Two‐stage approach for risk estimation of fetal trisomy 21 and other aneuploidies using computational intelligence systems

ABSTRACT Objective To estimate the risk of fetal trisomy 21 (T21) and other chromosomal abnormalities (OCA) at 11–13 weeks' gestation using computational intelligence classification methods. Methods As a first step, a training dataset consisting of 72 054 euploid pregnancies, 295 cases of T21 and 305 cases of OCA was used to train an artificial neural network. Then, a two‐stage approach was used for stratification of risk and diagnosis of cases of aneuploidy in the blind set. In Stage 1, using four markers, pregnancies in the blind set were classified into no risk and risk. No‐risk pregnancies were not examined further, whereas the risk pregnancies were forwarded to Stage 2 for further examination. In Stage 2, using seven markers, pregnancies were classified into three types of risk, namely no risk, moderate risk and high risk. Results Of 36 328 unknown to the system pregnancies (blind set), 17 512 euploid, two T21 and 18 OCA were classified as no risk in Stage 1. The remaining 18 796 cases were forwarded to Stage 2, of which 7895 euploid, two T21 and two OCA cases were classified as no risk, 10 464 euploid, 83 T21 and 61 OCA as moderate risk and 187 euploid, 50 T21 and 52 OCA as high risk. The sensitivity and the specificity for T21 in Stage 2 were 97.1% and 99.5%, respectively, and the false‐positive rate from Stage 1 to Stage 2 was reduced from 51.4% to ∼1%, assuming that the cell‐free DNA test could identify all euploid and aneuploid cases. Conclusion We propose a method for early diagnosis of chromosomal abnormalities that ensures that most T21 cases are classified as high risk at any stage. At the same time, the number of euploid cases subjected to invasive or cell‐free DNA examinations was minimized through a routine procedure offered in two stages. Our method is minimally invasive and of relatively low cost, highly effective at T21 identification and it performs better than do other existing statistical methods. Copyright © 2017 ISUOG. Published by John Wiley & Sons Ltd.