Machine Learning to Predict Pre-Eclampsia and Intrauterine Growth Restriction in Pregnant Women

• Published in: Electronics (Basel) 2022-10, Vol.11 (19), p.3240
• Main Authors: Gómez-Jemes, Lola, Oprescu, Andreea Madalina, Chimenea-Toscano, Ángel, García-Díaz, Lutgardo, Romero-Ternero, María del Carmen
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial intelligence; Biomarkers; Classification; Classifiers; Datasets; Decision trees; Diagnosis; Fetuses; Gestational age; Gestational diabetes; Gynecology; Health aspects; Health care facilities; Hypertension; Machine learning; Mathematical models; Mortality; Parameters; Preeclampsia; Pregnancy; Pregnancy, Complications of; Pregnant women; Recall; Ultrasonic imaging; Variables; Veins & arteries; Womens health
• ISSN: 2079-9292; 2079-9292
• DOI: 10.3390/electronics11193240

The use of artificial intelligence in healthcare in general and in obstetrics and gynecology in particular has great potential. Specifically, machine learning methods could help improve the health and well-being of pregnant women, closely monitoring their health parameters during pregnancy, or reducing maternal and perinatal morbidity and mortality with early detection of pathologies. In this work, we propose a machine learning model to predict risk events in pregnancy, in particular the prediction of pre-eclampsia and intrauterine growth restriction, using Doppler measures of the uterine artery, sFlt-1, and PlGF values. For this purpose, we used a public dataset from a study carried out by the University Medical Center of Ljubljana, in which data were collected from 95 pregnant women with pre-eclampsia and intrauterine growth restriction. We adopted a multi-label approach to accomplish the prediction task. Different classifiers were evaluated and compared. The performance of each model was tested in terms of accuracy, precision, recall, F1 score, Hamming loss, and AUC-ROC. On the basis of these parameters, a variation of the decision tree classifier was found to be the best performing model. Our model had a robust recall metric (0.89) and an AUC ROC metric (0.87), taking into account the size of the data and the unbalance of the class.