Model approach
A Microsoft Excel®-based lifetime Markov cohort model with monthly cycles comparing empagliflozin plus SoC vs. SoC alone was developed to simulate patients’ progression through health states based on the Kansas City Cardiomyopathy Questionnaire Clinical Summary Score (KCCQ-CSS; Fig. 1). The KCCQ score is an established and prognostically important patient-reported measure of health status in HFrEF [10–13]. Scores range from 0 to 100, with higher scores indicating better health status, lower symptom burden and better HRQoL. The clinical severity of HFrEF was modelled by using five health states, based on KCCQ-CSS quartiles, and death.
Several factors motivated the selection of a Markov cohort approach. Such models employ a simple structure and allow for rapid execution of analyses yet possess sufficient flexibility to account for patient heterogeneity through subgroup analyses and can capture both the short- (e.g., episodes of HHF) and long-term impacts of treatment (e.g., disease progression, as reflected in transitions between KCCQ quartiles). This approach also aligns with a recent health economic evaluation of dapagliflozin in HFrEF [14], which concluded that the KCCQ tool was a reasonable measure to classify disease severity and was appropriate for decision-making.
Model description
The modelled cohort was partitioned across the four KCCQ-CSS health states at baseline. Each cycle, patients could remain in the same state, progress to a lower (worse) KCCQ-CSS state, regress to a higher (better) KCCQ-CSS state, or die, and could experience HHF or treatment-related adverse events (AEs) as transient events while in a KCCQ-CSS state. The model monitored patients’ HRQoL and resource use over time as they transitioned between health states and tracked CV- and non-CV-related deaths. Treatment with empagliflozin plus SoC and SoC alone was assumed to impact patient outcomes by influencing the likelihood of transitions between KCCQ-CSS health states, which in turn affected the risk of mortality and HHF and the accrual of both costs and quality-adjusted life years (QALYs). Modelled SoC therapies represented treatments patients in the EMPEROR-Reduced trial received, including diuretics, inhibitors of the renin-angiotensin system and neprilysin, beta-blockers, mineralocorticoid receptor antagonists, and cardiac devices (when indicated). Patients for whom SoC was complemented by empagliflozin gradually discontinued empagliflozin therapy over time, after which they received SoC alone until death or the model horizon expired.
The model structure was the same for the UK, Spain and France, although the inputs used to inform each analysis were country-specific (Supplementary Appendix Section S1). Direct medical costs consisted of expenditures related to drug acquisition, clinical event management, and disease management. Utilities were accrued based on time spent in each KCCQ-CSS state, adjusted for disutilities associated with HHF and AEs. Future costs and benefits were discounted at annual rates of 3.5%, 3.0% and 2.5% for the UK [15], Spain [16] and France [17], respectively. The primary model outcome was the incremental cost-effectiveness ratio (ICER), expressed as cost per QALY gained, measuring the additional cost to gain one QALY by treating patients with empagliflozin plus SoC vs. SoC alone.
Population
The modelled cohort was representative of patients in the EMPEROR-Reduced trial [8]. Patients were initially distributed into KCCQ-CSS quartiles as follows: 0 to < 55 (24.3%), 55 to < 75 (25.1%), 75 to < 90 (27.2%) and 90 to 100 (23.4%). The mean starting age of the cohort was 67 years and 76% were male. Half the patients had history of type 2 diabetes (T2D), 48% had an estimated glomerular filtration rate (eGFR) less than 60 mL/min/1.73 m2, and 20% were receiving an angiotensin receptor-neprilysin inhibitor (ARNi) as part of SoC.
The model was equipped to perform a range of pre-specified subgroup analyses based on patient age at baseline (< 65 vs. ≥65 years), presence or absence of T2D, eGFR (< 60 vs. ≥60 mL/min/1.73 m2) and utilisation or non-utilisation of ARNi.
Disease and treatment effect
Disease progression or regression
EMPEROR-Reduced trial data were used to estimate transition probability matrices to capture improvement (ascent) or progression (descent) of disease through KCCQ-CSS quartiles. KCCQ-CSS quartiles were used to ensure enough patient data for robust statistical analysis and sufficient granularity in predicting movement between health states. Treatment-specific monthly transition probabilities between quartiles were computed based on transition count data. Independent transition matrices were developed based on the first three months of follow-up in EMPEROR-Reduced, from month 4 to 8, and from month 9 onward, recognising inflection points observed in the data (Supplementary Appendix Section S3). Transition probabilities were assumed to be equivalent for all patient subgroups.
Treatment discontinuation
Parametric survival analysis of EMPEROR-Reduced trial data was conducted to estimate time to empagliflozin treatment discontinuation (following an exponential distribution), with time-varying KCCQ-CSS as covariates (see Supplementary Appendix for statistical methods [Section S2] and results [Section S3]); alternative distributions were considered in sensitivity analyses. Discontinuation from empagliflozin was assumed to result in similar treatment effect on event risks, costs, and utilities as SoC alone.
Transient-event risks
Risk of first and subsequent HHF was estimated using a Poisson regression model fitted to patient-level EMPEROR-Reduced data with generalised estimating equations to account for the repeated measures on patients (Supplementary Appendix Section S2), including treatment allocation and time-varying KCCQ-CSS health states as covariates (Supplementary Appendix Section S3). Being on treatment with empagliflozin plus SoC or residing in higher KCCQ-CSS quartiles were associated with lower risk of HHF.
The model captured AEs experienced by patients in the EMPEROR-Reduced trial, including urinary tract infection, genital mycotic infection, acute renal failure, hepatic injury, volume depletion, hypotension, hypoglycaemic events and bone fracture. Patients were subject to ongoing risk of experiencing AEs, estimated from the trial data (Supplementary Appendix Section S4). A constant hazard was assumed for each AE.
Mortality
Parametric survival analyses were undertaken using the complete EMPEROR-Reduced dataset to capture all-cause and CV death as a function of treatment and time-varying KCCQ-CSS, both during and beyond the trial duration. Distributions chosen for all-cause and CV death (Weibull) were selected from potential distributions based on goodness of fit and clinical plausibility of long-term projections; alternative distributions were considered in sensitivity analyses. Details regarding the statistical methodology and results of the parametric survival analyses are presented in Supplementary Appendices Sections S2 and S3, respectively.
The risk of non-CV death was assumed to be at least as high as that observed in the population more generally. Accordingly, this was calculated as the difference between the parametric fits for all-cause and CV death, or between rates of age- and sex-specific all-cause mortality predicted from country-specific life tables [18–20] and rates of CV death recorded in national cause-of-death registries [21–23] — whichever was largest.
Utilities
The model applied utility scores ranging from 0 (death) to 1 (full health) to capture the effect of disease severity and clinical events on HRQoL. QALYs accrued during each model cycle, calculated as the difference between KCCQ-CSS health state utilities, and utility decrements attributable to transient events, including HHF and AEs (Table 1; also see Supplementary Appendix Section S5). The impact of transient events on HRQoL was captured as one-off decrements to the proportion of patients experiencing the event and assumed to apply during the month in which those events occurred.
Table 1
Summary of key case-case input parameters
Input parameter | UK | Spain | France |
Discount rate: cost | 3.5% | 3.0% | 2.5% |
Discount rate: health | 3.5% | 3.0% | 2.5% |
Distribution for clinical parameters | | | |
HHF | Poisson | Poisson | Poisson |
Time to CV death | Weibull | Weibull | Weibull |
Time to all-cause death | Weibull | Weibull | Weibull |
Time to empagliflozin treatment discontinuation | Exponential | Exponential | Exponential |
Utility parameters | | | |
Health state utility: KCCQ-CSS 0 to 55 (Q1) | 0.520 | 0.629 | 0.464 |
Health state utility: KCCQ-CSS 55 to < 75 (Q2) | 0.637 | 0.754 | 0.610 |
Health state utility: KCCQ-CSS 75 to < 90 (Q3) | 0.710 | 0.832 | 0.721 |
Health state utility: KCCQ-CSS 90 to 100 (Q4) | 0.774 | 0.891 | 0.810 |
Disutility: HHF | -0.246 | -0.291 | -0.240 |
Disutility: Urinary tract infection | -0.025 | -0.025 | -0.025 |
Disutility: Genital mycotic infection | -0.058 | -0.053 | -0.052 |
Disutility: Acute renal failure | -0.010 | -0.014 | -0.011 |
Disutility: Hepatic injury | -0.016 | -0.011 | -0.018 |
Disutility: Volume depletion | -0.018 | -0.015 | -0.017 |
Disutility: Hypotension | -0.025 | -0.025 | -0.025 |
Disutility: Hypoglycaemic event | -0.048 | -0.041 | -0.055 |
Disutility: Bone fracture | -0.165 | -0.170 | -0.148 |
Cost parameters | | | |
Monthly drug acquisition cost: empagliflozin + SoC | £79 | €157 | €129 |
Monthly drug acquisition cost: SoC | £40 | €100 | €84 |
Event management cost: HHF | £3,072 | €3,814 | €4,968 |
Event management cost: CV death | £4,146 | €6,276 | €3,764 |
Event management cost: Urinary tract infection | £40 | €57 | €36 |
Event management cost: Genital mycotic infection | £40 | €57 | €36 |
Event management cost: Acute renal failure | £1,906 | €4,243 | €3,739 |
Event management cost: Hepatic injury | £1,274 | €2,709 | €1,055 |
Event management cost: Volume depletion | £40 | €57 | €36 |
Event management cost: Hypotension | £40 | €57 | €36 |
Event management cost: Hypoglycaemic event | £627 | €1,461 | €2,319 |
Event management cost: Bone fracture | £2,710 | €5,042 | €3,352 |
Disease management cost: KCCQ-CSS 0 to 55 (Q1) | £77 | €55 | €48 |
Disease management cost: KCCQ-CSS 55 to < 75 (Q2) | £77 | €55 | €48 |
Disease management cost: KCCQ-CSS 75 to < 90 (Q3) | £77 | €55 | €48 |
Disease management cost: KCCQ-CSS 90 to 100 (Q4) | £77 | €55 | €48 |
Abbreviations: CV = cardiovascular; HHF = hospitalisation for worsening heart failure; KCCQ-CSS, Kansas City Cardiomyopathy Questionnaire Clinical Symptom Score; SoC = standard of care; UK = United Kingdom |
EQ-5D-5L responses collected in EMPEROR-Reduced were mapped to EQ-5D-3L using crosswalks formulated by van Hout et al. [24], which were then converted to index scores using published value sets for each of the three countries. Linear mixed-effects regression models (see Supplementary Appendix Section S3) were constructed to predict utility values from baseline demographics and medical history, time-varying KCCQ-CSS, and clinical events (HHF, AEs).
Utility decrements attributable to HHF were estimated as a weighted average of the relevant coefficients estimated from the regression models, each representing a different duration of time elapsed since the event occurred, and those durations themselves. Disutility estimates for most AEs were sourced directly from the regression model; however, urinary tract infection and hypotension in EMPEROR-Reduced could not be derived from the trial analysis due to low incidence and were obtained from published literature [25, 26].
An age adjustment was applied to KCCQ-CSS health state utility values to account for differences between EMPEROR-Reduced trial and country-specific population norms; in particular, the model assumed HRQoL in KCCQ-CSS quartile 4 would be similar to the population more generally, thereby ensuring utilities for patients in lower quartiles (i.e., whose conditions were more severe) would not be higher than this level.
Costs and resource utilisation
Direct reimbursable medical care costs were expressed in appropriate national currencies (i.e., 2021 British pounds for the UK and 2021 Euros for Spain and France), with adjustments for inflation. The model included costs associated with drug acquisition, management of clinical events, and disease management, sourced from local costing databases and published literature, as appropriate (Table 1).
Drug costs for empagliflozin 10 mg and SoC therapies were extracted from national databases in the three country settings — Monthly Index of Medical Specialties for the UK, BotPlus Web Database in Spain (retail plus value added tax for the base-case analysis and ex-factory price for sensitivity analysis) and the Official Journal in France (public prices including all taxes). For each SoC HF medication class, costs of individual therapies were used to compute an average cost, assuming treatments within each class were uniformly distributed (UK and Spain) or based on local market share (France). Then, the drug cost of SoC was computed as a weighted average cost based on these values and utilisation of HF medication classes at baseline in the empagliflozin plus SoC and SoC arms of EMPEROR-Reduced. Finally, a cost associated with each treatment regimen (i.e., empagliflozin plus SoC and SoC alone) was computed. Indicated strength and dosage for each drug were based on summaries of product characteristics. The cost of empagliflozin 10 mg was incurred until treatment discontinuation, whereas expenditures associated with SoC accrued until death or the model horizon expired.
Expenses for managing clinical events (HHF, CV death and AEs) were obtained from national databases or published literature and modelled as one-off costs. For AEs, these were calculated with reference to the distribution of visit types (i.e., outpatient or inpatient), reflecting their typical severity, and corresponding costs.
HF-related disease management expenditures accrued each cycle and were assumed similar across KCCQ quartiles, reflecting utilisation rates and local costs associated with primary care, cardiologist visits, and accident and emergency referrals. Detailed itemisation of costs incorporated into the model is supplied in Supplementary Appendix Section S6.
Model validation
Validation steps included assessment of face validity, technical validity, predictive validity and cross-validity. Thorough evaluation of face validity and technical validity was supported by an independent modelling expert using the TECHnical VERification checklist (TECH-VER) [27]. Meanwhile, predictive validity was assessed by comparing model-predicted rates of HHF, CV and all-cause death against observed rates from EMPEROR-Reduced (Supplementary Appendix Section S7), while cross-validity entailed identifying and explicating differences in analytical results relative to a recently published model with a similar decision problem [14].
Analyses
The base-case analysis predicted absolute and incremental lifetime cumulative events per 100 PYs, life-years (LY), QALYs and costs with empagliflozin plus SoC vs. SoC alone for the EMPEROR-Reduced trial intent-to-treat (ITT) population. Additionally, the model synthesised results for incremental costs and QALYs to compute the ICER for empagliflozin plus SoC vs. SoC alone. The ICERs for each setting were compared with country-specific, willingness-to-pay (WTP) thresholds — £20,000/QALY for the UK, €20,000/QALY for Spain and €30,000/QALY for France.
Scenarios were run for pre-specified subgroups that were considered clinically and/or economically meaningful (outlined above). Deterministic sensitivity analyses (DSA) were performed using plausible ranges or alternative values for key inputs (i.e., model horizon and discount rates; KCCQ-CSS utilities and utility decrements for AEs and HHF; costs associated with drug acquisition and clinical event/disease management; distributions and treatment effects for clinical outcomes; and whether to exclude treatment discontinuation or the life table adjustment for non-CV mortality) to test the robustness of model results to changes in the parameters. A probabilistic sensitivity analysis (PSA) was conducted using distributions reflecting parameter uncertainties, producing 1,000 pairs of incremental effectiveness and cost estimates. The parameters for sensitivity analyses are provided in Supplementary Appendix Section S8–S9.