Metformin Use and the Risk of Cancer in Patients with Diabetes: A Nationwide Sample Cohort Study

Cancer is one of the most important causes of death worldwide (1). From 2006 to 2016, there were 17.2 million cancer patients worldwide, 8.9 million deaths caused by cancer, and the global incidence of cancer increased by 28% (2). Furthermore, the incidence of cancer is expected to increase in the future, and the prevention of cancer is important to reduce the global burden of disease in the future (3).

Metformin is a biguanide (N-N-dimethylbiguanide hydrochloride), which is most commonly prescribed for the management of type 2 diabetes mellitus (4). Metformin is known to reduce glycogenesis through adenosine monophosphate-activated kinase signaling, and it increases glucose uptake in muscle cells, leading to a decrease in glucose levels (5). Recently, there have been some reports that metformin has an antitumor effect that might be beneficial in cancer prevention and treatment (6, 7). The antitumor activity of metformin has been explained by both direct and indirect molecular mechanisms in preclinical and in vitro studies (8–10). Against this background, recent clinical- and population-based cohort studies have reported that metformin therapy in patients with diabetes is beneficial in the prevention of various cancers (11–13). However, another recent nationwide population-based cohort study conducted in Israel reported that there was no significant association between metformin therapy and the incidence of major cancer (14), and the debate is ongoing.

Therefore, this study was designed to investigate the association of metformin therapy in patients with diabetes with the development of cancer, using a sample cohort from South Korea. We hypothesized that metformin therapy lowers the risk of cancer in patients with diabetes.

The “sample cohort database” of the National Health Insurance Service (NHIS) was developed to provide data for academic healthcare-related research among the general population in South Korea. The database included a stratified random sample of one million people who had registered with the NHIS since 2002. It was designed to be representative of the national population in terms of demographic and socioeconomic variables. The cohort was dynamic, and patients were followed up until the end of 2015. It was supplemented with additional data, including data on infants, to allow for attrition due to death and loss to follow up. The study protocol was approved by the Institutional Review Board of Seoul National University Bundang Hospital (X-1905-541-901) and the Health Insurance Review and Assessment Service (NHIS-2019-2-159).

We included all diabetic and adult patients (ages ≥18 years) in the 2010 cohort database of the NHIS. All subjects were registered with diagnoses of diabetes mellitus according to the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10; E10–E14) in the 2010 NHIS database. Next, we excluded individuals who died during 2010, individuals who emigrated between 2011 and 2015 due to inability to follow them up, and individuals who had a history of diagnosis of cancer in 2010, as our study focused on new diagnoses of cancer between 2011 and 2015.

Among patients with diabetes in the NHIS 2010 cohort, metformin users were defined as those who had been prescribed a continuous supply of oral metformin over a period of ≥90 days. All other individuals were classified as the control group. The classification of metformin use in the 2010 cohort was based on the metformin prescription data from October 2009 to December 2010 because we aimed to exclude immediate exposure to metformin (<90 days) before evaluating cancer development since January 2011 via a lag-time approach (15). Metformin users were divided into the following two groups: high daily dosage group (>1 g/day) and low daily dosage group (≤1 g/day).

According to the ICD-10 diagnostic system, the development of cancer in this study was defined as newly registered diagnoses of any malignancy (C00–C96) between 2011 and 2015. In detail, the development of cancer was divided into gastric cancer (C16), esophageal cancer (C15), colorectal cancer (C18–C20), gall bladder and biliary tract cancer (C23–C24), head and neck cancer (C00–C14), brain cancer (C71), liver cancer (C22), pancreatic cancer (C25), lung cancer (C34), bone and articular cartilage cancer (C40–C41), neoplasms of breast and genital organs (C50–C63), urinary tract cancer (C64–C68), thyroid cancer (C73), and lymphoma or leukemia (C81–C96). The time to cancer diagnosis was calculated from January 1 to the date of diagnosis of cancer, as registered officially in the ICD-10 system. In South Korea, all patients diagnosed with cancer of any C-code should be registered in the NHIS database to receive special financial coverage; 95% of the total cost for cancer treatment is covered by the NHIS. Therefore, there was no patient whose diagnosis of cancer was not registered in the sample cohort of 2010 in South Korea.

Data regarding the following variables were collected as confounders in this study: (i) demographic information (age and sex); (ii) socioeconomic information [income level in deciles and place of residence in 2010 (Seoul, metropolitan cities, or others)]; (iii) comorbidities that had been registered from 2009 to 2010 using the ICD-10 code diagnostic system in the NHIS database [hypertension (I10–I16), coronary artery disease (I20*–I25*), cerebrovascular disease (I60*–I69*), psychobehavioral disorder (F00–F99), musculoskeletal disorders (M00–M99), chronic kidney disease (N18*), dyslipidemia (E78.0), anemia (D64*), chronic obstructive pulmonary disease (J44*), arrhythmia (I49*), and liver cirrhosis (K74*)]; (iv) receipt of surgery in 2010; (v) total number of hospital visit days in 2010; and (vi) use of other antidiabetic medications (sulfonylureas, alpha-glucosidase inhibitors, thiazolidinediones, and insulin). The number of hospital visit days included the number of hospital outpatient clinic visits and days spent on admission, but did not include outpatient visits to primary care physicians. For example, an individual who visited a hospital outpatient clinic five times and was admitted to the hospital for 3 days would be considered to have eight hospital visit days. In the analysis, the number of hospital visit days was categorized into five groups (0, 1–7, 8–29, 30–90, and >90 days).

The primary endpoint of this study was the new development of cancer from January 1, 2011, to December 31, 2015, among patients with diabetes registered in the NHIS sample cohort of 2010 in South Korea.

The baseline characteristics of patients are presented as means with standard deviations for continuous variables or frequencies with percentages for categorical variables. First, because our study focused on time-dependent exposure to metformin in the 2010 cohort from 2011 to 2015, we investigated the proportion of patients with diabetes that went off metformin therapy or started using it during the follow-up period (2011–2015). Exposure to metformin in both the metformin and control groups in 2010 varied throughout the follow-up period (Supplementary Table S1). Therefore, we investigated the association between exposure to metformin and the new development of cancer using a time-dependent Cox regression model. In this model, exposure to metformin was considered as a time-dependent variable, and all other covariates were included in the time-dependent Cox regression model for multivariable adjustment. As a first sensitivity analysis, we performed a time-dependent Cox regression analysis of the development of cancer according to the daily dosage of metformin in the high daily dosage (>1 g/day) and low daily dosage (≤1 g/day) groups. In addition, we performed a time-dependent Cox regression analysis of the development of cancer in detail via the same method to investigate whether the association differed according to the type of cancer.

To enhance the robustness of our study findings, we performed propensity score (PS) matching, which is known to reduce confounders in cohort studies (16), using the nearest neighbor method with a 1:1 ratio, without replacement, and a caliper width of 0.2. Logistic regression analysis was performed to calculate PS values as a logistic model, and all covariates were included in the PS model. The absolute value of the standardized mean difference (ASD) was used to determine the balance between the metformin and control groups before and after PS matching. The ASDs between the two groups before and after PS matching were set to below 0.1 to determine if the two groups were well balanced.

After confirming adequate balance between the two groups after PS matching, we performed a time-dependent Cox regression analysis of the development of cancer between 2011 and 2015 in the PS-matched cohort. The results of the Cox regression models are presented as HRs with 95% confidence intervals (CI), and it was confirmed that there was no multicollinearity in all multivariable models of the entire cohort with a variance inflation factor of <2.0. All statistical analyses were performed using R version 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria), and P < 0.05 was considered statistically significant.

In total, 77,337 adult patients with diabetes were screened initially. Next, we excluded individuals who died in 2010 (n = 952), individuals who emigrated between 2011 and 2015 (n = 49) due to inability to follow them up, and individuals who had a history of cancer in 2010 (n = 9,710). Finally, 66,627 adult diabetes patients were included in the analysis. Among them, 7,079 (10.6%) patients (3,185 in the metformin group and 3,894 in the control group) were newly diagnosed with cancer between 2011 and 2015 (Fig. 1). Table 1 shows the characteristics of the entire study cohort.

The results of the time-dependent Cox regression analysis of the development of cancer after multivariable adjustment are presented in Table 2. There was no significant difference in the hazard for the development of cancer between the metformin and control groups (HR = 0.96; 95% CI, 0.89–1.03; P = 0.250). In the sensitivity analysis according to daily dosage, there was no significant difference in the hazard for the development of cancer in both the low daily dosage of metformin group (HR = 0.97; 95% CI, 0.92–1.03; P = 0.301) and high daily dosage of metformin group (HR = 1.02; 95% CI, 0.95–1.11; P = 0.497), compared with the control group. In addition, there was no significant difference in the hazard for the development of any cancer type between the metformin and control groups (all P > 0.05).

The results of the comparison of characteristics between the metformin and control groups before and after PS adjustment are shown in Table 3. The characteristics of both the metformin and control groups were well balanced as the ASDs of all variables were below 0.1 after PS matching. In the time-dependent Cox regression model of the PS-matched cohort, there was no significant difference in the hazard for the development of cancer between the metformin and control groups (HR = 0.96; 95% CI, 0.90–1.04; P = 0.320; Table 4). In addition, there were no differences in the hazards for the development of any cancer type between the metformin and control groups in the PS-matched cohort (all P > 0.05).

This population-based cohort study showed that metformin therapy was not significantly associated with a lower risk of cancer among patients with diabetes. This association did not significantly differ according to the daily dosage of metformin. Our finding is similar to that of a recent cohort study conducted by Dankner and colleagues (14), and suggested that the relationship between metformin therapy and the risk of cancer remains controversial.

Recently, some previous studies have reported that metformin therapy might have a potential benefit for the prevention of cancer development (6, 7, 17). This effect was supported by the in vitro evidence that metformin has a direct antitumor effect, and might suppress tumor proliferation, induce apoptosis, and arrest the division of cancer cells (8–10). Furthermore, metformin is known to confer an indirect antitumor effect by inhibiting protein phosphatase 2A (18). In this context, metformin therapy was expected to be beneficial in preventing cancer among patients with diabetes in population-based cohort and clinical studies (19–24). However, both our study and the study conducted by Dankner and colleagues (14) failed to show that metformin use yielded a potential benefit for the prevention of cancer among patients with diabetes.

Several observational studies have shown that metformin therapy confers a protective effect on cancer development or progression. Specifically, metformin therapy is reported to lower the risk of overall cancer (19–21), as well as colorectal cancer (22), esophageal cancer (23), and endometrial cancer (24). However, Suissa and colleagues reported that these observational studies mostly suffered from time-related bias such as immortal time bias, which might result in an overestimation of the effect of metformin on cancer development (25). Immortal time bias refers to a period of follow-up during which the study outcome cannot occur due to the study design, death, or the study outcome (26). In addition, Wei and colleagues reported that the association between metformin therapy and survival among patients with pancreatic cancer had been greatly exaggerated in previous cohort studies due to the wide presence of immortal time bias (27). Immortal time bias might be present in our study. For instance, the risk of cancer in the control group might be higher if patients in the metformin group died due to any other disease (e.g., cardiovascular disease) in 2011 before cancer development and the control group lived beyond 2011. However, immortal bias is known to be overcome in epidemiological studies via two statistical methods (28): time-dependent techniques and matching. We used both time-dependent Cox regression analysis and PS matching in this study. Furthermore, we considered as many comorbidities as possible with socioeconomic status-related confounders for PS matching, which might be closely related to death due to other diseases, before cancer development.

Second, there might be protopathic bias. This arises when a pharmaceutical agent is inadvertently prescribed for an early manifestation of a disease that has not yet been detected diagnostically (29). For example, the results might be biased due to inadequate duration of exposure for the event to occur if a patient received a prescription for metformin from December 2010 within 30 days. To minimize this bias, we defined metformin users using a lag-time approach as those who received prescriptions for metformin over 90 days until December 2010 (15, 30). By this approach, we excluded immediate exposure to metformin among patients with diabetes before evaluating the development of cancer from January 2011. Considering that protopathic bias may affect the association between exposure to metformin and the development of cancer, our study is notable with regard to using the lag time approach.

This study has several limitations. First, some important variables, such as body mass index, were not included in the statistical adjustment because they were not included in the NHIS data set. Second, we defined the comorbidities using ICD-10 codes registered in the NHIS database. The diseases specified by the ICD-10 codes might differ from the actual underlying diseases in all patients. Third, PS adjustment and multivariable adjustment could control only known confounders; thus, there might be residual confounders that could affect the study results. Fourth, we based the analysis on metformin prescription data and did not assess actual adherence or compliance among those classified as metformin users. Finally, we classified neoplasms of the breast and genital organs (C50–C63) as one group among specific cancer sites in this study, because the NHIS database contained one C-code (C_) for breast or genital organ cancer (C50–C63). Therefore, we could not evaluate the association between exposure to metformin and the development of cancer of the prostate, breast, or cervix; this could be a limitation of this study.

In conclusion, this population-based cohort study suggested a lack of association between metformin therapy and the risk of cancer among patients with diabetes, even in those receiving high daily doses (>1 g/day). However, there might be residual confounders or bias; thus, further prospective, large population-based cohort studies are needed to confirm these findings.

No potential conflicts of interest were disclosed.

Conception and design: T.K. Oh, I.-A. Song

Development of methodology: T.K. Oh

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): I.-A. Song

Writing, review, and/or revision of the manuscript: T.K. Oh, I.-A. Song

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.