Intensive care units (ICU's) are a dramatic and ever more expensive part of the health care system. However they may be used inefficiently. Physicians may admit acutely ill patients to ICU's who will never require their specialized services. In addition, doctors may fail to admit acutely ill patients whose condition rapidly deteriorates outside of these units, and who may have benefited from being in one. To improve the likelihood that the patients most in need of intensive care are the ones who receive it, several distinguished committees and researchers have proposed guidelines for admission to ICU's. In the absence of an immediately life- threatening condition that can only be treated within an ICU they suggest that decision-making should depend ont he probabilities of these outcomes: that the patient will soon develop such a condition, will survive short- term, and will make a """"""""reasonable recovery."""""""" However, physicians working with limited information may not be able to make accurate and discriminating predictions of these outcomes for acutely ill patients prior to the triage decision. There are no valid models available to predict these outcomes, although prior experience suggests that it would be feasible to develop them. Therefore, the major goals of this study are to: 1) assess the quality of physicians' predictions of these outcomes of acute illness that should be most relevant to their initial triage decisions, and to 2) develop multivariable models to predict these outcomes, and validate them by comparing their performance with that of physicians' estimates on an independent patient sample. Minor goals include assessing 3) whether the quality of physicians' predictions may be unfavorably influenced by cognitive biases or the inappropriate use of heuristics; 4) whether variability in physicians' decisions may result from variability in how they weight patient variables and institutional factors, both appropriate and inappropriate (e.g., socioeconomic status); 5) the rate of occurrence of other important results of acute illness. Accomplishment of these goals should pave the way for more discerning use of hospital resources, and potentially could lower costs while simultaneously improving the quality of patient care. To attain these goals, the study will be restricted to patients presenting with congestive heart failure, a common problem for which physicians' decisions are quite variable and probably often sub-optimal. It will be a cohort study of 1866 patients presenting to the emergency rooms of three hospitals of different types. Data collection will include predictions by the physicians who care for these patients in the ER, and in the hospital if they are admitted; physicians' actual triage decisions, and many clinical, physician, and institutional variables. Models will be developed for each outcome using logistic regression analyses of a three-quarters random sample of the data set. Each models' performance on the remaining data will be compared to that of individual and aggregated physicians' estimates. Criteria for performance will include receiver operating characteristic curves, calibration curves, and Brier scores.
