Neural networks to investigate the effects of smoking and alcohol abuse on the risk for preeclampsia

Abstract

Following the application of a large number of neural network schemes that have been applied to a large data base of pregnant women, aiming at generating a predictor for the risk of preeclampsia occurrence at an early stage, we investigated the importance of the parameters of smoking and alcohol intake on the classification yield. A number of feedforward neural structures, both standard multilayer and multi-slab, were tried for the prediction. The database was composed of 6838 cases of pregnant women in UK, provided by the Harris Birthright Research Centre for Fetal Medicine in London. For each subject, 24 parameters were measured or recorded. Out of these, 15 parameters were considered as the most influential at characterizing the risk of preeclampsia occurrence, including the characteristics on whether the pregnant woman was an active smoker or not, and on whether she was consuming alcohol. The same data were applied to the same neural architecture, after excluding the information on smoking and alcohol, in order to study the importance of these two parameters on the correct classification yield. It has been found that both information parameters, were needed in order to achieve a correct classification as high as 83.6% of preeclampsia cases in the whole dataset, and of 93.8% in the test set. The preeclampsia cases prediction, for the totally unknown verification test, was 100%. When information on smoking and alcohol intake were not used, the results deteriorated significantly