부산대학교 도서관

Background:Pulmonary arterial hypertension (PAH) is a fatal and difficult to treat disease due to patient inter-variability. Accurate risk stratification is necessary for guiding treatment but no PAH risk calculator has achieved "excellent performance" (receiver-operator AUC > 0.8). Conversion of REVEAL 2.0 to a Bayesian network has shown promising results (Pulmonary Hypertension Outcomes Risk Assessment or PHORA). In this study, a new Bayesian network model (PHORA 2.0) was developed with a novel network structure and optimal feature selection.Methods:Patient-level data had been previously aggregated and harmonized across six PAH clinical trials (AMBITION, PATENT-1/2, GRIPHON, SERAPHIN, FREEDOM-EV, ARIES-1/2); all patients were assessed at baseline. Forty-one variables were initially considered based on p-value ranking from previous meta-analyses, availability across trials, and expert opinion. Training data was created by random sampling of 80% of the harmonized dataset, dropping early censored patients (N = 2531), leaving 20% of the data as a test set (N = 626). Continuous variables were discretized through univariate decision trees using 10-fold cross-validation maximizing Brier score. Genetic search selected combinations of features that maximized ranked correlation (Kendall’s tau) with one-year survival, with increasing penalty for redundant features. Feature combinations were evaluated in augmented naïve Bayesian networks, the best model was selected by 10-fold cross-validation on training data. Final performance is reported as performance on the test set.Results:The final model achieved the best cross-validation AUC using 16 variables. Performance on the test set was an AUC = 0.85. The final model outperformed multiple risk calculators at test time (Figure 1).Conclusion:Bayesian network modeling coupled with genetic feature selection has discovered for the first time a one-year mortality model for PAH with excellent performance.