ABSTRACT
OBJECTIVE: There is a significant consumer demand for complementary and alternative medicine (CAM) therapies as possible alternatives to drugs in the treatment and prevention of chronic diseases. Expanding controlled vocabularies to include CAM treatment relations could help meet those needs by facilitating information retrieval from the published literature. The purpose of this study is to design and evaluate two methods to semi-automatically extract CAM treatment-related semantic predications (subject-predicate-object triplets) from the biomedical literature using the Semantic Medline database (SemMedDB). METHODS: Predications were retrieved from SemMedDB, a database of semantic predications extracted from article abstracts available in Medline. Predications were retrieved for 20 biologically-based and 3 mind-body CAM therapies. The first method (allMedline) retrieved predications from any Medline citation, while the second method (soundStudies) only retrieved predications from scientifically sound clinical studies. Filtering criteria were applied to identify the predications focusing on the treatment and prevention of medical disorders using various CAM modalities. The disorders were extracted for each CAM therapy and ranked by occurrence. A reference vocabulary, composed of 20 biologically-based and 3 mind-body CAM therapies, was developed to evaluate the performance of each method according to precision and recall of the top 100 ranked concepts as well as average precision and recall. RESULTS: The difference between allMedline and soundStudies in terms of median precision for the top 100 concepts ranked by occurrence was significant (21.0% versus 27.0%, p < .001). The soundStudies method had significantly higher precision (7.0% vs 11.5%, p < .001) and the allMedline had significantly higher recall (37.1% vs 25.6%, p < .001). CONCLUSION: The soundStudies method may be useful for extracting treatment-related predications from the biomedical literature for the highest ranked concepts. Additional work is needed to improve the algorithm as well as identify and report shortcomings for future enhancements of the tools used to populate SemMedDB.
