Abstract
In today's medicine, arguably the main objective is to provide the highest possible quality of medical care to patients. The provision of high-quality medical care is usually associated with the patients being monitored within medical establishments by specialist practitioners. However, there is a considerable number of medical conditions for which expert medical personnel is limited, or patients may be forced to be away from such experts for long periods of time. One of the most common cases where patients are forced to stay away from expert medical help for long periods of time is that of people suffering from spinal injury. Spinal injury affects tens of millions of people around the world, and has very severe consequences to patients; these include paralysis (of legs and, possibly, arms as well), loss of sensation, incontinence, etc. However, the medical centres around the world that can provide specialised care to such patients is very small, and they are concentrated in a very small number of areas around the globe. Therefore, providing efficient, quality care to patients suffering from spinal injury is a very important issue in the medical community. This observation has led us to the design and development of a Decision Support System that will act as a consultation and caring tool for such patients. The resulting system, called Dimitra-PRO, is based on probabilistic tools (more specifically, Bayesian Networks) to deliver a diagnosis about the possible conditions such patients may suffer from, given the symptoms they exhibit. Prior information is extracted from all possible sources of domain knowledge (literature, questioning experts and personal experience), and is modelled appropriately. In the context of this work, an important breakthrough in the applicability of Bayesian Networks (BNs) in real-world scenarios has also been made. So far, BNs have limited support for representing continuous random variables; if any are used, they are required to follow a Gaussian distribution in order to perform exact inference in a BN, otherwise discretisation is applied. In this thesis, we show that it is now possible to use any function that can be used as probability distribution function for nodes, without applying any kind of approximation, to perform exact inference in Bayesian Networks containing a random mixture of discrete and continuous nodes. BNs are now capable of modelling dependencies between discrete and continuous variables without the need to apply discretisation for continuous variables. By applying this paradigm in the case of Dimitra-PRO, it is experimentally demonstrated that the proposed method for representing continuous random variables in BNs outperforms discretisation of these nodes, thus enabling us to deliver a more accurate prognosis about the patients' condition-which, in turn, satisfies our demand for providing the highest possible quality of care to them.