Association between sodium–glucose cotransporter-2 inhibitors and arrhythmic outcomes in patients with diabetes and pre-existing atrial fibrillation

We conducted a retrospective, population-based cohort study of adults with DM and AF between 1 June 2014 and 31 March 2019, in Alberta, Canada, a province of ∼4.4 million people served by a single healthcare system,⁶ using a prevalent new-user design (see the Propensity score matching section).⁷ All Alberta residents are eligible for public health insurance, and >99% participate in the government-sponsored insurance plan, which covers physician visits and hospital-based care—but does not universally cover pharmaceuticals.⁸ Each resident is assigned a personal health number, which acts as a unique lifetime identifier that enables the linkage of administrative health data maintained by Alberta Health.⁹

We accessed de-identified data from the Interdisciplinary Chronic Disease Collaboration Data Repository,¹⁰ which contains linked administrative health, pharmacy, and laboratory patient-level data and includes demographic characteristics, vital statistics, physician claims, hospital admissions, emergency department and ambulatory care visits, laboratory results, and dispensed prescription data for the entire adult (≥18 years) population.

We identified a population-based cohort of adults (≥18 years) with both DM and AF as of 31 March 2017, leaving at least 2 years for outcome ascertainment. Eligible patients were identified if they had one hospitalization or two physician claims at least 30 days apart for both diabetes (excluding gestational diabetes) and AF, using previously validated administrative coding algorithms.¹¹ Patients without an indication for an SGLT2i were excluded: (i) Stages G4 and G5 CKD (estimated glomerular filtration rate [eGFR] < 30 mL/min/1.73 m²), (ii) prior lower limb amputation, and (iii) HbA1C < 7.5 who were taking only metformin (defined as taking metformin with no other agents ±365 days from the study start date). We also excluded individuals with Type I DM using a previously validated method based on the sole prescription of insulin monotherapy.¹²

The study cohort comprised adults with DM and AF who were dispensed SGLT2i (canagliflozin, dapagliflozin, or empagliflozin) or dipeptidyl peptidase-4 inhibitor (DPP4i) (alogliptin, linagliptin, saxagliptin, or sitagliptin) that was available in Canada during the study period. Patients were classified into one of two mutually exclusive groups: (i) patients using SGLT2i alone or in combination with other non-DPP4i diabetic drugs or (ii) patients using DPP4i alone or in combination with other non-SGLT2i drugs. Patients simultaneously starting SGLT2i and DPP4i on the same date were excluded. Dipeptidyl peptidase-4 inhibitor exerts hypoglycaemic effects by increasing incretin levels [glucagon-like peptide-1 (GLP-1) and gastric inhibitory peptide] and inhibits glucagon release, which, in turn, increase insulin secretion, decrease gastric emptying, and decrease blood glucose levels. Dipeptidyl peptidase-4 inhibitors were chosen as an active comparator agent to reduce the risk of immortal time bias in the study design,¹³ and since they were similarly considered a second-line or third-line agent for glucose-lowering during the study period, they demonstrated cardiovascular safety and were thought to have neutral cardiovascular effects.^14–16

This study used a prevalent new-user cohort design described in previous Canadian pharmaco-epidemiology studies of SGLT2i outcomes.^7,17,18 In brief, we included both prevalent new users of SGLT2i (i.e. people who previously used DPP4i) and incident new users of SGLT2i (i.e. people who had not previously used DPP4i) matched to DPP4i (Figure 1). The index date of outcome ascertainment was defined by the date when SGLT2i was dispensed or the corresponding date when DPP4i was dispensed in the matched exposure. Patients were followed up until death, emigration out of province, termination of registration with Alberta Health, or until the end of the study period (31 March 2019).

We first created exposure sets of potential matches defined by the level of diabetes therapy based on the number of diabetic medications, duration of DPP4i treatment for prevalent new users, and calendar time (DPP4i prescription within 180 days of the SGLT2i initiation). We matched incident SGLT2i users to incident DPP4i users who were initiated treatment in the same period, whereas we matched patients switching from a DPP4i to an SGLT2i (prevalent new users) to patients treated with DPP4i for a similar duration in their exposure sets.

Propensity scores were then constructed separately for incident and prevalent new users to estimate the propensity of receiving an SGLT2i vs. a DPP4i. We used logistic regression to model the association of the following covariates and receiving an SGLT2i vs. a DPP4i: age, sex, diabetes duration, AF duration, comorbidities (myocardial infarction, cirrhosis, CKD, HF, hypertension, thyroid disease, peripheral artery disease, stroke, cancer, and peripheral vascular disease), anti-hyperglycaemic medications (insulin, GLP-1 receptor agonists, alpha-glucosidase, meglitinides, metformin, sulfonylureas, thiazolidinediones), AF medications (oral anticoagulants, beta-blockers, calcium channel blockers, and anti-arrhythmics), CHADS₂ score, and health services utilization (number of diabetes medications, number of hospital visits, and number of outpatient consultations). Sodium–glucose cotransporter 2 inhibitor users were matched 1:1 using the nearest neighbour method with replacement to DPP4i users and using a caliper width of 0.2 SD of the log of the propensity score.

Baseline comorbidities were ascertained from administrative data based on previously validated algorithms over a 2-year look-back period.^11,19 We used the Pampalon Deprivation Index as a comprehensive indicator of socioeconomic status, which is a small-area–based composite index derived from census data that include employment status, income, education, marital status, single-parent status, or living alone.^20,21 The material index reflects deprivation of wealth, goods, and conveniences, and the social index reflects deprivation of relationships among individuals in the family, the workplace, and the community. Each index stratifies each dissemination area (i.e. smallest standard census area) into quintiles, from 1 (least deprived) to 5 (most deprived), and is assigned to individuals in the cohort based on postal code.²²

Medications were identified using Anatomical Therapeutic Chemical codes and included rate control medications for AF (i.e. beta-blockers excluding sotalol, non-dihydropyridine calcium channel blockers, and digoxin), anti-arrhythmic medications (i.e. Vaughan Williams Class IA and IC agents, amiodarone, and sotalol), anticoagulants (i.e. warfarin, rivaroxaban, apixaban, dabigatran, and edoxaban), and anti-hyperglycaemic medications (i.e. metformin, insulin, sulfonylureas, alpha-glucosidase inhibitors, meglitinides, GLP-1 agonists, SGLT2i, and DPP4i).

The primary outcome was an ‘AF event’ or a clinically relevant AF-related medical resource encounter. Atrial fibrillation events were defined as the first occurrence of an AF-related hospitalization, AF-related emergency department visit, synchronized electrical cardioversion, or catheter ablation. Secondary outcomes included all-cause mortality, all-cause hospitalization, HF hospitalization, and ischaemic stroke or transient ischaemic attack (TIA). Administrative definitions and data sources used to define each outcome are listed in Supplementary material online, Table S1.

Descriptive statistics were used to summarize covariates by the propensity-matched treatment group. Continuous variables were reported as means and standard deviations, and categorical variables were reported as proportions. The balance of covariates was assessed based on standardized differences with a threshold of 10%.

All analyses were conducted using an intention-to-treat principle. Kaplan–Meier curves were used to visualize the cumulative incidence of primary and secondary outcomes over time, and log-rank tests were used to compare the survival distribution by treatment groups. The relationship between treatment (SGLT2i vs. DPP4i comparator) and study outcomes was assessed using Cox regression models without competing risks to calculate hazard ratios (HRs) and 95% confidence intervals (95% CI). Cox proportional hazard modelling was chosen to facilitate clinical interpretability of HR.²³ Adjusted multivariable Cox proportional hazard regression was used to balance and control for covariates with a standardized difference >10% following matching.²⁴ Time zero was set to the date of the first SGLT2i prescription or the date of the DPP4i prescription in the corresponding time-conditional exposure sets. We tested the proportional hazards assumption for the Cox proportional hazards model by examining the Schoenfeld residuals and visual assessment of log–log plots. The proportional hazards assumption was met across models. The analyses were repeated within pre-defined subgroups including female sex, HF, CKD, and the baseline use of anti-arrhythmic medications.

As a secondary analysis, Fine–Gray models were used for non-fatal outcomes to account for the competing risk of death.²⁵ In addition to the time-to-event analyses, we conducted recurrent event analyses of the primary study endpoint using several approaches including (i) the Andersen and Gill model; (ii) the Prentice, Williams, and Petersen total time model; and (iii) a random effects approach using a frailty model.^26–28

To evaluate the potential for residual confounding, falsification analysis was conducted by ascertaining the relationship between the treatment group and falsification endpoints that a priori would not be expected to be associated with the effects of treatment. For this study, we considered incident rheumatoid arthritis, chronic obstructive pulmonary disease, and lymphoma. Statistical analyses were performed with SAS, version 9.4 (SAS Institute Inc.), and Stata/IC, version 14.2 (StataCorp LP).

Among 20 739 patients with concomitant DM and AF and without contraindications to SGLT2i, there were 1170 patients prescribed an SGLT2i and 1788 prescribed a DPP4i during our accrual period (Figure 2). Following 1:1 matching with replacement, the final propensity-matched cohort included a total of 2242 patients (1121 in each group) followed for a median of 3.0 years (inter-quartile range 2.1–4.1).

The mean age of the overall matched cohort was 66 years, 26% were female, and the mean CHADS₂ score was 2.3 (Table 1) with 53% on anticoagulation for stroke prevention. The most common comorbidities in both groups included hypertension (93%), CKD (83%), and HF (40%). The majority of baseline characteristics between the SGLT2i and DPP4i groups was balanced (i.e. standardized difference <10%) after propensity score matching. Notably, the mean duration of AF (i.e. time from AF diagnosis to the initiation of SGLT2i or DPP4i) was similar between groups (6.6 ± 4.8 years in the SGLT2i group vs. 6.9 ± 4.8 years in the DPP4i group; standardized difference 5.7%). The proportion of patients on anti-arrhythmic drugs at baseline were also similar between groups.

However, compared with the DPP4i group, patients in the SGLT2i group were younger (64.8 vs. 68.0 years). Additionally, the SGLT2i group had fewer patients with a history of HF (35.1% SGLT2i vs. 44.1% DPP4i), peripheral artery disease (3.7 vs. 6.9%), and ischaemic stroke/TIA (18.2 vs. 23.4%).

Over a median follow-up of 3.0 years, 9.3% (n = 209) of the matched cohort experienced the primary outcome of an AF event or clinically relevant AF-related medical resource encounter (Figure 3). Compared with the DPP4i group, there was a significantly lower risk of experiencing an AF event among patients in the SGLT2i group (adjusted HR 0.73, 95% CI 0.55–0.96; P = 0.03; Table 2).

In the competing risks model, a similar relationship between the primary outcome and the treatment group was observed (see Supplementary material online, Table S2). That is, SGLT2i decreased the cumulative incidence of the primary outcome of AF events compared with DPP4i, although this relationship was not statistically significant.

In the recurrent events analysis, the SGLT2i had fewer recurrent AF events compared with the DPP4i group (Andersen–Gill model–adjusted HR 0.61, 95% CI 0.50–0.74; P < 0.001), mainly driven by a reduction in AF-related emergency department visits (adjusted HR 0.61, 95% CI 0.49–0.78; P < 0.001) and synchronized electrical cardioversions (adjusted HR 0.55, 95% CI 0.40–0.76; P < 0.001). This relationship between the treatment group and recurrent events was consistent using the Prentice–William–Peterson total time and frailty models (see Supplementary material online, Table S3).

There were 455 (20.2%) deaths and 101 (4.5%) ischaemic strokes or TIAs observed among the matched cohort. There were a total of 1134 hospitalizations during the study follow-up period, of which 213 were primarily due to HF (Figure 4).

Compared with the DPP4i group, patients in the SGLT2i group had a significantly lower risk of all-cause mortality (adjusted HR 0.22, 95% CI 0.16–0.28; P < 0.001), all-cause hospitalization (adjusted HR 0.65, 95% CI 0.58–0.74; P < 0.001), and HF hospitalization (adjusted HR 0.53, 95% CI 0.40–0.71; P < 0.001). There was no difference in ischaemic stroke/TIA between the SGLT2i and DPP4i groups (adjusted HR 0.71, 95% CI 0.48–1.08; P = 0.11; Table 2). Similar relationships were observed in the competing risk models (see Supplementary material online, Table S2). That is, SGLT2i significantly decreased the cumulative incidence of all-cause hospitalization and HF hospitalization compared with DPP4i. Sodium–glucose cotransporter 2 inhibitor use had no significant effect on the subdistribution hazard function of ischaemic stroke/TIA.

The association between SGLT2i prescription and outcomes was assessed in several subgroups (Figure 5). In patients with HF at baseline (n = 887), the SGLT2i group had fewer AF events compared with the DPP4i group (adjusted HR 0.61, 95% CI 0.39–0.95; P = 0.03). A similar benefit was observed in females (n = 581; adjusted HR 0.59, 95% CI 0.36–0.98; P = 0.04), and there was no significant difference in the primary outcome between treatment groups among patients on anti-arrhythmic medications at baseline or patients with CKD.

In terms of secondary outcome measures, all subgroups (i.e. HF, CKD, female sex, and baseline anti-arrhythmic medications) prescribed an SGLT2i had a significant reduction in all-cause mortality, all-cause hospitalization, and HF hospitalization compared with patients prescribed a DPP4i (see Supplementary material online, Table S4). There were no differences between treatment groups in the rate of TIA/stroke.

The associated treatment–exposure relationships were neutral for the falsification endpoints of chronic obstructive pulmonary disease and lymphoma, which decreases the likelihood of persistent bias. The null hypothesis of neutral association for rheumatoid arthritis was rejected, indicating possible persistent bias; however, the association was in favour of the DPP4i group, which was associated with the lower incidence of rheumatoid arthritis (see Supplementary material online, Table S5).

Our findings need to be interpreted in the context of several study limitations. Given the non-randomized, retrospective design of the study, we were unable to exclude bias from unmeasured confounders that may influence the association between SGLT2i and study endpoints [such as symptom burden, baseline AF arrhythmic burden, AF subtype (paroxysmal, persistent), or imaging markers of AF recurrence such as left atrial size or left ventricular ejection fraction]. To mitigate bias from measured confounders as best as possible given the observational study design, we used time-conditional propensity score matching and double robust estimation to create comparator cohorts balanced over a large number of measurable variables. Additionally, we compared the SGLT2i cohort with an active comparator (i.e. DPP4i) to reduce the potential for immortal time bias. Nonetheless, we noted a greater-than-expected reduction in mortality associated with the SGLT2i group than previously reported,⁴³ suggesting the presence of residual confounders. Second, we did not specifically assess for differences in arrhythmia outcomes by individual SGLT2i (canagliflozin, dapagliflozin, or empagliflozin). Although SGLT2i benefit has been consistently observed in other cardiovascular outcomes across this drug class, we acknowledge the possibility that arrhythmic benefit may not be a medication class effect. Third, our study cohort (that included patients with filled SGLT2i prescriptions) represented a small proportion of the eligible patients for SGLT2is, which may reflect a clinical practice gap and delayed adoption of novel therapies. In this context, our study findings may be generalizable only to patients with similar baseline characteristics; for instance, our cohort was relatively young (mean age 64 years) compared with other reported AF cohorts.⁴⁴ Fourth, the primary study outcome of an AF event relied on administrative claims rather than the traditional definition of >30 s of atrial tachyarrhythmia or AF recurrence. While the current study outcome would not be sensitive to detect all episodes of AF (i.e. short duration or minimally symptomatic), AFs defined through healthcare encounters are clinically meaningful as they represent episodes requiring medical attention due to severity of symptoms or associated clinical decompensation. Our administrative definition of an AF event is consistent with that of prior studies.^45–47 Furthermore, we would not expect differential detection or diagnostic bias between treatment groups. However, we may have underestimated the potential benefits of SGLT2is on AF burden reduction if continuous electrocardiogram monitoring was used. Fewer AF events may have led to an underpowered analysis, which may account for the lack of statistical significance in the competing risks model evaluating the association between SGLT2i and primary endpoint. Lastly, we noted that overall anticoagulation use was relatively low in our cohort, which may have influenced the rate of stroke during the follow-up period. However, anticoagulation use was similar between the SGLT2i and the DPP4i groups, and the overall low rate of anticoagulation would not be expected to influence the study findings that there may be a benefit of SGLT2i on stroke/TIA.

The data underlying this article will be shared on reasonable request to the corresponding author.