A classification-tree hybrid method for studying prognostic models in intensive care

Health care effectiveness and effciency axe under constant scrutiny especially when treatment is quite costly as in Intensive Care (IC). At the heart of quality of care programs lie prognostic models whose predictions for a particular patient population may be used as a norm to which actual outcomes of that population can be compared. This paper motivates and suggests a method based on Machine Learning and Statistical ideas to study the behavior of current IC prognostic models for predicting in-hospital mortality. An application of this method to an exemplary logistic regression model developed on the IC data from the National Intensive Care Evaluation registry reveals the model's weaknesses and suggests ways for developing improved prognostic models