Predicting pregnancy outcomes using longitudinal biomarkers : analysis of urinary human chorionic gonadotrophin levels in normal and failing pregnancies

Early miscarriage affects approximately 25% of confirmed pregnancies and can adversely impact a woman's body and mind. Human chorionic gonadotrophin (hCG) is used to confirm pregnancy and as a triaging tool in cases of suspected pregnancy loss. Thought its current use may be limited, the potential of hCG in the context of miscarriage is greater than is currently acknowledged. Profiles of hCG for healthy and failing pregnancies have consistently been shown to be distinct, providing motivation for quantifying the association between repeatedly observed hCG and miscarriage. Naive approaches model this association via techniques typically reserved for modelling each outcome separately. Such methods fail to consider the continuous nature of the biomarker, measurement error and appropriate estimation of uncertainty. The joint longitudinal-survival model provides a framework to simultaneously model a longitudinally observed biomarker and time-to-event outcome. In its most common incarnation, the joint model consists of a linear mixed-effects model and a proportional hazards survival submodel. The dependency between the biomarker response and survival outcome is underpinned by shared random effects, laying the groundwork for subject-specific survival predictions. The aim was to use advanced statistical models to estimate the association between miscarriage and hCG, and to extend this to predict individual outcomes. These methods were applied to data collected by SPD Development Company Ltd. The study prospectively followed women as they tried to conceive. Volunteers collected daily urine samples from the start of their cycle to up to day 60 if they conceived. The application of cutting-edge methods to this unique dataset allowed the association to be estimated using complete longitudinal profiles of hCG. Analysis extended to diary data, to establish whether timing of intercourse can alter the rate of miscarriage. A key modelling assumption of the joint models fitted in this thesis were also assessed via a simulation study.