A comparative study of the feature selection influence on diagnosis in traditional Chinese medicine

As a complementary and alternative system to Western medicine, traditional Chinese medicine (TCM) forms an integrated and unique approach to treat diseases. In response to the subjectivity and fuzziness of TCM, quantitative methods are needed. In TCM, the symptoms are often high dimensional and the redundant and irrelevant symptoms may degrade the performance of classifiers. Therefore, a critical procedure in syndrome differentiation is identifying a representative set of features from which to construct a diagnostic model. Then one central problem in diagnosis is how the feature selection can influence the predictive accuracy. This problem was addressed in this work and a comparative analysis of seven important different feature selection methods is performed incorporating with learning algorithms. Two machine learning algorithms were used: decision tree (DT) and Bayesian networks (BNs). We utilize feature selection algorithm prior to the learning phase. The conclusions were evaluated by experiments on a clinical sample database for diagnosing apoplexy syndrome.