Predicting Who Dies Depends on How Severity Is Measured: Implications for Evaluating Patient Outcomes

OBJECTIVE: To determine whether assessments of illness severity, defined as risk for in-hospital death, varied across four severity measures. DESIGN: Retrospective cohort study. SETTING: 100 hospitals using the MedisGroups severity measure. PATIENTS: 11 880 adults managed medically for acute myocardial infarction; 1574 in-hospital deaths (13.2%). MEASUREMENTS: For each patient, probability of death was predicted four times, each time by using patient age and sex and one of four common severity measures: 1) admission MedisGroups scores for probability of death scores; 2) scores based on values for 17 physiologic variables at time of admission; 3) Disease Staging's probability-of-mortality model; and 4) All Patient Refined Diagnosis Related Groups (APR-DRGs). Patients were ranked according to probability of death as predicted by each severity measure, and rankings were compared across measures. The presence or absence of each of six clinical findings considered to indicate poor prognosis in patients with myocardial infarction (congestive heart failure, pulmonary edema, coma, low systolic blood pressure, low left ventricular ejection fraction, and high blood urea nitrogen level) was determined for patients ranked differently by different severity measures. RESULTS: MedisGroups and the physiology score gave 94.7% of patients similar rankings. Disease Staging, MedisGroups, and the physiology score gave only 78% of patients similar rankings. MedisGroups and APR-DRGs gave 80% of patients similar rankings. Patients whose illnesses were more severe according to MedisGroups and the physiology score were more likely to have the six clinical findings than were patients whose illnesses were more severe according to Disease Staging and APR-DRGs. CONCLUSIONS: Some pairs of severity measures assigned very different severity levels to more than 20% of patients. Evaluations of patient outcomes need to be sensitive to the severity measures used for risk adjustment.