Predictive contributions to five-year change in carotid-femoral pulse wave velocity using machine learning and penalized regression

Arterial stiffness, assessed by carotid-femoral pulse wave velocity (cf-PWV), is a validated biomarker of vascular aging and cardiovascular risk. However, the extent to which clinical and behavioral factors predict longitudinal changes in cf-PWV remains poorly understood. This study evaluated the extent to which selected baseline clinical and lifestyle variables could predict five-year change in cf-PWV in a population-based cohort. Data from the Influence of Different Risk Factors in Vascular Accelerated Aging (Estudio de Envejecimiento Vascular Acelerado, EVA study; ClinicalTrials.gov NCT02623894), a prospective cohort of adults free from cardiovascular disease, were analyzed. XGBoost was used to capture nonlinear interactions and rank predictor importance. SHAP (SHapley Additive exPlanations) analysis provided interpretability. Subsequently, Elastic Net regression was applied to derive a parsimonious, interpretable equation. Model tuning was performed using cross-validation within the training set, and performance was evaluated in the held-out test set. XGBoost identified smoking duration (years), adherence to the Mediterranean diet, HDL cholesterol, use of antidiabetic drugs, baseline age, and weekly sitting time (h/week) as key predictors. Mean absolute SHAP values were highest for smoking duration, baseline age, and weekly sitting time. The hierarchical Elastic Net model included baseline age (β = 0.31) and weekly sitting time (β = 0.02) as main effects and retained HDL-C (β = −0.13), use of antidiabetic drugs (β = −0.11), Mediterranean diet adherence (β = −0.008), and the age × weekly sitting time interaction (β = −0.24), with limited predictive performance (MAE = 1.27 m/s; RMSE = 1.70 m/s), which did not improve upon a mean-only benchmark (MAE = 1.27 m/s; RMSE = 1.68 m/s). This exploratory hybrid modeling approach characterized the predictive contributions of clinically relevant factors to five-year change in cf-PWV. Neither primary model improved upon the mean-only benchmark; therefore, the derived equation should be considered a hypothesis-generating research model requiring further development and independent external validation. ClinicalTrials.gov NCT02623894. Registered on 8 December 2015. Retrospectively registered. (https//classic.clinicaltrials.gov/ct2/show/NCT02623894).