This study utilized the medical record observation and assessment of drug safety (MOA) CDM, a standardized and distributed data network that allows for a multicenter data analysis using de-identified electronic healthcare record data collected from university hospitals (MOA CDM data partners) in South Korea [17]. The MOA CDM is coordinated by KIDS, a national organization affiliated with the Korea Ministry of Food and Drug Safety, and contains medical records of over 37 million patients from 30 data partners as of 2023 [17]. Patients are registered in the database on the day of their first visit to a hospital. For this study, we specifically collaborated with 20 data partners of which 10 are in the capital city (Seoul) and 7 in the nearby province (Gyeonggi-do).

We performed a retrospective cohort study with incident ARB users who were at least 18 years of age and had initiated ARB treatment between 1 January, 2018 and 30 June, 2021. The first prescription date of the ARB was defined as the index date. Patients were censored when they experienced the study outcome, switched to another ARB, or at 6 months from the treatment initiation, as the azilsartan-induced liver injury case submitted to KIDS occurred approximately 6 months after the first use of the drug [18].

We excluded patients who were younger than 18 years of age or were prescribed multiple ARBs at the index date. Additionally, patients were excluded if they were prescribed any ARBs 3 months before the index date, had a suspected DILI (study outcome) or clinically significant conditions that may interfere with the interpretation of the study results 1 year before the index date, or had a serious hepatobiliary condition or were pregnant, which is a contraindication for ARB use 1 year before the index date or during the study period (study figure available in the Electronic Supplementary Material [ESM]). The decision to exclude pregnant patients was based on the significant contraindication of ARBs during pregnancy, as discontinuation of ARBs is a common practice when planning pregnancy. The exclusion was applied to mitigate a potential serious mis-estimation of follow-up time in these cases. To further validate this approach, we examined the number of pregnant patients during the entire data period, which was 0.028% among ARB users and would not have a significant impact on the overall study results.

Patients were assigned to treatment groups according to the ARB prescribed at the index date. Nine ARBs were marketed in South Korea during the study period: azilsartan, eprosartan, telmisartan, fimasartan, valsartan, olmesartan, losartan, irbesartan, and candesartan.

We operationally defined DILI by adapting the clinical chemistry criteria provided by the International DILI Expert Working Group: alanine aminotransferase (ALT) ≥ 5× upper limit of normal (ULN), alkaline phosphatase (ALP) ≥ 2× ULN, or ALT ≥3× ULN and total bilirubin (TBL) > 2× ULN. After investigating the ULN standards of the data partners, the following values were selected: ALT, 40 U/L; ALP, 117 U/L; and TBL 1.2 mg/dL. Aspartate aminotransferase levels were not assessed as they may not specifically indicate liver injury [19].

The type of DILI was classified using the R ratio ([ALT/ALT ULN]/[ALP/ALP ULN]) as follows: hepatocellular (R ratio ≥ 5), mixed (R ratio 2–5), or cholestatic (R ratio ≤ 2) [19]. The severity of DILI was classified as follows: mild (ALT ≥ 5× ULN or ALP ≥2× ULN and TBL < 2× ULN), moderate-severe: (ALT ≥ 5× ULN or ALP ≥2× ULN and TBL ≥ 2× ULN), and fatal: any all-cause death within 1 year after the incident DILI [19]. The operational definitions for DILI were further reviewed by clinical experts and researchers with expertise in liver injury. As any patient identification information is pseudonymized in the CDM database, medical charts reviews were not feasible. Therefore, we inform that the cases detected in our study are all suspected cases for which no additional validation was conducted.

As the diagnosis of DILI mostly depends on the exclusion of alternative causes of liver injury, we listed clinical conditions that could be potential alternative causes of liver injury based on the previous literature [13]. With a review by clinical experts, the conditions were categorized as follows: to be excluded at baseline and adjusted for during the follow-up (clinically significant conditions), to be completely excluded at baseline and the follow-up (serious hepatobiliary conditions), and others to be adjusted for at baseline and the follow-up. In addition, we adjusted for hepatotoxic drugs by class at baseline and the follow-up, which were defined as drugs with a LiverTox DILI-likelihood score of A (well known) or B (known or highly likely) [ESM].

Patient demographics (sex, age, and enrollment year), encounter records (hospitalizations, days outpatient visits, and emergency room visits), Charlson Comorbidity Index, comorbidities, and prescription histories were included as baseline covariates. Additionally, predetermined potential alternative causes of liver injury and anti-hypertensive drugs class prescribed during the follow-up were also included as covariates.

Descriptive analyses were conducted to summarize the patient characteristics, treatment patterns, and characteristics of DILI. The number of patients by status over time was collected from each data partner to calculate pooled incidence rates. Cox proportional hazards models were used to derive hazard ratios (HRs) of DILI and to compare the risk among specific ARBs against valsartan. Subgroup analyses were conducted by treatment patterns and data partners to compare study populations. Valsartan was selected as the reference drug as it is the most used ARB in Korea [21]. To minimize selection bias, we applied propensity score-based inverse probability of treatment weighting (IPTW) and derived the average treatment effect [22, 23]. Additionally, covariates collected during the follow-up were included in the regression model to reduce confounding. From each data partner, coefficients, standard errors, and confidence intervals (CIs) from Cox proportional hazards models were collected to conduct the meta-analysis. We evaluated the average treatment effect by a random-effect model in the meta-analysis to consider population variance of each data source.

All statistical analyses were performed using R [24] and the level of statistical significance was set at p < 0.05 (two-sided). R packages ‘DatabaseConnector,’ ‘SqlRender,’ ‘plyr,’ ‘dplyr,’ ‘DBI,’ ‘odbc,’ ‘lubridate,’ and ‘nnet’ for database connection and preprocessing, ‘WeightIt’ and ‘cobalt’ for IPTW, ‘survival’ and ‘survminer’ for the survival analysis, and ‘meta’ for meta-analyses were used [25,26,27,28,29,30,31,32,33,34,35,36,37].