Prosthetic Valve Endocarditis After Aortic Valve Replacement With Bovine Versus Porcine Bioprostheses

Background Whether a bovine or porcine aortic valve bioprosthesis carries a higher risk of endocarditis after aortic valve replacement is unknown. The aim of this study was to compare the risk of prosthetic endocarditis in patients undergoing aortic valve replacement with a bovine versus porcine bioprosthesis. Methods and Results This nationwide, population‐based cohort study included all patients who underwent surgical aortic valve replacement with a bovine or porcine bioprosthesis in Sweden from 1997 to 2018. Regression standardization was used to account for intergroup differences. The primary outcome was prosthetic valve endocarditis, and the secondary outcomes were all‐cause mortality and early prosthetic valve endocarditis. During a maximum follow‐up time of 22 years, we included 21 022 patients, 16 603 with a bovine valve prosthesis and 4419 with a porcine valve prosthesis. The mean age was 73 years, and 61% of the patients were men. In total, 910 patients were hospitalized for infective endocarditis: 690 (4.2%) in the bovine group and 220 (5.0%) in the porcine group. The adjusted cumulative incidence of prosthetic valve endocarditis at 15 years was 9.5% (95% CI, 6.2%–14.4%) in the bovine group and 2.8% (95% CI, 1.4%–5.6%) in the porcine group. The absolute risk difference between the groups at 15 years was 6.7% (95% CI, 0.8%–12.5%). Conclusions The risk of endocarditis was higher in patients who received a bovine compared with a porcine valve prosthesis after surgical aortic valve replacement. This association should be considered in patients undergoing both surgical and transcatheter aortic valve replacement.

An alternative to this is regression standardization, also known as g-estimation using the parametric g-formula.This method instead models the outcome as the dependent variable and include the exposure and potential confounders as covariates.This model can then be used to predict the outcome for all individuals in the population.This step is done two times for a binary exposure, first setting all individuals exposure to the first level, and then setting all individuals exposure to the second level.Then, the mean of the outcome across the population is calculated.This can be interpreted as the expected outcome if the entire population would have treatment level X, versus the expected outcome if the entire population had treatment level Y, while all other covariates are held constant.While it is trivial to calculate the population average (although cumbersome in a survival setting with many different time points), the standard error calculations are bit more complex and benefit from computational support.This method has been implemented in the standsurv and stpm2 command in Stata, and in the rstpm2 and marginaleffects package in R, among others.
The interpretation of regression standardization is attractive when investigating the effect of xenograft material while accounting for other valve model effects.It would be hard to include valve model as a covariate when trying to create a model for the exposure, in this case xenograft material, as the valve model would be sufficient to predict the exposure rendering balance on the other covariates impossible.However, by modelling on the outcome and then including xenograft as exposure, and valve model as a covariate, regression standardization estimates what would happen if the entire population would have had a bovine valve (albeit the same valve model, with all other characteristics intact), versus if the entire population would have had a porcine valve (with the same valve model with all other characteristics intact).

Classification and regression tree imputation
The Classification and regression tree (CART) algorithm was first developed in the 1980s by Breiman and colleagues and has seen a revival as a popular machine learning algorithm.It is used to predict, estimate, or classify variables based on other variables, a characteristic that can be used to impute estimated values in case of missing data.
The algorithm works by classifying discrete variables using a classical classification tree based on Gini impurity and estimating continuous variables using a regression model and perform splits in the tree based on minimizing the residual sum of squares.
It is an attractive method as is can handle binary, categorical, and continuous variables both as predictors and as classification targets and is robust against outliers.The method has been implemented in R in the rpart, simputation, and mice packages.
For further reading, we recommend Buuren. 38

Model selection
Model selection for the regression standardization models were performed using clinical subject matter knowledge and was informed using the Akaike information criterion.Continuous variables were tested using polynomials and splines during the selection process.

Endocarditis
The final model for endocarditis used a 3 rd degree exponential B-spline with 2 interior knots at the 1/3 and 2/3 quantiles of the follow-up time for the baseline hazard.Age was included as a natural spline with four degrees of freedom.Year of surgery, body mass index and estimated glomerular filtration rate were included as natural splines with three degrees of freedom.The other included covariates were patient sex, operating hospital, left ventricular ejection fraction, emergent operating status, birth region, educational level, preoperative atrial fibrillation, preoperative cancer, preoperative chronic obstructive pulmonary disease, preoperative diabetes, previous endocarditis, preoperative heart failure, preoperative hypertension, preoperative hepatic disease, preoperative peripheral vascular disease, preoperative stroke, preoperative percutaneous coronary intervention, marital status, valve prosthesis model, disposable family income, and valve prosthesis size.

Mortality
The final model for mortality used a natural spline with three degrees of freedom to model the baseline hazard.Age was included as a natural spline with four degrees of freedom.Body mass index, year of surgery, and estimated glomerular filtration rate were included as natural splines with three degrees of freedom.The other included covariates were patient sex, operating hospital, left ventricular ejection fraction, emergent operating status, concomitant coronary artery bypass grafting, concomitant ascending aortic surgery, birth region, educational level, disposable family income, preoperative atrial fibrillation, preoperative alcohol dependency, preoperative myocardial infarction, preoperative cancer, preoperative chronic obstructive pulmonary disease, preoperative diabetes, previous endocarditis, preoperative heart failure, preoperative hyperlipidemia, preoperative hypertension, preoperative hepatic disease, preoperative peripheral vascular disease, preoperative stroke, preoperative major bleeding event, preoperative pacemaker, preoperative percutaneous coronary intervention, marital status, valve prosthesis size, valve prosthesis model, and preoperative drug abuse.