Novel Machine Learning Algorithms for Predicting Early Right Heart Failure Post Left Ventricular Assist Device Implantation

• Published in: The Journal of heart and lung transplantation Vol. 41; no. 4; p. S25
• Main Authors: Veen, K., Wang, X., Soliman, O., de By, T.M., Mokhles, M.M., Caliskan, K., Takkenberg, J.J.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.04.2022
• ISSN: 1053-2498; 1557-3117
• DOI: 10.1016/j.healun.2022.01.054

Right heart failure is a major contributor to mortality and morbidity after left ventricular assist device implantation (LVAD). Previously our group developed a logistic regression-based model to predict right heart failure (RHF)after LVAD implantation, the EUROMACS RHF Risk Score, with currently the best but yet moderate discriminative ability. However, identification of patients prone to develop RHF after LVAD implantation using traditional statistical models has important limitations. Due to the exponential increase of computational power in the last decade, novel machine learning algorithms have emerged as a powerful tool in prediction modelling. Large international multicenter registries containing thousands of patients like the European Registry for patients with Mechanical Circulatory Support (the EUROMACS Registry) database are particularly suitable for these novel machine learning models. In this study, we aim to use the EUROMACS Registry database, to train and test novel machine learning algorithms predicting RHF after LVAD implantation, holding our previous model as benchmark. The results of this study may facilitate more accurate prediction of early RHF post LVAD implantation. The EUROMACS registry will be used to develop different machine learning models, including Random Forests, AdaBoost, XGBoost, Neural networks and Support Vector Machines, to predict postoperative RHF after LVAD implantation. Baseline variables collected in EUROMACS will serve as potential predictors. Both wrapper methods and filter methods will be used for predictor selection. The whole dataset will be used as training set and internal performance of the models will be assessed using resampling methods based upon k-fold cross validation or bootstrapping to limit overfitting. Performance measures that will be used include Area Under the Curve (discrimination), calibration intercepts/slope (calibration), Brier score (discrimination/calibration) and Net benefit curves. The primary endpoint of this study is early postoperative RHF using the following definitions: the need for temporary or durable right ventricular assist device, or need of inotropic support for ≥14 days, and/or NO ventilation for ≥48 hours.