Doctoral Defense

Gaussian Process-based Models for Health Prediction of Fetal Well-being from Cardiotocographic Signals

Liu Yang 

February 8, 2024
12:00 PM
Light Engineering, Room 250
Advisor:  Petar M. Djuric

Cardiotocography, an important instrument for fetal surveillance, continuously monitors fetal heart rate (FHR) and uterine activity (UA) during delivery, aiming to assist obstetricians in reducing the risk of fetal hypoxia and acidosis through prompt surgical interventions. Despite continuous development over the past few decades, electronic fetal monitoring and computer-aided fetal health prediction have not effectively lowered the incidence of severe brain damage or infant mortality, but rather associated with increased cesarean delivery rates. In addressing the challenges in practical application, this dissertation proposes solutions through novel methodologies in both supervised and unsupervised machine learning, using Gaussian processes (GP). 

The initial section of this dissertation focuses on identifying uterine contractions (UC) from tocodynamometer recordings, a fundamental step for FHR interpretation. Following this, an ensemble-based hierarchical model is introduced to tackle skewness in class-imbalanced classification. GP classifiers serve as base classifiers, and a GP latent variable model (GPLVM) is employed to nonlinearly synthesize the outputs from GP classifiers. This method is applied in FHR classification with known labels. Considering the absence of reliable fetal health labels in practical scenarios, innovative unsupervised approaches are developed. These include a first attempt that draws inspiration from phase space reconstruction and manifold analysis, revealing distinctions between the manifolds of healthy and pathological FHR segments. To gather more information beyond FHR, we explore FHR responses to UC. 

The FHR features during and after contractions are extracted. We hypothesize that the relationship between them encodes vital fetal health information. By applying GPLVM, we project the high-dimensional features into a visual space and propose an interpretation approach based on the trajectory of UC- dependent FHR patterns. In addition, we investigate a novel method for sequentially detecting abnormal outputs of an unknown function by employing Yule-Simon processes to model system dynamics and GP for learning the latent function. This method is applied in fetal health prediction, where features during contractions are considered as system inputs and features after contractions as system outputs. The results from unsupervised approaches highlight enhanced contraction tolerance in healthy fetuses despite temporary hypoxia induced by contraction pressure and consecutive abnormal FHR responses to UC in unhealthy fetuses.