Machine learning and translational approaches to personalised care for women with high-risk pregnancies

Thesis

This thesis examines how machine learning and translational approaches may support more personalised care for women with high-risk pregnancies, with a particular focus on gestational diabetes (GDM), long-term prevention of type 2 diabetes (T2DM), and the use of digital tools to improve continuity of care and frontline maternal healthcare.  

First, I conducted a systematic review to assess the methodological quality and clinical readiness of prediction models for adverse outcomes following GDM. This review found that most published models were at high risk of bias and rarely externally validated. Second, I externally validated an existing prognostic model for T2DM complications using a UK cohort of 167,798 women. The model showed moderate discrimination in the validation cohort but poor calibration and overestimation of risk, limiting transportability to a new setting. 

In two further studies, I examined postpartum prevention pathways in primary care. This work showed that automated search tools in primary care could improve identification of women eligible for referral to the NHS Diabetes Prevention Programme. However, this translational work also revealed substantial coding gaps between hospital and primary care records, with many women with hospital-recorded GDM lacking a corresponding primary care code. These missing cases were socially patterned and associated with greater future risk of T2DM. 

Finally, I explored the development and evaluation of SMARThealth GPT (SHP GPT), a large language model (LLM)-based support tool for community health workers in India. This work showed that LLM-based tools may support frontline maternal healthcare when developed through participatory design, grounded in trusted guidelines, and iteratively refined through expert and user feedback.  

Taken together, the studies in this thesis suggest that personalised care in high-risk pregnancy depends not only on improved computational methods, but also on validating them rigorously, engaging intended users throughout development and evaluation, and integrating them into routine care pathways.

Authors: Al Ghadban, Y

• DOI: 10.5287/ora-mvqpox0br
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English