Predictions of hospital mortality rates: A comparison of data sources

Background: Comparing hospital mortality rates requires accurate adjustment for patients' intrinsic differences. Commercial severity systems require either administrative data that omit vital clinical facts about patients' conditions at hospital admission or costly, time-consuming abstraction of medical records. The validity of supplementing administrative data with laboratory data has not been assessed. Objective: To compare risk-adjusted mortality predictions using administrative data alone; administrative data plus laboratory values; and the combination of administrative, laboratory, and clinical data. Design: Retrospective cohort study. Setting: 30 acute care hospitals. Patients: 46 769 patients hospitalized with acute myocardial infarction, cerebrovascular accident, congestive heart failure, or pneumonia. Measurements: Each patient's probability of dying was estimated by using administrative data only (unrestricted administrative models), administrative data restricted to secondary diagnoses that are unlikely to be hospital-acquired complications (restricted administrative models), restricted administrative data plus laboratory data (laboratory models), and restricted administrative data plus laboratory and abstracted clinical data (clinical models). Results: The unrestricted administrative models predicted death better than the restricted administrative models (average areas under the receiver-operating characteristic [ROC] curves, 0.87 and 0.75, respectively) and as well as the laboratory models and the clinical models (average areas under the ROC curves, 0.86 and 0.87, respectively). The good mortality predictions obtained by using the unrestricted administrative models result from inclusion of hospital-acquired complications that commonly precede death. The laboratory models ranked 93% of patients and 95% of hospitals in a manner similar to the clinical models; in comparison, rankings provided by the laboratory models were similar to those provided for 75% of patients and 69% of hospitals by the unrestricted administrative models and for 72% of patients and 77% of hospitals by the restricted administrative models. Conclusions: Adding laboratory data (often available electronically) to restricted administrative data sets can provide accurate predictions of inpatient death from acute myocardial infarction, cerebrovascular accident, congestive heart failure, or pneumonia. This alternative avoids the cost of data abstraction and the serious errors associated with using administrative data alone.