Using Clinical Variables to Estimate the Risk of Patient Mortality

Abstract

The Health Care Financing Administration (HCFA) uses information from hospital bills, such as age, sex, and diagnoses, to estimate statistical models for the probability, or risk, of death during and after hospital stays. The average risk estimates (expected death rates) are compared with the actual death rates to identify potentially poor quality of care. However, the methods have been criticized as inadequate and an often cited reason is the failure to incorporate risk factors for mortality that are known from clinical research. This hypothesis was tested using a stratified, random sample of 41,963 Medicare patients in 84 hospitals. Many clinical measurements were abstracted for testing as possible risk factors, and a few (26) were identified as useful predictors of death using logistic regression. The estimated regressions accounted for 39% of the variation in mortality, a standard severity classification accounted for 29%, and a relatively simple classification of patients into 17 groups, based on diagnoses, accounted for 17%. The logistic regressions yielded more accurate estimated mortality rates than the severity classification, which in turn was superior to the estimation methods used by HCFA. The HCFA methods were found to be biased in identifying outlier hospitals and this bias can be removed or ameliorated by using clinical risk factors to predict mortality. It is possible to estimate the risk of death more accurately using clinical risk factors and to measure the quality of care.