Effectiveness of a hybrid pattern classifier for medical applications

Abstract

This paper compares two pattern classifiers with applications in medicine: the first is an artificial neural network with weight-elimination (ANN-we); the second is a hybrid classifier consisting of a decision-tree (DT) to eliminate variables which have little impact on predicting the outcome of interest, then processing the remaining variables through an artificial neural network with weight elimination (ANN-we). A small database of adult intensive care unit patients was used to compare the performance of the two pattern classifiers. The hybrid classifier performed better than the ANN-we alone as measured by the resulting sensitivity, specificity, and area under the receiver operating characteristic curve (ROC). The second part of the paper describes the application of the better classifier to the problem of predicting pre-term birth, using a very large and complex medical database of mothers and newborns. The hybrid classifier was able to estimate pre-term birth with an accuracy as high as the invasive and expensive fibronectin test, using only 19 variables available in North America before 23 weeks of gestation for parous women (who had a previous child). Additionally, the classifier was able to predict pre-term birth in nulliparous women (who had no previous children) with slightly less accuracy, but higher than any found in the literature today for this population.