부산대학교 도서관

Introduction: PAH is a deadly disease without cure. Formalized risk stratification allows therapeutic adjustments that optimize drug utilization. Risk scores, like REVEAL 2.0 recommended by PAH guidelines only offer good discrimination. Our goal was to create risk models with excellent discrimination (C-Index over 0.8), using modern statistical techniques and expanded variable pools including imaging and genomics. The AIM of this study was to demonstrate the improved performance with the addition of cardiac MRI variables.Methods: PAH patients from the ASPIRE cardiac MRI database were analyzed. Imaging variable (IMV) selection was performed using three machine learning methods: logistic regression, Lasso, and Random Forest. Rankings of the IMVs from these sources were aggregated to arrive at a consensus list. The selected IMVs were added to the set of variables for deriving the REVEAL 2.0 composite score. Bayesian networks (BN) were then built to predict 1-year survival based on the Tree-Augmented naïve Bayes (TAN) algorithm. Five-fold cross-validation was performed to assess the improvement in survival prediction from adding the selected imaging variables.Results: A total of 343 PAH previously diagnosed subjects were included in this analysis. The rank aggregation algorithm identified several IMV, including LVSVI, RVCO, LVEDVI, and RVESVI, that were predictive of survival but not the REVEAL 2.0 composite score. Adding these IMV to the REVEAL 2.0 variables, we built a BN model that depicts the non-linear relationships among the predictors and one-year survival (Fig 1). We obtained an average AUC of 0.83 over the five cross-validation test sets, an improvement over the AUC of 0.78 using only the REVEAL 2.0 variables. A Mann-Whitney non-parametric test shows the improvement is statistically significant at the 0.1 level.Conclusion: Advanced statistical models that include cardiac IMVs improve performance of PAH risk assessment models.