Comparative risk evaluation for cardiovascular events associated with dapagliflozin vs. empagliflozin in real-world type 2 diabetes patients: a multi-institutional cohort study

This retrospective cohort study analyzed data from Chang Gung Research Database (CGRD), the largest multi-institutional electronic medical records database in Taiwan [18]. The CGRD was established for research purposes in 2016, and currently covers 1.3 million patients (6% of the population of Taiwan). The CGRD includes clinical data for all patients who had outpatient treatment or were hospitalized in one of the 7 Chang Gung Memorial Hospitals from northern to southern Taiwan. The CGRD identifies diseases based on the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) before 2016, and ICD-10-CM afterwards. Unlike administrative data, the CGRD contains laboratory data which can provide a valid estimate of association between outcomes and exposures. The structures of CGRD have been described in another publication [18], and the accuracy and validity of the diagnostic codes in CGRD are well established [19‐21].

From the CGRD we identified type 2 diabetes patients newly treated with a SGLT2 inhibitor, either dapagliflozin or empagliflozin, during 2016–2017. The index date was defined as the date of SGLT2 inhibitor initiation. Because SGLT2 inhibitors were approved for type 2 diabetes management starting in 2016, all patients in this study could be considered as new users. We defined the baseline period as 1 year prior to the index date. Patients were excluded if they were less than 18 years old. Patients also needed to have had at least one clinical visit before the index date, in order to establish their baseline information. To ensure that cardiovascular events represented new cases, we excluded patients who had diagnoses of myocardial infarction, ischemic stroke and heart failure within 1 year prior to the index date. We excluded patients without laboratory examinations including body mass index (BMI), blood sugar controls (e.g., HbA1c), renal functions (e.g., estimated glomerular filtration rate; eGFR and urine albumin–creatinine ratio; UACR) or lipid profiles (e.g., low-density lipoprotein; LDL) before the index date, because these patients may not receive routine medical care in our hospitals. Finally, we also excluded patients who incurred cardiovascular events of interest within 30 days after the index date, considering that such a short-term exposure to SGLT2 inhibitors was unlikely to be the cause of cardiovascular events.

The primary outcome was the composite event of cardiovascular mortality, myocardial infarction, ischemic stroke and heart failure in the diagnosis of hospitalization and outpatient data. As secondary outcomes we evaluated the risk of the aforementioned individual events. Although studies have indicated that these cardiovascular outcome diagnosis codes have high positive predictive values (91.5–97.9%) in Taiwan’s healthcare settings [20, 22, 23], we also reviewed the electronic medical records from 3 Chang Gung Memorial Hospitals to assess the validity of the diagnosis codes corresponding to the cardiovascular events of interest. Specifically, we confirmed the heart failure diagnoses by laboratory examination data (e.g., B-type Natriuretic Peptide, BNP), left ventricular ejection fraction (LVEF) from echocardiogram as well as the symptoms of heart failure (e.g., dyspnea). We reviewed the medical charts in 222 cardiovascular events, and found the positive predictive values of 75.0%, 95.8%, 91.9% and 89.8% for cardiovascular mortality, myocardial infarction, ischemic stroke and heart failure, respectively. We performed intention-to-treat analysis and followed patients until the occurrence of the cardiovascular outcomes, death or last date of clinical records in the CGRD or December 31, 2018.

Approximately 40 potential confounding covariates based on previous studies and experts’ opinions were included in the regression models [24, 25]. We collected data on individual co-morbidities (i.e., hypertension and dyslipidemia, coronary heart diseases and microvascular complications) and computed composite scores of co-morbidities (i.e., Charlson comorbidity index [26]) at baseline. We also collected data of concomitant medications (i.e., statin, antiplatelet agents, anticoagulant agents, antihypertensive agents and antihyperglycemia agents) within 3 months before the index date. We included the most recent laboratory data before SGLT2 inhibitor initiation to evaluate the baseline cardiovascular risk. For the missing values of laboratory data, we performed multiple imputations using the Markov chain Monte Carlo method with a combination of 10 simulations [27, 28]. The details of co-morbidity and concomitant medications are described in Additional file 1: Table S1, Additional file 2: Table S2.

Negative control outcome could detect residual confounding and bias due to unobserved confounders [29, 30]. We examined the association of SGLT2 inhibitors with incident atrial fibrillation as a negative control outcome in this study population. Because there is currently no plausible mechanism by which SGLT2 inhibitors are associated with an altered risk of atrial fibrillation [25], any observed significant associations in this falsification test should be due to bias.

We defined the patients with empagliflozin as the reference group, and used multivariable Cox proportional hazards models to estimate the hazard ratios (HR) and 95% confidence intervals (CI) of outcomes. The variables used in the Cox proportional hazards models included age, sex, laboratory data, co-morbidities and concomitant medications (Table 1). All variables significantly related to outcomes from the prior uni-variate Cox regression model at alpha level of 0.1 were included by stepwise selection in the multivariate analyses. All analyses were two-tailed with a type I error level at 0.05, and were conducted using SAS Enterprise (Version 5.1; SAS Institute Inc., Cary, NC, USA).

We performed several sensitivity and subgroup analyses to create more homogenous groups for comparisons to test robustness of results from the main analyses. First, to evaluate the effects from non-adherence to the SGLT2 inhibitors, we performed per-protocol design and followed patients until the occurrence of cardiovascular outcome or discontinuation of the initial SGLT2 inhibitors (e.g., grace period over 90 days or switching to another SGLT2 inhibitor), whichever came first. Second, because patients who died from causes other than cardiovascular events (e.g., cancer) were no longer at risk of the study outcomes, this non-cardiovascular mortality was regarded as the major competing risk in the analyses. To address the possible competing risk events, we specifically compared rates of non-cardiovascular mortality between the empagliflozin and dapagliflozin groups. We found the rates of mortality were similar between the two groups (5.0 vs. 4.5 per 1000 person-years for empagliflozin vs. dapagliflozin, respectively), which implied that the effect of competing risk was minor. Nevertheless, we used the proportional subdistribution hazard models raised by Fine and Gray et al. [31] to manage possible competing risk of mortality. Third, we used three propensity score methods, including propensity score matching, stabilized inverse probability of treatment weighting (SIPTW) and standardized mortality ratio weighting (SMRW), to balance potential baseline differences between the empagliflozin and dapagliflozin groups. We describe patient characteristics before and after propensity score methods in Additional file 3: Table S3. Fourth, we increased washout periods up to 5 years for selection of patients without cardiovascular diseases to test the original analyses. Fifth, we excluded patients who incurred cardiovascular events of interest up to 1 year after the index date to avoid protopathic bias. Sixth, we repeated the analyses based on the specific SGLT2 inhibitor dose (i.e., low dose: dapagliflozin < 10 mg and empagliflozin < 25 mg; full dose: dapagliflozin 10 mg and empagliflozin 25 mg). Additionally, we conducted stratification analyses based on baseline risks for cardiovascular diseases, such as patients’ BMI levels (≥ 30 or < 30 kg/m²), blood sugar levels (HbA1c: ≥ 8.5 or < 8.5%), renal functions (eGFR: ≥ 90 or < 90 mL/min/1.73 m²; UACR: < 30 or ≥ 30 mg/g) and lipid profiles (LDL: ≥ 100 or < 100 mg/dL). Finally, we conducted additional analyses based on the specific concomitant medications which may modify the cardiovascular disease risk, including anti-platelet drugs, angiotensin-converting enzyme inhibitors (ACEI)/angiotensin receptor blockers (ARB) and loop diuretics.

We included a total of 12,681 patients newly initiating SGLT2 inhibitors, of which 5812 received dapagliflozin and 6869 received empagliflozin (Fig. 1). Patients newly receiving empagliflozin had worse renal functions at baseline (e.g., eGFR: 93.0 ± 31.6 vs. 97.7 ± 28.6 mL/min/1.73 m²; UACR: 242.9 ± 723.3 vs. 171.5 ± 502.7 mg/g). The mean baseline HbA1c levels (8.9 ± 1.6 vs. 8.8 ± 1.7%) and LDL levels (81.8 ± 23.0 vs. 81.5 ± 23.0 mg/dL) were comparable between dapagliflozin and empagliflozin new users. There were similar rates of co-morbidity between the SGLT2 inhibitors, but patients with empagliflozin had higher rates of hypertension (67.1% vs. 64.0%) and coronary heart disease (17.6% vs. 14.3%) at baseline. Greater use of anti-platelet drugs (30.2% vs. 26.7%), ACEI/ARB (60.1% vs. 55.7%), beta-blockers (24.6% vs. 21.5%), calcium channel blockers (40.4% vs. 35.4%) and loop diuretics (4.1% vs. 2.6%) was found in patients initiating empagliflozin than in those initiating dapagliflozin (Table 1).

We followed a total of 10,442 and 12,096 person-years with mean follow-up time of 656 days and 643 days for the dapagliflozin and empagliflozin group, respectively. There were 128 and 197 composite cardiovascular events in the dapagliflozin (incidence rate: 12.3 per 1000 person-years) and empagliflozin (incidence rate: 16.3 per 1000 person-years) groups, respectively. We found a similar risk of composite cardiovascular outcome (adjusted HR: 0.91; 95% CI 0.73–1.14), cardiovascular mortality (adjusted HR: 0.54; 95% CI 0.14–2.12), myocardial infarction (adjusted HR: 0.77; 95% CI 0.49–1.19) and ischemic stroke (adjusted HR: 1.15; 95% CI 0.80–1.65) between dapagliflozin and empagliflozin (Table 2). However, dapagliflozin use was associated with a lower risk of heart failure than empagliflozin (adjusted HR: 0.68; 95% CI 0.49–0.95).

In falsification testing, there were neutral associations between SGLT2 inhibitors and incident atrial fibrillation (adjusted HR: 1.08; 95% CI 0.73–1.60) (Additional file 4: Table S4).

The trends of results from sensitivity and subgroup analyses were consistent with the main analysis (Fig. 2). Specifically, we found the lower risk of heart failure associated with dapagliflozin was also statistically significant in several sensitivity analyses. For example, the observed effects were still found after we used SMRW analyses (adjusted HR: 0.69; 95% CI 0.49–0.98). We found the results of sensitivity analyses in the subgroups of patients receiving low SGLT2 inhibitor dose (adjusted HR: 0.91; 95% CI 0.53–1.55) or full SGLT2 inhibitor dose (adjusted HR: 0.73; 95% CI 0.41–1.28) and different baseline cardiovascular risks based on the clinical data, such as BMI ≥ 30 kg/m² (adjusted HR: 0.71; 95% CI 0.41–1.24) or BMI < 30 kg/m² (adjusted HR: 0.62; 95% CI 0.41–0.95), were consistent with the main analysis, in that patients receiving dapagliflozin had lower risk of heart failure compared to empagliflozin. Dapagliflozin reduced the risk of heart failure compared to empagliflozin regardless of whether or not anti-platelet drugs, ACEI/ARB and loop diuretics were used concomitantly.