PhD

dissertationliad is a medical diagnostic decision support system with a very large knowledge base (KB) focused on internal medicine diseases. It uses a special knowledge representation (KR) (the Iliad-KR) for flexible and efficient encoding of medical knowledge. Due to the heuristic nature of the Iliad-KR, probabilities generated by the system have been found to be less than sound. In this dissertation, I proposed a probabilistic KR named Bayesian networks as an alternative to the Iliad-KR and describe a set of algorithms that can transform any KB in Iliad-KR form into a Bayesian network automatically. A two-part experiment was conducted to evaluate the performance of the Iliad-KR and the Bayesian network alternative. The first part was a feasibility. Here I transformed a small KB into Bayesian network form and used a set of statistical performance indices to evaluate the probabilities generated by the Iliad-KR and the Bayesian network model respectively. This study demonstrated the feasibility of such transformation and also suggested that the Bayesian network model is more reliable and discriminative than the Iliad-KR model. The second part of the experiment was a behavioral study. Here the complete Iliad KB for internal medicine was transformed into a large Bayesian network. Twenty patients from two domains of internal medicine were evaluated by four subspecialists from these domains (two in each domain). Diagnostic suggestions generated by the Iliad-KR and the Bayesian network models were given to these physicians and the impact of these suggestions on the physicians' diagnostic decision was measured. The results suggested that computerized diagnostic systems can affect the behavior of physicians and that the Bayesian network model had more positive influence on physicians' diagnosis than the Iliad-KR model