Selecting a patient characteristics index for the prediction of medical outcomes using administrative claims data

Recently, there has been a great deal of discussion regarding the use of administrative databases to study outcomes of medical care. A major issue in this discussion is how to classify patients in terms of characteristics such as disease-severity, comorbidities, resource needs, stability, etc. Different indices have been developed in an attempt to provide a common classification scheme in terms of these characteristics. In this paper, we examine the utility of four indices in the prediction of length of stay and 30-day mortality for patients undergoing total knee replacement surgery between 1985 and 1989. The indices that we compare are the Deyo-adapted Charlson Index, the Relative Intensity Score derived from Patient Management Categories (PMCs), the Patient Severity Level derived from PMCs, and the number of diagnoses (up to five) listed in the Medicare claims data. The first three of these indices represent measures of comorbidity, resource use, and severity of illness, respectively. The number of diagnoses is likely to capture aspects of each of these characteristics. We find that all of the indices improve (in terms of model fit) over the baseline (no index) models of length of stay and mortality, and the Relative Intensity Score and Patient and Severity Level result in the greatest improvement in measures of model fit. We found, however, that these two indices have a non-monotonic relationship with length of stay and mortality. The Deyo-adapted Charlson Index performed least well of the four indices in terms of explanatory ability. The number of diagnoses performed well, and does not suffer from problems associated with miscoding on claims data.