Machine learning in cardiovascular disease Modelling, prediction, and phenotyping

Machine learning (ML) holds promise for advancing cardiovascular medicine, yet its clinical value depends on rigorous validation and careful evaluation of its limitations. This thesis examines clinical prediction model development and validation, temporal trend analysis, and data-driven patient phenotyping across two complementary settings: a nationwide hospital registry of transcatheter aortic valve implantation (TAVI) procedures (Netherlands Heart Registration), and the Academic Network of General Practice Amsterdam, North-Holland and Almere (AWH-ANHA).
For TAVI, geographic validation revealed substantial variation in 30-day mortality model performance across all 16 Dutch TAVI centres (AUC 0.59–0.79), largely attributable to patient case-mix differences. Temporal validation showed that frozen models degraded in calibration within a year of deployment, with regular refitting improving but not fully resolving instability. Federated learning approaches matched centralized model predictive performance while preserving patient privacy, demonstrating their viability for multicenter model development. Trend analyses over 2013–2022 showed a tripling of TAVI volume alongside declining adverse event rates, with persistent sex- and age-specific differences in outcomes — patterns that should inform both patient selection and future model development.
For heart failure in primary care, multimodal phenotyping of approximately 394,000 patients identified five high-risk phenotypes with incidence rates three to eight times the population baseline. Incorporating phenotype membership improved capabilities of an established risk model, and free-text data captured additional cases missed by structured data alone.
Together, these findings underscore that robust validation, regular model updating, and thoughtful integration of diverse health data are essential to realising the potential of ML-driven cardiovascular care.