Early risk prediction of pregnancy complications:Integrating clinical and physiological data

Pregnancy involves profound systemic adaptations, including progressive insulin resistance, increased blood volume, and immune modulation. While essential for fetal development, these changes increase predisposition to complications such as gestational diabetes mellitus (GDM), preeclampsia (PE), macrosomia, preterm birth (PTB),and postpartum depression (PPD). Contemporary risk stratification often depends on late-gestation screening and conventional risk factors, resulting in delayed diagnosis, modest discrimination, and substantial costs. This thesis describes the development and evaluation of integrative, early-pregnancy risk-prediction models that leverage routinely collected electronic health record (EHR) data together with physiological signals that can be obtained outside clinical settings. Part One (Chapters 2–3). We target early prediction of GDM as one example pregnancy complication and develop predictive models that incorporate EHR and physiological signals. In Chapter 2, we trained prediction models on demographic and clinical variables derived from EHR data from cohorts in the Netherlands and the United States. Maternal age, pre-pregnancy body mass index, ethnicity, family history of diabetes, and prior obstetric history remained independently predictive of GDM. Prior GDM was strongly associated with risk in the current pregnancy, whereas lack of prior GDM did not confer protection. Chapter 3 augments clinical models with nocturnal heart rate variability (HRV), a non-invasive marker of autonomic function that could be obtained from wearable sensors during sleep. Early-pregnancy models combining HRV with clinical risk factors consistently outperformed the model using only clinical risk factors, underscoring the potential value of wearable-accessible physiological signals in early risk assessment. Part Two (Chapters 4–6). We extend the focus from GDM to other pregnancy complications and introduce approaches for uncertainty quantification. This part includes prediction models for fetal macrosomia, analyses of cardiorespiratory coupling (CRC) as an early physiological marker, and the application of conformal prediction to improve the interpretability and reliability of classification models. Chapter 4 develops trimester-specific models for fetal macrosomia using maternal demographics, anthropometry, early dietary intake, and (when available) fetal ultrasound. In the analyzed cohort, first-trimester models that excluded ultrasound achieved performance comparable to standard third-trimester growth assessments. Moderate periconceptional alcohol intake was an independent risk factor, underscoring the contribution of early lifestyle factors to risk estimation. Chapter 5 investigates CRC measured during first-trimester sleep using bivariate phase-rectified signal averaging. Early alterations in CRC dynamics, including reduced coupling strength and changes in signal complexity, were associated with later GDM, macrosomia, and PTB. These findings suggest that CRC may serve as a promising early biomarker for risk stratification across multiple complications. Chapter 6 develops trimester-specific classification models for multiple pregnancy complications, including GDM, PTB, PE, and macrosomia, using an ensemble of machine-learning models. Each model is augmented with Mondrian Conformal Prediction (MCP) to provide individual-level calibrated confidence and enable stratification into high-,low-, and uncertain-risk groups. Model interpretability is further enhanced by Shapley Additive exPlanations (SHAP) values, which clarify feature contributions at both cohort and individual levels. In summary, this thesis advances early, individualized risk prediction for major pregnancy complications by integrating EHR-based clinical factors with physiological signals. Across complications, markers such as HRV and CRC complement clinical risk factors, while uncertainty-aware predictions and explainability enhance clinical usability. These contributions support earlier stratification and proactive care pathways, with implications for scalable, cost-conscious maternal health.