Dziopa, Katarzyna; (2023) Data-driven approaches to cardiovascular risk prediction for people with type 2 diabetes. Doctoral thesis (Ph.D), UCL (University College London).

	Text PhD_Thesis_KDziopa-28-07-2023.pdf - Accepted Version Access restricted to UCL open access staff until 1 August 2024. Download (14MB)
	Archive Dziopa_10174337_Thesis_SuppM.zip Access restricted to UCL open access staff until 1 August 2024. Download (203kB)

Abstract

Cardiovascular disease (CVD) is a leading cause of mortality and morbidity across Europe and the world. People with type 2 diabetes (T2DM) have an increased CVD risk compared to people without diabetes. Initiation and intensification of CVD treatment are commonly guided by risk prediction algorithms combining multiple predictors to estimate an individual’s risk of developing the disease. Many existing prediction scores have not been validated in people with T2DM and have not considered diabetes-specific predictors, therefore it is unclear which of these scores should be recommended. This thesis addresses the problem of predicting CVD risk in people with type 2 diabetes. Chapter 2 provides a general overview of the literature considering clinicaland genetic-based models. In Chapter 3, I performed an evaluation of 22 existing CVD prediction models in people with T2DM utilizing Clinical Practice Research Datalink (CPRD) data of 168,871 people with diabetes. These models had limited predictive ability in people with T2DM, with c-statistics below 0.70. Furthermore, scores derived explicitly for people with T2DM did not perform better than those from the general population. Given the poor performance of the currently existing CVD prediction models in people with T2DM, the potential benefit of including polygenetic information was explored. Specifically, data was sourced from UK Biobank and stratified into three groups (representing group with distinct CVD risks) based on T2DM and CVD histories (without T2DM or CVD; with T2DM and without CVD; with T2DM and CVD). The analysis showed that combining PGS scores with traditional predictors improved performance, resulting in improved discriminative ability for coronary heart disease (CHD), atrial fibrillation (AF), and heart failure (HF) in people with type 2 diabetes wihtout history of CVD at enrolment. When considering individuals with newly diagnosed T2DM, the discriminative performance of the models was substantially increased (from below 0.75 to c-statistic for CHD 0.83 and for HF 0.84). Given that incorporating genetic data meaningfully improved performance in people with diabetes, I subsequently attempted to identify additional novel predictors to predict CVD outcomes in T2DM. First, I created a feature engineering pipeline to map 593 UKB data fields to create a cleaned dataset. Through a purpose-built feature selection pipeline, I was able to identify novel features for the prediction of CVD related to kidney disease, glycaemic control, anthropometrics, and life events related to mental health and or familial diseases histories. Additionally, in people without diabetes, I found that glycemic measurements such as HbA1c as well as information about life events are as important as canonical CVD risk factors such as LDL-C or HDL-C. Translating these discoveries into clinical practice may be beneficial in reducing the burden of CVD in individuals with or without T2DM.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item