Understanding potential outcomes that might happen during, or the end of a pregnancy is vital to help support the healthcare decisions of mothers and their unborn child. Clinical prediction models (CPMs) are mathematical models/tools that use information from before and during a pregnancy to estimate the chance/risk of a poor outcome. Currently CPMs are developed to predict the risk of just one outcome (e.g., the risk of developing high blood pressure during pregnancy). While this offers some information, predicting the risks of more than one outcome is of more interest (e.g., the chance of developing high blood pressure and having a very small baby at birth). This joint risk would give health professionals an overview of both mother and child’s health to help direct the level of care and monitoring they need. Estimating joint risk is not supported by the ways CPMs are currently developed in pregnancy research. This study will develop a CPM that can predict the risk of multiple events during or immediately after a pregnancy and compare how well these models work to that of models that predict the risk of one single outcome. Comparing how well these different types of models work, the best approach can be used during early antenatal appointments to test which mothers have a higher risk of developing any (not just one) complication. This will ensure at risk mothers receive the extra monitoring they need, helping to improve the overall health of mum and baby throughout pregnancy and prevent poor outcomes.
Understanding potential outcomes that might occur during or the end of pregnancy is crucial to help direct the level of healthcare support given to mothers and the unborn child. As such, there is increasing interest in developing clinical prediction models (CPMs) to predict a single event or outcome using information observed before and during pregnancy. While this offers some information, health professionals are always considering a holistic view of the health of mother and child, meaning that predicting the risks of more than one outcome is of benefit.
Primary care, linked HES and the pregnancy register will be used to develop and validate novel CPMs for individual outcomes (logistic regression) and multiple outcome (multinominal regression) of pregnancy. The multiple outcome model will be extended to predict the risk of any combination of outcome events co-occurring. Such a model would allow end users to the joint risk of different outcomes. The predictive performance of each modelling approach will be compared using three main comparisons: (i) the miscalibration in predicted risk for each outcome of interest, comparing discrimination and calibration metrics, (ii) the level of calibration drift by running on a subset of the data (by calendar year), to identify which modelling technique is subjected to calibration drift, and to what degree, (iii) with input from clinical and patient stakeholders to inform the potential gain and cost thresholds, a decision curve analysis will identify the ‘net benefit’ for each modelling approach compared to a ‘treating all’ or ‘treating none’ approach.
This information will inform which modelling approach will provide the most benefit for predicting the risk and could therefore be implemented for future use in maternity services to effectively identify pregnancies at high risk of complication and allow for more targeted care to mothers and babies that need it the most.
This project will consider the development of risk prediction models for several common and commonly occurring adverse events during pregnancy. These include:
• gestational hypertension,
• gestational diabetes,
• pre-eclampsia,
• pre-term birth,
• fetal macrosomia,
• fetal growth restriction (FGR), and
• stillbirth or neonatal death
This set of outcomes were chosen in collaboration with members of the project team with clinical experience. The operational definition of these outcomes will be the first occurrence of each outcome during a pregnancy episode, identified by either Read, SNOMED CT or ICD-10 codes.
This project will develop risk models for each outcome individually and compare this with developing a single model that can predict the (joint) risk of these outcomes (e.g., the risk of a mother developing gestational hypertension and gestational diabetes). Specifically, for the individual outcome prediction models, the main outcome will be the onset of one of the above outcomes (i.e., one model per outcome); for example, a model predicting the risk of developing hypertensive disorders of pregnancy, the outcome of interest is the first record of a hypertension disorder during that unique pregnancy episode. For the multiple outcome model, the main outcome will be defined as a polytomous outcome, being the onset of the outcomes listed above. The levels of the polytomous outcome will be each possible combination of the above outcomes.
These models will predict the risk of an adverse event during a pregnancy episode. We intend for this project to be an exemplar and have hence, selected a set of adverse outcomes that represent common complications in pregnancy, so that the methodology can be developed and compared; it is not intended to be an exclusive list of all adverse outcomes in pregnancy.
Early pregnancy loss -defined as a pregnancy loss within the first three months, although common -affecting 1 in 4 pregnancies in the UK [1], will not be selected for this study as an outcome of interest. The intended operational use of a clinical risk prediction algorithm in pregnancy would occur during a booking consultation (approximately 8-12 weeks), which often occurs after many early pregnancy losses. Additionally, many co-occurring events (the main interest of this analysis) would develop later in gestation after an early pregnancy loss occurred, meaning that pregnancy episode could not be included in the composite models.
Victoria Palin - Chief Investigator - University of Manchester
Victoria Palin - Corresponding Applicant - University of Manchester
Darren Ashcroft - Collaborator - University of Manchester
Glen Martin - Collaborator - University of Manchester
Holly Hope - Collaborator - University of Manchester
Jenny Myers - Collaborator - University of Manchester
Tjeerd van Staa - Collaborator - University of Manchester
Yiran Zhang - Collaborator - University of Manchester
HES Admitted Patient Care;ONS Death Registration Data;Patient Level Index of Multiple Deprivation;CPRD Aurum Ethnicity Record;CPRD Aurum Pregnancy Register