Empagliflozin is associated with lower cardiovascular risk compared with dipeptidyl peptidase-4 inhibitors in adults with and without cardiovascular disease: EMPagliflozin compaRative effectIveness and SafEty (EMPRISE) study results from Europe and Asia

The methods used in this study including outcomes and analyses are described in detail in the full study protocol registered in the EU PAS Register (EUPAS27606) [16].

Data sources

The data for this retrospective cohort study was obtained from electronically recorded longitudinal data sources in Denmark, Finland, Germany, Israel, Japan, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom (UK). The different types of data sources included nationwide healthcare registers, regional quality registers, regional high-quality medical health records, and other health claims data. The data sources used in the study are described in more detail in Additional file 1: “Data sources” section. The study included locally adapted versions of International Classification of Diseases 9 (ICD-9) and 10 revision (ICD-10), Anatomical Therapeutic Chemical (ATC) codes (Additional file 1: Tables S1a and S2a).

Study population

In each country, adults (≥ 18 years) diagnosed with T2D that initiated empagliflozin or any DPP-4i during the study period were considered eligible for the study. The study period started from the date of market authorisation of empagliflozin in the respective countries (Additional file 1: Supplemental Materials – Table S3a) until the end of data availability (i.e., December 2018 for all countries except Germany [December 2019], Japan [April 2018], South Korea and Taiwan [December 2017]). Eligible patients who initiated (had date of prescription/dispensation) the study drugs within the study period were included. The date of drug initiation was the cohort entry date (index date). Patients who were < 18 years old, had secondary or gestational diabetes mellitus, or end-stage renal disease before study entry [≤ 12 months of data before index date (i.e., baseline period)], or incomplete data on age or sex were excluded from the study cohorts in each country.

Exposure

Patients were defined as empagliflozin initiators if they had a record of prescription/dispensation of empagliflozin during the study period and no record of prescription/dispensation of any SGLT-2i or any DPP-4i during the preceding 12 months (6 months for Germany) before the cohort entry date, i.e., in the baseline period. A similar approach was used to define initiators of any DPP-4i.

A list of all study drugs prescribed/dispensed during the study period in the countries included in this study is provided in Additional file 1: Table S2a. The duration of drug exposure and date of treatment discontinuation were determined based on the information available in each country and therefore defined separately in each country. Drug initiation was assumed to begin on the date of a prescription/dispensation (index date). The duration of exposure for each drug was extracted directly from the days’ supply information (when data were available) or derived from the dispensed amount and the daily dose. In most countries, a grace period of 100% of the calculated duration of drug exposure was applied to address the uncertainty of the actual duration of exposure [17]. Further, drug exposures overlapping in time were handled by moving the subsequent exposure by a maximum of 14 days. Periods of overlapping supplies and grace periods were combined into exposure periods. The exposures were defined using an ‘as-treated ‘(AT) approach, therefore, the follow-up was censored at discontinuation (defined as the end date of the last grace period), switch to other study drug, or concomitant use.

Patients were followed from index date to occurrence of any of the study outcomes, death, discontinuation of the initial study drug (defined as end of grace period), switch to any other study drug, initiation of concomitant use of study drugs (either as free or fixed-dose combinations), end of data availability, or end of study (31 December 2018 for all countries except Germany [31 December 2019], Japan [April 2018], South Korea and Taiwan [December 2017]), whichever occurred first.

Outcomes

The primary study outcomes were HHF (available in all countries except in UK THIN [The Health Improvement Network, also known as IMRD [IQVIA Medical Research Data]), MI (available in all countries) and stroke (available in all countries). Secondary cardiovascular effectiveness outcomes included CVM (available in Finland, Norway, Sweden, Taiwan and UK CPRD [Clinical Practice Research Datalink]) and coronary revascularisation procedures (available in all countries except Germany and Spain). Two composite outcomes were also assessed: (1) HHF or CVM (available in Finland, Norway, Sweden and the UK CPRD); (2) MI, stroke, or CVM (i.e., 3-point MACE; available in Finland, Norway, Sweden, Taiwan and UK CPRD).

Two approaches were used to identify HHF, including use of a broad HHF definition (any diagnosis of HF associated with hospitalisations, specialist outpatient and primary care records, and/or a dispensation/record of high-ceiling or loop diuretics [ATC: C03C]) applied to data from Israel, Japan, South Korea, Spain, Taiwan and UK CPRD, and a specific HHF definition (a diagnosis of HF during hospitalisation or diagnosis of HF that led to hospitalisation, required most healthcare resources, or was coded as the main disease in hospital claims) applied to data from Denmark, Finland, Germany, Israel, Japan, Norway, Sweden and Taiwan (Additional file 1: Table S4a). The composite outcome including HHF or CVM was based only on data using the specific HHF definition (except where the broad HHF definition, diagnosis of HF in any position of hospitalisation, was applied in the UK CPRD since this definition was similar to the specific HHF definition used in other countries). All other study outcomes were generally defined as having a primary diagnosis (or procedures) of the condition of interest during hospitalisations, specialist outpatient, or primary care visits.

Statistical analysis

Using country-level data, patients were matched by creating propensity score (PS) models between the exposure group (empagliflozin) and the comparator group (DPP-4i) using logistic regression based on available variables in each country. Covariates included sociodemographic, lifestyle characteristics, diabetic complications, comorbidities, comedications, and healthcare resource utilization in the logistic models to indicate the predicted treatment probability. Sociodemographic characteristics and lifestyle variables were measured at index date, healthcare resource utilization variables during ≤ 12 months before the index date, whereas the other covariates were measured in all available data for Nordic countries and UK THIN and during ≤ 12 months before the index date for all other countries (except Germany ≤ 6 months). The matching was done without replacement using a ratio of 1:1 and caliper width of 0.2 of the standard deviation of the logit of the PS. In case of multiple potential matching comparators, the first comparator in an ascending order of absolute difference in the logit of the PS was chosen; in case of a tie, comparators were chosen randomly. The matching process was evaluated by observing the absolute standardised differences (ASD). Any covariate that remained unbalanced (ASD > 0.1) [18] during the matching process was included in the outcome models [19].

Each subgroup analysis was performed separately as follows: (a) Subcohorts (e.g., CVD/No CVD) were created from the main study cohorts, (b) PS-matching was performed separately for each pair of subcohorts (empagliflozin vs DPP-4i) using baseline patient characteristics, and (c) analyses were performed on the separately PS-matched cohort. Within each country, analyses were conducted using an ‘as-treated’ approach comparing empagliflozin initiators with an active comparator group of DPP-4i initiators. The risks were compared between treatment groups with stratification of patients based on presence/absence of pre-existing CVD and HF.

All outcomes were analysed using Cox proportional hazards regression models and Hazard Ratios (HR) with 95% Confidence Intervals (CI) were presented. Overall and country-specific results were pooled via random-effects meta-analysis methods. Across the estimates, heterogeneity was assessed using the estimated total heterogeneity, Chi-square test for heterogeneity (significance level: 0.1, null hypothesis: no heterogeneity) and the I² statistic (0% to 40%: may not be important; 30% to 60%: moderate heterogeneity; 50% to 90%: substantial heterogeneity; 75% to 100%: considerable heterogeneity) [20]. If high levels of heterogeneity existed (I² ≥ 50% or p < 0.1), the possible reasons for heterogeneity were investigated and discussed. Additionally, sensitivity analyses were conducted using alternative outcome definitions i.e., broad, or specific definitions of HHF. Sensitivity analyses were also conducted using fixed-effect meta-analysis models.

The analyses conducted in this study were done in accordance with local laws and regulations and approvals from respective scientific/ethics/data protection committees were obtained. The country-specific analyses were conducted by independent academic/statistical groups within each country using a predefined statistical analysis plan, while the meta-analyses were conducted by IQVIA using the R language [21].