A new hybrid probability estimation tree algorithm for improving probability-based ranking for diagnosis of chronic hepatitis in traditional Chinese medicine

Accurate probability estimation is much desired in medical diagnosis applications. However traditional learning algorithms aim only at high classification accuracy, such as C4.5. Although several probability estimation tree (PET) algorithms have been proposed and applied in many areas, which assume that each attribute of the samples plays a uniform contribution for classification analysis. In addition, the researches and applications of probability estimation tree models are still blanks in data mining area of traditional Chinese medicine (TCM). In this paper, a new hybrid PET method is proposed, which consists of two phases. One is attribute selection and weighting phase. The critical attributes are filtered out from the original attribute sets, and a weight to be assigned for each attribute. The other is modeling and prediction phase by weighting the attribute of samples based on PET algorithms. Then we compare our method with several representative methods, for example traditional decision trees C4.5, C4.4, NBTree and CLLTree, on chronic hepatitis B (CHB) of TCM sample set. From our experiments, 25 critical symptoms of CHB are selected from original 123 symptoms, and provide an objective standard of symptoms to CHB in TCM diagnosis. The proposed hybrid PET algorithm achieves better performance results than those representative models do. The experimental results show that our proposed method performs well in the field of TCM diagnosis. © 2012 American Scientific Publishers. All rights reserved.