Stroke risk evaluation in a contemporary population of 0.5 million Chinese adults

Thesis

Stroke is a leading cause of death and disability globally, with China bearing the largest stroke burden in the world. Global strokes can be halved through primary prevention strategies including the use of risk prediction models to identify individuals with high risk of stroke. However, the models advocated for in current international guidelines have been derived mainly in high-income countries, and their clinical utility in a contemporary Chinese population is uncertain. Moreover, these models generally do not account for the heterogeneity between ischaemic stroke (IS) and haemorrhagic stroke (HS) pathological types as well as further aetiological subtypes of IS. The investigations in this thesis describe the assessment of existing risk prediction methods and development of novel models for evaluating primary stroke risk in the China Kadoorie Biobank – a prospective cohort study of 0.5 million Chinese adults recruited from 10 areas in China in 2004-2008. First, we evaluated the 2017 Framingham Stroke Risk Profile in CKB, and used a similar Cox modelling approach to develop novel models for 9-year risk prediction of total stroke that yielded a 6%-8% increase in area under the receiver operating characteristic curve (AUC) and >90% reduction in the calibration χ^2 test statistic. Risk prediction models for IS and HS were developed with similar results. Second, machine learning approaches including logistic regression, support vector machines, random survival forests, gradient boosted trees (GBT), and multilayer perceptrons were compared for 9-year risk prediction of total stroke. We also developed an ensemble approach combining Cox and GBT predictions, which identified individuals at high-risk of stroke with higher accuracy, specificity, and precision than any single-model approach. Third, we explored the utility of single versus sequential measurements of risk factors for prediction of stroke, and did not find any substantial improvements from sequential risk factor measurements. Fourth, we developed models to assess differential risk of IS and HS and estimated reductions in events and fatalities using optimised statin therapy policies informed by such models. Finally, we characterised the heterogeneity of IS aetiological subtypes and developed models to classify 24% of CKB’s primary IS events that were previously characterised as having undetermined aetiology by the existing Causative Classification System (CCS) for Ischemic Stroke. The contributions of this thesis advance the understanding and prevention of stroke in China.

Authors: Chun, M

• DOI: 10.5287/ora-4jb0rkxm9
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; risk prediction; cardiovascular disease; China; stroke
• Subjects: Machine learning; Health; Cerebrovascular disease; Risk assessment