Abstract
As the population ages, the incidence and mortality of valvular heart disease (VHD) are rising. Current diagnostic approaches depend on expert heuristics, which may miss complex phenotypes. Unsupervised machine learning (ML) offers a scalable, data-driven alternative capable of identifying hidden patterns in large, multivariable datasets which may improve phenotyping, inform prognosis, and guide therapeutic decisions. We systematically searched PubMed for eligible studies evaluating the use of unsupervised ML on aortic stenosis, mitral regurgitation, and tricuspid regurgitation and extracted data on study population, algorithmic input parameters, ML algorithm, goals and outcome of study. Across VHD categories, we identified that unsupervised learning provides more detailed insights than traditional guidelines-based severity classes in understanding patient phenotypes and outcome prediction. These insights can be personalized to guide management with transcatheter and pharmacologic approaches for asymptomatic or early-stage VHD. Prospective studies are needed to validate these novel unsupervised ML approaches.