Machine Learning Techniques for Medical Analysis

Student: Newton Lim-Tung Cheung

Supervisor: Prof. Tom Downs

Category: Computer Systems Engineering Thesis Project

Recent research in machine learning has shown that it incorporates a small amount of dependencies among attributes involved in a database and represents them graphically in the form of a network, known as Tree Augmented Naive Bayes (TAN). This is competitive with the state-of-the-art classifiers such as the C4.5 decision tree and naive Bayes classifier. This thesis explores approaches for constructing Bayesian networks, based on the theory of naive Bayes classifier and existing dependencies among attributes.

One of the proposed approaches for finding the dependencies, is to apply naive Bayes classifier between the attributes itself, then using the attributes with dependencies to construct Bayesian network. This proposed approach is called Bayesian Network with Naive Dependence (BNND). The prediction performance from this network is competitively better than the C4.5 decision tree and naive Bayes classifier. However, if there are too many dependencies found, then the prediction performance of the network is significantly reduced. Therefore, this paper also explores the uses of feature selection to select well-related attributes, in order to reduce the chance of over-fitting in Bayesian network and to increase the computational time.

These Bayesian networks are experimented using medical databases from the machine learning repository in the University of California, Irvine. The results are compared with those from a C4.5 decision tree, naive Bayes classifier and TAN in terms of classification performance, sensitivity, specificity and computational time.

Figure 1: Bayesian network for breast cancer

Poster Presentation (PDF)

Thesis Document (PDF)

Other Related Files