Cardiovascular Outcomes with SGLT2 Inhibitors versus DPP4 Inhibitors and GLP-1 Receptor Agonists in Patients with Heart Failure with Reduced and Preserved Ejection Fraction

The study was approved by the Rutgers University Institutional Review Board, and the appropriate data use agreements were in place.

Data sources

Study subjects were drawn from Medicare insurance claims, a US federal program that provides healthcare to US citizens over 65 years of age. More specifically, we utilized a 50% random sample of Medicare fee-for-service beneficiaries enrolled in Part D from March 2013—coinciding with the approval of SGLT2i in the US—to December 2019. Data elements of interest included patient demographics, medical and pharmacy monthly enrollment status, inpatient and outpatient medical service use (International Classification of Disease [ICD], Ninth and Tenth Revisions; Current Procedural Terminology codes, Fourth Edition), and outpatient pharmacy dispensing data (drug name and strength, units dispensed, and days’ supply).

Study population and exposure definition

Within the database, a separate cohort was created for each pairwise comparison of SGLT2i versus an alternative non-gliflozin class. Cohort membership required patients to be new users of the study medications of interest (defined as no use of the medications in the 365-day washout period preceding medication initiation), be older than 65 years of age at cohort entry and have no evidence of gestational or type 1 diabetes (T1D), cancer, end-stage renal disease, or human immunodeficiency virus infection. With the sole exception of heart failure phenotype (see below), all baseline covariates including eligibility criteria and patient characteristics were assessed in the 365 days prior to the date of medication initiation.

The study cohort was further restricted to patients with the presence of HHF with ICD codes corresponding to HFrEF (ICD-9: 428.2× or ICD-10: I50.2×) or HFpEF (ICD-9: 428.3 × or ICD-10: I50.3×) in either the first or second position of the inpatient discharge diagnosis using all available lookback. The positive predictive value for this approach for identifying patients with HFrEF is 72% and 90% using ejection fraction [EF] thresholds of ≤ 40% and ≤ 50%, respectively, and 92% for HFpEF for an EF threshold of > 50% [19]. Patients with evidence of both or neither HF subtypes were excluded from analyses.

The study was comprised of four pairwise comparison cohorts, which included patients with: (1a) HFrEF initiating SGLT2i versus DPP4i; (1b) HFrEF initiating SGLT2i versus GLP-1RA; (2a) HFpEF initiating SGLT2i or DPP4i; and (2b) HFpEF initiating SGLT2i or GLP-1RA. For SGLT2i versus DPP4i comparisons, patients using combination empagliflozin–linagliptin therapy were excluded from analysis. Further, individuals initiating SGLT2i and the comparator on the same day were also excluded from analyses. Patients meeting the inclusion and exclusion criteria could contribute to each cohort only once, but the same patient could be included in more than one cohort.

Follow-up and study end points

Separately for each study outcome, patients began contributing to follow-up time on the day after cohort entry (i.e., medication initiation) up until the first occurrence of one of the following: end of pharmacy or health care eligibility, medication discontinuation defined as 60-day gap in treatment, medication switching (e.g., patients in SGLT2i arm initiating non-gliflozin therapy and vice versa), end of study data (December 2019), or the occurrence of the outcome.

The two primary outcomes of interest were (1) hospitalization for heart failure (HHF) (positive predictive value [PPV]: > 90%) [20], and (2) MI (PPV = 94%) or stroke (PPV = 85%) hospitalizations [21, 22]. Analysis for each of the two primary outcomes was conducted independently of the other.

Baseline covariates and inverse probability of treatment weighting

To mitigate risk of confounding, we assessed and adjusted for > 30 baseline covariates that were assessed in the 12-month period prior to and including the index date. These covariates included patient sociodemographics (e.g., age at medication initiation, biological sex, and race, calendar year), complications of diabetes (e.g., diabetic neuropathy, nephropathy, retinopathy), oral and injectable glucose lowering therapies (e.g., metformin, sulfonylureas, insulin), diagnosis of cardiovascular conditions (e.g., myocardial infarction, stroke, HF), and cardiovascular medication use (e.g., dispensing of β-blockers, loop diuretics, statins). Frailty status was ascertained using the claims based frailty index, and using a threshold of ≥ 0.25 to define frailty [23].

Propensity scores were estimated using a logistic regression that modelled the probability of initiating SGLT2i (exposure) versus a non-gliflozin medication (control) conditional on the baseline covariates. These propensity scores were then used to estimate stabilized inverse probability of treatment weights (IPTW) to account for imbalances in patient characteristics [24].

Statistical analysis

We assessed the performance of propensity scores based IPTW to control for confounding by examining the distribution of baseline covariates prior and after IPT weighting, and using a threshold of 10% in standardized difference as a metric for a meaningful imbalance [25]. Using an as-treated approach, where patients were censored on treatment discontinuation or switching, we estimated the rates of the primary outcomes among patients using SGLT2i (exposure) or non-gliflozin medications (control) by calculating the number of events and incidence rates (IRs). Adjusted incidence-rate differences (RD) and hazard ratios (HR) along with their 95% confidence intervals (CIs) were modelled through weighted Cox and Poisson regressions respectively.

Sensitivity and secondary analyses were conducted to assess the robustness of the study findings. First, we examined several secondary outcomes including a composite of the two primary outcomes (i.e., HF, MI or stroke hospitalizations), as well as individually examined MI hospitalizations, stroke hospitalizations, and all-cause mortality. Second, we conducted sensitivity analyses varying exposure-related censoring criteria, where instead of censoring patients at the time of treatment switching or discontinuation, we carried the index exposure forward to mimic an intention-to-treat approach with a maximum follow up truncated to 2 years.

Third, as our primary definitions to identify HF subtypes prioritize positive predictive values at the possible cost of lowered sensitivity (i.e., under-detection of patients with HF), we also employed alternative-more sensitive-HF definitions to identify HFrEF and HFpEF patients. More specifically, we allowed patients to be included in the study if they had presence of relevant HF codes in (1) any position of the inpatient discharge diagnosis, or (2) any inpatient or outpatient diagnoses fields. Fourth, we conducted sensitivity analyses where we excluded patients with a recent hospitalization (i.e., 30-days prior to the index date). Finally, to assess impact of the study estimates across calendar time, we also estimated stratified results before and after 2016. Other eligibility criteria (e.g., no evidence of T1D) were similar for all cohorts. For all cohorts, pairwise comparisons, and sensitivity analyses, the propensity scores were re-estimated, and stabilized inverse probability of treatment weights were re-calculated. All analyses were performed using SAS 9.4 (SAS Institute Inc, Cary, NC).