Statistical approaches to development and validation of predictive instruments

Abstract

Concepts underlying the development of prediction models for intensive care unit (ICU) outcome are outlined, special emphasis is placed on multivariate logistic regression analysis. A short exposition to experimental study designs and techniques for predictor validation is given. Two statistical procedures for assessing the degree of agreement between predicted and observed mortality risks are discussed. The use of receiver operating characteristic (ROC) analysis is thoroughly reviewed as a major tool for describing and comparing, graphically and statistically, the accuracy of predictors for classifying patients prospectively into survivors and nonsurvivors.