Methods
Model approach and description
An individual patient-level discrete-event simulation model was developed in Microsoft Excel to track patients’ risk of CV and renal events and adverse events (AEs) over their lifetimes when treated with empagliflozin, canagliflozin, dapagliflozin or SoC (figure 1). This approach was chosen based on a systematic literature review of approaches in modelling hard end points from clinical trials.21 Multiple events for each patient can be captured, with the risk of events changing over time dependent on the type of events previously experienced by the patient and their clinical characteristics (eg, age, hemoglobin A1c (HbA1c)).
Figure 1Diagram of the simulation model process.
The simulation began by generating a cohort of patients with T2DM and established CVD. Patients were duplicated and assigned to each therapy arm. Based on patients’ CV risk profiles, the model simulated nine possible CV or renal events that corresponded to end points in EMPA-REG OUTCOME and published data from CANVAS and DECLARE-TIMI 58: CV death, non-fatal MI, non-fatal stroke (the primary outcome in EMPA-REG OUTCOME, CANVAS, and DECLARE-TIMI 58 was a composite of these three events), HHF, progression of albuminuria (a primary outcome in CANVAS-R), a composite renal outcome (defined as a 40% reduction in estimated glomerular filtration rate, renal replacement therapy, or renal death), hospitalization for unstable angina (UA), transient ischemic attack (TIA), and revascularization (online supplemental table OS1).14 15 Recurrent non-fatal CV events were permitted in the model (eg, a simulated patient may experience more than one non-fatal MI), but renal events were considered non-recurring. Selected AEs in the model were GMI, AKI, LLA, bone fracture, and major hypoglycemic event.
For each simulated patient, the time-to-event for CV, renal events, and AEs were estimated. Then the model compared the timing of all events, and the earliest time determined which event happened first. When any non-fatal event occurred, the patient remained in the model and their treatment history, risk of future events, and time to next event were updated. The model process repeated to identify the next event. Non-fatal events could recur and influence the patient’s risk (or experience) of future events. If a fatal event occurred or the end of the time horizon was reached, the simulation of the patient ended, and the model moved to the next patient. For each patient, cumulative events per 100 patients-years (PYs), cumulative costs of management, life years, and quality-adjusted life years (QALYs) were tracked. Once all patients had been simulated on all treatments, the individual patient outcomes were aggregated to compute the mean population outcomes.
Population baseline characteristics
Individual patient profiles were created (see online supplemental file 1) based on the EMPA-REG OUTCOME trial population baseline characteristics previously published.14 Each sampled profile was duplicated, and identical copies were simulated for empagliflozin and each comparator (canagliflozin, dapagliflozin, and SoC) so that treatment comparisons captured observed incremental treatment effects, and were not influenced by differences in patient characteristics.
Risk equations
Time-dependent parametric survival analyses of the EMPA-REG OUTCOME trial data were conducted to characterise CV and renal event rates over time under SoC and empagliflozin. An individual patient-level risk equation was developed for each CV and renal event in the model using a systematic two-stage analysis. First, event-free survival (EFS) curves were fit to the trial data to describe the population-level occurrence of each CV and renal event. Second, individual patient-level estimates of risk were generated by testing baseline and time-dependent patient characteristics as potential predictors of the outcomes in parametric proportional-hazards regression analyses. Details on statistical analyses and risk equations included in the economic model are provided in online supplemental table OS2. To validate that the derived risk equations reproduced the overall event rates in the EMPA-REG OUTCOME trial when treated as competing events, the model was run for a 3-year time horizon to match the mean trial follow-up duration. Predicted 3-year HRs for empagliflozin versus SoC were congruent with the trial data (online supplemental table OS3).
US life table data were used to predict risk of non-cardiac death in simulated patients. An exponential-shaped EFS curve was assumed to estimate risk of AEs from published data.
Relative treatment effects
Head-to-head trial data were not available, thus treatment effects of SGLT2 inhibitors against the common placebo comparator were used to derive indirect estimates of the relative effect of canagliflozin versus empagliflozin and dapagliflozin versus empagliflozin using the indirect treatment comparison (ITC) method previously described by Bucher et al.22
The publications for EMPA-REG OUTCOME,14 CANVAS Program,15 and DECLARE-TIMI 5816 were used for the ITC. Outcomes from the CREDENCE trial (canagliflozin) were not used for comparison due to population differences.17 A standard process was followed to assess whether an ITC was feasible in terms of CV and renal outcomes (details in online supplemental file 1).
The feasibility analysis concluded the control arms could serve as the common comparator. In all CVOTs, use of SoC therapies was encouraged in line with local treatment guidelines, and not restricted to a specific type of SoC. Some differences were identified across the CVOTs with regard to inclusion criteria, baseline demographic and clinical characteristics, concomitant CV medications, history of CVD and outcome definitions (online supplemental table OS4 and table OS5). Mean age, percentage female, and most clinical characteristics (eg, HbA1c, BMI, SBP) were homogeneous across the CVOTs. There was heterogeneity across the CVOTs with respect to renal function, particularly between the EMPA-REG OUTCOME and DECLARE-TIMI 58 trials. Clinical history (prior PAD, MI, stroke, HF) was not consistently reported and showed some heterogeneity across the trials. Concomitant CV medications were generally similar between EMPA-REG OUTCOME and the CANVAS Program; some differences between EMPA-REG OUTCOME and DECLARE-TIMI 58 were observed in baseline treatment with beta-blockers and lipid-lowering therapy. The proportion of patients with established CVD at baseline varied from 100% in EMPA-REG OUTCOME, 65.6% in the CANVAS Program, and 40.6% in DECLARE-TIMI 58. Published subpopulation data were available from the CANVAS Program23 and DECLARE-TIMI 58 trial16 for patients with established CVD at baseline. Thus, it was possible to reduce the heterogeneity between the EMPA-REG OUTCOME trial and the CANVAS Program and DECLARE-TIMI 58 trial populations by using this subpopulation data for patients with baseline CVD to inform the ITC. Intent-to-treat (ITT) population data were used to derive relative efficacy parameters for scenario analyses. The definitions of clinical events were not identical, but the differences were modest and considered not to preclude the feasibility of an ITC.
HRs with 95% CIs for canagliflozin versus empagliflozin and dapagliflozin versus empagliflozin are shown in figure 2A,B, respectively. The survival functions for each of the CV and renal events were used to estimate risk of clinical events for empagliflozin, and this risk was adjusted for canagliflozin and dapagliflozin using the HRs. The AE rates for empagliflozin were similarly adjusted.
Figure 2HRs of event rates for sodium-glucose co-transporter-2 therapies versus empagliflozin. (A) canagliflozin versus empagliflozin, (B) dapagliflozin versus empagliflozin. Studies included in the indirect treatment comparison: EMPA-REG OUTCOME, CANVAS Program, and DECLARE-TIMI 58. CV, cardiovascular; CVD, cardiovascular disease; HF, heart failure; ITT, intent-to-treat; MI, myocardial infarction.
Quality of life
Published health utility scores were obtained from studies of patients with T2DM.24–26 Health-related utilities were computed by applying permanent event disutilities to a baseline utility value (online supplemental table OS6). As patients accumulated multiple clinical events, the total combined utility decrement was adjusted based on the number of events experienced to account for overlapping effects.26 QALYs consisted of the number of life years (ie, length of survival from model initiation to death or until the time horizon expires) weighted by the utility score associated with each of those years.
Costs and perspective
Direct costs were accrued in 2020 US$ (online supplemental table OS7). The model simulated commercially insured and Medicare populations separately and overall.
Treatment costs were based on published27 wholesale acquisition costs (WAC) of empagliflozin, canagliflozin, and dapagliflozin. Costs to the health plan were computed net of a US$35 patient co-pay28 and rebate (assumed to be 50% in commercially insured patients, 53% in Medicare patients, or 51% overall weighted based on patients in EMPA-REG OUTCOME). Pharmacy costs for SoC therapies and all other regular disease management and monitoring costs were assumed to be the same across regimens and were therefore not included in the model.
Acute costs of care for each clinical event were identified for commercial29–32 and Medicare31–33 payers and inflated to 2020 prices using the medical component of the US consumer price index.29 31 32 34 For each event, the model used an average of the commercial and Medicare costs, weighted by the per cent of patients below age 65 years at baseline. Non-CV death events were assumed to incur no costs.
Model assumptions
A few key additional modeling assumptions were made. First, changes in the risk of clinical events due to changes in treatment were implicitly captured in event rate trajectories. The statistical analyses of EMPA-REG OUTCOME data quantified associations among time-dependent risk factors. Event histories were predictors across other events, creating coupled, time-dependent risk equations (ie, as events accumulate, they can alter the risk of future events). Second, regardless of changes in event or treatment history, a constant treatment effect was assumed for each event. Proportional-hazards models assumed that the effect of the covariates on the hazard rate was the same at all times. Third, unmodeled comorbidities were assumed to not significantly influence the shapes of the statistical extrapolations or the role of specific risk predictors. The role of any baseline confounders not influenced by empagliflozin was minimized by the trial randomization process, which insured balance between treatment arms.
Model analyses
In the base case, a lifetime horizon was selected to fully capture costs and QoL associated with each treatment. Future costs and QALYs were discounted at a 3.0% annual rate. Relative clinical effects of canagliflozin and dapagliflozin versus empagliflozin for patients with baseline CVD in the CANVAS Program and DECLARE-TIMI 58 trial, respectively, were used. The analysis for empagliflozin versus canagliflozin excluded hospitalization for UA, TIA, and revascularization, because these were not published outcomes of the CANVAS Program, but included GMI, AKI, LLA, and bone fracture AEs. For empagliflozin versus dapagliflozin, the analysis excluded hospitalization for UA, TIA, revascularization, and progression of albuminuria, as these were not published outcomes in DECLARE-TIMI 58, but included GMI, AKI, and major hypoglycemic event AEs. All nine CV and renal events from EMPA-REG OUTCOME were included in the empagliflozin versus SoC analysis, plus GMI and AKI AEs. EMPA-REG OUTCOME data indicated that GMI and AKI occurred at significantly different rates (p<0.05) between treatment arms.
Deterministic sensitivity analyses were conducted to evaluate the robustness of the model inputs and assumptions. The model varied discount rates, empagliflozin treatment effect, and relative efficacy of comparators (notably, comparator HRs vs empagliflozin using their 95% CIs and ITT population data), utilities, and costs. A probabilistic sensitivity analysis was performed using distributions reflecting parameter uncertainties (online supplemental table OS8).35 Risk equation coefficients derived from EMPA-REG OUTCOME were varied using Cholesky decomposition, and the comparator HRs versus empagliflozin derived from ITCs were varied over their 95% CIs using a lognormal distribution. The model produced 1000 pairs of incremental effectiveness and cost estimates. Scenario analyses assessed the impact of shorter time horizons (1, 3, 5, and 10 years).