Improving case identification in administrative databases such as workers compensation data: use of narrative text

Abstract

To evaluate if case identification in administrative data can be improved using narrative text searches. We analyzed workers compensation (WC) claims data to identify both the nature and cause of injuries resulting from ladder use. As part of an ongoing study, we identified 9826 potential ladder-related injury cases using a computerized search of 535,605 worker's compensation (WC) claims from a large insurer. Cases were identified using index search terms (e.g., “ladder”, “ladd”) and by specific codes indicating ladders. Among these cases we used comparable index terms such as “frac” to identify all fracture cases, which were compared with the cases that were coded as fractures. Manual review of injury mechanism and nature of injury text was used to validate the ability of predetermined index term searches to identify cases. Index term searches alone identified 98.3% of potential ladder cases. Fracture was assigned as the nature of injury code for 589 cases, of which only 34 did not contain appropriate fracture index terms (4 were miscoded as fractures). An additional 141 potential fractures were identified by fracture index terms, of which 113 were true fractures. A similar review of dislocations found an additional 7 true fracture cases. Of the 705 true fracture cases, 120 (17%) were not coded as fractures—all but 2 of whom were identified by index term searches of free text. Of the 705 potential ladder injuries identified involving fractures, 589 (83.5%) were falls from ladders, 101 (14.3%) were other injuries involving ladders (e.g., tripped on ladder), and only 15 (2.1%) were incorrectly identified as ladder injuries. Narrative text analysis using index term searches provides a valuable adjunct to case identification in administrative databases that often have incomplete nature and cause of injury coding. These searches also provide a useful check on the validity of injury coding, with 17% of all fractures being identified only by text searches.