Automating concept identification in the electronic medical record: an experiment in extracting dosage information

We discuss the development and evaluation of an automated procedure for extracting drug-dosage information from clinical narratives. The process was developed rapidly using existing technology and resources, including categories of terms from UMLS96. Evaluations over a large training and smaller test set of medical records demonstrate an approximately 80% rate of exact and partial matches' on target phrases, with few false positives and a modest rate of false negatives. The results suggest a strategy for automating general concept identification in electronic medical records.