Improvement of classification performance for triple test using data mining approaches

The triple test is a screening test used to calculate the probability of a pregnant woman having a fetus that has an aneuploidy. AFP (Alpha-Fetoprotein), hCG (Human Chorionic Gonadotropin), and uE3 (Unconjugated Estriol) values of pregnant women are computed and compared with the similar records where the outputs (healthy baby or having a disease) are actually known. Bayes theorem is combined with a prior probability derived from maternal age at expected date of delivery is used to calculate the likelihood ratio of a fetus to have diseases like Down syndrome. Current approaches to the calculation of likelihood are known to produce high bias. In this paper, a data mining analysis has been performed to find the best model that is capable of explaining the likelihood of a fetus to have an aneuploidy. 81 triple test records of actually completed pregnancies have been analyzed. 76 of the 81 singleton pregnancies were detected unaffected and 5 of them associated with Down syndrome. The number of 5 pregnancies were increased to 50 pregnancies with the over-sampling technique SMOTE (Synthetic Minority Over-sampling Technique). The Multilayer Perceptron model provided the least false positive rate (13%) and the best detection rate (94%) among several modeling alternatives with the proposed approach. It has been seen that performance of the triple screening test has been significantly improved when compared to the conventional risk assessment.