Bayesian Networks-CBR Hybrid Model And Application

Early detection and diagnosis of cancer, heart disease, diabetes, and other majoremergencies disease is the key to successful treatment of the disease. Currently, thediagnoses of these diseases rely mainly on the clinical experience of the doctors.There are large diagnostic level differences between different areas and medicalinstitutions, for the lack of medical resources. Therefore the patients may not gettimely treatment, and the quality of medical care is difficult to effectively guarantee.For this reason, the case-based reasoning and Bayesian networks, which are moremature technology of artificial intelligence, are introduced to automate diagnosis.Through the analysis of the clinical data in the past, the experiences of doctordiagnoses become objective and specific to improve the diagnostic accuracy ofmedical diagnosis. Bayesian networks and case-based reasoning have their respectiveadvantages and disadvantages, the effective integration models of both methodswhich are used in medical diagnostic are relatively rare.In this paper, we discuss a series of Bayesian networks-CBR hybrid architectures,and then improve a new BCBR model which is combined with the characteristics ofmedical diagnosis. The model of case-based reasoning is used to structure the basicreasoning framework, the Bayesian network is used as a data mining tool to get thediagnostic properties associated knowledge, and its features can simplify the processof case retrieval. Firstly, by calculating the information gain ratio, we can screen theoptimal subset of attributes from a large number of diagnostic properties and assignweights to each attribute. Secondly, the data is divided into each subset, Bayesiannetworks are established respectively for each subset. The subsets are clustered toreduce the search space, and then we use the frame to express cases, thereby the initialcase base can be obtained. Thirdly, the case retrieval is divided into two sub processes,the initial match and secondary retrieve. The initial match can screen a series of relatively matching cases, and the secondary retrieve use the distance method tocalculate the similarity, so as to get the best match case. Finally, the new case isdiagnosed by the Bayesian networks from the best match case.The BCBR model is applied to the example of breast cancer diagnosis, and usethe breast cancer Wisconsin data set as the training and testing sets, which are fromthe UCI database. The diagnosis results are compared with the naive Bayes classifier,case-based reasoning classifier performances. The results show that the BCBR modelof breast cancer diagnosis has better performances in overall accuracy, sensitivity andspecificity. The results also prove the effectiveness of BCBR when applied to medicaldiagnosis.