A Medication Extraction Framework for Electronic Health Records

This thesis addresses the problem of concept and relation extraction in medical docu-ments. We present a medical concept and relation extraction system (medNERR) that incorporates hand-built rules and constrained conditional models. We focus on two con-cept types (i.e., medications and medical conditions) and the pairwise administered-for relation between these two concepts. For medication extraction, we design a rule-based baseline medNERRr"eedy that identifies medications using the UMLS dictionary. We en-hance medNERR9r S ' with information from topic models and additional corpus-derived heuristics, and show that the final medication extraction system outperforms the baseline and improves on state-of-the-art systems. For medical conditions extraction we design a Hidden Markov Model with conditional constraints. The conditional constraints frame world knowledge into a probabilistic model and help support model decisions. We ap-proach relation extraction as a sequence labeling task, where we label the context between the medications and the medical concepts that are involved in an administered-for rela-tion. We use a Hidden Markov Model with conditional constraints for labeling the relatio