Statistical Models For Predicting Intensive Care Unit Mortality; Comparison And Choice

Abstract

The most important clinical outcome measure from the patient’s perspective is ICU mortality. ICU mortality provides a global impression of the ICU performance but is affected by a large number of factors such as severity of illness and age of patients. Special efforts can be made by using ICU prediction model to prevent the mortality risk factors and to improve critical care medicine in Pakistan. Mortality rate in medical intensive care unit was found 88.33% and a high mortality rate was observed for all age groups indicating poor socio-economic and health conditions. Five predictors unconsciousness, history of pre-existent chronic disease, creatinine, SPO2 and temperature were obtained by using both best subset logistic regression and stepwise logistic regression. The stepwise discriminant analysis is also applied and selected five predictors of logistic model and one additional predictor mechanical ventilation, although the assumption of normality and equal variances and covariance are violated. The correct classification rate (CCR) for logistic models and discriminant model were 90% and 80.7% respectively. The area under the receiver operating curve (ROC) for logistic models and discriminant model were 90% and 88.6% respectively. The logistic discrimination reported mechanical ventilation as insignificant factor and its CCR was equal to 90%, which was about 9.3% higher than discriminant model. The ROC curves for logistic model and logistic discrimination model cover the same area. The comparison of selected models indicated the superiority of logistic model over discriminant model. This study shows that logistic model has greater classification and discriminating power than discriminant model when most of the predictors are attribute type. Thus when the assumption of multivariate normality and equal variance and co-variance are not met, and the purpose is classification then discriminant analysis can provide accurate results in terms of variable selected or the signs of the coefficients, but final model must be based on maximum likelihood estimators of the coefficients