Aspirin, Statins, Non-aspirin NSAIDs, Metformin, and the Risk of Biliary Cancer: A Swedish Population-Based Cohort Study

Biliary tract cancers (BTC), the second most common type of hepatobiliary cancer, include cholangiocarcinoma and gallbladder cancer. Cholangiocarcinomas comprise intrahepatic cholangiocarcinoma (iCCA), perihilar cholangiocarcinoma (pCCA), and distal cholangiocarcinoma (dCCA). The latter two are designated extrahepatic cholangiocarcinoma (eCCA). Recent estimates from the Global Burden of Disease Study are 184,000 incident BTC cases worldwide in 2016, with 108,000 (59%) in women and 76,000 (41%) in men (1). Only 10% of patients with BTC present with early stage disease amenable to surgical resection. Almost all patients with late stage disease will progress on systemic treatment. Consequently, the median survival with advanced cholangiocarcinoma is 11 to 12 months (2). With the increase in global BTC incidence, its associated morbidity and mortality, and limited therapies, there is increased interest in strategies for prevention of BTC.

A number of FDA-approved drugs have been identified as chemopreventive against cancer, including low-dose aspirin, statins, NSAIDs, and metformin (3–7). The US Preventive Services Task Force recommends low-dose aspirin for the primary prevention of cardiovascular disease (CVD) and colorectal cancer in adults aged 50 to 59 years with a 10% or greater 10-year CVD risk who are not at increased risk for bleeding, have a life expectancy of at least 10 years, and are willing to take low-dose aspirin for at least 10 years. The recommendation for adults 60 to 69 years should be individualized. However, for adults 70 and older or younger than age 50, the evidence is insufficient to make a recommendation (8). The recommendation in adults 50 to 59 years has generated interest in the prevention of other cancers, including BTC. Case–control studies have found aspirin decreases cholangiocarcinoma and gallbladder cancer risk by 66% and 63%, respectively (9, 10). However, population-based cohort studies are needed for validation.

Given the incidence of pCCA is higher in men than in women, while the incidence of gallbladder cancer is higher in women than men, it is possible that the associations of chemopreventive agents with risk of BTC and its subtypes vary by sex (11, 12).

We conducted a population-based cohort study of Swedish adults to address the following key questions: (i) Is the use of low-dose aspirin, statins, non-aspirin NSAIDs, or metformin associated with lower risk of BTC or its subtypes iCCA, eCCA, or gallbladder cancer? (ii) Is combined use of low-dose aspirin and statins associated with a lower risk? (iii) Are the associations the same in women and men?

Due to restrictions established by National Board of Health and Welfare, access to the entire Swedish population was infeasible. To delineate a cohort as representative as possible of the population, Swedish residents who received at least one prescription of commonly prescribed drugs (i.e., sex hormones, drugs for peptic ulcers and gastroesophageal reflux disease, acetylsalicylic acid, NSAIDs, HMG CoA reductase inhibitors, drugs affecting bone structure and mineralization, and antibiotics) between July 1, 2005 and December 31, 2012 were identified through the Swedish Prescribed Drug Registry using the Anatomical Therapeutic Chemical Classification codes (Supplementary Table S1). Subjects who did not receive any commonly prescribed drugs were excluded. This approach captured 85% of the entire Swedish population (13). Patients younger than 18 years of age and those who had a cancer diagnosis other than nonmelanoma skin cancer, assessed at the beginning of the study period, were excluded. This cohort has been extensively described elsewhere (13–16).

Factors associated with increased risk of BTC were abstracted from the Swedish National Patient Registry using International Classification of Diseases (ICD) 7–10 codes (Supplementary Table S2). These included cholangitis–including primary sclerosing cholangitis (PSC), inflammatory bowel disease (IBD), gallstone disease, cirrhosis, hepatitis B virus (HBV) and hepatitis B virus (HCV), as well as other less consistently associated factors: autoimmune diseases (celiac disease, hypothyroidism, and rheumatoid arthritis), CVD and metabolic diseases (obesity, diabetes and hyperlipidemia), and proxies for lifestyle habits (smoking-related diseases and alcohol-related diseases). The Inpatient and Outpatient Registries only include discharge/outpatient diagnoses. Therefore, patients hospitalized for reasons other than diabetes were not always recorded as diabetic. To overcome this limitation, metformin use was employed as a proxy to identify these patients through the Prescribed Drug Registry.

Exposures of interest included low-dose aspirin, statins, non-aspirin NSAIDs, and metformin. Drug consumption was expressed as the defined daily dosage (DDD), the assumed average maintenance dose per day for a drug used for its main indication in adults. Estimated use was based on the average DDD per package (17). Individuals were considered exposed if they accumulated a DDD equivalent to 6 months of use or greater.

Low-dose aspirin was defined as use of 75 mg of aspirin daily. Those who used higher doses (n = 8,957, 0.16% of entire cohort and 0.89% of the low-dose aspirin group) represented less than 1% of the low-dose aspirin group. Consequently, they were excluded from the analyses.

We followed patients from July 1, 2005 until BTC diagnosis, death, or date of censoring on December 31, 2012, whichever occurred first. Outcomes were identified from the Swedish Cancer Registry using ICD-10 codes (Supplementary Table S3). Subjects developing other malignancies after study onset were censored at cancer diagnosis.

Cox proportional hazard models, with age as the time-scale (18), were used to estimate BTC risk associated with exposure to: (i) low-dose aspirin, (ii) statins, (iii) non-aspirin NSAIDs, and (iv) metformin, compared with subjects of the same age unexposed to the medication. To account for the time-varying nature of medication use, exposure to aspirin, statins, and/or NSAIDs was considered a time-dependent covariate. Since the definition of diabetes was based partly on metformin use, we did not model metformin use as a time-dependent covariate. All models were adjusted for BTC risk factors, as time-fixed covariates (18). The covariates were cholangitis–including PSC, IBD, gallstone disease, cirrhosis, HBV and HCV, as well as other less consistently associated factors: autoimmune diseases (celiac disease, hypothyroidism, and rheumatoid arthritis), CVD and metabolic diseases (obesity, diabetes and hyperlipidemia), and proxies for lifestyle habits (smoking-related diseases and alcohol-related diseases).

Initially, we analyzed use of a single drug, without regard to use of the other drugs, with subgroup analyses stratified by BTC subtype and sex. Subsequent analysis combined modeling of drugs that were found to decrease BTC risk in the initial analyses (i.e., low-dose aspirin and statins).

To account for inherent differences that are likely to separate users from nonusers and that might introduce bias, we repeated the analysis using inverse probability weights (IPW). We estimated each subject's probability of exposure to the medication of interest conditional on the observed confounders. Normalized weights were then calculated as the inverse of the estimated conditional probabilities of exposure. Checks of positivity, covariate balance, and overall distribution showed the weights to be well behaved.

We examined the differences between women and men using a test for interaction between medication use and sex. To address reverse causation, a landmark analysis was performed that excludes the first year of follow-up. We also analyzed variation in the association of drug use and BTC risk during follow-up.

Analyses were performed using SAS, version 9.4. This study was approved by the Stockholm Regional Ethical Review Board (2014/1291–31/4). It was exempt from human subjects review by the Committee on Human Research at Mayo Clinic and at Karolinska Institutet because the study dataset did not contain identifiers. The exception included a waiver of the requirement for informed consent of participating subjects.

After excluding individuals younger than 18 years and individuals with cancers other than nonmelanoma skin cancer, 5,760,482 subjects were included. Eighteen percent used low-dose aspirin, 17% statins, 10% both low-dose aspirin and statins, 58% non-aspirin NSAIDs, and 6% metformin (Fig. 1). The mean age was 48.8 years (SD 18.1) and 55% were women (Table 1 and Supplementary Table S4). During up to 7.5 years of follow-up, 2,160 (0.04%) subjects developed BTC, with 609 iCCA, 543 eCCA, and 1,008 gallbladder cancer. Event counts per exposure stratum and person-time are presented in Supplementary Table S5.

Table 2 shows the results of the multivariable-adjusted Cox models described in detail below.

Use of low-dose aspirin alone was not associated with risk of BTC [adjusted hazard ratio (HR), 0.93; 95% confidence interval (CI), 0.81–1.07]. Statins were associated with a lower risk of BTC (HR, 0.66; 95% CI, 0.56–0.78). Combined use of low-dose aspirin and statins was not associated with lower risk than statins alone (HR, 0.77; 95% CI, 0.66–0.89; Fig.2). The associations for low-dose aspirin did not change with IPW analysis. With IPW, statins were associated with a lower risk (IPW HR, 0.50; 95% CI, 0.41–0.61) Risk estimates for combined use of low-dose aspirin and statins did not change after performing IPW analysis (Supplementary Fig. S1).

A test of interaction between medication use and sex revealed there were no significant differences between women and men. The P values for the interaction test between medication use and sex were the following: BTC P value = 0.24, iCCA P value = 0.93, eCCA P value = 0.67, gallbladder cancer P value = 0.24. Given that there are no significant differences between women and men, we present the overall estimates obtained for the entire cohort.

Low-dose aspirin was not associated with iCCA risk (HR, 1.21; 95% CI, 0.93–1.57). Statins were associated with a lower risk of iCCA (HR, 0.69; 95% CI, 0.50–0.95). Combined use of low-dose aspirin and statins was not associated with a lower risk of iCCA (HR, 0.90; 95% CI, 0.68–1.18; Fig. 2).

Low-dose aspirin was not associated with risk of eCCA (HR, 0.80; 95% CI, 0.60–1.07). Statins were associated with a lower risk of eCCA (HR, 0.54; 95% CI, 0.38–0.76). Combined use of low-dose aspirin and statins did not lower the risk beyond statins alone (HR, 0.70; 95% CI, 0.52–0.93; Fig. 2).

Low-dose aspirin use was not associated with risk of gallbladder cancer (HR, 0.87; 95% CI, 0.71–1.06). Statins were associated with a lower risk of gallbladder cancer (HR, 0.72; 95% CI, 0.57–0.91). Combined use low-dose aspirin and statins did not lower the risk beyond statins alone (HR, 0.74; 95% CI, 0.59–0.92; Fig. 2).

Non-aspirin NSAIDs were associated with a higher risk of BTC (HR, 1.44; 95% CI, 1.33–1.57), iCCA (HR, 1.47; 95% CI, 1.25–1.72), eCCA (HR, 1.32; 95% CI, 1.11–1.56), and gallbladder cancer (HR, 1.49; 95% CI, 1.32–1.69; Fig. 3A).

Metformin was not associated with risk of BTC or its subtypes (Fig. 3B).

Landmark analysis was performed excluding the first year of follow-up (Supplementary Table S6). We also analyzed variation in association of drug use and risk of BTC during follow-up (Supplementary Fig. S2). Low-dose aspirin was associated with a relatively higher risk of BTC in the first years of follow-up, which decreased during successive years. Statins were associated with a relatively lower risk of BTC during the first two years of follow-up. The risk subsequently increased and then flattened out. These trends suggest reverse causation is not present.

In this Swedish nationwide cohort study comprising 5.7 million individuals, we assessed the associations between risk of BTC and its subtypes and use of low-dose aspirin, statins, NSAIDs, and metformin. Stratified analyses by sex showed no difference between the sexes. Considering BTC overall, statins, but not low-dose aspirin, were associated with a lower risk. Combined use of low-dose aspirin and statins was not associated with a further decrease in risk beyond that observed for statins alone. Adding IPW to the BTC model sharpened the estimates even more, strengthening the associations obtained by adjusting for the covariates without the weights. Non-aspirin NSAID use was associated with a higher risk of BTC overall, while metformin use was not associated with a change in BTC risk overall.

For BTC subtypes, low-dose aspirin was not associated with risk of iCCA, eCCA, or gallbladder cancer while statins were associated with a lower risk of all three subtypes, iCCA, eCCA, and gallbladder cancer. Non-aspirin NSAID use was associated with a higher risk of all three BTC subtypes. Metformin was not associated with risk of BTC or its subtypes.

Prior studies of aspirin use and BTC risk have yielded mixed results (9, 10, 19). The inconsistencies are likely related to the methods used to define exposure. A Mayo Clinic case–control study found that aspirin reduced the risk of iCCA and the Liver Cancer Pooling Project confirmed a lower risk only in men (9, 19). These results contrast with our finding of a lack of association between the use of low-dose aspirin and iCCA risk in both sexes. This discrepancy may be due to measurement errors. The Liver Cancer Pooling Project included only 4 iCCA participants who used low-dose aspirin, defined as <163 mg, and relied on self-reports through baseline interviews. While Choi and colleagues did evaluate the use of low-dose aspirin, they used medication lists and abstraction of physician notes in the medical record to assess use of the medication (9). The subjects in both cohorts had access to low-dose aspirin purchased over-the-counter (OTC), making it very challenging to accurately measure exposure since it relies on self-reports. In contrast, even though low-dose aspirin is inexpensive, OTC access is virtually nonexistent in Sweden. The Swedish population used in this cohort study could only obtain low-dose aspirin with a prescription and did not have access to OTC low-dose aspirin (20). The need for a prescription allowed prospective data collection via the Prescribed Drug Registry. Therefore, our approach was able to overcome the disadvantage of self-reported data commonly used in epidemiology studies that assess the use of low-dose aspirin.

For eCCA, one study suggested that aspirin was associated with lower risk of the extrahepatic subtypes (i.e., pCCA and dCCA; ref. 9), while another observed a lower risk of eCCA (10). These results are different from our findings of no change in risk of eCCA. It is also noteworthy that the previous studies did not take into consideration concurrent use of statins with low-dose aspirin, which are commonly co-prescribed, and may have confounded the results. Our results also differ from those of a case–control study in China that reported lower risk of gallbladder cancer in aspirin users (10).

A case–control study using the UK Clinical Practice Research Datalink was consistent with our finding of lower risk of BTC with statin use (21). The risk was lower in subjects with the highest number of prescriptions and cumulative dose of statin. The UK study was conducted using a large database with long-term follow-up and, similar to ours, evaluated the impact of important confounders. However, it did not evaluate the association of low-dose aspirin use in depth. Further, the case–control matching design prohibited evaluation of the impact of sex, which was important given the sex differences previously reported in BTC.

The strengths of our study are the following. First, we assembled the largest cohort for evaluating the risk and protective factors for BTC from the Swedish population-based registries, allowing stratified analyses by sex and subtype (22). The universal healthcare system in Sweden provides equal access to healthcare, reducing bias due to differences in access (23). Second, registry data are prospectively collected and have virtually complete follow-up, allowing highly precise estimates. Third, the Swedish Cancer Registry has high overall validity, with 99% of the cancers morphologically verified, compulsory reporting since 1983, and cross-linking with the Swedish Prescribed Drug Registry and the Swedish National Registries (24, 25). Using the Prescribed Drug Registry avoids primary noncompliance by providing information on dispensed, not prescribed medications (26). For our analysis, once dispensed the equivalent of 6 months of drugs, the individual was considered exposed for the remainder of the study. An advantage of requiring 6 months of use from the date of first prescription is that it increased the likelihood of sustained exposure, as patients would have needed to refill their medication. This minimized inclusion of transient or noncompliant users. However, some nonusers may have been categorized as users, which would bias our results towards the null, making our estimates conservative. Fourth, we considered drug use as a time-dependent covariate, which is more accurate for defining drug exposure status than time-fixed covariates because it accounts for variation in timing of drug initiation and the period of nonexposure to the drug (27). As a result, our estimates were less prone to overestimation (28). We also repeated the analysis of BTC risk, including IPW to minimize bias. The estimates obtained with the normalized weights are similar to estimates obtained without weighting. Lastly, our models were computed with age as the time-scale rather than time since enrollment. Since age is a strong driver of cancer risk, comparing users of drugs with nonusers by age provides more accurate estimates.

Our study has limitations. First, the use of ICD codes to measure covariates and outcomes limited our ability to obtain information on types of cholangitis and eCCA subtypes. Current ICD codes group all cholangitis types under one code, without distinguishing primary from secondary cholangitis. Since PSC has one of the strongest risks for cholangiocarcinoma, adjusting for PSC, instead of all cholangitis, would have strengthened our models. Second, we could not investigate associations with the extrahepatic subtypes pCCA and dCCA separately, as they have the same code (C24.0). pCCA and dCCA are separate biological entities and are analyzed separately when possible. Assessment of the confounding effect of adiposity, alcohol and smoking was limited by the lack of precise ICD codes. Third, our results may suffer from indication bias, particularly the association between NSAIDs and BTC risk. Since NSAID use is initiated after diagnosis, unlike low-dose aspirin, which is used for primary prevention, it is possible the indication for use drives the increased risk, not the drug itself. Two previous studies did not find an association between NSAID use and cholangiocarcinoma (29, 30). Thus, additional studies that control for indication bias are needed. The same limitation applies for metformin. In addition, our analysis of metformin use was limited by the fact that we used it to define diabetes. This precluded us from analyzing the association between metformin use and risk of BTC in a time-dependent fashion. Instead, use of metformin was ever-use defined at the beginning of the study period. Fourth, the external validity of our study is limited because our cohort strictly comprises people who were prescribed commonly used drugs. However, as the cohort includes 85% of the eligible Swedish population, the findings could be generalized to the entire population. Further, this could be an advantage since users of commonly prescribed drugs may be most likely to be compliant with prolonged use of chemopreventive drugs. Lastly, additional studies are required to determine whether longer use of these medications have a more beneficial effect. Lack of knowledge of medication use prior to the years our study covers limited our ability to determine the effect of longer use.

The strongest evidence of the beneficial effect of low-dose aspirin and statins for cancer prevention is obtained from randomized clinical trials (31). Given the relatively low incidence of BTC, it would be challenging to conduct a randomized trial of low-dose aspirin and/or statins for BTC prevention. However, a randomized study in a well-defined high-risk population, such as patients with PSC, might be feasible and potentially practice-changing (9). Moreover, genetic variation may influence the chemopreventative effects of low-dose aspirin and statins. It will be interesting to study the relationship between genetic variation and the chemopreventive effect of these drugs (9).

In conclusion, we have undertaken a comprehensive epidemiologic investigation of the association of low-dose aspirin, statins, non-aspirin NSAIDs, and metformin with BTC risk. Given the well-known difference in risk of BTC between women and men, we examined the associations of the agents with BTC risk separately by sex, but found no significant interaction between sex and BTC risk. Further, in recognition of the increasing evidence that the BTC subtypes represent distinct disorders, we used the opportunity of the large Swedish cohort to examine association with risk of BTC subtypes.

While the absolute risk reductions are small and our study does not provide sufficient justification for the use of statins for the reduction of BTC risk alone in average risk persons, the recognition of the salutary benefits of these drugs further strengthens the evidence for the public health benefits of statins, not only for CVD but also for colorectal cancer and now also for BTC (8). For individuals at high risk of BTC, such as patients with PSC, the number needed to treat to achieve benefit may be within the range that justifies the use of statins in cancer prevention. Further studies are required to confirm this.

T.M. Therneau reports grants from NIH during the conduct of the study. L.R. Roberts reports grants from NIH, Mayo Clinic, The Cholangiocarcinoma Foundation, Bayer, Boston Scientific, Exact Sciences, Fujifilm Medical Sciences, Gilead Sciences, Glycotest Inc, Redhill Biopharma, TARGET PharmaSolutions, and Karolinska Institutet during the conduct of the study as well as other support from MedEd Design LLC, Pontifax, Global Life Science Consulting, The Lynx Group, AstraZeneca, Bayer, Eisai, Exact Sciences, and GRAIL Inc outside the submitted work. No disclosures were reported by the other authors.

L. Marcano-Bonilla: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. C.D. Schleck: Data curation, formal analysis, validation. W.S. Harmsen: Data curation, formal analysis, validation, methodology. O. Sadr-Azodi: Conceptualization, formal analysis, validation, investigation, methodology. M.J. Borad: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. T. Patel: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. G.M. Petersen: Conceptualization, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. T.M. Therneau: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. L.R. Roberts: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. N. Brusselaers: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, project administration, writing–review and editing.

This work was supported by the NIH through the Mayo Clinic Center for Clinical and Translational Science (UL1 TR000135), the NCI through the Mayo Clinic Hepatobiliary SPORE Grant (P50 CA210964), the Women's Health Research Center at Mayo Clinic, the Karolinska Institutet-Mayo Clinic Collaborative Travel Grant, The Cholangiocarcinoma Foundation, and the Strategic Funding in Epidemiology (SFO, Karolinska Institutet, Young Scholar Grant).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.