Cardiovascular outcomes and mortality after initiation of canagliflozin: Analyses from the EASEL Study

The study included 2 comparator cohorts: new users of canagliflozin or new users of non‐SGLT2i on top of standard‐of‐care therapy. The non‐SGLT2i cohort included dipeptidyl peptidase‐4 (DPP‐4) inhibitors, GLP‐1 receptor agonists, thiazolidinediones, sulfonylureas, insulin, and other AHAs (acarbose, bromocriptine, miglitol, nateglinide and repaglinide) and excluded metformin. Patients with any exposure to any other SGLT2i (ie, empagliflozin or dapagliflozin) were excluded. New users were defined as patients whose first exposure to a non‐metformin AHA during the study period from 4/1/2013 to 12/31/2016 occurred ≥365 days after the start of observation in the database, with no prior exposure to any medication within the same AHA medication class in the prior 365 days, and the date of the first dispensing of the therapy of interest was considered the index date. Eligible patients with T2DM were required to have ≥1 year of observation before the index date, with established CV disease (including coronary artery disease, heart failure, cerebrovascular disease and peripheral artery disease), and be ≥18 years of age. Patients with type 1 diabetes mellitus or secondary diabetes mellitus were excluded from this study. Patients were followed from the index date until the first occurrence of any of the following: (a) outcome of interest, (b) death, (c) disenrollment from the DoD or (d) last observation in the database.

The above analytical design was prespecified in the study protocol, noting that patients who met the new user criteria for both treatment arms were eligible for inclusion in both cohorts, as of the date of earliest initiation of each treatment, specifically addressing the study design issues raised by others.,For these patients, baseline characteristics were independently assessed as of each index date, and patients were available for potential matching in both instances to an eligible subject from the other treatment arm. Patients who initiated canagliflozin and a non‐SGLT2i on the same day were excluded from the analysis. 31 32

Exposure propensity score (EPS) matching was used to reduce confounding due to imbalance in baseline covariates. A regularized logistic regression model was used to estimate the predicted probability of patients receiving canagliflozin, and canagliflozin new users were EPS‐matched to new users of non‐SGLT2i in a 1:1 ratio. Approximately 1000 variables were considered for inclusion in the model, including patient demographics and characteristics, duration of diabetes, baseline comorbidities and medication use, comprehensive diagnoses and procedures mapped to respective Clinical Classifications Software categories, a calculated Charlson Comorbidity Index (CCI) score and various healthcare resource utilization measures. No missing data imputation methods were applied in any calculation of prevalence rates for baseline covariates or incidence rates for the outcomes of interest. If a medical condition was not observed in the patient's record, then this condition was assumed not present. Baseline measures were assessed over 2 periods, the full pre‐index period spanning back to 1 April 2008, and a 1‐year pre‐index period, with the ability for all variables across both periods to be included in the final model. The number of unique baseline AHA medications was included in the EPS model to factor in differences in background AHA therapy. By design, the new use of other non‐SGLT2i defined the control group and necessitated specific prescriptions of these drugs before the index date not to be included in the EPS estimation to avoid multicollinearity. Procedure and diagnostic codes used to identify comorbidities have been validated in previous studies.,,,,,,,,,,,,,,Additional details of EPS matching have been published. 3 4 6 9 11 12 13 16 19 20 21 27 33 34 35 38

The primary outcome of the study was the composite of ACM and HHF. In addition, a composite of MACE (CV death, nonfatal MI and nonfatal stroke), an expanded MACE outcome that included HHF, a modified MACE that included non‐CV death (ACM, nonfatal MI and nonfatal stroke), and a composite of modified MACE + HHF, as well as the individual components of the composite end‐points, were evaluated. MI and stroke events were considered nonfatal if patients did not die during hospitalization for the index event. BKLE amputation was assessed as a safety end‐point and includes both minor (digits, partial foot and ankle disarticulation) and major (below‐knee) amputations.

ACM was defined as any record of death regardless of cause. To differentiate the cause of death, patients who died were linked with the National Death Index (NDI), which utilizes coroner records and other available sources to determine cause of death.22 CV death was defined using the standard recommended by the American Heart Association (International Classification of Diseases, Tenth Revision [ICD‐10] diagnostic codes for diseases of the circulatory system [I00‐I99] and congenital malformations of the circulatory system [Q20‐Q28]).2 MI, stroke and HHF were ascertained based on International Classification of Diseases, Ninth Revision (ICD‐9) and ICD‐10 diagnosis codes, and BKLE amputation was ascertained based on ICD‐9 and ICD‐10 procedure codes, consistent with our prior work (Table S1).38 Patients with a history of BKLE amputation events before the index exposure were excluded from comparative analyses of BKLE amputation to avoid confounding due to inherent intrasubject risk, potential for reverse causation and potential for immortal time bias in the situation in which such patients may no longer be at risk for future BKLE amputation events at the location of a prior amputation.

The statistical methods employed in this study were consistent with those described previously.Specifically, conditional Cox proportional hazards regression based on time to first event was used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs), comparing the treatment effect of canagliflozin against non‐SGLT2i (reference group) in relation to each study end‐point using both intent‐to‐treat (ITT) and on‐treatment approaches. For the ITT analysis, time at risk was calculated from the index date until the occurrence of an outcome of interest or the end of observation, whichever occurred first. It is worth noting for the on‐treatment analyses of patients in both cohorts that follow‐up was censored for one arm at the time of crossover from or to SGLT2i exposure unless follow‐up time had already been censored for another reason described above. Kaplan‐Meier plots were generated to characterize the contour of risk over time for each outcome. Because the results were generally consistent between both approaches, for the purpose of this reporting, we primarily focused on the ITT results, unless otherwise specified. Although the formal statistical analyses focused on the comparison of canagliflozin new users versus non‐SGLT2i new users, additional descriptive data (eg, event rates) were summarized based on individual non‐SGLT2i therapeutic classes (ie, DPP‐4 inhibitors, GLP‐1 receptor agonists, thiazolidinediones, sulfonylureas, insulin and other AHAs). 38

Due to the potential heterogeneity of non‐SGLT2i new users (eg, insulin use may represent an advanced stage of T2DM and sulfonylureas may be associated with heart failure‐related outcomes), sensitivity analyses were conducted to assess whether the study findings were driven by any particular subset of patients. As part of sensitivity analyses, patients receiving insulin, sulfonylureas and thiazolidinediones were removed (individually and collectively) from the non‐SGLT2i cohort along with their canagliflozin matching pairs to further evaluate treatment effect, as done previously.,Several subgroup analyses were prespecified, including sex, age, insulin use, GLP‐1 receptor agonist use, history of heart failure, recent HHF (past 12 months), number of CV risk factors (ie, CV disease, coronary artery disease, peripheral vascular disease), renal disease by CCI score and chronic renal disease. 7 14

The study protocol was reviewed and approved by the DoD Institutional Review Board, and all analyses were performed by a research organizationusing SAS V9.4 (SAS Institute Inc). 10

Overall, 7713 new users of canagliflozin and 102 516 new users of a non‐SGLT2i with T2DM and established CV disease were identified during the study period (Figure). There were 99 (1.3%) patients who started canagliflozin and a non‐SGLT2i on the same day and were excluded. After EPS matching, 7697 (99.8% of the total eligible) new users of canagliflozin were matched 1:1 with 7697 new users of a non‐SGLT2i, for a total of 15 394 patients. In this PS‐matched cohort, 888 new users were eligible for and assigned to both cohorts (ie, 521 were new users of non‐SGLT2i before canagliflozin and 367 were new users of non‐SGLT2i after canagliflozin). S1

Key clinical characteristics among new users of canagliflozin and non‐SGLT2i before and after matching are shown in Table 1. Before matching, patients in the canagliflozin cohort were younger and had longer durations of T2DM compared with the non‐SGLT2i cohort. By design and as expected, the canagliflozin cohort had greater usage of baseline AHA medications compared with the non‐SGLT2i cohort, and differences in non‐AHA medication use were observed between cohorts. Differences were also observed among most comorbidities of interest, with the canagliflozin cohort having lower prevalence of many baseline comorbidities, as well as lower CCI scores, compared to the non‐SGLT2i cohort.

After EPS matching, all baseline patient characteristics included in the EPS model were well balanced (standardized differences <0.1 for all baseline characteristics after propensity matching; Figure). Among the matched cohort, the mean age was 65.6 (standard deviation [SD], 9.3) years, 44.3% were female, the mean duration of T2DM was 5.4 (SD, 1.8) years, and the mean duration of CV disease was 4.2 (SD, 2.1) years. Histories of hypertension (86.2%), hyperlipidemia (75.4%), chronic obstructive pulmonary disease (22.3%), peripheral vascular disease (15.5%) and cerebrovascular disease (14.7%) were fairly prevalent, and 81.2% of patients were treated with metformin and 22.2% with insulin at baseline. S2

The median ITT follow‐up time was 2.03 years (interquartile range, 1.29, 2.82), which was similar between cohorts (2.00 and 2.08 years with canagliflozin and non‐SGLT2i, respectively). The median on‐treatment follow‐up time was 0.71 (interquartile range, 0.25, 1.49; 0.68; and 0.74 years with canagliflozin and non‐SGLT2i, respectively).

Among the 555 patients with an ACM outcome in this analysis, the cause of death for 552 (99.5%) patients was ascertained based on the NDI file. The remaining 3 patients, for whom a cause of death could not be determined, were excluded, along with their respective match, from comparative analyses that included CV and non‐CV death.

The primary composite outcome of ACM and HHF and secondary CV outcomes for patients in the EPS‐matched ITT cohort are shown in Figure 1. The incidence rate of the primary outcome was 1.79 versus 2.88 per 100 person‐years among new users of canagliflozin and non‐SGLT2i, respectively (HR, 0.61; 95% CI, 0.53‐0.71; P < .0001; Figure 1). Similarly, initiation of canagliflozin was associated with a lower rate of ACM (1.38 vs 2.15 per 100 person‐years; HR, 0.63; 95% CI, 0.53‐0.75; P < .0001) and HHF (0.51 vs 0.90 per 100 person‐years; HR, 0.57; 95% CI, 0.43‐0.74; P < .0001) compared with non‐SGLT2i. For these outcomes, the treatment benefit associated with canagliflozin started early and persisted over the study period (Figure 2).

Based on the NDI cause of death data, CV death was analysed as part of a composite outcome of MACE (CV death, nonfatal MI and nonfatal stroke). The rate of MACE was lower in new users of canagliflozin compared with new users of non‐SGLT2i (1.81 vs 2.41 per 100 patient‐years; HR, 0.74; 95% CI, 0.63‐0.86; P = .0001). The rates of the individual MACE components of nonfatal MI (0.56 vs 0.68 per 100 patient‐years; HR, 0.81; 95% CI, 0.61‐1.08; P = .16) and nonfatal stroke (0.51 vs 0.56 per 100 patient‐years; HR, 0.88; 95% CI, 0.64‐1.19; P = .40) were not significantly different. The rate of CV death was higher than developing either nonfatal atherosclerotic event in each treatment group. The rate of CV death was lower in new users of canagliflozin compared with new users of non‐SGLT2i (0.83 vs 1.31 per 100 patient‐years; HR, 0.63; 95% CI, 0.50‐0.78; P < .0001), and similar reductions were seen for non‐CV death (0.55 vs 0.83 per 100 patient‐years; HR, 0.65; 95% CI, 0.49‐0.85; P = .002) compared with non‐SGLT2i. In addition, the rate of a modified MACE outcome substituting ACM for CV death (ie, ACM, nonfatal MI and nonfatal stroke) was lower in new users of canagliflozin compared with new users of non‐SGLT2i (2.36 vs 3.27 per 100 patient‐years; HR, 0.71; 95% CI, 0.62‐0.82; P < .0001). Furthermore, the rate of the composite of MACE and HHF was significantly lower among new users of canagliflozin compared with new users of a non‐SGLT2i (2.19 vs 3.12 per 100 patient‐years; HR, 0.70; 95% CI, 0.60‐0.80; P < .0001). Consistent results were observed in a composite of modified MACE and HHF (2.75 vs 3.98 per 100 patient‐years with canagliflozin and non‐SGLT2i, respectively; HR, 0.68; 95% CI, 0.60‐0.77; P < .0001). In the on‐treatment analyses, lower event rates were generally seen among active canagliflozin patients (Figure S3).

Analysis of the primary outcome in prespecified subgroups showed a consistent benefit of canagliflozin treatment compared with non‐SGLT2i among each of the subgroups based on sex, age, insulin or GLP‐1 receptor agonist use in the past 12 months, history of heart failure, HHF in the past 12 months, number of cerebrovascular risk factors and renal disease, with no between‐subgroup heterogeneity detected (Figure). Results of sensitivity analyses that removed patients treated with insulin, sulfonylureas and thiazolidinediones at baseline, individually and in combination, were generally quantitatively consistent with the overall study results, suggesting that none of these medications were disproportionally impacting the final results (Figure). S4 S5

Excluding patients with previous BKLE amputation events (n = 6) and their respective matches, a total of 50 new BKLE amputation events were observed in the ITT cohort and 14 events in the on‐treatment cohort. In the ITT analysis, the incidence rate of BKLE amputation was not significantly different, with 29 and 21 events in new users of canagliflozin and non‐SGLT2i, respectively (0.18 vs 0.13 per 100 person‐years; HR, 1.44; 95% CI, 0.82‐2.52; P = .20; Figure 1). Similar results were observed in the on‐treatment analysis with 7 BKLE amputation events in each cohort (0.10 vs 0.08 per 100 person‐years with canagliflozin and non‐SGLT2i, respectively; HR, 1.26; 95% CI, 0.44‐3.55; P = .67; Figure S3). Because the number of events was relatively limited, the CI for the HR is quite wide and contains the point estimate that was observed in the CANVAS Program.23 Generally consistent results were observed among all prespecified subgroups (Figure S6).