Cardiovascular disease risk prediction models: does one-score-fit-all?

Abstract

Cardiovascular disease (CVD) remains the leading cause of death globally and is a major contributor to the global burden of disability and impaired quality of life. Individuals with existing or prior health conditions, including diabetes, cancer, depression, and severe COVID-19 are particularly at elevated risk of developing CVD, highlighting the need for precise and equitable risk prediction tools. The changing landscape of CVD, partly due to the COVID-19 pandemic disrupting patterns of CVD incidence and care delivery, and partly due to the rising prevalence of obesity, unhealthy lifestyle factors and chronic conditions, including amongst young adults, raises concerns about the continued applicability of existing risk prediction models derived in the general population. Therefore, the overall aim of this thesis was to assess whether the “one-score-fits-all” approach to CVD risk prediction, whereby one risk prediction model is derived and applied in the general population, remains appropriate for individuals with pre-existing health conditions or during periods of geographical-temporal fluctuations in CVD incidence such as the COVID-19 pandemic. Four specific objectives were addressed: 1) To assess whether current CVD risk prediction models require adaptation in the aftermath of the COVID-19 pandemic; 2) To outline practical recommendations for applying multiple imputation in the external validation of risk prediction models in population-wide electronic health records; 3) To evaluate the prognostic performance and clinical utility of CVD risk prediction models in young adults with pre-existing health conditions; and 4) To compare the predictive value of polygenic risk scores (PRSs) in individuals with and without pre-existing health conditions. Key finding 1: Using population-wide electronic health records (EHRs) from the CVD-COVID-UK/ COVID-IMPACT consortium and NHS England’s Secure Data environment, the SCORE2 family-of-models demonstrated good calibration and discrimination, with no strong evidence indicating the need for adaptation following the COVID-19 pandemic. Some variation in calibration was found across geographical regions of England and populations of different ethnic groups suggesting that country-specific adaptation and the consideration of ethnicity may be beneficial to allow for differences in CVD risk. Individuals previously hospitalised with COVID-19 presented higher estimates of 10-year CVD risk. However, studies with extended follow-up are warranted to determine whether individuals with a COVID-19 history would have their risk underpredicted in the long-term. Key finding 2: While outlining multiple imputation plans for the external validation of CVD risk prediction models in population-wide EHRs, we formulated practical recommendations for efficiently conducting multiple imputation for this purpose with limited computational resources. Four key recommendations emerged: (1) Conduct imputation and analysis using efficient sampling strategies, and carefully consider the trade-off between sampling fraction, number of imputations and number of iterations; (2) Carefully select methods and imputation model specifications to optimise compatibility of imputation and model validation analysis models; (3) Scale and centre variables to minimise the number of variables removed from the imputation model as part of the automatic quality control; (4) Understand the purpose of implementing multiple imputation to ensure appropriate use when externally validating a risk prediction model. Nevertheless, recommendations were very specific to the illustrative example and more robust simulation studies are needed to provide more generalisable and widely applicable guidance. Key finding 3: The SCORE2 family-of-models demonstrated reasonable discrimination in both young and middle-aged adults, with minor overprediction of 10-year predicted risks in young adults. SCORE2 family-of-models displayed limited clinical utility for CVD prevention in young adults, however, prioritising high-risk young adults such as those with type 2 diabetes, a cancer history or a COVID-19 hospitalisation history can yield similar benefit to screening middle-aged adults. With appropriate recalibration and threshold adaptations, existing CVD risk prediction models have the potential to be valuable tools in CVD primary prevention in young adults. Key finding 4: The addition of PRSs to a conventional CVD risk prediction model improved model performance in the general population and in individuals with pre-existing health conditions, adding similar prognostic value in individual with type 2 diabetes, a cancer history or depression compared to the general population. The relative contribution of PRSs showed minor variations between subgroups but were generally comparable to SBP and HDL cholesterol. While these findings support the potential utility of PRSs in diverse clinical subgroups, further research is required to evaluate their cost-effectiveness, clinical acceptability, and implementation feasibility, particularly in younger adults and those with pre-existing health conditions. Overall, this thesis evaluated different aspects of the “one-score-fits-all” approach of CVD risk prediction. It demonstrated that existing CVD risk prediction models remained robust to recent shifts in CVD epidemiology. It highlighted their potential to be applied to high-risk young adults, where with adaptations, it could provide similar clinical utility to middle-aged adults. Finally, it quantified the prognostic value of novel PRSs in subgroups with different pre-existing health conditions.