APPLICATION OF A BAYESIAN NETWORK MODEL TO PREDICT OUTCOMES IN PULMONARY ARTERIAL HYPERTENSION

Abstract

SESSION TITLE: Comorbidities and the Management of Pulmonary Hypertension SESSION TYPE: Original Investigations PRESENTED ON: 10/10/2018 08:45 AM - 09:45 AM PURPOSE: In spite of emerging new therapeutic options, pulmonary arterial hypertension (PAH) remains a highly morbid and fatal disease. The Registry to Evaluate Early and Long-Term PAH Disease Management (REVEAL) developed a multivariable, weighted risk formula to predict the risk of one-year survival using multiple clinical variables, which was recently updated (REVEAL 2.0). This study aimed to develop an improved prognostic model by employing dynamic machine learning tools and Bayesian modeling on the REVEAL data and compare the results to the COMPERA and French registry risk assessment when applied on the same data set. METHODS: Data sets from the REVEAL registry, comprised of 54 US sites and 2,964 adult patients with PAH, were used to develop a Bayesian network (BN) risk model to predict 1-year survival. The model was learned from 80% (n=2347) of the patients using the Tree Augmented Naive Bayes (TAN) algorithm and validated in the remaining 20% (n=617). We used the same variables found in the REVEAL 2.0 calculator (e.g. BNP, PVR, DLCO, NYHA class, 6MWD etc.) and the same discretization cut points. We then applied the same methods for the discretized variable cutoffs published in the COMPERA and French registry to the patients in the REVEAL database to compare their relative performance as measured by areas under the ROC curve (AUC). RESULTS: The Bayesian model demonstrated an accuracy of 90% with an AUC of 0.83 for predicting one-year survival. This was an improvement to the existing AUC of 0.76 for the REVEAL 2.0 calculator. It also significantly outperformed the risk stratification in the COMPERA and French registry (AUC of 0.67 and 0.63 respectively). Bayesian analytic algorithms also allowed the demonstration of the impact of each of these variables on each other, as well as the final outcome (survival), which is a unique feature of Bayesian analysis. CONCLUSIONS: A Bayesian model for PAH demonstrated a modest improvement in accuracy over the existing multi-variate model of the REVEAL 2.0 calculator. It also outperformed the risk predictions based on COMPERA and French registry. CLINICAL IMPLICATIONS: Our long-term goal is to generate a personalized risk assessment tool to aid in bedside decision making for managing patients with PAH. DISCLOSURES: No relevant relationships by James Antaki, source=Web Response Consultant relationship with Bayer, actelion, arena, Please note: $1001 - $5000 Added 02/28/2018 by Raymond Benza, source=Web Response, value=Grant/Research Support Owner/Founder relationship with BayesFusion, LLC Please note: $1-$1000 Added 03/02/2018 by Marek Druzdzel, source=Web Response, value=Ownership interest No relevant relationships by Manreet Kanwar, source=Web Response No relevant relationships by Jidapa Kraisangka, source=Web Response No relevant relationships by Lisa Lohmueller, source=Web Response Employee relationship with Actelion Pharmaceuticals US, Inc Please note: >$100000 Added 02/20/2018 by Mona Selej, source=Web Response, value=Salary No relevant relationships by Judith Speck, source=Web Response No relevant relationships by Carol Zhao, source=Web Response