Development of prediction models for cardiovascular disease risk in China

Background: Cardiovascular diseases (CVD) are the leading causes of death in China. Since population CVD incidence and risk factor levels vary considerably across regions in China, prevention of CVD taking into account heterogeneity in risk factors and disease rated across China could be advantageous. Risk prediction models are an integral part of CVD prevention guidelines and can be used to help guide intervention. However, there is no model generalizable to the various incidence, risk-factor levels, and composition of CVD (proportion of CHD and stroke) in different regions of China. Recalibration, an approach to adapt risk scores to local contemporary circumstances, would be of potential benefit to reflect the diversity of CVD epidemiology across China. This thesis aimed to construct a CVD risk estimation system, which is calibrated to CVD risk in different regions in China and can be regularly updated in the future in response to changing trends in CVD rates. Methods: The project involved several interlinked steps. First, I conducted a thorough review of the current epidemiological features of CVD across regions and the presented implications for the prevention of CVD in China; and reviewed and summarised the qualities of CVD risk scores recommended by primary prevention guidelines from a global and China perspective, to provide a benchmark for new score development and reveal where improvements may be warranted specifically for the Chinese population; Secondly, I explored different methods for using aggregate and individual data for recalibration of risk prediction scores to select an optimal approach for use in China. Different methods were illustrated and assessed using 185 222 males aged 40-69 years without previous CVD at baseline and no missing data on measurements of risk predictors from UK Biobank (UKB). Thirdly, I compared the performance of three risk scores recommended by national and international primary prevention guidelines (i.e., World Health Organization [WHO] CVD risk charts, Pooled Cohorts Equations (PCE), and Prediction for Atherosclerotic Cardiovascular Disease Risk in China (China-PAR) models) in 11,169 participants without CVD at baseline survey during 2008-2010 from the Fangshan cohort (FSC) in rural Beijing. The original and recalibrated models were assessed for calibration, discrimination, and reclassification. Finally, the WHO CVD score was recalibrated to China as a whole using aggregate level country, sex- and age group-specific risk factors averages and mortality estimates. And the China-specific score was further recalibrated to each providence with province-specific estimates. Risk factor values were obtained from the China Chronic Disease and Risk Factors Surveillance (CCDRFS), a nationally and provincially representative cross-sectional survey of 145 268 participants aged 40-80 years old. The Chinese Center for Disease Control and Prevention (China CDC) provided mortality rates, estimated using the Global Burden of Diseases 2017 study methodology and data from published scientific reports, disease registries, and health system administrative records. The benefits of the province- versus whole country-specific recalibration were explored by comparing calibration and potential public health impact. Furthermore, preliminary recalibration to urban and rural-specific areas within each province was explored by using mortality statistics from the national death surveillance points (DSP) system which included 605 surveillance points covering more than 80% of the deaths nationally. To avoid the limitations in predicting only fatal CVD risk and develop a risk prediction system calibrated to estimate province-specific total (fatal and non-fatal) CVD risk, I further explored possibility of the developing of a total (fatal plus non-fatal) CVD risk prediction system recalibrated to each province respecting the different rates in urban and rural differences in CVD rates. To explore the translation of mortality rates to total event rates, multipliers (the ratio of total incidence to mortality) were estimated using data of 512,726 participants, aged 30-79 years, enrolled from 5 urban and 5 rural areas in 2004-2008 from China Kadoorie Biobank (CKB). Main results: There are significant regional variations in the mortality rate, composition (proportion of IHD and stroke), and epidemiological patterns of CVD in China. More importantly, the disparity in the epidemiology of CVD among Chinese provinces widened over time. Meanwhile, the CVD prevention recommendations depend more and more on models of CVD risk. There is significant unmet need for a better calibrated risk score taking into account the variations in risk across province and urban/rural environment. A general practical approach was compared with model-specific replacement method for recalibration of risk models of various types, using various alternative sources of data. The example in UKB showed that the general regression method could effectively calibrate the risk model with aggregate data and provide greater simplicity when recalibration needs to be done in an age-specific way. For countries/regions with limited resources, the general regression methods with aggregate data can facilitate regular update of CVD risk score to the CVD rates and risk factor levels in contemporary population. Thus, the general regression approach with aggregate level population data was chosen for recalibration of CVD risk score in China. The validation in FSC showed, the WHO CVD risk score, PCE and China-PAR score all discriminated risk well, with C-indices of 0.735 (95% CI: 0.715, 0.755), 0.731 (95% CI: 0.712, 0.750) and 0.732 (95% CI: 0.712, 0.751) respectively, but performed variously in relation to calibration. However, a simple recalibration could significantly improve and equalized their calibration. This led to selection of the WHO risk score for further province specific recalibration due to its simplicity and advantages in its adaptability to different stroke and CHD rates. After recalibration with risk factor values from CCDRFS and mortality rates from China CDC to the China as a whole, although the China-specific WHO score performed well in China on average, it still overestimated or underestimated mortality risk when used in each province. Province-specific models provided more accuracy prediction of CVD risk in each province. Accordingly, using the province-specific scores for an individual with the same combination of risk factors, the 10-year risk of CVD mortality differed substantially across provinces. Consequently, the proportions of individuals classified at high risk (5-year fatal CVD risk>5%) were strikingly more varied across provinces when using the province- versus China-specific models. For example, the proportion of men classified as high risk ranged from 9% to 39% across provinces. Based on the national DSP system, the raw mortality of CVD was higher in rural versus urban areas within 28 of 31 provinces by 1.03 to 1.96-fold. Thus, a preliminary urban and rural-specific risk prediction system within each province was developed to further avoid over or underestimation using the same recalibration methods. Results in CKB showed that while mortality rates were higher in rural areas, estimated multipliers were lower compared with urban areas. Multipliers were more extreme for stroke versus CHD events, higher for women, and decreased with age. Further research is needed to assess the appropriate application of multipliers from CKB to all 31 provinces in the development of a province-specific total CVD risk estimation system. Conclusion: A province-specific CVD mortality risk estimation system that can be regularly recalibrated in the future using routinely available information was constructed. The application of the recalibrated CVD risk score should help accurately estimate CVD risk in individuals from China and assist policymakers in making more appropriate decisions about the allocation of preventative resources. Further studies are warranted to determine the value of urbanization in the CVD risk prediction and the appropriate method to develop a total CVD risk estimation system with more reliable data