Laboratory data predicts survival post hospitalization.

Abstract

From a database of 93,077 in-patient admissions, patients assigned to catastrophic, very severe, moderately severe, and average 30-day mortality risk categories (as defined in Medicare Hospital Mortality Information, 1989 release, from the Health Care Financing Administration (HCFA] were selected for study. These admissions account for 30% of all admissions, but 70% of all deaths up to 1 year post admission. To determine whether laboratory information adds to the predictive power of the information used by HCFA, we compare the performance of 1 year survival predictors (Cox model) that use only diagnostic, demographic, and comorbidity information, with the performance of predictors that also include laboratory information. Using a separate set of patients not used for model definition, we find that laboratory data contain significant prognostic information independent of that already available in non-laboratory data. In HCFA's catastrophic disorders for example, non-laboratory information reduces the average risk of predicting a wrong outcome by 17% relative to considering only catastrophic group membership, and adding laboratory data reduces this risk by a further 21%. These improvements result primarily from considering the outcomes of a small set of routine laboratory tests (maximum BUN, AST, and WBC, and minimum CO2, hematocrit, and sodium).