Evaluating genetic liability in hypertension and stroke using machine learning and traditional statistical models: Insights from UK biobank studies

Abstract

Stroke is one of the leading causes of death and disability worldwide, with hypertension being a major risk factor for stroke. According to the World Stroke Organisation Global Stroke Fact Sheet, hypertension alone is responsible for over half of all stroke-related deaths and disability-adjusted life years (DALYs). While conventional risk factors for both diseases are well established, the added prediction value of genetic liability remains less clear. The traditional risk prediction models for hypertension and stroke, such as the Framingham Hypertension Risk Score (FHRS) and the Framingham Stroke Risk Score (FSRS), typically rely on clinical and demographic factors, often assume linear effects, and typically overlook genetic liability and complex interactions between predictors. In this thesis, three complementary studies were conducted to examine whether genetic liability could enhance the classification of hypertension and the prediction of strokes via both machine learning (ML) and traditional modelling techniques using data from more than 116,000 participants with European ancestry in the UK Biobank. Genetic variants and their effects obtained from genome-wide association studies were used to construct genetic liabilities for selected cardiovascular disease (CVD) risk factors and stroke, respectively. Multiple predictive models, such as Cox proportional hazards, penalized regression models (both logistic and Cox), tree-based algorithms (random forest, gradient boosting, and decision trees), and neural networks, were assessed after participants were randomly divided into training and testing sets. Discrimination (AUC), calibration, and reclassification indices (NRI, IDI, and Brier score) were used to evaluate the models. Incorporating the genetic liabilities resulted in modest but steady improvements across the studies. The genetic liabilities associated with lipids improved the classification of hypertension best with AUC using random forest. Additionally, stroke genetic liability enhanced stroke prediction with the Cox models, outperforming machine learning models. Among hypertensive individuals, the model's predictive performance (AUC) was higher in men and older adults than in women or younger adults. The Cox models outperformed all the machine learning models. Though ML methods allow for the investigation of non-linearities and interactions. Overall, genetic liability slightly enhances classification and risk prediction.