0.47-0.88 mg/dL) in men and 0.53 mg/dL (IQR, 0.41-0.70 mg/dL) in women (Table 1). A subset of 188 736 patients had at least 2 bilirubin measurements with a median difference of ±0.12 mg/dL (IQR, ±1.70-0.23 mg/dL). The percentage of variability in bilirubin levels explained by practice-level variation in the multivariable model was statistically significant but low at 6% (P < .001). Factors significantly associated with higher bilirubin levels in men were younger age, nonsmoking status, moderate or high alcohol consumption, lower BMI, higher systolic blood pressure, and lower levels of social deprivation (Table 2). Similar associations were observed for women (Table 2). In the subset of patients who were current smokers during follow-up (n = 160 502), both intensity and duration of smoking were inversely associated with bilirubin levels (eTable 3 and eAppendix).

Outcome incidence rates per 10 000 person-years were 2.5 for lung cancer, 11.9 for COPD, and 42.5 for mortality. Approximately inverse linear associations with bilirubin were observed for the crude incidence of all outcomes across both sexes (Table 3). The incidence rate difference per 10 000 person-years in men for the first (0.18-0.34 mg/dL) compared with the fifth (0.58-0.63 mg/dL) decile categories of bilirubin was 2.1 (95% confidence interval [CI], 0.9-3.2) for lung cancer, 5.1 (95% CI, 2.6-7.7) for COPD, and 13.2 (95% CI, 9.1-17.3) for mortality. For women, the incidence rate difference per 10 000 person-years for the first (0.18-0.28 mg/dL) compared with the fifth (0.47-0.52 mg/dL) decile categories of bilirubin was 1.4 (95% CI, 0.7-2.1) for lung cancer, 2.6 (95% CI, 0.8-4.4) for COPD, and 11.7 (95% CI, 8.5-14.9) for mortality.

The unadjusted associations between bilirubin levels and outcomes were all statistically significant (Table 4). Adjusting for other factors slightly weakened the relationships, but these remained significant (Table 4). The linear spline model was a significantly better fit for the mortality outcome than assuming a simple linear relationship (log-likelihood test, P < .001) (Table 4). There was no evidence of interactions between bilirubin and smoking status for any of the outcomes (log-likelihood test, P = .05). The associations remained significant in a range of post hoc sensitivity and subgroup analyses including adjustment for comorbidity, drug prescriptions, local pollution levels, total cholesterol level, and smoking intensity and duration (eTable 4, eTable 5, eTable 6, eTable 7, and eAppendix). Increasing the minimum time between the bilirubin test and cohort exit from 90 days to 5 years also had a negligible effect on the direction of the associations (eTable 4, eTable 5, eTable 6, eTable 7, and eAppendix).

To our knowledge, this is the largest longitudinal analysis of the association between bilirubin levels and lung cancer and the first to examine the relationship with COPD. After accounting for other important health indicators, there was an inverse association between bilirubin levels and the incidences of lung cancer, COPD, and all-cause mortality. Whereas earlier studies have lacked statistical power to demonstrate associations for clinical outcomes in women,^13,21 we were able to show that the relationships were similar for both sexes.

Based on our findings, bilirubin levels within the normal range appear to capture information about patients that may reflect a combination of environmental and genetically determined susceptibility to respiratory diseases.⁵ The strength of the associations seemed to weaken above 1 mg/dL for all outcomes, although significantly so only for mortality, which could reflect a real dose-response relationship with only a marginal physiological benefit beyond levels typically associated with Gilbert syndrome. However, the slight increase in mortality rates for women beyond 1 mg/dL could also indicate inclusion of some hepatobiliary disease cases.

The genetic variation of UGT1A1 underlying Gilbert syndrome was found to be strongly protective against cardiovascular disease in the Framingham Offspring Study, which supports a protective effect of moderately increased bilirubin.^10,22 Similar assessment of the association among UGT1A1 variants, bilirubin levels, and respiratory outcomes in large prospective cohort studies will help determine whether the observed inverse associations with COPD and lung cancer reflect causal relationships. Such research may also help inform the evolutionary theories on the high prevalence of alleles conferring mild hyperbilirubinemia across human populations.^3,23- 24

The major strength of this study is the sample size that enables examination of outcomes separately for men and women. The longitudinal analysis means that the results are less susceptible to reverse causality compared with cross-sectional analyses. Although the cohort was selected based on records of laboratory testing, the patient characteristics were broadly similar to all the patients within the THIN data set, except for being slightly older. There is no clear reason to believe the results would differ for the entire UK population.

Although bilirubin levels are strongly genetic² and there are large amounts of data supporting a cytoprotective function,^5,10 some level of residual confounding by unmeasured environmental exposures could also be reflected in the inverse associations with disease outcomes. Consistent with earlier studies,^13- 14,25 current smoking status, duration of smoking, higher BMI, and higher levels of social deprivation were all associated with moderately lower levels of bilirubin. Although it would have been interesting to examine the association between bilirubin and race/ethnicity, only a small proportion of patients have this information recorded in their primary care records. Although lower bilirubin levels have been reported for African Americans,²⁶ the proportion of the UK population defining themselves as “black” or “black British” is only 2%²⁷; therefore, race/ethnicity is unlikely to have had a major confounding effect. The magnitude of the association between bilirubin and smoking variables was similar to a large cohort study of Europeans.^13,25 This study also identified an inverse association between bilirubin and smoking intensity that was not statistically significant in the European cohort.¹³ Bilirubin in human plasma appears to be depleted on exposure to reactive oxidative species such as those found in cigarette smoke,²⁸ which could explain these findings.¹³ Thus, passive smoking or cigarette type (eg, unfiltered) could have slightly confounded the results, but this seems unlikely given that adjusting for smoking intensity and duration had almost no effect on the associations. It is possible that adjustment for social deprivation had already removed some of these confounding effects. Although we cannot infer causality using these epidemiological data, bilirubin levels appear to capture some objective and quantifiable information about a patient's risk of respiratory disease independent of his or her self-reported smoking status. The study of plasma molecules as objective markers for environmental exposures that are difficult to measure and quantify in routine health care is a field of research interest.^29- 30

Epidemiologic analyses of bilirubin levels have reported independent inverse associations with cardiovascular disease, cancer, and mortality.^13- 14,21 Studies of cardiovascular disease have reported risk rate ratios per 0.1-mg/dL increase in bilirubin of 0.84 (95% CI, 0.73-0.97)³¹ and 0.91 (95% CI, 0.84-0.96).¹⁰ The only longitudinal study examining bilirubin and respiratory outcomes was an analysis of 10 000 Belgian individuals with a mean follow-up of 10 years.¹³ The incidence of lung cancer mortality in men was about 60% lower in the highest bilirubin category (>0.6 mg/dL) compared with the lowest category (<0.2 mg/dL) after adjusting for age, smoking status, education level, BMI, total cholesterol, and dietary factors.¹³ These results for men seem similar in magnitude to our data for lung cancer but were not statistically significant in the Belgian study, possibly due to the small number of events (n = 96). Lung cancer was not examined in women, presumably because of the low number of events. Adjusted all-cause mortality rates were also similar to our data at about 27% lower for the Belgian men and 13% lower for women in the highest bilirubin categories within the normal range compared with the lowest, but this was statistically significant only for men.

Moderately higher levels of bilirubin within the range considered normal were associated with reduced risk of respiratory disease and all-cause mortality. Further research is needed to investigate causal associations between bilirubin levels and respiratory outcomes. A fuller understanding of these mechanisms may lead to the potential use of targeted clinical treatments that mildly suppress UGT1A1 activity and moderately increase bilirubin levels.