| Abstract [eng] |
This doctoral dissertation examines cardiovascular risk prediction by comparing conventional risk calculators with machine-learning models that integrate vascular ageing biomarkers. The study conducts a head-to-head evaluation of nine risk prediction models in a Lithuanian primary prevention cohort, assessing differences in patient risk stratification, statin therapy eligibility, and overall prognostic accuracy. The study also presents a detailed analysis of inter-model agreement and assesses each model’s discrimination and calibration across multiple endpoints. Results revealed substantial heterogeneity among conventional models: patient risk classification and treatment recommendations varied substantially across different algorithms, and no model was consistently superior across all endpoints. Several models – particularly SCORE2 – exhibited poor calibration in this cohort, systematically overestimating absolute risk, which underscores the need for local recalibration. In addition, a machine-learning model integrating traditional risk factors with vascular ageing biomarkers – most notably carotid–femoral pulse wave velocity – achieved accuracy comparable to established models while yielding additional prognostic information. Overall, biomarker-guided ML prediction models that combine traditional risk factors with selected vascular biomarkers can refine established risk assessment and support more personalised prevention and guideline refinement. |
