Metformin and Risk of Hepatocellular Carcinoma in Patients With Type 2 Diabetes

Chin-Hsiao Tseng


Liver International. 2018;38(11):2018-2027. 

In This Article

Materials and Methods

Taiwan's National Health Insurance (NHI) program was implemented in March 1995. It currently covers more than 99% of the population and has contracts with 93% of medical settings and with all hospitals nationwide. The reimbursement database of the NHI keeps all records of diseases diagnosed, medications prescribed and procedures performed. Researchers may use the database for academic purposes after approval by an ethics review board. This study was granted approval number 99274 by the Institutional Review Board of the National Health Research Institutes. Informed consent was not required according to local regulations because all personal data were de-identified before the release of the database for analyses.

The methods used in this study have been described in detail elsewhere.[16] Disease diagnoses were coded using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD–9–CM) during the study period. Diabetes was coded 250.XX and HCC 155 (excluding 155.1).

The unmatched original cohort and the matched cohort were created step-by-step following the procedures illustrated in Figure 1. In brief, 423 949 patients who were newly diagnosed as having diabetes mellitus during 1999–2005 and who had been followed up in outpatient clinics with ≥2 prescriptions of antidiabetic drugs were first identified. Exclusion criteria were the following: (i) use of other antidiabetic drugs before metformin was initiated among metformin users (n = 183 837); (ii) type 1 diabetes mellitus (n = 2062); (iii) missing data (n = 423); (iv) diagnosis of any cancer before entry or within 6 months of diabetes diagnosis (n = 26 637); and (v) follow-up duration of <6 months (n = 15 173). In the end, 173 917 ever-users and 21 900 never-users of metformin were identified in the unmatched original cohort. A PS matched-pair cohort (the matched cohort) of ever– and never-users of metformin was created using the Greedy 8→1 digit match algorithm.[17] All baseline characteristics listed in Table 1 and the date of entry were used to create the PS by logistic regression.

Figure 1.

The procedures in creating the unmatched original cohort and the cohort of 1:1 matched-pair of ever– and never-users of metformin using the reimbursement database of the National Health Insurance (HCC: hepatocellular carcinoma)

Cumulative duration of metformin use (months) was calculated for each patient, and its tertiles were used for analyses. Potential confounders were categorized into demographic data (age, sex, occupation and residential area), major comorbidities (hypertension, dyslipidaemia and obesity), diabetes-related complications (nephropathy, eye disease, stroke, ischaemic heart disease and peripheral arterial disease), use of antidiabetic drugs (insulin, sulfonylurea, meglitinide, acarbose, rosiglitazone and pioglitazone), potential risk factors for cancer (chronic obstructive pulmonary disease, tobacco abuse, alcohol-related diagnoses, gallstone, history of Helicobacter pylori infection, Epstein-Barr virus-related diagnoses, hepatitis B virus infection, hepatitis C virus infection, cirrhosis and other chronic non-alcoholic liver diseases) and medications that are commonly used among diabetes patients or that may affect cancer risk (angiotensin-converting enzyme inhibitors/angiotensin receptor blockers, calcium channel blockers, statins, fibrates and aspirin). The patient's residential area and occupation were classified as detailed elsewhere.[18] Diagnoses of cirrhosis and other chronic non-alcoholic liver diseases were based on ICD–9–CM 571.5 and 571.8 respectively. The ICD–9–CM codes for other diagnoses have been described in previously published papers.[16,18]

Both the unmatched original cohort and the matched cohort were analyzed. Student's t test was used to compare the differences in age between the never– and ever-users, and the Chi-square test was used to compare the differences in other variables. Standardized difference was calculated for each covariate, as described by Austin and Stuart, who proposed a value >10% as an indication of potential confounding from the variable.[19]

Incidence was calculated for each subgroup of metformin exposure, that is never-users and ever-users, and for each tertile of cumulative duration of therapy. The numerator was the case number of newly diagnosed HCC identified during follow-up. The denominator was the person-years of follow-up, which ended at the time of HCC diagnosis or on the date of death, the last reimbursement record or December 31, 2011.

Hazard ratios and their 95% confidence intervals for ever-users and tertiles of cumulative duration versus never-users were estimated by Cox regression incorporated with the inverse probability of treatment weighting (IPTW) using the PS, as recommended by Austin, to reduce confounding from the differences in characteristics.[20] To further examine the consistency of the findings, models were also created after excluding patients aged <25 or >75 years in the unmatched cohort as sensitivity analyses, because HCC rarely occurs in patients younger than 25 years, and patients older than 75 years may represent a group of healthy survivors. Subgroup analyses were conducted for patients with and without liver diseases, including alcohol-related diagnoses, hepatitis B virus infection, hepatitis C virus infection, cirrhosis, other chronic non-alcoholic liver diseases and any of the above.

To investigate the interactions between metformin and aspirin or statin, hazard ratios and their 95% confidence intervals and P–values for interaction were estimated. Users of aspirin or statin for <2 years were excluded to allow a potential incubation period.

SAS statistical software (version 9.3, SAS Institute, Cary, NC) was used for statistical analyses. P < .05 was considered statistically significant.