Prediction of the risk of adverse outcome for women with pre-eclampsia

Student thesis: Doctoral Thesis

Abstract

Preeclampsia is a leading cause of potentially life-threatening, -altering, and -ending complications during pregnancy globally. The sole method of initiating maternal recovery from preeclampsia is delivery of the placenta. Hence, to optimise maternal outcomes in preeclampsia, we need objective, time-of-disease maternal risk assessment to inform decision-making. Clinical decisions are made at different points of a woman’s journey with preeclampsia, first on admission, then continuously during expectant care, and in vastly different settings with different resources. This thesis describes the development of predictive tools to address these issues: a static model for day of admission, a dynamic model for up to 2 weeks after admission, and hierarchical multistep classification tool for low resource settings with missing data. Methods: Using data from published studies, various static and dynamic modelling methods were tested with variable selection to fit a static and a dynamic model, each balancing model complexity and performance, resulting in the static random forest PIERS-ML model, and the binary mixed effects random forest based dynamic PIERS model. Internal validation was carried out using 25% data withheld from model development, assessing performance via area-under-the-receiver-operator characteristic (AUROC) and defining risk strata with likelihood ratios (LR- and LR+). The PIERS-ML was externally validated in an additional cohort. For the panPIERS hierarchical multistep classification tool, predictor variables were divided into variable groups, and models were fitted to each variable group combination using logistic regression, LASSO and random forest methods. Models were ranked on their ability to classify into the very low or very high risk strata. Generative adversarial imputation nets were used to impute missing values within variable groups when making predictions, and the best model with all variables available or imputed was used. Results: Of 8843 participants in the static data, 590 (6·7%) developed the composite adverse maternal outcome within 2 days. PIERS-ML was accurate (AUROC 0·80 [95% CI 0·76–0·84]) and categorised women into very low risk (8, 0% outcome), low risk (321, 2% outcome), moderate risk (979, 5% outcome), high risk (87, 26% outcome), and very high risk (11, 91% outcome). The external validation in a cohort of 2901 women confirmed the model’s utility. The dynamic PIERS model exhibited robust performance up to 7 days (daily AUROC>0.70) and acceptable accuracy up to 14 days (daily AUROC>0.60), effectively stratifying for the first 7 days into very low (4-10% of patients, 0-1% outcome rate), low (35-40% of patients, 0.5-3% outcome), moderate (46-52% of patients, 2-7% outcome), high (3-6% of patients, 20-35% outcome) and very high (0.5-1% of patients, 70-100% outcome). Of the 128 possible available variable group combinations, the panPIERS tool could rule in the outcome with a LR+≥10 in 112 combinations using a collection of 12 models, while we could rule out the outcome with LR-≤0.1 in all combinations using a collection of 15 models. Of 11472 patients, the model currently used in clinical practice could be used for only 29.3% of patients, while the panPIERS classification tool could be used for all. Conclusion: This thesis presents models and tools optimized for preeclampsia risk assessment, using data from admission, temporal information, and accommodating missing values. These tools provide user-friendly outputs, enhancing the identification of women at varying risks of adverse maternal outcomes within a crucial two-day window. These advancements support informed clinical decision-making, especially in settings with limited resources.
Date of Award6 Sept 2024
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorKimberley Kavanagh (Supervisor) & Paul Murray (Supervisor)

Cite this

'