Unsupervised method for extracting machine understandable medical knowledge from a large free text collection

Abstract

Definitions of medical concepts (e.g diseases, drugs) are essential background knowledge for researchers, clinicians and health care consumers. However, the rapid growth of biomedical research requires that such knowledge continually needs updating. To address this problem, we have developed an unsupervised pattern learning approach that extracts disease and drug definitions from automatically structured randomized clinical trial (RCT) abstracts. In addition, each extracted definition is semantically classified without relying on external medical knowledge. When used to identify definitions from 100 manually annotated RCT abstracts, our medical definition knowledge base has precision of 0.97, recall of 0.93, F1 of 0.94 and semantic classification accuracy of 0.96.