Risk Prediction of Advanced Adenoma and Adverse Pregnancy Outcomes
Open Access DepositedDownloadable Content
Accurate and generalizable predictions of disease risk in the presence of complicated medical data, such as interval-censored screening data, informative screening times, correlated disease outcomes, longitudinal clinical information collected at different times, multi-modal prognostic features, and heterogeneous source and target populations, are of great importance in current biomedical research. In this dissertation, we focus on novel likelihood-based and deep learning models in three settings for the prediction of advanced adenoma and adverse pregnancy outcomes. First, we address the prediction of advanced adenoma using interval-censored colorectal cancer screening data in the presence of informative screening times through joint frailty and marginal models. Second, we develop a deep learning model, GRU-D-Static, to predict multiple adverse maternal and neonatal outcomes using longitudinal, multi-modal clinical data collected in the PRISMA Maternal and Newborn Health Study. The GRU-D-Static model is designed to accommodate irregular visit schedules, missingness, and heterogeneous clinical inputs. Third, we propose a transfer prediction framework that integrates summary-level scientific knowledge reflecting association measurements between disease outcomes and input features from the target population with individual-level covariates from the source population using soft labels. This approach enables improved prediction of source-trained models in settings with limited individual-level data availability. Together, these modeling strategies provide flexible and interpretable frameworks for risk prediction across data complexities commonly encountered in large-scale screening studies and global health research.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.