Bayesian networks from ontological formalisms in radiation oncology

Abstract

Bayesian networks (BNs) are compact, powerful representations of probabilistic knowledge well suited to applications of reasoning under uncertainty in medical domains. Traditional development of BN topology requires that modeling experts establish relevant dependency links between domain concepts by searching and translating published literature, querying domain experts, or applying machine learning algorithms on data. For initial network development, these methods are time-intensive, and this cost hinders the growth of BN applications in medical decision making. In addition, they result in networks with inconsistent and incompatible topologies, and these characteristics make it difficult for researchers to update old BNs with new knowledge, to merge BNs that share concepts, or to explore the space of possible BN models in any simple intuitive way. My research alleviates the challenges surrounding BN modeling by leveraging a hub and spoke system for BN construction. I implement the hub and spoke system by developing 1) an ontology of knowledge in radiation oncology (the hub) which includes dependency semantics similar to BN relations and 2) a software tool that operates on ontological semantics using deductive reasoning to create BN topologies. I demonstrate that network topologies built using my software are terminologically consistent and topologically compatible by updating a BN model for prostate cancer prediction with new knowledge, exploring the space of other dependent concepts surrounding prostate cancer radiotherapy, and merging the updated BN with a different prostate cancer BN containing cross terms with the original model. I also produce a BN to aid in error detection in radiation oncology, showing the extent to which Bayes nets are clinically impactful. Moreover, I show that the methodology developed in this research is applicable to medical domains outside radiation oncology by extracting a BN from a description logic version of the Disease Ontology. By translating medical domain literature into ontological formalisms and developing a software tool to operate on those formalisms, I establish a novel, feasible, and useful methodology that advances and improves the creation of clinically viable Bayesian network models. In sum, my research represents a foundational component of a larger framework of automation and innovation that contributes to further application of BNs in medical decision support roles.