Abstract
The paper employed Bayesian network (BN) modelling approach to discover causal dependencies among different data features of Breast Cancer Wisconsin Dataset (BCWD) derived from openly sourced UCI repository. K2 learning algorithm and k-fold cross validation were used to construct and optimize BN structure. Compared to Na‹ve Bayes (NB), the obtained BN presented better performance for breast cancer diagnosis based on fine needle aspiration cytology (FNAC) examination. It also showed that, among the available features, bare nuclei most strongly influences diagnosis due to the highest strength of the influence (0.806), followed by uniformity of cell size, then normal nucleoli. The discovered causal dependencies among data features could provide clinicians to make an accurate decision for breast cancer diagnosis, especially when some features might be missing for specific patients. The approach can be potentially applied to other disease diagnosis.