Prediction of the development of pregnancy-induced hypertensive disorders in high-risk pregnant women by artificial neural networks

Pregnancy-induced hypertensive disorders (PIHD) are common complications of pregnancy and are associated with increased maternal and fetal morbidity. In this study, artificial neural networks (aNN) and multivariate logistic r egression (MLR) were applied to a set of clinical and laboratory data (urea , creatinine, uric acid, total proteins, hematocrit, iron and ferritin) col lected at 16 and 20 weeks of gestation. The efficacy of the two approaches in predicting the development of PIHD in 303 consecutive normotensive pregn ant women at high risk of pre-eclampsia and intrauterine fetal growth retar dation was then compared. The aNN were trained with a randomly selected set of 187 patient records and evaluated on the remainder (n = 116). MLR analy sis was done with the same 116 patients. The performance of each model was assessed using receiver operator characteristic (ROC) curves. Pregnancies h ad a normal physiological course in 227 cases, whereas 76 (25.1%) women dev eloped PIHD during the third trimester. The best aNN at 20 weeks yielded an area under the ROC curve of 0.952, the sensitivity of 86.2%, the specifici ty of 95.4%, the positive predictive value of 86.2% and the negative predic tive value of 95.5% for PIHD. The corresponding values for the MLR at 20 we eks were 0.962, 79.3%, 97.7%, 92% and 93.4%, respectively. The computer-aid ed integrated use of these conventional tests seems to provide a useful mea ns for and early prediction of PIHD development.