Data at ICU admission and after 24 h in the ICU were collected on 755 patients, to derive multiple logistic regression models for predicting hospital mortality. The derived models contained relatively few and easily obtained variables. The weight associated with each variable was determined objectively. There were seven admission variables, none of which were treatment dependent, and seven 24-h variables reflecting treatments and patients' conditions in the ICU. Predicted outcomes using these two models were closely correlated with actual outcome. Theoretically, a predictive model would be useful to physicians for triage decisions as well as determining aggressiveness of care through discussions with families, determining utilization of ICU facilities, and objectively comparing different ICUs. This research represents an initial attempt to develop models that are not based on subjectively determined weights.
Division of Public Health, University of Massachusetts, Amherst (Drs. Lemeshow and Pastides, and Ms. Avrunin), and the Department of Medicine/Division of Surgery, Baystate Medical Center, Springfield, MA (Drs. Teres and Steingrub).