Objective: Since 1989, data have been reported to the Society of Thoracic Surgeons National Adult Cardiac Surgery Database for quality improvement. This information is also data mined for national quality indicators, policy initiatives, and research. Such use has important limitations, because data elements cannot be verified for accuracy. We determined variability of disease etiology and operative data database elements when abstracted by untrained physician abstractors.
Methods: We selected 30 patients who underwent cardiovascular surgery from January to December 2005 (10 each of coronary artery bypass grafting, mitral valve repairs, and aortic valve and associated aortic procedures). Four abstractors (2 cardiothoracic residents and 2 fellows) abstracted 28 variables. Results were compared with abstraction performed by a professional abstractor.
Results: Median percentage agreement among all cases was 89% (range, 42%-100%). Agreements were 94% (28%-100%) for mitral valve, 84% (48%-100%) for aortic valve, and 93% (35%-100%) for coronary artery bypass grafting. Among the aortic valve group, etiology of aortic valve disease had poor agreement (68%) because of cases in which multiple definitions could apply. Degree of valvular regurgitation also had poor agreement (median, 67%; range, 28%-95%). Number of internal thoracic artery grafts and absence of significant valvular disease were reported consistently. Agreements between types of aortic valve procedure and between methods of mitral valve repair (65% and 83%, respectively) were less than expected.
Conclusions: We found variable agreement among untrained data abstractors. This has important implications regarding interpretation of database studies with de-identified data. Without good quality control and consistent standardized definitions, aggregate data in clinical databases may be suspect.
Copyright 2010 The American Association for Thoracic Surgery. Published by Mosby, Inc. All rights reserved.