A method to enhance the efficiency of Markov blanket for BN in medical diagnosis

Although successfully used in medical diagnosis, Bayesian network is facing great challenge due to the relatively small amount of diagnosed data and the large dimension of features. To address this issue, this paper presents an effective method for creating Markov blanket when building Bayesian network models. The proposed approach consists of two stages. In the first stage, a clustering based method is introduced to rebuild a representative training data by exploiting the undiagnosed data. In the second stage for feature selection, Markov blanket is built up with the consideration of feature interaction. To evaluate its performance, eight disease datasets from UCI machine learning database are chosen and four off-the-shelf classification algorithms are used for comparison. The test result showed that our approach has better classification accuracy than other traditional methods. Furthermore, two stages in our approaches are isolated in experiment to check their relative efficiency.