Refining Cardiovascular Disease Risk Prediction

Cardiovascular disease (CVD) affects over 500 million people worldwide despite advances in prevention. Newer, more intensive treatments are available but carry higher costs and risks, necessitating careful patient selection. The 2021 ESC Prevention Guidelines recommend using risk prediction models to identify patients who are at high risk of (recurrent) CVD events or who may have high treatment benefits. With this central role in clinical decision-making, model performance has become increasingly important. This thesis refines CVD risk prediction through integration of additional risk factors, development of contemporary models, validation in diverse populations, and exploration of approaches for further personalisation. The 2021 ESC Guidelines organise preventive treatments into a stepwise approach. Chapter 2 assessed the potential impact of this approach in 27,094 patients with established atherosclerotic CVD (ASCVD) across Europe. Following standard Step 1 therapy, substantial residual CVD burden remained, ranging from 22% 10-year risk in low-risk European countries to 60% in very high-risk regions. Intensive Step 2 treatments could prevent 198-245 recurrent CVD events per 1,000 patients treated. Implementation of Step 2 substantially increased the proportion of patients achieving residual risk <10%, from 0.5% to 12% in very high-risk regions and from 20% to 63% in low-risk regions. Chapter 3 validated the SCORE2 model in 1,622 cancer patients. The model underestimated CVD risk, but simple recalibration achieved reasonable agreement between predicted and observed risk. Following recalibration, approximately half of cancer patients would be eligible for preventive treatment according to current thresholds. Chapters 4 and 5 investigated integrating additional risk factors beyond the SMART2 model. Psychological factors (depression, anxiety, insomnia) showed no added predictive value in 20,050 patients across five cohorts. However, several other factors demonstrated added value in 179,382 patients across 11 cohorts, including heart failure, atrial fibrillation, multivessel disease, troponin I, NT-proBNP, albuminuria, employment, and education. A flexible methodology allows integration of available additional factors whilst maintaining SMART2's core utility. Chapter 6 developed SMART-REACH2 for estimating lifetime CVD risk in patients with ASCVD. Built on data from 8,708 patients and validated in over 2 million patients across 54 countries, the model incorporates sex-specific development and recalibration to four European risk regions, demonstrating adequate performance (C-statistic 0.68). Chapter 7 implemented secure multi-party computation to recalibrate SMART2 to a hospital's outpatient clinic by linking electronic health record data with insurance data whilst preserving privacy. This represents one of few real-world implementations of privacy-enhancing technologies in healthcare. Chapter 8 compared models developed from electronic health records versus cohort data from the same outpatient clinic. Cohort-derived models showed superior discrimination (C-statistic 0.670 vs 0.621-0.625) and calibration when validated in 129,705 patients across five external data sources. Chapter 9 identified four distinct clinical phenotypes in coronary artery disease patients using latent class analysis in 88,894 patients: "young & metabolic", "smokers with few traditional risk factors", "elderly with few comorbidities", and "polyvascular & comorbidity". These phenotypes, validated in 5,506 patients, showed marked differences in recurrent CVD risk (hazard ratios 0.97-4.38), providing a framework for future research into pathophysiology and potential treatment response differences.