Artificial neural networks to investigate the significance of PAPPA and b-hCG for the prediction of chromosomal abnormalities

Abstract

A systematic approach has been done, to investigate different neural network structures for the appraisal of the significance of the free b-human chorionic gonadotrophin (b-hCG) and the pregnancy associated plasma protein-A (PAP-PA) as important parameters for the prediction of the existence of chromosomal abnormalities in fetuses. The database that has been used was highly unbalanced. It was composed of 35,687 cases of pregnant women. In the vast majority of cases (35,058) there had not been any chromosomal abnormalities, while in the remaining 629 (1.76%) some kind of chromosomal defect had been confirmed. 8,181 cases were kept as a totally unknown database that was used only for the verification of the predictability of each network, and for evaluating the importance of PAPP-A and b-hCG as significant predicting factors. In this unknown data set, there were 76 cases of chromosomal defects. The system was trained by using 8 input parameters that were considered to be the most influential at characterizing the risk of occurrence of these types of chromosomal anomalies. Then, the PAPP-A and the b-hCG were removed from the inputs in order to ascertain their contributory effects. The best results were obtained when using a multilayer neural structure having an input, an output and two hidden layers. It was found that both of PAPP-A and b-hCG are needed in order to achieve high correct classifications and high sensitivity of 88.2% in the totally unknown verification data set. When both the b-hCG and PAPP-A were excluded from the training, the diagnostic yield dropped down to 65%.