A retrospective evaluation of a data mining approach to predict fetal asphyxia at delivery in a hospital database

Prenatal care is one of the most frequently used health services. Therefore care takers and researchers have become more interested in managing large administrative databases designed to monitor health practices and guide policy debates. This approach would aid in determining which particular components of prenatal care are key aspects of care leading to optimal birth outcomes. Data mining tools providing simple and effective methods of extracting knowledge from general medical information were used to identify high risk patients, define the most important factors (37 reproductive variables) leading to fetal asphyxia at delivery and to low birth weight infants, and thereafter build a multivariate relationship model to show the relationship between any two variables in a way that such relationships are easy to view. The first part of this study tested the value of the data mining (SAS Enterprise Miner) approach in predicting the occurrence of fetal asphyxia (2%) and low birth weight infants in an unselected, general obstetric population of 23,284 women. Three data mining models, logistic regression, neural networks and decision tree induction, were used in the analysis. In the second part of the study, results were compared with each other in assessment node of EM. Neural networks yielded with the first data set quite a similar and with the second data set greater percentage of correct classifications (AUC=.52 in the first, AUC 0.66 in the second) with regard to the outcomes. Collectively, in predicting low birth weight infants neural network and tree methods appeared to be more efficient than logistic regression. Our evaluation showed that the data mining approach had a promising predictive value in identifying early signals of fetal asphyxia, and with a very large amount of information to evaluate, such an approach is less open to human error 3 than the former statistical methods. However, further studies with more numerous variables are needed to improve the sensitivity for clinical purposes.