Prenatal diagnosis of aneuploidy using artificial neural networks in relation to health economics

Abstract

The early detection of fetal chromosomal abnormalities such as aneuploidies, has been an important subject in medicine over the last thirty years. A pregnant woman is advised by the doctor to perform an amniocentesis test, after the identification of increased risk for fetal aneuploidy. Even though the amniocentesis test is almost perfectly accurate, it has several drawbacks. It is an invasive test with around 1% risk for miscarriage; it is financially expensive and requires laboratories and special equipment. In this work we propose a non-invasive method for aneuploidy detection using a dataset with pre-natal examinations of pregnant women and artificial neural networks. We have used a dataset with 50,517 euploid and 691 aneuploid cases. Biological markers of the mother such as the age, blood proteins and ultrasonographic information from the fetus are used as input to the networks. A training set is used to construct neural networks and a test set is used for validation. Each unknown case is assigned into a class between “euploid” and “aneuploid” using a cut-off value on the network output. We create a ROC curve by computing the sensitivity and the specificity for a set of different cut-off values. From the ROC curve, we indicate the importance of the cut-off values in terms of health economics and social affection. It is shown that by increasing the cut-off value, the false positive rate reduces with the cost of an increased false negative rate.