Reducing Error in Mortality Models for ICU Patients. Abstract, Executive Summaryand Final Report

摘要

Using multiple logistic regression modeling, four new Mortality ProbabilityModels were developed for characteristics at time of admission and at 24, 48, and 72 hours postadmission. The models performed well, in terms of calibration and discrimination, in the developmental and validation samples of a study cohort alone and combined with an international multisite cohort (19,124 patients). Model performance deteriorated with time from admission. In computer simulations varying the frequency of patient-mix variables, the more commonly occurring variables affected performance negatively more rapidly than the rarer ones. Customizing the logit was more effective than customizing the coefficients in adjusting for deteriorating model performance.