Who Will Die in the ICU? APACHE II, ROC Curve Analysis, and, of Course, Cleone

To the Editor. —  The article by Kruse et al¹ important contribution to efforts at predicting the survival of patients admitted to medical intensive care units. However, the authors' use of the area under the receiver operating characteristic (ROC) curve to evaluate the predictions is potentially misleading.It is hard to imagine that the ability of clinicians to estimate intermediate probabilities (ie, a 60% mortality) with precision influences clinical action. Only extremely high or extremely low estimates of the probability of dying provide information that might change clinicians' decisions in the intensive care unit.² Since a ROC curve is established by many intermediate predictions, the estimates of a mediocre predictor can still generate a relatively large area under the ROC curve and appear to have "good" predictive accuracy, but offer little practical clinical value. The relevance of any predictor of mortality is best evaluated by evaluating the correctness