Supervised and Unsupervised Learning of Fetal Heart Rate Tracings with Deep Gaussian Processes

Cardiotocography (CTG) comprises of fetal heart rate (FHR) and uterine activity (UA) monitoring during pregnancy. It is used in hospitals on a regular basis because FHR and UA tracings contain important information about fetal well-being. Despite the CTG's long history of use (of almost 50 years), the benefits it brings to the daily practice remain unsatisfying. The interpretation of CTG recordings by obstetricians suffer from high inter- and intra-variability, while their computerized analysis still remains difficult. In this paper, we propose both supervised and unsupervised learning by deep Gaussian processes (DGPs) for classification of FHR tracings. In working with real FHR signals, we obtained promising results which demonstrate the potential of the DGPs methodology. Further, we showed that the performance of the DGPs was improved by utilizing corresponding UA signals.