Case-mix adjustment using objective measures of severity: the case for laboratory data

We evaluate the use of routinely gathered laboratory data to subclassify surgical and nonsurgical major diagnostic categories into groups homogeneous with respect to length of stay (LOS).

The source of data is the Combined Patient Experience database (COPE), created by merging data from computerized sources at the University of California San Francisco (UCSF) Medical Center and Stanford University Medical Center for a total sample size of 73,117 patient admissions.

The study is cross-sectional and retrospective. All data were extracted from COPE consecutive admissions; the unit of analysis is an admission. The outcome variable LOS proxies hospital resource utilization for an inpatient stay. Nine (candidate) predictor variables were derived from seven lab tests (WBC, Na, K, C02, BUN, ALB, HCT) by recording the whole-stay minimum or maximum test result.

Patient groups were formed by first assigning to major diagnostic categories (MDCs) all 73,117 admissions. Each MDC was then partitioned into medical and surgical subgroups (sub-MDCs). The 13 sub-MDCs selected for study define a study population of 32,599 patients that represents approximately 45 percent of inpatients. Within each of the 13 sub-MDCs, patients were randomly assigned to one of two data sets in a ratio of 2:1. The first set was used to create, the second to validate, three different LOS predictors. Predictive accuracies of individual DRG classes were compared with those of two alternative classification schemes, one formed by recursive partitioning (the sub-MDC) using only lab test results, the other by partitioning with both lab test results and individual DRGs.

For the eight largest sub-MDCs (81 percent of study population), individual DRGs explained 23 percent of the within sub-MDC variance in LOS, laboratory data classes explained 31 percent, and classes derived by considering individual DRGs and laboratory data explained 37 percent. (Each result is a weighted average R2. The average number of LOS classes into which the eight largest sub-MDCs were partitioned were 20, 10, and 10, respectively. Within six of the eight, partitioning on the basis of laboratory data alone explained more within sub-MDC variance than did partitioning into individual DRGs.

Routine lab test data improve the accuracy of LOS prediction over that possible using DRG classes. We note that the improvements do not result from overfitting the data, since the numbers of LOS classes we use to predict LOS are considerably fewer than the numbers of individual DRGs.