What's New

Poster at ERS International Congress 2022: Pulmonary Arterial Hypertension risk assessment model using Bayesian Network Analysis

Fauvel, C.Liu, Z., Lin, S., Correa-Jaque, P., Webb, A., Vanderpool, & Benza, R. L. “Pulmonary Arterial Hypertension risk assessment model using Bayesian Network Analysis”

BACKGROUND. Usual pulmonary arterial hypertension (PAH) risk assessment (i.e., used to predict survival) models are based on statistical analysis which do not consider the relationship between variables, assuming their linear association with outcome.  

AIMS. To be free of this bias, we aimed to build a PHORA2.0 invasive and noninvasive risk assessment model using machine learning method.  

METHODS. Clinical data set from 7 PAH trials (GRIPHON, SERAPHIN, EARLY, COMPASS-2 and 3, MAESTRO, TRANSIT-1) with 147 variables (clinical, biological, hemodynamic) were harmonized. Random Forrest (RF) was used to identify variable importance in predicting outcomes. To build the Bayesian Network (BN), variables were selected based on both clinical knowledge, RF minimal depth (selected when <8). The outcome was 1-year survival. Both noninvasive and invasive BN were then validated using 10-fold cross validation.  

RESULTS. 2,870 patients were included (mean age 43 years old, 77% female, 50% idiopathic or heritable PAH) with a 1-year mortality of 14%. The invasive and noninvasive BN included 13 and 12 variables (8 overlaps). AUC were 0.83 and 0.80, respectively. BN shows the existence (arrow direction) and weight (arrow thickness) of the relationships between these variables and with the outcome.  

CONCLUSION. BN provides a powerful method to build PAH risk assessment model. Although important, invasive parameters provided minimal discrimination increase. Future addition of genomic and imaging data may continue to enhance noninvasive models.