Comparison of Different Statistical Prediction Methods for Fetal Growth Abnormalities

Abstract

The decision tree method has become very popular because it can efficiently accommodate a large amount of data with a mixture of different types of variables, missing data and many irrelevant predictors. Trees also can be graphically presented and easily explained. However, the weaknesses of the decision tree model are high variance and lower predictive accuracy. These problems have been substantially improved by the tree ensemble-based methods: random forests and boosting trees. In this study, tree and tree ensemble-based methods, as well as logistic regression are reviewed and are applied to the Nova Scotia Atlee Perinatal Database to predict fetal growth abnormalities such as infants with birth weight small for gestational age or large for gestational age. It was found that predictive accuracy of the boosted tree model is better than both random forests and decision trees, but this model does not show much improvement over logistic regression.