A framework for Bayesian Network revisions applied to oncology

Valk, de, Thijme

A framework for Bayesian Network revisions applied to oncology

Files

Valk, de, T. s-4798902-MSc-MKI92-Thesis-2022.pdf (1.94 MB)

Authors

Valk, de, Thijme

Issue Date

2022-08-12

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16360

Abstract

Maintainability is a requirement for the implementation of Bayesian Networks (BNs) for decision support in dynamic environments. This thesis proposes a new framework categorising possible revisions to BNs. The framework provides scenarios demonstrating relevancy, as well as a computational level formalizations. Within the framework, existing literature is discussed, and proposed algorithmic solutions are highlighted. Besides proposing a new framework, this thesis contributes to the field by proposing a new algorithmic solution for revisions to the included set of variables in a BN. Revisions to structure and parameterization and revisions to the included set of variables were illustrated in a casus. The casus subject was revising BNs for treatment selection in lung cancer. This domain is particularly dynamic, due to the development of new treatments and changing perspectives on the relevant outcome variables in treatment selection. In this casus, a BN was constructed from data using score-based methods. Domain knowledge was elicited from experts though a conceptual model and evidence framework, validated with a domain expert. After construction, the initial network was revised using the proposed algorithmic solutions. Structural changes were highlighted, and changes to parameterization were illustrated through inference. This thesis finds that revisions can have meaningful effects on inference results. Moreover, this work explains how revision can produce different BNs than fully retraining, which poses a challenge to revisions as a simple replacement of fully retraining. Future work could quantitatively evaluate the performance of the proposed algorithmic solutions to determine whether they should play a role in maintenance.