Skip to Main Content
Right ventricular failure is a significant complication following implantation of a left ventricular assist device (LVAD), which increases morbidity and mortality. Consequently, researchers have sought predictors that may identify patients at risk. However, they have lacked sensitivity and/or specificity. This study investigated the use of a decision tree technology to explore the preoperative data space for combinatorial relationships that may be more accurate and precise. We retrospectively analyzed the records of 183 patients with initial LVAD implantation at the Artificial Heart Program, University of Pittsburgh Medical Center, between May 1996 and October 2009. Among those patients, 27 later required a right ventricular assist device (RVAD+) and 156 remained on LVAD (RVAD-) until the time of transplantation or death. A synthetic minority oversampling technique (SMOTE) was applied to the RVAD+ group to compensate for the disparity of sample size. Twenty-one resampling levels were evaluated, with decision tree model built for each. Among these models, the top six predictors of the need for an RVAD were transpulmonary gradient (TPG), age, international normalized ratio (INR), heart rate (HR), aspartate aminotransferase (AST), prothrombin time, and right ventricular systolic pressure. TPG was identified to be the most predictive variable in 15 out of 21 models, and constituted the first splitting node with 7 mmHg as the breakpoint. Oversampling was shown to improve the senstivity of the models monotonically, although asymptotically, while the specificity was diminished to a lesser degree. The model built upon 5X synthetic RVAD+ oversampling was found to provide the best compromise between sensitivity and specificity, included TPG (layer 1), age (layer 2), right atrial pressure (layer 3), HR (layer 4,7), INR (layer 4, 9), alanine aminotransferase (layer 5), white blood cell count (layer 5,6, &7), the number of inotrope agents (layer 6), creatinine (layer 8), A- T (layer 9, 10), and cardiac output (layer 9). It exhibited 85% sensitivity, 83% specificity, and 0.87 area under the receiver operating characteristic curve (RoC), which was found to be greatly improved compared to previously published studies.