Creating Domain Specific Resources For Building Semantic Role Labling System For Operative Notes

Abstract

Operative notes contain rich information about techniques, instruments, and materials used in surgeries. With widespread electronic health record (EHR) system adoption throughout healthcare, operative reports are increasingly accessible in electronic format and are potential information sources which may be valuable for a wide variety of secondary functions including new medical knowledge development, decision support, and clinical research. But manual review of large number of reports is time consuming and limits our ability to provide timely evidence-based guide in clinical environment. Automatic extraction of techniques, instruments, materials, and other factors surrounding operative procedures from operative notes can provide an efficient way for physicians to acquire valuable information distilled from diverse experiences reported by clinicians and decide optimal technique approach for patients. To automate the representation and extraction of the rich information from operative notes, the goal of this research is to create domain specific resources needed for creating a semantic role labeling (SRL) system to extract information from operative notes. The coverage of existing domain-specific resources and general English resources for building a SRL system for operative notes were evaluated on a corpus obtained from the Fairview Health Services and the sublanguage used to describe surgical actions in operative notes was investigated. The results from the study show that general English resources are not adequate for building a SRL system for clinical text. Also the study shows some sublanguage characters of operative notes that can be used for parser adaption. Next, an existing unlexicalized probabilistic context-free grammar (PCFG) parser, the Stanford PCFG parser, was adapted to clinical text for better syntactic parsing performance. Finally, domain specific predicate structure (PAS) frames were created for operative notes, as existing semantic frames for general English are not enough for operative notes. The domain specific resource created in this research can be used to build a SRL system for automatically extracting detailed information from operative notes.