Machine Learning-based Detection of In-Utero Fetal Presentation from Non-Invasive Fetal ECG

Preterm births account for more than 10 % of all newborns. An adverse fetal presentation is a risk factor for intrapartum and neonatal mortality. To date, no technology enables a longitudinal, ubiquitous, and unobtrusive monitoring of fetal presentation. This study presents a first approach to fetal orientation detection based on non-invasive fetal electrocardiography (NI-fECG) using the non-invasive multi-modal foetal ECG-Doppler data set for antenatal cardiology research. The data set contains 60 recordings from 39 pregnant women (21–27 weeks), including NI-fECG and ultrasound position ground truth. We evaluated both handcrafted and generic features for five different classifiers (k-Nearest-Neighbor, Decision Tree Classifier, Support Vector Classification, AdaBoost Classifier, and Multilayer Perceptron) using cross-validation on subject splits on a cleaned subset. Best results for the distinction between vertex (head down) and breech (head up) were achieved using an AdaBoost classifier with a balanced accuracy of 86.5 ± 15.0 %. With this work, we take a first step towards longitudinal fetal presentation monitoring, which contributes to a better understanding of reduced fetal movements and extends the potential applications of NI-fECG in prenatal care. In future work, we will expand our classification system to detect more detailed fetal presentations using a newly created data set.