PURPOSE

: To introduce probabilistic modelling using Bayesian networks and demonstrate their use for HEOR data analysis, prediction of multiple correlated outcomes and decision-analytic modelling with a focus on interpretability and incorporating existing knowledge.

DESCRIPTION

: Machine learning methods are increasingly being used in healthcare and clinical research, particularly for analyzing real-world data. Although proof-of-concept studies have demonstrated the usefulness of these methods, there are several barriers to their translation to widespread use in HEOR and clinical research. "Black box” machine learning models, which boast a high predictive performance, are difficult to interpret and troubleshoot, and often fail in unpredictable ways. It can also be challenging to incorporate domain knowledge, handle large amounts of missing data and multiple correlated outcomes.

This workshop will introduce Bayesian networks for probabilistic modelling of HEOR data. Bayesian networks are advantageous over commonly used machine learning models in cases where interpretability is important, predictions need to be made using missing covariates or large amounts of missing data, and where robust domain knowledge is available for modelling. The audience will be led through a simplified description of the technical details of Bayesian networks, followed by a simple demonstration of (1) learning Bayesian networks from data, and (2) using them for making predictions and for HEOR decision analysis.