Strategies for improving comorbidity measures based on Medicare and Medicaid claims data

Claims-based measures of comorbid illness severity have generally relied on the diagnoses listed for a single hospitalization. Unfortunately, such dia gnostic information is often limited because patients have not been hospita lized during periods of interest, because of incomplete coding of diagnoses on claims forms, or because listed diagnoses represent complications of th e hospitalization rather than preexisting comorbid conditions. To address t hese limitations, we developed and tested four comorbidity index scores for patients with breast cancer, each based on different sources of health ser vices claims from Medicare and Medicaid: hospitalization for breast cancer surgery; outpatient care prior to the hospitalization; other inpatient care prior to the hospitalization; and all sources combined. Varying the number and type of sources of diagnostic information yielded only very small impr ovements in the prediction of mortality at 1 and 3 years. Surprisingly, eve n simpler measures of comorbidity (crude number of diagnoses) and of prior health care utilization (total days spent in the hospital) performed at lea st as well in predicting mortality as did the more complex index scores whi ch assigned points and weights for specific conditions. The greatest improv ement in explanatory power was observed when another source of clinical inf ormation (cancer stage derived from a population-based cancer registry) was used to supplement claims information. Expanding the source of claims diag noses and focusing on time periods prior to an index hospitalization are in sufficient for substantially improving the explanatory power of claims-base d comorbidity indices. Other improvements suggested by our results should i nclude: increasing the completeness and accuracy of claims diagnoses; suppl ementing diagnoses with health care utilization information in claims data; and supplementing claims data with other sources of clinical information. (C) 2000 Elsevier Science Inc. All rights reserved.