Explainable Clinical Decision Support for the Prediction of Complications in Pregnancy

Pregnancy care plays an essential role in health outcomes of both pregnant women and their offspring. Machine learning-based clinical decision support systems (CDSSs) could be helpful in improving pregnancy healthcare delivery. They could provide an objective second opinion on clinical decisions, and they could also support resource management in maternity hospitals and clinics. The successful adoption of machine learning techniques in clinical settings is sparse in the literature. A combination of factors has been identified that could hinder the use of machine learning-based CDSSs, and one of the key factors is the lack of explainability in black-box machine learning models, because it could conceal algorithmic biases and undermine user trust. This has raised concerns in the medical field. Studies have shown the benefits of providing explanations in CDSSs, yet to date, explainability has not been widely incorporated into machine learning-based CDSSs. Large-for-gestational-age (LGA) births and gestational diabetes mellitus (GDM) are pregnancy complications associated with maternal overweight and obesity, and they are detrimental to the health of both pregnant women and their offspring. Antenatal lifestyle interventions could be effective in preventing these complications. Therefore, early risk prediction tools for LGA or GDM could help to identify women at risk and allow targeted interventions. This thesis describes the development of explainable machine learning-based CDSSs to predict the risk of LGA and GDM early in pregnancy. It has been conducted through a workflow of using data collected from pregnant women, applying state-of-the-art machine learning and explainable artificial intelligence techniques, developing a web-based prototype, and user testing. In this thesis, multiple early risk prediction models were developed for use in different scenarios to enhance clinical usability, post-hoc explainability was incorporated to explain black-box models, and a prototyping CDSS was developed and tested with healthcare practitioners to understand the impact of different explanations in a CDSS. This research helps to bridge the gap between machine learning studies and clinical adoption of CDSSs, and provides guidelines for future CDSS developers.