New heart failure and heart attack cases in users of SGLT2 inhibitors compared to DPP-4 inhibitors

This study was approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster and from The Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee. It included type 2 diabetes mellitus patients with SGLT2I or DPP4I prescriptions from 1 January 2015 to 31 December 2020. Patients who received both DPP4I and SGLT2I, in addition to patients who discontinued the medication during the study, were excluded. The exclusion criteria for the HF study cohort were as follows: patients with prior HF diagnosis or with the use of medications for HF (e.g. diuretics for HF and beta‐blockers for HF). For the MI study cohort, patients with prior old MI or MI diagnosis were excluded. The patients were identified from the Clinical Data Analysis and Reporting System (CDARS), a territory‐wide database that centralizes patient information from individual local hospitals to establish comprehensive medical data, including clinical characteristics, disease diagnosis, laboratory results, and drug treatment details. The system has been previously used by both our team and other teams in Hong Kong to conduct population‐based cohort studies,18, 19 including those on diabetes mellitus.20, 21

Clinical and biochemical data were extracted from CDARS for the present study. Patients' demographics include sex and age of initial drug use (baseline). Prior comorbidities before initial drug use were extracted, including diabetes with chronic complication, diabetes without chronic complication, gout, hypertension, ischaemic heart disease, liver diseases, peripheral vascular disease, renal diseases, stroke/transient ischaemic attack, atrial fibrillation, ventricular tachycardia (VT)/ventricular fibrillation (VF)/aborted sudden cardiac death (SCD), anaemia, overweight, and cancer. Charlson's standard comorbidity index was also calculated. Mortality was recorded using the International Classification of Diseases Tenth Edition (ICD‐10) coding, whilst the study outcomes and comorbidities were documented in CDARS under ICD‐9 codes. The ICD codes used to search for diagnoses and outcomes are shown in Supporting Information, Table S1.

Non‐SGLT2I/DPP4I medications were also extracted, including metformin, sulphonylurea, insulin, acarbose, thiazolidinedione, glucagon‐like peptide‐1 receptor agonists, and statins and fibrates. A limited number of enrolled patients have been prescribed calcium channel blockers; thus, they were not considered. Baseline laboratory data were extracted. Subclinical biomarkers were calculated accordingly, including neutrophil‐to‐lymphocyte ratio, platelet‐to‐lymphocyte ratio, neutrophil‐to‐high‐density lipoprotein ratio, lymphocyte‐to‐high‐density lipoprotein ratio, lymphocyte‐to‐low‐density lipoprotein ratio, low‐density lipoprotein ratio‐to‐high‐density lipoprotein ratio, total cholesterol‐to‐high‐density lipoprotein ratio, triglyceride‐glucose index, bilirubin‐to‐albumin ratio, protein‐to‐creatinine ratio, and prognostic nutritional index.

Standard deviation (SD) was calculated for glycaemic and lipid profile parameters once there are at least three examinations for each patient since initial drug exposure of SGLT2I or DPP4I. We also calculated more specific variability measures for HbA1c and fasting glucose profiles including SD, SD/initial, coefficient of variation (CV), and variability independent of mean as listed in Supporting Information, TableS2.

The study outcomes are new‐onset HF, and new‐onset MI, cardiovascular mortality, and all‐cause mortality as defined by the first incidence of ICD‐9 codes of these adverse events (Supporting Information, Table S1). Mortality data were obtained from the Hong Kong Death Registry, a population‐based official government registry with the registered death records of all Hong Kong citizens linked to CDARS. ICD‐10 codes I00–I09, I11, I13, and I20–I51 were used to identify cardiovascular mortality. Descriptive statistics are used to summarize baseline clinical and biochemical characteristics of patients with SGLT2I and DPP4I use. For baseline clinical characteristics, the continuous variables were presented as median (95% confidence interval [CI]/interquartile range [IQR]) or mean (SD) and the categorical variables were presented as total number (percentage). Continuous variables were compared using the two‐tailed Mann–Whitney U test, whilst the two‐tailed χ² test with Yates' correction was used to test 2 × 2 contingency data. Univariable Cox regression was used to identify significant predictors for the primary and secondary outcomes. Propensity score matching was performed to generate control of SGLT2I users to compare against DPP4I users in a 1:1 ratio based on baseline age, sex, prior comorbidities, and non‐SGLT2I/DPP4I medications using nearest neighbour matching strategy.

Multivariable Cox models adjusting for significant risk factors of demographics, past comorbidities, non‐SGLT2I/DPP4I medications, subclinical biomarkers, HbA1c, and fasting glucose to identify the treatment effects of SGLT2I vs. DPP4I on the mentioned adverse outcomes. Cause‐specific and subdistribution hazard models were conducted to consider possible competing risks. Lastly, subgroup analyses were done on age (≤65 and >65 years) and sex on drug exposure effects. A standardized mean difference (SMD) of no <0.2 between the treatment groups post‐weighting was considered negligible. The hazard ratio (HR), 95% CI, and P‐value were reported. Statistical significance is defined as P‐value < 0.05. The statistical analysis was performed with RStudio software (Version 1.1.456) and Python (Version 3.6).

In this study, patients with type 2 diabetes mellitus and use of either SGLT2I or DPP4I from 1 January 2015 to 31 December 2020 were included (Table1). Patients with the use of both classes, or with prior HF diagnoses or admissions due to HF or with anti‐HF drugs (e.g. beta‐blockers for HF and diuretics for HF), were excluded. After exclusion, 41 994 patients (58.89% males, median admission age at 58 years old, IQR: 51.2–65.3) fulfilled the eligibility criteria in the study cohort for subsequent analysis (Figure1). The study cohort has a median follow‐up duration of 5.6 years (IQR: 5.32–5.82). Propensity score matching (1:1) between SGLT2I and DPP4I users using the nearest neighbour search strategy with a 0.1 caliper was performed (Supporting Information, Figure S1). Bootstrapping procedures were performed for propensity matching estimates, and the estimations of bootstrapped standard error (replications = 50) were <0.001. Together, these indicated no significant confounding characteristics remained significant after propensity matching.

The cumulative incidence curves for new‐onset HF, MI, cardiovascular mortality, and all‐cause mortality stratified by SGLT2I or DPP4I use for the matched cohort are shown in Figure2. Lower incidences of all of these outcomes were observed for SGLT2I users compared with DPP4I users. Univariable Cox regression was applied to identify significant predictors of the study outcomes (Supporting Information, TablesS2 and S3). In the matched cohort, SGLT2I use was associated with significantly lower risks of new‐onset HF (HR: 0.52, 95% CI: [0.48, 0.58], P < 0.0001), new‐onset MI (HR: 0.60, 95% CI: [0.54, 0.66], P < 0.0001), cardiovascular mortality (HR: 0.23, 95% CI: [0.18, 0.28], P < 0.0001), and all‐cause mortality (HR: 0.23, 95% CI: [0.21, 0.26], P < 0.0001). Multivariable Cox models were developed adjusting for significant demographics, past comorbidities, and medications (Table2). SGLT2I use remained a significant predictor of all four study outcomes (HR < 1, P < 0.001).

To evaluate the predictiveness of the models, different sensitivity analyses were performed. Firstly, a 1 year lag time between treatment initiation and study outcomes was applied (Supporting Information, TableS4). Secondly, competing risk analyses using cause‐specific hazard models and subdistribution hazard models were applied (Supporting Information, TableS5). Thirdly, different propensity score approaches were used to evaluate the effects of the matching approach on the analysis, including propensity score stratification, inverse probability of treatment weighting (IPTW), and stable inverse probability of treatment weighting (SIPTW) (Supporting Information, TableS6). All of these analyses demonstrated that SGLT2I use was associated with lower risks of new‐onset HF, MI, cardiovascular mortality, and all‐cause mortality.

There are several limitations to the present study. Firstly, inherent information bias with a risk of under‐coding and coding errors should be noted, given its observational and retrospective nature. However, the difference in patient characteristics, past comorbidities, and other medication usages between SGLT4I/DPP4I users and controls was addressed through matching using propensity scores, although residual bias may remain. There are also patients with missing data for the laboratory parameters because not all blood tests were routinely performed for all. Moreover, we were unable to access important lifestyle predictors for cardiovascular adverse events, such as body mass index, smoking, and alcoholism. Thirdly, coding for clinical diagnoses of HF was used but echocardiographic data are not coded in the administrative database, and therefore, different types of HF based on ejection fraction could not be examined. Finally, DPP4I use is associated with an increased risk of HF compared with placebo, and therefore, this study could not distinguish between whether gliptins cause HF and whether SGLT2I reduce HF.