Evaluating the completeness of RadLex in the chest radiography domain

RadLex was developed to create a unified language for radiologists. Despite the large number of terms, little research has evaluated the degree to which RadLex contains terms frequently used in clinical practice. The purposes of this project are to estimate the completeness of RadLex in the chest radiography domain and to characterize the absent terms. We chose chest radiography because it is a common exam generating a large number of reports, and the terms used represent a relatively well-circumscribed set of terms compared to other anatomic regions and modalities.

We collected a random sample of 100 chest radiograph reports from 1 month of routine clinical practice of three board-certified radiologists. We parsed each report's findings and impression sections into individual objects. An "object" was defined as any discrete physical object, body part, observation, descriptive modifier, diagnosis, or procedure. Objects were compared to RadLex by entering the object into the RadLex Term Browser. We calculated descriptive statistics and compared the match rate across RadLex categories.

We identified 339 unique objects, with an overall match rate of 62%. The match rate for each category was anatomic object, 77%; physiological condition, 73%; physical object, 65%; imaging observation, 47%; procedure, 0%; other, 41% (P < .0005).

Our study shows that despite the large number of terms in RadLex, terms are still absent and complexities in the definitions of terms exist. However, increasing the completeness and refining the definitions in RadLex is easily surmountable, possibly using manual methods.