Automatic knowledge acquisition from MEDLINE

Abstract

Construction of medical knowledge bases for use in expert systems is an arduous task. We propose a procedure for obtaining medical knowledge via automated analysis of citations found in the National Library of Medicine's MEDLINE database. In this method, simple pattern of keywords and subheading co-occurrences are detected in the keyword descriptor portion of the citations. Each pattern corresponds to a fact, expressed as a semantic relationship between medical concepts. We have constructed a set of 504 pattern-matching rules and applied it to a set of 673 MEDLINE citations to produce 2,795 such facts. The results are presented of an analysis of the syntactic and semantic features of these facts to understand the kinds of knowledge than can be obtained through our method and speculate on the potential uses and pitfalls for knowledge of this type.