Back to search

The University of Edinburgh

DNA methylation and electronic health record prediction of incident type 2 diabetes

Abstract

dc:description.abstract

Type 2 diabetes mellitus (T2D) is a highly prevalent disease that presents a substantial burden, both to public health and cost to healthcare systems. Prediction of disease onset years in advance could help inform early intervention and preventative treatments. While standard risk factors have shown good predictive performance, there is increasing interest in incorporating molecular biomarkers and electronic health records (EHRs) in statistical machine learning models to improve the ability to predict future disease onset. This thesis explores the use of DNA methylation (DNAm) and EHR data in prediction models for incident T2D and explores a range of challenges related to using these data types. Moreover, the methods applied in these studies are disease agnostic, with potential application to other incident diseases of interest. Firstly, the potential for blood-based DNAm data to augment 10-year T2D risk prediction beyond standard risk factors is investigated, utilising data from a large population cohort, Generation Scotland (GS) (Chapter 3). This study incorporated DNAm and time-to-event information from 14,613 GS participants to train and test a range of statistical and machine learning models. The addition of DNAm data resulted in a significant improvement in predictive performance (AUC=0.872) compared to using standard risk factors only (AUC=0.839). A similar discrimination performance increase was observed in an external validation cohort, the German-based KORA study. Secondly, the challenge of feature pre-selection in the development of predictive models using array-based DNAm data is explored (Chapter 4). Typically, penalised regression models are used with hundreds of thousands of CpG sites in this context. Feature pre-selection is an important step in this process as in such ultra-high dimensional data, the effectiveness of these models may be reduced. This study introduced Related Trait-based Feature Screening (RTFS), a method for utilising associations between CpG sites and health-related continuous traits to pre-select relevant CpG sites for incident disease prediction. This was compared with commonly-utilised pre-selection methods as well as dimensionality reduction with principal components analysis (PCA). RTFS was competitive with the highest performing method which incorporated CpG sites previously identified in incident T2D epigenome-wide association studies. Finally, while the first two studies utilised DNAm, the final (Chapter 5) investigated the use of EHRs as predictors. With the increasing availability of EHRs for research and development purposes in recent years, there has been growing interest in utilising this data type for predictive modeling. However, a limited number of studies have focused on the transferability of the resulting models when applied to data generated within a different healthcare system or country. This study utilised a Finland-based cohort, FinnGen, to develop EHR-based risk prediction models for incident T2D and tested a set of feature engineering methods for improving model transferability. These were evaluated in two UK-based cohorts, GS and UK Biobank. Additionally, a large language model embedding-based imputation approach was assessed for addressing the challenge of missing codes in the data when transferring models to other cohorts. Overall, this work explored a range of challenges in utilising DNAm and EHR data for incident T2D risk prediction and shows improvements in specific aspects key to model performance and generalisability such as CpG pre-selection and transferability of EHR-based models across healthcare systems. The thesis demonstrated that DNAm data can augment incident T2D risk prediction compared to risk factors used in current risk prediction tools and developed a method for pre-selecting relevant CpG sites to help address challenges of using such ultra high-dimensional data. Additionally, this work showed the potential for large language model-based representations of EHR codes to improve transferability of diagnosis code-based risk predictions scores.

Degree

thesis:*
Grantor dc:publisher
The University of Edinburgh
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Cheng, Yipeng
Advisors dc:contributor.advisor
  • Marioni, Riccardo
  • McCartney, Daniel
  • Vallejos, Catalina

Subjects

dc:subject × 7

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
University of Edinburgh
Base URL
era.ed.ac.uk/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Cheng, Yipeng. DNA methylation and electronic health record prediction of incident type 2 diabetes. The University of Edinburgh, 2025. https://hdl.handle.net/1842/43773