First trimester diagnosis of trisomy-21 using artificial neural networks

Abstract

Langdon Down in 1866 reported on a syndrome in which individuals have skin appearing to be too large for the body, a nose small and a flat face. This is a chromosomal disorder caused by the presence of all or part of an extra 21st chromosome, and is known as the Down syndrome, or trisomy 21, or trisomy G. In the last fifteen years it has become possible to observe these features by ultrasound examination in the third month of intrauterine life. About 75% of trisomy 21 fetuses have absent nasal bone. In the present work, neural network schemes that have been applied to a large data base of findings from ultrasounds of fetuses, aiming at generating a predictor for the risk of Down syndrome are reported. A good number of feed forward neural structures, both standard multilayer and multi-slab, were tried for the prediction. The database was composed of 23513 cases of fetuses in UK, provided by the Fetal Medicine Foundation in London. For each subject, 19 parameters were measured or recorded. Out of these, 19 parameters were considered as the most influential at characterizing the risk for this type of chromosomal defect. The best results obtained were with a multi-slab neural structure. In the training set there was a correct classification of the 98.9% cases of trisomy 21 and in the test set 100%. The prediction for the totally unknown verification test set was 93.3%.