A Cascade Approach to Extracting Medication Events

A Cascade Approach to Extracting Medication Events Jon Patrick Min Li School of IT School of IT The University of Sydney The University of Sydney Sydney, NSW 2006, Australia Sydney, NSW 2006, Australia jonpat@it.usyd.edu.au mili9528@uni.sydney.edu.au Abstract Information Extraction, from the electronic clinical record is a comparatively new topic for computational linguists. In order to utilize the records to improve the efficiency and quality of health care, the knowledge content should be automatically encoded; however this poses a number of challenges for Natural Language Processing (NLP). In this paper, we present a cascade approach to discover the medication- related information (MEDICATION, DOSAGE, MODE, FREQUENCY, DURATION, REASON, and CONTEXT) from narrative patient records. The prototype of this system was used to par- ticipate the i2b2 2009 medication extraction challenge. The results show better than 90% accuracy on 5 out of 7 entities used in the study. 1 Introduction Electronic records are widely used in the health care domain since we believe they can provide more advantages than the traditional paper record (Sujansky, 1998). However, the value of electronic clinical records depends significantly on our ability to discover and utilize the specific content found in them. Once this content can be detected, the poten- tial benefits for individual clinicians and healthcare organizations are considerable. In this study we focus on discharge summaries, which have their own challenges. This kind of clin- ical record includes several sections. The average word count of these reports is around 1500 words per record. This paper presents a method to extract all the medication related information, and connect the relative entities together to build medication entries using a cascaded approach based on two machine learners. 2 Related Work In this paper, we focus on both NER and RC tasks to extract the medications and their related information (DOSAGE, MODE, FREQUENCY, DURATION, REASON, and CONTEXT) from free-text clinical records. At this time, it’s difficult to com- pare our system with other systems which partici- pated the i2b2 2009 medication extraction challenge, since these publications are unavailable now. Consequently, we can only compare our sys- tem with some similar studies in the literature. In the previous work, only three published studies ad- dress this issue (see the performance comparison in the final section) and these studies do not have a comprehensive and precise definition of medication information. The closest research for medication event extraction relies on parsing rules written as a set of regular expressions and a user-configurable drug lexicon. It includes the event for DRUG, DOSAGE, ROUTE, FREQUENCY, CONTEXT and NECESSITY (Gold et al. 2008). The basic work flow for their system starts by discovering drug names based on a drug dictionary, and the rest of the process uses the MERKI parser. The CLARIT NLP system (Evans et al. 1996) can extract DRUG-DOSAGE information from clinical narratives. This system is based on the rule-based method and five main steps are included, such as tokenization, stemming, syntactic category assign- ment, semantic category assignment and pattern matching. Another system focuses on …

Content truncated to 500 words