Background: Identifying candidates for left ventricular assist device surgery at risk of right ventricular failure remains difficult. The aim was to identify the most accurate predictors of right ventricular failure among clinical, biological, and imaging markers, assessed by agreement of different supervised machine learning algorithms.
Methods: Seventy-four patients, referred to HeartWare left ventricular assist device since 2010 in two Italian centers, were recruited. Biomarkers, right ventricular standard, and strain echocardiography, as well as cath-lab measures, were compared among patients who did not develop right ventricular failure (N = 56), those with acute-right ventricular failure (N = 8, 11%) or chronic-right ventricular failure (N = 10, 14%). Logistic regression, penalized logistic regression, linear support vector machines, and naïve Bayes algorithms with leave-one-out validation were used to evaluate the efficiency of any combination of three collected variables in an "all-subsets" approach.
Results: Michigan risk score combined with central venous pressure assessed invasively and apical longitudinal systolic strain of the right ventricular-free wall were the most significant predictors of acute-right ventricular failure (maximum receiver operating characteristic-area under the curve = 0.95, 95% confidence interval = 0.91-1.00, by the naïve Bayes), while the right ventricular-free wall systolic strain of the middle segment, right atrial strain (QRS-synced), and tricuspid annular plane systolic excursion were the most significant predictors of Chronic-RVF (receiver operating characteristic-area under the curve = 0.97, 95% confidence interval = 0.91-1.00, according to naïve Bayes).
Conclusion: Apical right ventricular strain as well as right atrial strain provides complementary information, both critical to predict acute-right ventricular failure and chronic-right ventricular failure, respectively.
Keywords: Right ventricle; echocardiography; heart failure; machine learning; strain imaging.