Bayesian Networks for Health Care Support.
Publisher
Metadata
Show full item recordAbstract
Bayesian Networks (BNs) have been considered as a potentially useful technique in the
health service domain since they were invented. Many authors have presented BNs for
managing health care and waiting time, predicting outcomes, improving treatment
recommendation process and many more. Despite all these development effort, BNs
have been rarely applied to provide support in any of these clinical areas. This thesis
investigates the use of BNs for analysing clinical evidence data from observational
studies, currently considered the type of study proving the weakest evidence.
It begins by investigating challenges around the analysis of data and evidence faced by
health professionals in health service. It then discusses the importance of observational
studies to understand how disease, treatments and other clinical factors interact with
each other. Further it describes the various techniques, such as using statistical inference
methods and clinical judgements, available to justify any discovered interactions. In
contrary to Frequentist approaches, Bayesian Networks can combine knowledge and
data to derive evidence of relationships between different factors.
This thesis proposes a novel way to combine knowledge and observational data in
Bayesian Networks to derive evidence for clinical queries. Firstly, it shows how to
construct and refine a Bayesian Network model by performing hypothesis tests to check
which out of a number of experts’ judged causal relations between a set of domain
variables are plausible for the available observational data. Secondly, it proposes
techniques to evaluate the strength of all plausible relations/associations. Finally, it
shows how these techniques are combined into a novel data analysis method for
answering clinical queries by combining knowledge with data. In order to illustrate this
method this thesis uses a case study and data about the operation of a multidisciplinary
team (MDT) that provided treatment recommendations to cancer patients, at Barts and
the London HPB Centre over five years. In summary, the case study shows the potential
for the method and allows us to propose ways to present results in a comprehensible
format.
Authors
Pauran, NargisCollections
- Theses [4223]