The data source of this research is the full population file of Taiwan’s National Health Insurance Research Database (NHIRD). The features of the NHIRD have been described in our previous study [14]. The diagnosis of diseases in the NHIRD is based on the International Classification of Diseases Ninth/Tenth Revision, Clinical Modification (ICD-9/10-CM). The NHIRD has links to the National Death Registry dataset for death information. All patient and health care information was encrypted before release to protect the privacy of individuals. The Research Ethics Committee of China Medical University and Hospital (CMUH104-REC2-115-CR4) approved this study and waived informed consent from patients.

We identified patients diagnosed with type 2 diabetes from the National Health Insurance Database between January 1, 2008, and December 31, 2018, who received follow-ups till December 31, 2019 (Additional file 1: Fig. S1). T2D was diagnosed according to the ICD codes (Additional file 1: Table S1) with antidiabetic drug use and at least two outpatient visits or one hospitalization within one year for T2D. The algorithm for defining T2D using ICD codes was validated as 74.6% accurate [15]. Exclusion criteria were as follows: missing age or gender data, age below 18 or above 80 years, diagnosis of type 1 diabetes, hepatitis B virus infection, hepatitis C virus infection, alcohol-related disorders, dialysis, liver cirrhosis, esophageal varices with bleeding, ascites, hepatic encephalopathy, jaundice, hepatic failure, hepatocellular carcinoma (HCC), and liver transplant before the index date. The study also excluded patients who died or were lost to follow-up and diagnosed with HCC within 6 months of the index date to avoid latent morbidity or mortality.

We defined patients who had received GLP-1 RAs after T2D diagnosis as GLP-1 RA users and those who had never received GLP-1 RAs during the study period as nonusers. The first date of GLP-1 RA use was set as the index date of the study group. We recorded the index date for the control cases with the same time interval from T2D diagnosis to the index date of the study group. Since GLP-1 RAs were launched in Taiwan in 2011, the index dates for the study and control groups were recorded after 2011. Some clinically relevant variates assessed were as follows: age, sex, family income, obesity, smoking, hypertension, dyslipidemia, stroke, coronary artery disease, heart failure, chronic kidney disease, chronic obstructive pulmonary disease (COPD, diagnosed one year before the index date), and number of oral antidiabetic drugs, insulin, antihypertensive drugs, aspirin, statin, fibrates use (drug use one year prior to index date), and duration of diabetes. We also counted the scores of the Charlson Comorbidity Index (CCI) [16] and Diabetes Complications Severity Index (DCSI) [17] to evaluate the disease burden and diabetes complications in these patients.

This study compared the use or no-use of GLP-1 RAs in the following outcomes: liver cirrhosis development [18], hepatic failure (coagulopathy, hepatic coma, with or without other organ failures) [19], hepatocellular carcinoma, major adverse cardiovascular events (MACE, a composite outcome of admitted stroke, coronary artery disease, and heart failure), liver-related death (death due to liver cirrhosis, decompensated cirrhosis, hepatic failure, and HCC), cardiovascular death, and all-cause mortality (the confirmation and cause of death were from the link with the National Death Registry). For the outcomes of interest, we followed up with the patients until the occurrence of outcomes, death, or the end of the study period on December 31, 2019, whichever appeared first.

We performed 1:1 propensity score matching to match and balance the variation in the study and control groups [20]. Non-parsimonious multivariable logistic regression was used to estimate the propensity score for every person who received GLP-1 RAs. The GLP-1 RA use was treated as a dependent variable, and 37 critical variables (including age, sex, income, obesity, smoking, comorbidities, medications and duration of diabetes) were used as independent variables (Table 1). We used greedy nearest-neighbor matching to perform optimal matching and matched the control group without replacement. The nearest-neighbor algorithm was used to identify matched pairs with a width of less than 0.001. We assumed a standardized mean difference (SMD) of less than 0.1 as a negligible difference between the study and control groups.

Crude and multivariate-adjusted Cox proportional hazards models were used to compare the endpoints between GLP-1 RA users and nonusers. We presented the results as hazard ratios (HR) and 95% confidence interval (CI). The Schoenfeld residuals and complementary log–log plots were used to check the proportional-hazards assumption. We used a stepwise approach to adjust for the variables in the Cox models, as follows: Model 1: adjusted for sex and age; Model 2: adjusted for sex, age, income, and obesity; Model 3: adjusted for sex, age, income, obesity, and comorbidities; Model 4: adjusted for sex, age, income, obesity comorbidities, medications and duration of diabetes as shown in Table 1. Kaplan–Meier method was used to describe the cumulative incidence of outcomes between GLP-1 RA use and no-use over the follow-up time. We performed the subgroup analysis, including the subgroups of sex, age, comorbidities, CCI, DCSI, and medications, of GLP-1 RA use versus no use in the outcomes of all-cause death, cardiovascular diseases, cardiovascular death, and liver-related death. We also performed dose–response analysis on the cumulative duration of < 182, 182–364, > 364 days of GLP-1 RA use versus no use in the outcomes of all-cause death, cardiovascular diseases, cardiovascular death, and liver-related death.

We considered the two-tailed p-value < 0.05 statistically significant and used SAS version 9.4 and Stata SE version 11.0 for analysis.