Measuring diagnoses: ICD code accuracy

Health Services Research, Oct, 2005 by Kimberly J. O'Malley, Karon F. Cook, Matt D. Price, Kimberly Raiford Wildes, John F. Hurdle, Carol M. Ashton

Nosology (the systematic classification of diseases) has always fascinated the sick and their would-be healers. Western societies developed an interest in nosology in the seventeenth and eighteenth centuries when they began to track the causes of sickness and death among their citizens. In the twentieth century, when medical insurance programs made payers other than patients responsible for medical care, nosology became a matter of great interest to those public and private payers. The most commonly used nosologies include International Classification of Diseases (ICD), the American Medical Association's Current Procedural Terminology, 4th Edition (CPT-4); the Health Care Financing Administration (HCFA, now known as the Centers for Medicare and Medicaid Services) Health Care Common Procedural Coding System (HCPCS); the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders, 4th Revision (DSM-IV); Europe's Classification of Surgical Operations and Procedures, 4th Revision (OPCS-4); and the Agency for Healthcare Research and Quality's Clinical Classification Software (CCS).

This paper focuses on the International Classification of Diseases, now in its ninth and soon to be tenth iteration; the most widely used classification of diseases. Beginning in 1900 with the ICD-1 version, this nosology has evolved from 179 to over 120,000 total codes in ICD-10-CM (ICD-10 2003; ICD-10CM 2003). The use of codes has expanded from classifying morbidity and mortality information for statistical purposes to diverse sets of applications, including reimbursement, administration, epidemiology, and health services research. Since October 1 1983, when Medicare's Prospective Payment System (PPS) was enacted, diagnosis-related groups (DRGs) based on ICD codes emerged as the basis for hospital reimbursement for acute-care stays of Medicare beneficiaries (U.S. Congress 1985). Today the use of ICD coding for reimbursement is a vital part of health care operations. Health care facilities use ICD codes for workload and length-of-stay tracking as well as to assess quality of care. The Veterans Health Administration uses ICD codes to set capitation rates and allocate resources to medical centers caring for its 6 million beneficiaries. Medical research uses ICD codes for many purposes. By grouping patients according to their diagnoses, clinical epidemiologists use ICD codes to study patterns of disease, patterns of care, and outcomes of disease. Health services researchers use the codes to study risk-adjusted, cross-sectional, and temporal variations in access to care, quality of care, costs of care, and effectiveness of care. Medical and health services researchers commonly use ICD codes as inclusion and exclusion criteria to define sampling frames, to document the comorbidities of patients, report the incidence of complications, track utilization rates, and determine the case fatality and morbidity rates (see Calle et al. 2003 for a recent example) (Steinman, Landefeld, and Gonzales 2000; Calle et al. 2003; Charbonneau et al. 2003; Jackson et al. 2003; Martin et al. 2003; Studdert and Gresenz 2003). The widespread and diverse use of ICD codes demonstrates the central role nosology plays in health care.

Increased attention to code accuracy has occurred both as a result of the application of ICD codes for purposes other than those for which the classifications were originally designed as well as because of the widespread use for making important funding, clinical, and research decisions. Code accuracy, defined as the extent to which the ICD nosologic code reflects the underlying patient's disease, directly impacts the quality of decisions that are based on codes, and therefore code accuracy is of great importance to code users. Accuracy is a complicated issue, however, as it influences each code application differently. Using the codes for reporting case fatality rates in persons hospitalized for influenza, for example, might require a different level of accuracy than using codes as the basis for reimbursing hospitals for providing expensive surgical services to insured persons. Therefore, users of disease classifications, just as users of any measure, must consider the accuracy of the classifications within their unique situations. An appreciation of the measurement context in which disease classifications take place will improve the accuracy of those classifications and will strengthen research and health care decisions based on those classifications.

Researchers studying errors in the code assignment process have reported a wide range of errors. Studies in the 1970s found substantial errors in diagnostic and procedure coding. These error rates ranged from 20 to 80 percent (Institute of Medicine 1977; Corn 1981; Doremus and Michenzi 1983; Johnson and Appel 1984; Hsia et al. 1988). Studies in the 1980s reported slightly increased accuracy with average error rates around 20 percent, and most below 50 percent (Lloyd and Rissing 1985; Fischer et al. 1992; Jolis et al. 1993). Studies in the 1990s found rates similar to those of the 1980 studies, with error rates ranging from 0 to 70 percent (Benesch et al. 1997; Faciszewski, Broste, and Fardon 1997; Goldstein 1998). The inconsistency in the error rates and wide range of reported amounts of error is due largely to differences across study methods (i.e., different data sets, versions of the ICD classifications, conditions studied, number of digits compared, codes examined, etc.) (Bossuyt et al. 2004). However, variation in error rates is also influenced by the many different sources of errors that influence code accuracy (Green and Winfield 1993). By clearly specifying the code process and the types of errors and coding inconsistencies that occur in each study, researchers can begin to understand which errors are most common and most important in their situation. They can then institute steps for reducing those errors.