Errors and bias in using predictive scoring systems

Abstract

Scoring systems used to predict clinical outcomes for critically ill patients have been refined in the past decade, yet even the most recently developed systems contain flaws that limit their application. In general, prediction rules are derived by defining an association between a number of clinical variables and a particular outcome in a reference patient population. By systematically examining the qualities of the independent variables and the size and scope of the derivation data set, potential sources of error and bias can be identified. Existing and future predictive systems must be validated on large groups of patients and continuously updated to keep pace with new approaches to the practice of critical care medicine.