Memoona Hasnain1, Hirotaka Onishi, Arthur S Elstein. 1. Department of Family Medicine, College of Medicine, University of Illinois at Chicago, 1919 W. Taylor Street, Chicago, IL 60612-7309, USA. memoona@uic.edu
Abstract
CONTEXT AND OBJECTIVE: Inter-rater agreement is essential in rating clinical performance of doctors and other health professionals. The purpose of this study was to establish inter-rater agreement in categorising errors in the diagnostic process made by clinicians using computerised decision support systems. METHODS: Eight possible error categories were developed for coding errors in diagnostic hypotheses and plans for next steps in the work-up. Two independent doctor judges rated 54 work-ups (representing 2 cases, each worked-up by 27 doctors). Inter-rater agreement between the 2 raters was computed using the kappa coefficient. RESULTS: High inter-rater agreement was achieved in all categories except where the manual was not sufficiently specific and raters had to use their judgement. As is typical of the kappa coefficient, however, agreement corrected for chance fell markedly into the "poor" range when the percentages of expected and observed agreement were about the same. CONCLUSION: Raters can achieve good agreement in categorising errors provided they are given explicit scoring rules and do not rely solely upon clinical judgement. The kappa coefficient has limitations in cases where the expected agreement between judges is high and variability is low. The use of 2 indices to assess agreement, analogous to test sensitivity and specificity, is recommended.
CONTEXT AND OBJECTIVE: Inter-rater agreement is essential in rating clinical performance of doctors and other health professionals. The purpose of this study was to establish inter-rater agreement in categorising errors in the diagnostic process made by clinicians using computerised decision support systems. METHODS: Eight possible error categories were developed for coding errors in diagnostic hypotheses and plans for next steps in the work-up. Two independent doctor judges rated 54 work-ups (representing 2 cases, each worked-up by 27 doctors). Inter-rater agreement between the 2 raters was computed using the kappa coefficient. RESULTS: High inter-rater agreement was achieved in all categories except where the manual was not sufficiently specific and raters had to use their judgement. As is typical of the kappa coefficient, however, agreement corrected for chance fell markedly into the "poor" range when the percentages of expected and observed agreement were about the same. CONCLUSION: Raters can achieve good agreement in categorising errors provided they are given explicit scoring rules and do not rely solely upon clinical judgement. The kappa coefficient has limitations in cases where the expected agreement between judges is high and variability is low. The use of 2 indices to assess agreement, analogous to test sensitivity and specificity, is recommended.