Ryan Woods MD, MPH, Presenter: Author,Amirsys, Inc

John Eng MD, Abstract Co-Author: Nothing to Disclose

RadLex is a lexicon developed to unify radiologists’ language and is rich in both disease and anatomic terms due to association with other ontologies such as SNOMED and the Foundational Model of Anatomy. Despite the large number of terms in RadLex, less work has addressed imaging observations, and no research has addressed potentially missing terms. The purpose of our study was to estimate RadLex completeness in chest radiography, a common exam with a relatively well-defined group of entities.

We collected a random sample of 51 chest radiograph reports from one month of routine clinical practice of 3 board certified radiologists, each with at least 10 years experience. We parsed each report’s findings and impression sections into individual entities. We defined an entity as a word or phrase identifying an anatomic location, disease, radiographic observation (e.g., hyperinflation), or other entity (e.g., infusion port). We manually compared entities to RadLex by entering the term or a synonym into the online RadLex Term Browser. A match was considered positive when the RadLex term was an exact match, a synonym, or its definition was sufficiently similar to the chest radiograph entity. We calculated frequencies for each category and compared them using a chi-square test.

Among the surveyed reports we found 949 entities, 240 of which were unique. Of the 240 entities, RadLex contained 144 matching terms (60%). The match rate was 71% (45/63) for anatomic entities, 72% (13/18) for disease entities, 61% (65/106) for radiological observation entities, and 40% (21/53) for other entities. The match rate differences were statistically significant (p<0.003). Among unmatched other entities, 9 (28%) described surgical devices or procedures and 11 (34%) described change or imaging recommendations.

While RadLex already contains a large number of terms covering a wide spectrum of anatomy and disease entities, deficiencies remain. We found a lower match rate for imaging observations and other entities, suggesting more terms are particularly needed in these areas to cover routine radiology practice.  

Understanding the completeness of RadLex for radiology report generation is important in developing a more robust, clinically relevant ontology.

Woods, R, Eng, J, Evaluating the Completeness of RadLex in the Chest Radiography Domain.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11012954.html