Early cost-utility analysis of tissue-engineered heart valves compared to bioprostheses in the aortic position in elderly patients

Objectives Aortic valve disease is the most frequent indication for heart valve replacement with the highest prevalence in elderly. Tissue-engineered heart valves (TEHV) are foreseen to have important advantages over currently used bioprosthetic heart valve substitutes, most importantly reducing valve degeneration with subsequent reduction of re-intervention. We performed early Health Technology Assessment of hypothetical TEHV in elderly patients (≥ 70 years) requiring surgical (SAVR) or transcatheter aortic valve implantation (TAVI) to assess the potential of TEHV and to inform future development decisions. Methods Using a patient-level simulation model, the potential cost-effectiveness of TEHV compared with bioprostheses was predicted from a societal perspective. Anticipated, but currently hypothetical improvements in performance of TEHV, divided in durability, thrombogenicity, and infection resistance, were explored in scenario analyses to estimate quality-adjusted life-year (QALY) gain, cost reduction, headroom, and budget impact. Results Durability of TEHV had the highest impact on QALY gain and costs, followed by infection resistance. Improved TEHV performance (− 50% prosthetic valve-related events) resulted in lifetime QALY gains of 0.131 and 0.043, lifetime cost reductions of €639 and €368, translating to headrooms of €3255 and €2498 per hypothetical TEHV compared to SAVR and TAVI, respectively. National savings in the first decade after implementation varied between €2.8 and €11.2 million (SAVR) and €3.2–€12.8 million (TAVI) for TEHV substitution rates of 25–100%. Conclusions Despite the relatively short life expectancy of elderly patients undergoing SAVR/TAVI, hypothetical TEHV are predicted to be cost-effective compared to bioprostheses, commercially viable and result in national cost savings when biomedical engineers succeed in realising improved durability and/or infection resistance of TEHV. Electronic supplementary material The online version of this article (10.1007/s10198-020-01159-y) contains supplementary material, which is available to authorized users.

The ACSD only includes data on COPD and diabetes mellitus. The remaining patients were randomly assigned to one of the comorbidity subgroups based on age group specific proportions derived from a healthcare insurance claims database.
[1] This database contains the healthcare expenditures of all the insured who underwent heart valve replacement between 2010 and 2013 in the Netherlands. The randomly assigned comorbidity variable will only be used for costs calculations, not for determining clinical outcomes or utilities.
The following concomitant procedures included in the regression formula for intervention costs derived from the health insurance claims database were not available in the ACSD: correction of Tetralogy of Fallot, hypertrophic obstructive cardiomyopathy (HOCM), and left ventricle repair.
In total 35,258 patients did not have missing values in the required variables (Table S2). The patient and intervention characteristics of the 15,405 patients above 70 years old who underwent aortic valve replacement we have selected for this analysis are presented in Table S1.
Transcatheter aortic valve implantation (TAVI) Unfortunately we did not have access to a clinical dataset such as the ACSD for patients after TAVI. Instead we have simulated a patient population using patient characteristic frequencies from the health insurance claims database. [1] The Vektis database is a health insurance claims database which contains the healthcare expenditures of all the insured in the Netherlands. The patient and intervention characteristics of the 809 patients ≥70 years old who underwent TAVI between 2010-2013 are presented in Table S1.  QALY: quality adjusted life year. 1 We assumed all patients in the starting population did not have previous valve replacement or CVA because it was not available or there were many missing values in the database. 2 Not available in the database, but considering the high age of these patients we assume they all received a bioprosthesis. The background of this study and the development of the conceptual model is described before. [2] This supplement describes the changes that were made to the conceptual model described by Huygens et al. [2] To be in line with the definition in the ACSD and published literature, the definition of cerebrovascular accident was changed to stroke for both SAVR and TAVI. In addition, atrial fibrillation and acute kidney injury were changed into arrhythmias and renal failure for SAVR patients to be in line with the definition used in the ACSD. Finally, conversion to another approach (transcatheter to surgical valve implantation and vice versa) was excluded from the final decision analytic model.
Emergent conversion from TAVI to SAVR occurs rarely (1.2%-2.1%) and according to expert opinion conversion from SAVR to TAVI occurs even less. [3,4] In addition, since the causes of conversion of approach are not related to the prosthetic heart valve itself, the conversion rate is likely to be comparable for TEHV and existing heart valve substitutes. [4] Figure S1. Flowchart patient level simulation model Preoperative endocarditis* At the moment of the heart valve replacement the patient is being treated with antibiotics for endocarditis.

Instable angina pectoris
Angina pectoris that requires intravenous nitrate therapy until entering the operation theatre. Pulmonary hypertension Condition of increased blood pressure within the arteries of the lungs. Co-morbidity categories in costanalyses Co-morbidities were based on Pharmacy Cost Groups, which is an outpatient morbidity measure based on prior use of prescribed drugs as marker for chronic conditions. -COPD, diabetes mellitus (DM), kidney disease and/or heart failure (HF) Patients with COPD, DM, kidney disease and/or HF.

-Hypertension
Patients without COPD, DM, kidney disease and/or HF, but with hypertension. -Other co-morbidities Patients with other co-morbidities than COPD, DM, kidney disease, HF or hypertension. -No co-morbidities Patients without co-morbidities.
Socioeconomic status Based on status scores reflecting the SES of a district (defined by postal code) based on characteristics of its residents: education, income, and position on the labor market. The status scores were divided in four groups based on percentiles, with lower percentiles representing lower SES.

Procedure related
Emergency procedure Unplanned intervention that cannot wait until the beginning of the next working day due to medical reasons.

Concomitant procedures
Procedures that are performed at the same time of the valve replacement. -Coronary artery bypass grafting (CABG) Coronary artery bypass grafting.
-Other valve replacement Replacement of more than one valve.
-Other valve repair Repair of another valve than the valve being replaced.
-Aortic root procedure Intervention on the aortic root (part of the aorta from the aortic valve until the sinotubular junction) only. -Aorta ascendens procedure Intervention on the aorta ascendens (part of the aorta from the aortic valve until the arteria anonyma) only.
-Bentall procedure Intervention involving composite graft replacement of the aortic valve, aortic root and ascending aorta, with re-implantation of the coronary arteries into the graft. -Aortic arch procedure Intervention on the aortic arch (part of the aorta beyond the arteria anonyma until the arteria subclavia sinistra). -Aorta descendens procedure Intervention on the aorta (part of the aorta beyond the subclavian sinister artery until beyond the diaphragm. -Thoracic aorta surgery Surgical intervention on the aorta ascendens, arch or descendens. -Maze procedure Surgical treatment for atrial fibrillation.

Stroke
Stroke with or without residual injury. Reintervention Any surgical or transcatheter procedure that repairs, otherwise alters or adjusts, or replaces a previously implanted prosthesis.
[6] *Not stable over time. † Events are only included during the first thirty days after the intervention. Other healthcare costs adults = time since intervention + death + age group at intervention + male gender + SES class + AF + AKI + stroke + TIA + endocarditis + MI + PI + reintervention.
Other healthcare costs children = time since intervention + male gender + SES class.

Societal costs
9 Productivity costs of unpaid work Generalized linear models with binominal family (glm function in R) Probability of unpaid work after SAVR = age + male + years since intervention + biological valve (compared to mechanical valve) + concomitant CABG + multiple valve replacement.
Probability of unpaid work after TAVI = age + male + years since intervention Probability of less unpaid work after SAVR = age + male + years since intervention + biological valve (compared to mechanical valve) + concomitant CABG + multiple valve replacement.
Probability of less unpaid work after TAVI = age + male + years since intervention + biological valve (compared to mechanical valve) + concomitant CABG + multiple valve replacement.
Productivity costs unpaid work last four weeks after SAVR = age + male + years since intervention + biological valve (compared to mechanical valve) + concomitant CABG + multiple valve replacement.

Generalized linear model with gamma family and log link (glm function in R)
Estimated productivity costs last four weeks = probability of unpaid work * probability of less unpaid work * estimated productivity costs of less unpaid work 10 Informal care costs Probability of using informal care after SAVR = age + male + years since intervention + biological valve (compared to mechanical valve) + concomitant CABG + multiple valve replacement.
Probability of using informal care after TAVI = age + male + years since intervention

Supplement 3 -Excess mortality
Excess mortality is expressed as a hazard ratio of the additional excess mortality not directly resulting from valve-related events relative to background mortality. The estimation of this hazard ratio in elderly patients after SAVR was reported previously. [8] For the estimation of this hazard ratio in TAVI patients, the model containing only background mortality and mortality due to valve-related events (excluding early mortality) was run for 10,000 iterations at the mean age and proportion of males of the UK TAVI registry. [9] Subsequently, the hazard ratio was estimated by fitting the survival output of this simulation model to the survival observed in the UK TAVI registry (excluding early mortality) using varying values for the hazard ratio of excess mortality. The best fit was determined by using the least squares method (Table S4). 2.0 2341 Bold print indicates the selected model. 1 Hazard ratio of background mortality + excess mortality relative to background mortality. 2 Sum of squared residuals between microsimulation-based survival and survival observed in our metaanalysis of Kaplan-Meier freedom from all-cause mortality.

Surgical aortic valve replacement (SAVR) Early mortality
The data on early mortality and events after SAVR was derived from the Adult Cardiac Surgery Database (ACSD) from The Netherlands Association for Cardio-Thoracic Surgery (NVT). This database includes patient and intervention characteristics, early mortality (i.e. death within 30 days after the intervention), and several peri-and postoperative complications (CVA, renal failure, vascular complications, rhythm problems and myocardial infarction (only perioperative)). For the logistic regression analysis of early mortality we extracted the records of aortic valve replacements (AVR) from 1 January 2007 until 31 December 2015. In total there were 35,732 (isolated or combined) AVR procedures in the Netherlands.
Previously, we identified the potentially relevant predictors of in-hospital events and mortality in patients with heart valve disease [2]: age, gender, symptomatic status (New York Heart Association class), left ventricular ejection fraction (3 categories), pulmonary artery systolic pressure, creatinine (< or > 200), chronic pulmonary disease, extracardiac arteriopathy/peripheral vascular disease, neurological impairment affecting daily activity, concomitant coronary artery disease, concomitant coronary artery bypass surgery, type of valve surgery, concomitant surgery of the ascending aorta, redo cardiac surgery, emergency surgery, frailty, major organ system dysfunction, and procedurespecific impediments. Some of these predictors were not available in the ACSD: concomitant coronary artery disease, frailty, major organ system dysfunction and procedure specific impediments. Further, NYHA class and pulmonary artery systolic pressure were available in the ACSD but the proportion of missing values was very high (>55%) and therefore these predictors were not included in the regression analysis to predict early mortality and early events.
There were 32,345 (90.5%) complete case and 3,387 uncomplete cases. Table S5 compares the patient-and procedure related risk factors, peri-and postoperative complications, and early mortality of complete cases with cases with at least one missing value. The cases with one or more missing value are younger, more often have a serum creatinine level above 200 µmol/l, higher LV ejection fraction, less peripheral vascular disease, and less concomitant procedures. Early mortality risk is slightly lower in the uncomplete cases.
Variables with missing values were completed by multiple imputation (MI). We assume these values were Missing at Random (MAR). MI was performed with the Amelia package. We constructed 50 imputed datasets. [10] We included all available variables in the imputation model, except for the variables that had too many missing values (>80%) within the subgroup of incomplete cases, unless these variables were to be included in the logistic regression analysis of the data after imputation (previous cardiac surgery) (Table S6).
Subsequently, logistic regression analyses were performed with the glm function in R for every imputed dataset. Table S7 shows the pooled estimates of these analyses compared to the logistic regression analysis of only the complete cases. In the final model we used the equation based on the imputed cases.

Early events
Table S8-11 show the logistic regression formulas for the risk on early events specific for patients who survive and those who do not survive the first 30 days after SAVR. Since there were many missing values in the occurrence of early events (>50%), the logistic regression analyses are based on the complete cases. The area under the curve of these formulas are low, but we decided to implement these formulas in the model anyway to make use of the available data.
The occurrence of the other early events (vascular complication, bleeding, pacemaker implantation, prosthetic valve dysfunction, valve thrombosis and endocarditis) was not available in the ACSD or only a small number of events occurred. The risk on reexploration for bleeding (4.2%) and pacemaker implantation (8.1%) was derived from our metaanalysis outcomes after AVR with bioprostheses in elderly patients. [8] We assume this risk is equal for all patients. The other early event risks were assumed to be zero for all patients.
Stroke (Table S8) The following possible predictors for stroke after AVR were identified from previous studies investigating predictors of stroke after cardiac surgery: age, gender, previous CVA, previous cardiac surgery, isolated vs. Concomitant other valve replacement, concomitant CABG, concomitant aorta ascendens surgery, serum creatinine level > 200 µmol/l, LV function, COPD, diabetes, extra cardiac arterial vascular pathology, neurological dysfunction, preoperative endocarditis, emergency procedure, instable angina pectoris, and pulmonary hypertension. [11][12][13] Unfortunately, for diabetes the proportion of missing values in the Adult Cardiac Surgery Databases was very high (>55%) and therefore this variable was excluded from the analyses. The tables below show the results of the logistic regression analyses of stroke for patients that do and do not survive the first 30 days.
Renal failure (Table S9) Renal failure was registered in the Adult Cardiac Surgery Database if one or more of the following criteria were fulfilled during the postoperative period: renal replacement treatment (dialysis, CVVH) not existing before procedure and/or highest postoperative serum creatinine level > 177 μmol/L and doubled preoperative level. This narrow definition does not include acute kidney injury stage 1 as defined by the AKIN classification in VARC-2. [9] The following variables were identified from previous studies investigating predictors of renal failure after cardiac surgery: age, gender, preoperative renal insufficiency, prior cardiac surgery, NYHA class, congestive heart failure, LV function, peripheral vascular disease, COPD, diabetes, preoperative intra-aortic balloon pump, emergency procedure, concomitant CABG, and increased cardiopulmonary bypass time. [10] Perioperative renal insufficiency was operationalized with the dummy variable serum creatinine level above or below 200 µmol/l. NYHA class and diabetes were available in the Adult Cardiac Surgery Database but the proportion of missing values was very high (>55%) and therefore these variables were excluded from the analyses. Congestive heart failure, preoperative intra-aortic balloon Myocardial infarction (Table S11) MI are registered according to the definition used in the STS Adult Cardiac Surgery Database. [5] Possible predictors of MI after cardiac surgery in the literature are: age, gender, renal failure, diabetes, peripheral vascular disease, emergency surgery, redo surgery, LV dysfunction, perioperative MI, and concomitant CABG. [17,18] The tables below show the results of the logistic regression analyses of myocardial infarctions for patients that do and do not survive the first 30 days.

Late events
The occurrence of late events after SAVR was based on our previously published systematic review and meta-analysis.
For more details regarding the methods and results of this study we refer to the publication in the Journal of Thoracic and Cardiovascular Surgery. [8] 2

.3.2 Transcatheter aortic valve implantation (TAVI) in elderly Early mortality and events
The clinical outcomes after TAVI are derived from a systematic review performed by Gargiulo et al. (2016). [19] This study included five randomized trials and 31 observational matched studies comparing outcomes after TAVI or SAVR.
Gargiulo et al. [19] only report odds ratios of clinical outcomes after TAVI compared to SAVR instead of the mortality and event risks and rates we needed as input for our patient level simulation model. Therefore, we have pooled the extracted data reported in Gargiulo et al. [19] with the use of the inverse variance method in a random-effects model, on a logarithmic scale, as the Shapiro-Wilk test revealed a significantly skewed distribution among the included studies in the outcome measures. In total 7,726 TAVI patients from 36 studies were included in the meta-analysis. Their mean age was 80.9 years, 50.3% were males, and the mean STS score was 6.7%. [19] The risks on mortality and events during the first 30 days after TAVI are reported in Table 1 in the main manuscript. We assume these risks are equal for all patients, whether they die within 30 days or not. The other early event risks (i.e. prosthetic valve thrombosis and endocarditis) were assumed to be zero for all patients.

Late events
The rates of valve-related events after TAVI are derived from the systematic review of Gargiulo et al. on outcomes after TAVI. [19] The long term event occurrence was not reported in the meta-analyses of Gargiulo et al. Therefore we extracted and pooled the relevant data from the included studies ourselves (Table 1 in  time-to-event data was then extrapolated from the digitized curve coordinates, assuming a constant rate of censorship between each time point at which the number of patients at risk were specified. [20] The occurrence rate of SVD after AVR with bioprostheses and TAVI was modelled by fitting a gompertz or lognormal distribution to our pooled time-toevent data, respectivelly, showing an increasing occurrence rate of SVD over time. These distributions had the best fit according to visual comparison, log-likelihood and/or Akaike information criterion (Table S12).        mechanical aortic valve replacement and deteriorates to almost zero after six years. [22] The higher occurrence in the early phase may be caused by suboptimal anticoagulation treatment in the first post-intervention period. Since, the mean follow-up of the Bern TAVI Registry was only one year, it is likely that the occurrence rate of valve thrombosis after TAVI found in this study will not remain constant but will reduce over time. Therefore we recalculated the linearized occurrence rate, assuming that it will be zero from year 7 onwards. The adjusted linearized occurrence rate of VT is then 0.24%/patient-year.
The pooled linearized occurrence rate of endocarditis (0.54%/patient-year) was based on studies with a mean followup of 1.9 years, since this rate is comparable to the rate of endocarditis after SAVR, we assume this will remain constant over time.
The linearized occurrence rates of stroke (2.03%/patient-year) and bleeding (3.34%/patient-year) were relatively high after TAVI. However, this is probably caused by the short mean follow-up duration of the studies that reported long term stroke and bleeding (1.8 and 2.5 years, respectively), since bleedings and strokes mainly occur in the earliest period after TAVI. [23] In patients after SAVR we have seen that (probably due to higher use of anticoagulation) the occurrence of strokes is lower (HR: 0.7) and of bleeding is higher (HR: 5.3) than in the age and sex matched general population.
The late occurrence of stroke and bleedings was comparable between SAVR and TAVI in two large randomized controlled trials. [24,25] Therefore we applied the same hazard ratio of strokes of SAVR patients compared to the general population to the general population with comparable age and sex as the TAVI population. Therefore we applied the hazard ratios determined for SAVR patients to the occurrence of stroke and bleeding in the age and sex matched general population for TAVI resulting in a linearized occurrence rate of strokes of 0.955%/patient-year and bleeding of 0.954%/patient-year.  .017 †The exponentiated coefficient is the factor by which the arithmetic mean outcome on the original scale is multiplied. N.B. The results should be interpreted as follows, for example for sex: males are more likely to have unpaid work than females (logistic regression model 1), are less likely to be unable to perform unpaid work (logistic regression model 2) and when all other variables are equal, the mean estimated productivity costs of males is almost equal to females(GLM). 1 Age excluded because number of events (n=21) was too small for three predictors. 2 The productivity costs of patients who performed less unpaid work after TAVI was only reported for 16 of the 21 patients with less unpaid work. Therefore we apply the average productivity costs of unpaid work in these patients to all patients instead of using a GLM. The results should be interpreted as follows, for example in the models for SAVR patients: males are less likely to use informal care than females (logistic regression model) and when all other variables are equal, the mean estimated informal care costs of males is 0.001 times (i.e. 0.1%) lower than for females (GLM). Prosthetic valve dysfunction without re-intervention 8.67 days [27] Valve thrombosis 10 days

Supplement 6 -Probabilistic sensitivity analysis
To estimate the numbers of patients and simulations required in our probabilistic sensitivity analyses (PSA), we used the approach described in O'Hagan et al. as recommended by the NICE DSU guidelines on patient-level modelling. [28] In this approach the number of PSA runs (outer loop = N) and patients per PSA run (inner loop = n) needed to achieve accurate cost-effectiveness estimates while keeping the number of runs as small as possible can be estimated. The cost-effectiveness measure used in this estimation was the number of undiscounted QALYs. translating to a value for c of approximately 0.12 (i.e. almost twice more accurate). -Subgroup patients aged >80 years 6.6 3.8 Decreased durability (50% more events) but improvements in thrombogenicity and infection resistance (50% less events) 9.0 4.5