Computational Intelligent Diagnostic System in Predicting Chromosomal Abnormalities of the Fetus

Abstract Title

Computational Intelligent Diagnostic System in Predicting Chromosomal Abnormalities of the Fetus

Abstract Authors

Kypros H. Nicolaides1, Christos N. Schizas2, Andreas Neocleous2, Natasa Schiza2,Costas Neocleous3

1Fetal Medicine Foundation, UK, 2University of Cyprus, Cyprus, 3Cyprus University of Technology, Cyprus

Abstract

Objective:An artificial neural network (ANN) is proposed for predicting chromosomal abnormalities in the fetus by providing as input to the ANN the following parameter values for each case,that are considered by the expert doctor to be the most significant features. Effective first-trimester screening for fetal trisomy 21 and most other major chromosomal abnormalities is achieved by a combination of certain maternal and feto-placental characteristics. The risk for aneuploidies firstly, increases with maternal age and is higher in women with previous affected pregnancies, secondly increases with fetal nuchal translucency (NT) thickness and is higher in those with absence of the fetal nasal bone and abnormal flow through the ductus venosus and across the tricuspid valve, and thirdly, is related to the maternal serum concentration of the placental products free ß-human chorionic gonadotrophin (ß-hCG) and pregnancy associated plasma protein-A (PAPP-A), plus other data such as the previous medical history and ethnicity of the mother.

Method:A suitable database was provided by the Fetal Medicine Foundation (FMF) composed of 34,182 cases from which 33,792 are normal, meaning that the born baby had no trisomy abnormality, where the remaining 390 cases had some sort of abnormality; 213 Down syndrome(T21); 97 trisomy 18(T18); 27 trisomy 13(T13); 18 triploidy; and 35 Turner. The database was divided into two parts; a Training part, and an Evaluation part. For the Training part 25,741 and 198 cases were randomly selected from the normal and the T21 cases respectively. These cases were used to train various multilayer artificial neural networks (ANN) with different architectures in an effort to find a suitable architecture, able to correctly classify all the cases in the training set. The remaining 8,032 normal cases, 15 T21, 97 T18, 27 T13, 18 triploidy, and 35 Turner cases were kept unknown to the ANNs for evaluating the performance of the system. The doctor pointed out that it would have been very important to be able to separate first the T21 from the normal cases with high accuracy.

Results:The ANNs were trained by using the above 9 features and the results have shown that the system was able to correctly classify 100% of the unknown T21 cases and 99.9% of the unknown normal cases. An algorithm was subsequently developed that was able to classify with high accuracy (81%) the other abnormalities. The same exercise was repeated by taking 25,741 and 152 T21 normal and T21 cases respectively and formed a training set, and the remaining 8,032 normal cases, 60 T21, 97 T18, 27 T13, 18 triploidy, and 35 Turner cases were kept unknown to the ANNs for evaluating its performance.The results have shown that the system was able to correctly classify 98.3% of the unknown T21 cases and 99.9% of the unknown normal cases. An algorithm was subsequently developed that was able to classify with high accuracy (89.27%) the other abnormalities.

Conclusion:The proposed methodology produced very good results.