Abstract: According to the problems existing which include the much larger numbers of dimensions of the sample，much heavier data features and redundant attribution of TCM（Traditional Chinese Medicine） diagnosis，the paper provid the GENRED_GROWTH TCM diagnosis algorithm combining the frequency with the importance of attribution on the basic of research and study of the basic theory and attribution reduction of Rough Sets.Meanwhile the algorithm is applied in the prototype data mining experiment on TCM diagnosis.The results show that the algorithm can be used to reduce redundant attributes in data mining on TCM diagnosis，and the classification accuracy is higher.And it performs accurately on TCM diagnosis datasets.The useful information for the diagnosis can be extracted in order to help to provide some decision-making for the assistant diagnoses.