Automatic symptom name normalization in clinical records of traditional Chinese medicine
- Equal contributors
1 Department of Computer Science, Sichuan University, Chengdu, Sichuan, PR China
2 College of Preclinical Medicine, Chengdu University of TCM, Chengdu, Sichuan, PR China
BMC Bioinformatics 2010, 11:40 doi:10.1186/1471-2105-11-40Published: 20 January 2010
In recent years, Data Mining technology has been applied more than ever before in the field of traditional Chinese medicine (TCM) to discover regularities from the experience accumulated in the past thousands of years in China. Electronic medical records (or clinical records) of TCM, containing larger amount of information than well-structured data of prescriptions extracted manually from TCM literature such as information related to medical treatment process, could be an important source for discovering valuable regularities of TCM. However, they are collected by TCM doctors on a day to day basis without the support of authoritative editorial board, and owing to different experience and background of TCM doctors, the same concept might be described in several different terms. Therefore, clinical records of TCM cannot be used directly to Data Mining and Knowledge Discovery. This paper focuses its attention on the phenomena of "one symptom with different names" and investigates a series of metrics for automatically normalizing symptom names in clinical records of TCM.
A series of extensive experiments were performed to validate the metrics proposed, and they have shown that the hybrid similarity metrics integrating literal similarity and remedy-based similarity are more accurate than the others which are based on literal similarity or remedy-based similarity alone, and the highest F-Measure (65.62%) of all the metrics is achieved by hybrid similarity metric VSM+TFIDF+SWD.
Automatic symptom name normalization is an essential task for discovering knowledge from clinical data of TCM. The problem is introduced for the first time by this paper. The results have verified that the investigated metrics are reasonable and accurate, and the hybrid similarity metrics are much better than the metrics based on literal similarity or remedy-based similarity alone.